Levenshtein Transformer paper code

Summary:
Code for our NeurIPS paper [Levenshtein Transformer](https://arxiv.org/abs/1905.11006)
* Added Levenshtein Transformer model, task and criterion class
* Added iterative NAT Transformer, insertion Transformer and CMLM Transformer model class for baselines
* Add an option for prepending BOS to dictionary class and translation task class

Reviewed By: myleott

Differential Revision: D17297372

fbshipit-source-id: 54eca60831ae95dc721c2c34e882e1810ee575c7

README.md

@@ -6,6 +6,7 @@ modeling and other text generation tasks.
 
 ### What's New:
 
+- September 2019: [Nonautoregressive translation code released](examples/nonautoregressive_translation/README.md)
 - August 2019: [WMT'19 models released](examples/wmt19/README.md)
 - July 2019: fairseq relicensed under MIT license
 - July 2019: [RoBERTa models and code released](examples/roberta/README.md)
@@ -32,6 +33,13 @@ Fairseq provides reference implementations of various sequence-to-sequence model
   - [Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)](examples/translation_moe/README.md)
   - [RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)](examples/roberta/README.md)
   - [Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)](examples/wmt19/README.md)
+- **Non-autoregressive Transformers**
+  - Non-Autoregressive Neural Machine Translation (Gu et al., 2017)
+  - Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement (Lee et al. 2018)
+  - Insertion Transformer: Flexible Sequence Generation via Insertion Operations (Stern et al. 2019)
+  - Mask-Predict: Parallel Decoding of Conditional Masked Language Models (Ghazvininejad et al., 2019)
+  - [Levenshtein Transformer (Gu et al., 2019)](examples/nonautoregressive_translation/README.md)
+
 
 **Additionally:**
 - multi-GPU (distributed) training on one machine or across multiple machines
@@ -50,7 +58,7 @@ translation and language modeling datasets.
 
 # Requirements and Installation
 
-* [PyTorch](http://pytorch.org/) version >= 1.1.0
+* [PyTorch](http://pytorch.org/) version >= 1.2.0
 * Python version >= 3.5
 * For training new models, you'll also need an NVIDIA GPU and [NCCL](https://github.com/NVIDIA/nccl)
 * **For faster training** install NVIDIA's [apex](https://github.com/NVIDIA/apex) library with the `--cuda_ext` option
@@ -92,6 +100,7 @@ as well as example training and evaluation commands.
 - [Language Modeling](examples/language_model/README.md): convolutional and transformer models are available
 
 We also have more detailed READMEs to reproduce results from specific papers:
+- [Levenshtein Transformer (Gu et al., 2019)](examples/nonautoregressive_translation/README.md)
 - [Facebook FAIR's WMT19 News Translation Task Submission (Ng et al., 2019)](examples/wmt19/README.md)
 - [RoBERTa: A Robustly Optimized BERT Pretraining Approach (Liu et al., 2019)](examples/roberta/README.md)
 - [wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)](examples/wav2vec/README.md)

examples/nonautoregressive_translation/README.md

@@ -0,0 +1,90 @@
+# Non-autoregressive Neural Machine Translation (NAT)
+
+This page mainly includes instructions for reproducing results from the paper
+* [Levenshtein Transformer (Gu et al., 2019)](https://arxiv.org/abs/1905.11006).
+
+We also provided our own implementations for several popular non-autoregressive-based models as reference:<br>
+* [Non-Autoregressive Neural Machine Translation (Gu et al., 2017)](https://arxiv.org/abs/1711.02281)<br>
+* [Deterministic Non-Autoregressive Neural Sequence Modeling by Iterative Refinement (Lee et al. 2018)](https://arxiv.org/abs/1802.06901)<br>
+* [Insertion Transformer: Flexible Sequence Generation via Insertion Operations (Stern et al. 2019)](https://arxiv.org/abs/1902.03249)<br>
+* [Mask-Predict: Parallel Decoding of Conditional Masked Language Models (Ghazvininejad et al., 2019)](https://arxiv.org/abs/1904.09324v2)
+
+## Dataset
+
+First, follow the [instructions to download and preprocess the WMT'14 En-De dataset](../translation#prepare-wmt14en2desh).
+Make sure to learn a joint vocabulary by passing the `--joined-dictionary` option to `fairseq-preprocess`.
+
+### Knowledge Distillation
+Following [Gu et al. 2019](https://arxiv.org/abs/1905.11006), [knowledge distillation](https://arxiv.org/abs/1606.07947) from an autoregressive model can effectively simplify the training data distribution, which is sometimes essential for NAT-based models to learn good translations.
+The easiest way of performing distillation is to follow the [instructions of training a standard transformer model](../translation) on the same data, and then decode the training set to produce a distillation dataset for NAT.
+
+### Download
+We also provided the preprocessed [original](http://dl.fbaipublicfiles.com/nat/original_dataset.zip) and [distillation](http://dl.fbaipublicfiles.com/nat/distill_dataset.zip) datasets. Please build the binarized dataset on your own.
+
+
+## Train a model
+
+Then we can train a nonautoregressive model using the `translation_lev` task and a new criterion `nat_loss`.
+Use the `--noise` flag to specify the input noise used on the target sentences.
+In default, we run the task for *Levenshtein Transformer*, with `--noise='random_delete'`. Full scripts to run other models can also be found [here](./scripts.md).
+
+The following command will train a *Levenshtein Transformer* on the binarized dataset.
+
+```bash
+fairseq-train \
+    data-bin/wmt14_en_de_distill \
+    --save-dir checkpoints \
+    --ddp-backend=no_c10d \
+    --task translation_lev \
+    --criterion nat_loss \
+    --arch levenshtein_transformer \
+    --noise random_delete \
+    --share-all-embeddings \
+    --optimizer adam --adam-betas '(0.9,0.98)' \
+    --lr 0.0005 --lr-scheduler inverse_sqrt \
+    --min-lr '1e-09' --warmup-updates 10000 \
+    --warmup-init-lr '1e-07' --label-smoothing 0.1 \
+    --dropout 0.3 --weight-decay 0.01 \
+    --decoder-learned-pos \
+    --encoder-learned-pos \
+    --apply-bert-init \
+    --log-format 'simple' --log-interval 100 \
+    --fixed-validation-seed 7 \
+    --max-tokens 8000 \
+    --save-interval-updates 10000 \
+    --max-update 300000
+```
+
+## Translate
+
+Once a model is trained, we can generate translations using an `iterative_refinement_generator` which will based on the model's initial output and iteratively read and greedily refine the translation until (1) the model predicts the same translations for two consecutive iterations; or (2) the generator reaches the maximum iterations (`--iter-decode-max-iter`). Use `--print-step` to check the actual # of iteration for each sentence.
+
+For *Levenshtein Transformer*, it sometimes helps to apply a `--iter-decode-eos-penalty` (typically, 0~3) to penalize the model finishing generation too early and generating too short translations.
+
+
+For example, to generate with `--iter-decode-max-iter=9`:
+```bash
+fairseq-generate \
+    data-bin/wmt14_en_de_distill \
+    --gen-subset test \
+    --task translation_lev \
+    --path checkpoints/checkpoint_best.pt \
+    --iter-decode-max-iter 9 \
+    --iter-decode-eos-penalty 0 \
+    --beam 1 --remove-bpe \
+    --print-step \
+    --batch-size 400
+```
+In the end of the generation, we can see the tokenized BLEU score for the translation.
+
+
+## Citation
+
+```bibtex
+@article{gu2019levenshtein,
+  title={Levenshtein Transformer},
+  author={Gu, Jiatao and Wang, Changhan and Zhao, Jake},
+  journal={arXiv preprint arXiv:1905.11006},
+  year={2019}
+}
+```

examples/nonautoregressive_translation/scripts.md

@@ -0,0 +1,148 @@
+# Examples of Training scripts for Non-autoregressive Machine Translation models
+
+### Non-autoregressive Transformer (NAT, Gu et al., 2017)
+Note that we need to have an additional module to perform "length prediction" (`--length-loss-factor`) before generating the whole sequence.
+```bash
+fairseq-train \
+    data-bin/wmt14_en_de_distill \
+    --save-dir checkpoints \
+    --ddp-backend=no_c10d \
+    --task translation_lev \
+    --criterion nat_loss \
+    --arch nonautoregressive_transformer \
+    --noise full_mask \
+    --share-all-embeddings \
+    --optimizer adam --adam-betas '(0.9,0.98)' \
+    --lr 0.0005 --lr-scheduler inverse_sqrt \
+    --min-lr '1e-09' --warmup-updates 10000 \
+    --warmup-init-lr '1e-07' --label-smoothing 0.1 \
+    --dropout 0.3 --weight-decay 0.01 \
+    --decoder-learned-pos \
+    --encoder-learned-pos \
+    --pred-length-offset \
+    --length-loss-factor 0.1 \
+    --apply-bert-init \
+    --log-format 'simple' --log-interval 100 \
+    --fixed-validation-seed 7 \
+    --max-tokens 8000 \
+    --save-interval-updates 10000 \
+    --max-update 300000
+```
+
+### Non-autoregressive Transformer with Iterative Refinement (iNAT, Lee et al., 2018)
+Note that `--train-step` means how many iterations of refinement we used during training, and `--dae-ratio` controls the ratio of denoising auto-encoder training described in the original paper.
+```bash
+fairseq-train \
+    data-bin/wmt14_en_de_distill \
+    --save-dir checkpoints \
+    --ddp-backend=no_c10d \
+    --task translation_lev \
+    --criterion nat_loss \
+    --arch nonautoregressive_transformer \
+    --noise full_mask \
+    --share-all-embeddings \
+    --optimizer adam --adam-betas '(0.9,0.98)' \
+    --lr 0.0005 --lr-scheduler inverse_sqrt \
+    --min-lr '1e-09' --warmup-updates 10000 \
+    --warmup-init-lr '1e-07' --label-smoothing 0.1 \
+    --dropout 0.3 --weight-decay 0.01 \
+    --decoder-learned-pos \
+    --encoder-learned-pos \
+    --pred-length-offset \
+    --length-loss-factor 0.1 \
+    --train-step 4 \
+    --dae-ratio 0.5 \
+    --stochastic-approx \
+    --apply-bert-init \
+    --log-format 'simple' --log-interval 100 \
+    --fixed-validation-seed 7 \
+    --max-tokens 8000 \
+    --save-interval-updates 10000 \
+    --max-update 300000
+```
+
+### Insertion Transformer (InsT, Stern et al., 2019)
+Note that we need to specify the "slot-loss" (uniform or balanced tree) described in the original paper. Here we use `--label-tau` to control the temperature.
+
+```bash
+fairseq-train \
+    data-bin/wmt14_en_de_distill \
+    --save-dir checkpoints \
+    --ddp-backend=no_c10d \
+    --task translation_lev \
+    --criterion nat_loss \
+    --arch insertion_transformer \
+    --noise random_delete \
+    --share-all-embeddings \
+    --optimizer adam --adam-betas '(0.9,0.98)' \
+    --lr 0.0005 --lr-scheduler inverse_sqrt \
+    --min-lr '1e-09' --warmup-updates 10000 \
+    --warmup-init-lr '1e-07' --label-smoothing 0.1 \
+    --dropout 0.3 --weight-decay 0.01 \
+    --decoder-learned-pos \
+    --encoder-learned-pos \
+    --pred-length-offset \
+    --length-loss-factor 0.1 \
+    --apply-bert-init \
+    --log-format 'simple' --log-interval 100 \
+    --fixed-validation-seed 7 \
+    --max-tokens 8000 \
+    --save-interval-updates 10000 \
+    --max-update 300000
+```
+
+
+### Mask Predict (CMLM, Ghazvininejad et al., 2019)
+```bash
+fairseq-train \
+    data-bin/wmt14_en_de_distill \
+    --save-dir checkpoints \
+    --ddp-backend=no_c10d \
+    --task translation_lev \
+    --criterion nat_loss \
+    --arch cmlm_transformer \
+    --noise random_mask \
+    --share-all-embeddings \
+    --optimizer adam --adam-betas '(0.9,0.98)' \
+    --lr 0.0005 --lr-scheduler inverse_sqrt \
+    --min-lr '1e-09' --warmup-updates 10000 \
+    --warmup-init-lr '1e-07' --label-smoothing 0.1 \
+    --dropout 0.3 --weight-decay 0.01 \
+    --decoder-learned-pos \
+    --encoder-learned-pos \
+    --apply-bert-init \
+    --log-format 'simple' --log-interval 100 \
+    --fixed-validation-seed 7 \
+    --max-tokens 8000 \
+    --save-interval-updates 10000 \
+    --max-update 300000
+```
+
+
+
+
+### Levenshtein Transformer (LevT, Gu et al., 2019)
+```bash
+fairseq-train \
+    data-bin/wmt14_en_de_distill \
+    --save-dir checkpoints \
+    --ddp-backend=no_c10d \
+    --task translation_lev \
+    --criterion nat_loss \
+    --arch levenshtein_transformer \
+    --noise random_delete \
+    --share-all-embeddings \
+    --optimizer adam --adam-betas '(0.9,0.98)' \
+    --lr 0.0005 --lr-scheduler inverse_sqrt \
+    --min-lr '1e-09' --warmup-updates 10000 \
+    --warmup-init-lr '1e-07' --label-smoothing 0.1 \
+    --dropout 0.3 --weight-decay 0.01 \
+    --decoder-learned-pos \
+    --encoder-learned-pos \
+    --apply-bert-init \
+    --log-format 'simple' --log-interval 100 \
+    --fixed-validation-seed 7 \
+    --max-tokens 8000 \
+    --save-interval-updates 10000 \
+    --max-update 300000
+```

fairseq/clib/libnat/edit_dist.cpp

@@ -0,0 +1,222 @@
+/**
+ * Copyright 2017-present, Facebook, Inc.
+ * All rights reserved.
+ *
+ * This source code is licensed under the license found in the
+ * LICENSE file in the root directory of this source tree.
+ */
+
+#include <torch/torch.h> // @manual=//caffe2:torch_extension
+#include <vector>
+
+using namespace ::std;
+
+vector<vector<uint32_t>> edit_distance2_with_dp(
+    vector<uint32_t>& x,
+    vector<uint32_t>& y) {
+  uint32_t lx = x.size();
+  uint32_t ly = y.size();
+  vector<vector<uint32_t>> d(lx + 1, vector<uint32_t>(ly + 1));
+  for (uint32_t i = 0; i < lx + 1; i++) {
+    d[i][0] = i;
+  }
+  for (uint32_t j = 0; j < ly + 1; j++) {
+    d[0][j] = j;
+  }
+  for (uint32_t i = 1; i < lx + 1; i++) {
+    for (uint32_t j = 1; j < ly + 1; j++) {
+      d[i][j] =
+          min(min(d[i - 1][j], d[i][j - 1]) + 1,
+              d[i - 1][j - 1] + 2 * (x.at(i - 1) == y.at(j - 1) ? 0 : 1));
+    }
+  }
+  return d;
+}
+
+vector<vector<uint32_t>> edit_distance2_backtracking(
+    vector<vector<uint32_t>>& d,
+    vector<uint32_t>& x,
+    vector<uint32_t>& y,
+    uint32_t terminal_symbol) {
+  vector<uint32_t> seq;
+  vector<vector<uint32_t>> edit_seqs(x.size() + 2, vector<uint32_t>());
+  /*
+  edit_seqs:
+  0~x.size() cell is the insertion sequences
+  last cell is the delete sequence
+  */
+
+  if (x.size() == 0) {
+    edit_seqs.at(0) = y;
+    return edit_seqs;
+  }
+
+  uint32_t i = d.size() - 1;
+  uint32_t j = d.at(0).size() - 1;
+
+  while ((i >= 0) && (j >= 0)) {
+    if ((i == 0) && (j == 0)) {
+      break;
+    }
+
+    if ((j > 0) && (d.at(i).at(j - 1) < d.at(i).at(j))) {
+      seq.push_back(1); // insert
+      seq.push_back(y.at(j - 1));
+      j--;
+    } else if ((i > 0) && (d.at(i - 1).at(j) < d.at(i).at(j))) {
+      seq.push_back(2); // delete
+      seq.push_back(x.at(i - 1));
+      i--;
+    } else {
+      seq.push_back(3); // keep
+      seq.push_back(x.at(i - 1));
+      i--;
+      j--;
+    }
+  }
+
+  uint32_t prev_op, op, s, word;
+  prev_op = 0, s = 0;
+  for (uint32_t k = 0; k < seq.size() / 2; k++) {
+    op = seq.at(seq.size() - 2 * k - 2);
+    word = seq.at(seq.size() - 2 * k - 1);
+    if (prev_op != 1) {
+      s++;
+    }
+    if (op == 1) // insert
+    {
+      edit_seqs.at(s - 1).push_back(word);
+    } else if (op == 2) // delete
+    {
+      edit_seqs.at(x.size() + 1).push_back(1);
+    } else {
+      edit_seqs.at(x.size() + 1).push_back(0);
+    }
+
+    prev_op = op;
+  }
+
+  for (uint32_t k = 0; k < edit_seqs.size(); k++) {
+    if (edit_seqs[k].size() == 0) {
+      edit_seqs[k].push_back(terminal_symbol);
+    }
+  }
+  return edit_seqs;
+}
+
+vector<vector<uint32_t>> edit_distance2_backtracking_with_delete(
+    vector<vector<uint32_t>>& d,
+    vector<uint32_t>& x,
+    vector<uint32_t>& y,
+    uint32_t terminal_symbol,
+    uint32_t deletion_symbol) {
+  vector<uint32_t> seq;
+  vector<vector<uint32_t>> edit_seqs(x.size() + 1, vector<uint32_t>());
+  /*
+  edit_seqs:
+  0~x.size() cell is the insertion sequences
+  last cell is the delete sequence
+  */
+
+  if (x.size() == 0) {
+    edit_seqs.at(0) = y;
+    return edit_seqs;
+  }
+
+  uint32_t i = d.size() - 1;
+  uint32_t j = d.at(0).size() - 1;
+
+  while ((i >= 0) && (j >= 0)) {
+    if ((i == 0) && (j == 0)) {
+      break;
+    }
+
+    if ((j > 0) && (d.at(i).at(j - 1) < d.at(i).at(j))) {
+      seq.push_back(1); // insert
+      seq.push_back(y.at(j - 1));
+      j--;
+    } else if ((i > 0) && (d.at(i - 1).at(j) < d.at(i).at(j))) {
+      seq.push_back(2); // delete
+      seq.push_back(x.at(i - 1));
+      i--;
+    } else {
+      seq.push_back(3); // keep
+      seq.push_back(x.at(i - 1));
+      i--;
+      j--;
+    }
+  }
+
+  uint32_t prev_op, op, s, word;
+  prev_op = 0, s = 0;
+  for (uint32_t k = 0; k < seq.size() / 2; k++) {
+    op = seq.at(seq.size() - 2 * k - 2);
+    word = seq.at(seq.size() - 2 * k - 1);
+    if (prev_op != 1) {
+      s++;
+    }
+    if (op == 1) // insert
+    {
+      edit_seqs.at(s - 1).push_back(word);
+    } else if (op == 2) // delete
+    {
+      edit_seqs.at(s - 1).push_back(deletion_symbol);
+    }
+
+    prev_op = op;
+  }
+
+  for (uint32_t k = 0; k < edit_seqs.size(); k++) {
+    if (edit_seqs.at(k).size() == 0) {
+      edit_seqs.at(k).push_back(terminal_symbol);
+    }
+  }
+  return edit_seqs;
+}
+
+vector<uint32_t> compute_ed2(
+    vector<vector<uint32_t>>& xs,
+    vector<vector<uint32_t>>& ys) {
+  vector<uint32_t> distances(xs.size());
+  for (uint32_t i = 0; i < xs.size(); i++) {
+    vector<vector<uint32_t>> d = edit_distance2_with_dp(xs.at(i), ys.at(i));
+    distances.at(i) = d.at(xs.at(i).size()).at(ys.at(i).size());
+  }
+  return distances;
+}
+
+vector<vector<vector<uint32_t>>> suggested_ed2_path(
+    vector<vector<uint32_t>>& xs,
+    vector<vector<uint32_t>>& ys,
+    uint32_t terminal_symbol) {
+  vector<vector<vector<uint32_t>>> seq(xs.size());
+  for (uint32_t i = 0; i < xs.size(); i++) {
+    vector<vector<uint32_t>> d = edit_distance2_with_dp(xs.at(i), ys.at(i));
+    seq.at(i) =
+        edit_distance2_backtracking(d, xs.at(i), ys.at(i), terminal_symbol);
+  }
+  return seq;
+}
+
+vector<vector<vector<uint32_t>>> suggested_ed2_path_with_delete(
+    vector<vector<uint32_t>>& xs,
+    vector<vector<uint32_t>>& ys,
+    uint32_t terminal_symbol,
+    uint32_t deletion_symbol) {
+  vector<vector<vector<uint32_t>>> seq(xs.size());
+  for (uint32_t i = 0; i < xs.size(); i++) {
+    vector<vector<uint32_t>> d = edit_distance2_with_dp(xs.at(i), ys.at(i));
+    seq.at(i) = edit_distance2_backtracking_with_delete(
+        d, xs.at(i), ys.at(i), terminal_symbol, deletion_symbol);
+  }
+  return seq;
+}
+
+PYBIND11_MODULE(libnat, m) {
+  m.def("compute_ed2", &compute_ed2, "compute_ed2");
+  m.def("suggested_ed2_path", &suggested_ed2_path, "suggested_ed2_path");
+  m.def(
+      "suggested_ed2_path_with_delete",
+      &suggested_ed2_path_with_delete,
+      "suggested_ed2_path_with_delete");
+}

fairseq/criterions/nat_loss.py

@@ -0,0 +1,190 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+
+import torch.nn.functional as F
+from fairseq import utils
+from torch import Tensor
+
+from . import FairseqCriterion, register_criterion
+
+
+@register_criterion("nat_loss")
+class LabelSmoothedDualImitationCriterion(FairseqCriterion):
+    @staticmethod
+    def add_args(parser):
+        """Add criterion-specific arguments to the parser."""
+        # fmt: off
+        parser.add_argument(
+            '--label-smoothing',
+            default=0.,
+            type=float,
+            metavar='D',
+            help='epsilon for label smoothing, 0 means no label smoothing')
+        # fmt: on
+
+    def _compute_loss(
+        self, outputs, targets, masks=None, label_smoothing=0.0, name="loss", factor=1.0
+    ):
+        """
+            outputs: batch x len x d_model
+            targets: batch x len
+            masks:   batch x len
+
+            policy_logprob: if there is some policy
+                depends on the likelihood score as rewards.
+        """
+
+        def mean_ds(x: Tensor, dim=None) -> Tensor:
+            return (
+                x.float().mean().type_as(x)
+                if dim is None
+                else x.float().mean(dim).type_as(x)
+            )
+
+        if masks is not None:
+            outputs, targets = outputs[masks], targets[masks]
+
+        logits = F.log_softmax(outputs, dim=-1)
+        if targets.dim() == 1:
+            losses = F.nll_loss(logits, targets, reduction="none")
+
+        else:  # soft-labels
+            losses = F.kl_div(logits, targets, reduction="none")
+            losses = losses.float().sum(-1).type_as(losses)
+
+        nll_loss = mean_ds(losses)
+        if label_smoothing > 0:
+            loss = nll_loss * (1 - label_smoothing) - mean_ds(logits) * label_smoothing
+        else:
+            loss = nll_loss
+
+        loss = loss * factor
+        return {"name": name, "loss": loss, "nll_loss": nll_loss, "factor": factor}
+
+    def _custom_loss(self, loss, name="loss"):
+        return {"name": name, "loss": loss, "factor": 1}
+
+    def forward(self, model, sample, reduce=True):
+        """Compute the loss for the given sample.
+        Returns a tuple with three elements:
+        1) the loss
+        2) the sample size, which is used as the denominator for the gradient
+        3) logging outputs to display while training
+        """
+        nsentences, ntokens = sample["nsentences"], sample["ntokens"]
+
+        # B x T
+        src_tokens, src_lengths = (
+            sample["net_input"]["src_tokens"],
+            sample["net_input"]["src_lengths"],
+        )
+        tgt_tokens, prev_output_tokens = sample["target"], sample["prev_target"]
+
+        outputs = model(src_tokens, src_lengths, prev_output_tokens, tgt_tokens)
+        losses = []
+        if "mask_ins_out" in outputs:
+            mask_ins_losses = self._compute_loss(
+                outputs["mask_ins_out"],
+                outputs["mask_ins_tgt"],
+                outputs["mask_ins_mask"],
+                name="m_ins-loss",
+                factor=1 if "mask_ins_w" not in outputs else outputs["mask_ins_w"],
+            )
+            losses += [mask_ins_losses]
+
+        if "word_ins_out" in outputs:
+            word_ins_losses = self._compute_loss(
+                outputs["word_ins_out"],
+                outputs["word_ins_tgt"],
+                outputs["word_ins_mask"],
+                self.args.label_smoothing,
+                name="w_ins-loss",
+                factor=1 if "word_ins_w" not in outputs else outputs["word_ins_w"],
+            )
+
+            losses += [word_ins_losses]
+            nll_loss = word_ins_losses["nll_loss"]
+
+        if "word_del_out" in outputs:
+            word_del_losses = self._compute_loss(
+                outputs["word_del_out"],
+                outputs["word_del_tgt"],
+                outputs["word_del_mask"],
+                0.01,
+                name="w_del-loss",
+                factor=1 if "word_del_w" not in outputs else outputs["word_del_w"],
+            )
+
+            losses += [word_del_losses]
+
+        if "length_out" in outputs:
+            length_losses = self._compute_loss(
+                outputs["length_out"],
+                outputs["length_tgt"],
+                name="len-loss",
+                factor=1 if "length_w" not in outputs else outputs["length_w"],
+            )
+
+            losses += [length_losses]
+
+        for w in outputs:
+            if "-loss" in w:
+                losses += [self._custom_loss(outputs[w], w)]
+
+        loss = sum(l["loss"] for l in losses)
+
+        # NOTE: as we are summing up per token mlm loss and per sentence nsp loss
+        # we don't need to use sample_size as denominator for the gradient
+        # here sample_size is just used for logging
+        sample_size = 1
+        logging_output = {
+            "loss": utils.item(loss.data) if reduce else loss.data,
+            "nll_loss": utils.item(nll_loss.data) if reduce else nll_loss.data,
+            "ntokens": ntokens,
+            "nsentences": nsentences,
+            "sample_size": sample_size,
+        }
+
+        for l in losses:
+            logging_output[l["name"]] = (
+                utils.item(l["loss"].data / l["factor"])
+                if reduce
+                else l[["loss"]].data / l["factor"]
+            )
+
+        return loss, sample_size, logging_output
+
+    @staticmethod
+    def aggregate_logging_outputs(logging_outputs):
+        """Aggregate logging outputs from data parallel training."""
+        ntokens = sum(log.get("ntokens", 0) for log in logging_outputs)
+        nsentences = sum(log.get("nsentences", 0) for log in logging_outputs)
+        sample_size = sum(log.get("sample_size", 0) for log in logging_outputs)
+        loss = sum(log.get("loss", 0) for log in logging_outputs)
+        nll_loss = sum(log.get("nll_loss", 0) for log in logging_outputs)
+
+        results = {
+            "loss": loss / sample_size / math.log(2) if sample_size > 0 else 0.0,
+            "nll_loss": nll_loss / sample_size / math.log(2)
+            if sample_size > 0
+            else 0.0,
+            "ntokens": ntokens,
+            "nsentences": nsentences,
+            "sample_size": sample_size,
+        }
+
+        for key in logging_outputs[0]:
+            if key[-5:] == "-loss":
+                results[key[:-5]] = (
+                    sum(log.get(key, 0) for log in logging_outputs)
+                    / sample_size
+                    / math.log(2)
+                    if sample_size > 0
+                    else 0.0
+                )
+
+        return results

fairseq/data/dictionary.py

@@ -74,7 +74,10 @@ class Dictionary(object):
             else:
                 return self[i]
 
-        sent = ' '.join(token_string(i) for i in tensor if i != self.eos())
+        if hasattr(self, 'bos_index'):
+            sent = ' '.join(token_string(i) for i in tensor if (i != self.eos()) and (i != self.bos()))
+        else:
+            sent = ' '.join(token_string(i) for i in tensor if i != self.eos())
         return data_utils.process_bpe_symbol(sent, bpe_symbol)
 
     def unk_string(self, escape=False):

fairseq/iterative_refinement_generator.py

@@ -0,0 +1,154 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from fairseq.models.model_utils import skip_tensors as _skip
+
+
+class IterativeRefinementGenerator(object):
+    def __init__(self,
+                 tgt_dict,
+                 eos_penalty=0.,
+                 max_iter=10,
+                 max_ratio=2,
+                 decoding_format=None,
+                 retain_dropout=False,
+                 adaptive=True):
+        """
+        Generates translations based on iterative refinement.
+
+        Args:
+            tgt_dict: target dictionary
+            eos_penalty: if > 0.0, it penalized early-stopping in decoding
+            max_iter: maximum number of refinement iterations
+            max_ratio: generate sequences of maximum length ax, where x is the source length
+            decoding_format: decoding mode in {'unigram', 'ensemble', 'vote', 'dp', 'bs'}
+            retain_dropout: retaining dropout in the inference
+            adaptive: decoding with early stop
+        """
+        self.bos = tgt_dict.bos()
+        self.pad = tgt_dict.pad()
+        self.unk = tgt_dict.unk()
+        self.eos = tgt_dict.eos()
+        self.vocab_size = len(tgt_dict)
+        self.eos_penalty = eos_penalty
+        self.max_iter = max_iter
+        self.max_ratio = max_ratio
+        self.decoding_format = decoding_format
+        self.retain_dropout = retain_dropout
+        self.adaptive = adaptive
+
+    @torch.no_grad()
+    def generate(self, models, sample, prefix_tokens=None):
+
+        # TODO: model ensemble
+        assert len(models) == 1, 'only support single model'
+        model = models[0]
+        if not self.retain_dropout:
+            model.eval()
+
+        # TODO: better encoder inputs?
+        src_tokens = sample['net_input']['src_tokens']
+        src_lengths = sample['net_input']['src_lengths']
+        bsz, src_len = src_tokens.size()
+        sent_idxs = torch.arange(bsz, device=src_tokens.device)
+
+        # encoding
+        encoder_out = model.forward_encoder([src_tokens, src_lengths])
+
+        # initialize buffers (very model specific, with length prediction or not)
+        prev_decoder_out = model.initialize_output_tokens(
+            encoder_out, src_tokens)
+        prev_out_tokens = prev_decoder_out['output_tokens'].clone()
+
+        finalized = [[] for _ in range(bsz)]
+
+        def is_a_loop(x, y, s, a):
+            b, l_x, l_y = x.size(0), x.size(1), y.size(1)
+            if l_x > l_y:
+                y = torch.cat([y, x.new_zeros(b, l_x - l_y).fill_(self.pad)], 1)
+                s = torch.cat([s, s.new_zeros(b, l_x - l_y)], 1)
+                if a is not None:
+                    a = torch.cat([a, a.new_zeros(b, l_x - l_y, a.size(2))], 1)
+            elif l_x < l_y:
+                x = torch.cat([x, y.new_zeros(b, l_y - l_x).fill_(self.pad)], 1)
+            return (x == y).all(1), y, s, a
+
+        def finalized_hypos(step, prev_out_token, prev_out_score, prev_out_attn):
+            cutoff = prev_out_token.ne(self.pad)
+            tokens = prev_out_token[cutoff]
+            scores = prev_out_score[cutoff]
+            if prev_out_attn is None:
+                hypo_attn, alignment = None, None
+            else:
+                hypo_attn = prev_out_attn[cutoff]
+                alignment = hypo_attn.max(dim=1)[1]
+            return {
+                'steps': step,
+                'tokens': tokens,
+                'positional_scores': scores,
+                'score': scores.mean(),
+                'hypo_attn': hypo_attn,
+                'alignment': alignment,
+            }
+
+        for step in range(self.max_iter + 1):
+
+            decoder_options = {
+                'eos_penalty': self.eos_penalty,
+                'max_ratio': self.max_ratio,
+                'decoding_format': self.decoding_format
+            }
+            prev_decoder_out['step'] = step
+            prev_decoder_out['max_step'] = self.max_iter + 1
+
+            decoder_out = model.forward_decoder(
+                prev_decoder_out, encoder_out, **decoder_options
+            )
+
+            if self.adaptive:
+                # terminate if there is a loop
+                terminated, out_tokens, out_scores, out_attn = is_a_loop(
+                    prev_out_tokens, decoder_out['output_tokens'],
+                    decoder_out['output_scores'], decoder_out['attn'])
+                decoder_out['output_tokens'] = out_tokens
+                decoder_out['output_scores'] = out_scores
+                decoder_out['attn'] = out_attn
+
+            else:
+                terminated = decoder_out['output_tokens'].new_zeros(
+                    decoder_out['output_tokens'].size(0)).bool()
+
+            if step == self.max_iter:  # reach last iteration, terminate
+                terminated.fill_(1)
+
+            # collect finalized sentences
+            finalized_idxs = sent_idxs[terminated]
+            finalized_tokens = decoder_out['output_tokens'][terminated]
+            finalized_scores = decoder_out['output_scores'][terminated]
+            finalized_attn = None if decoder_out['attn'] is None else decoder_out['attn'][terminated]
+
+            for i in range(finalized_idxs.size(0)):
+                finalized[finalized_idxs[i]] = [
+                    finalized_hypos(
+                        step,
+                        finalized_tokens[i],
+                        finalized_scores[i],
+                        None if finalized_attn is None else finalized_attn[i]
+                    )
+                ]
+            # check if all terminated
+            if terminated.sum() == terminated.size(0):
+                break
+
+            # for next step
+            prev_decoder_out = _skip(decoder_out, ~terminated)
+            encoder_out = _skip(encoder_out, ~terminated)
+            sent_idxs = _skip(sent_idxs, ~terminated)
+
+            prev_out_tokens = prev_decoder_out['output_tokens'].clone()
+
+        return finalized

fairseq/models/cmlm_transformer.py

@@ -0,0 +1,136 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""
+This file implements:
+Ghazvininejad, Marjan, et al.
+"Constant-time machine translation with conditional masked language models."
+arXiv preprint arXiv:1904.09324 (2019).
+"""
+
+import torch
+from fairseq.models import register_model, register_model_architecture
+from fairseq.models.nonautoregressive_transformer import NATransformerModel
+
+
+def _skeptical_unmasking(output_scores, output_masks, p):
+    sorted_index = output_scores.sort(-1)[1]
+    boundary_len = (
+        (output_masks.sum(1, keepdim=True).type_as(output_scores) - 2) * p
+    ).long()
+    skeptical_mask = (
+        torch.arange(output_masks.size(1), device=output_masks.device)[None, :]
+        < boundary_len
+    )
+    return skeptical_mask.scatter(1, sorted_index, skeptical_mask)
+
+
+@register_model("cmlm_transformer")
+class CMLMNATransformerModel(NATransformerModel):
+    @staticmethod
+    def add_args(parser):
+        NATransformerModel.add_args(parser)
+
+    def forward(
+        self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs
+    ):
+        assert not self.decoder.src_embedding_copy, "do not support embedding copy."
+
+        encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs)
+        length_out, length_tgt = self.decoder.forward_length_prediction(
+            encoder_out, tgt_tokens
+        )
+
+        word_ins_out, word_ins_tgt, _ = self.decoder(
+            prev_output_tokens, encoder_out=encoder_out, tgt_tokens=tgt_tokens
+        )
+        word_ins_mask = prev_output_tokens.eq(self.unk)
+        return {
+            "word_ins_out": word_ins_out,
+            "word_ins_tgt": word_ins_tgt,
+            "word_ins_mask": word_ins_mask,
+            "length_out": length_out,
+            "length_tgt": length_tgt,
+            "length_w": self.decoder.length_loss_factor,
+        }
+
+    def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs):
+
+        step = decoder_out["step"]
+        max_step = decoder_out["max_step"]
+
+        output_tokens = decoder_out["output_tokens"]
+        output_scores = decoder_out["output_scores"]
+
+        # execute the decoder
+        output_masks = output_tokens.eq(self.unk)
+        _scores, _tokens = self.decoder(
+            output_tokens, encoder_out=encoder_out, decoding_format=decoding_format
+        )
+        output_tokens.masked_scatter_(output_masks, _tokens[output_masks])
+        output_scores.masked_scatter_(output_masks, _scores[output_masks])
+
+        # skeptical decoding (depend on the maximum decoding steps.)
+        if (step + 1) < max_step:
+            skeptical_mask = _skeptical_unmasking(
+                output_scores, output_tokens.ne(self.pad), 1 - (step + 1) / max_step
+            )
+
+            output_tokens.masked_fill_(skeptical_mask, self.unk)
+            output_scores.masked_fill_(skeptical_mask, 0.0)
+
+        return {"output_tokens": output_tokens, "output_scores": output_scores}
+
+
+@register_model_architecture("cmlm_transformer", "cmlm_transformer")
+def base_architecture(args):
+    args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
+    args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+    args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048)
+    args.encoder_layers = getattr(args, "encoder_layers", 6)
+    args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8)
+    args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+    args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False)
+    args.decoder_embed_path = getattr(args, "decoder_embed_path", None)
+    args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim)
+    args.decoder_ffn_embed_dim = getattr(
+        args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim
+    )
+    args.decoder_layers = getattr(args, "decoder_layers", 6)
+    args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8)
+    args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False)
+    args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False)
+    args.attention_dropout = getattr(args, "attention_dropout", 0.0)
+    args.activation_dropout = getattr(args, "activation_dropout", 0.0)
+    args.activation_fn = getattr(args, "activation_fn", "relu")
+    args.dropout = getattr(args, "dropout", 0.1)
+    args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None)
+    args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0)
+    args.share_decoder_input_output_embed = getattr(
+        args, "share_decoder_input_output_embed", False
+    )
+    args.share_all_embeddings = getattr(args, "share_all_embeddings", True)
+    args.no_token_positional_embeddings = getattr(
+        args, "no_token_positional_embeddings", False
+    )
+    args.adaptive_input = getattr(args, "adaptive_input", False)
+    args.apply_bert_init = getattr(args, "apply_bert_init", False)
+
+    args.decoder_output_dim = getattr(
+        args, "decoder_output_dim", args.decoder_embed_dim
+    )
+    args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim)
+
+    # --- special arguments ---
+    args.sg_length_pred = getattr(args, "sg_length_pred", False)
+    args.pred_length_offset = getattr(args, "pred_length_offset", False)
+    args.length_loss_factor = getattr(args, "length_loss_factor", 0.1)
+    args.ngram_predictor = getattr(args, "ngram_predictor", 1)
+    args.src_embedding_copy = getattr(args, "src_embedding_copy", False)
+
+
+@register_model_architecture("cmlm_transformer", "cmlm_transformer_wmt_en_de")
+def iter_nat_wmt_en_de(args):
+    base_architecture(args)

fairseq/models/insertion_transformer.py

@@ -0,0 +1,259 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import numpy as np
+import torch
+import torch.nn.functional as F
+from fairseq import libnat
+from fairseq.models import register_model, register_model_architecture
+from fairseq.models.levenshtein_transformer import (
+    LevenshteinTransformerDecoder,
+    LevenshteinTransformerModel,
+)
+from fairseq.models.transformer import Linear, TransformerModel
+from fairseq.modules.transformer_sentence_encoder import init_bert_params
+
+
+class NegativeDistanceScore(object):
+    def __init__(self):
+
+        # pre-compute some values
+        self.scores = {}
+
+        self.scores[0.5] = self.compute_score_full(50, 0.5)
+        self.scores[1.0] = self.compute_score_full(50, 1.0)
+        self.scores[2.0] = self.compute_score_full(50, 2.0)
+
+    def __call__(self, i, L, tau):
+        if (tau is None) or (tau > 1000):
+            return 1 / L
+
+        if tau in self.scores:
+            if L < self.scores[tau].shape[0]:
+                return self.scores[tau][L - 1, i]
+        return self.compute_score(L, tau)[i]
+
+    def compute_score(self, L, tau):
+        s = np.array([-abs(L / 2 - i) / tau for i in range(L)])
+        s = np.exp(s - s.max())
+        return s / s.sum()
+
+    def compute_score_full(self, L, tau):
+        s = -abs(np.arange(0, L - 1)[:, None] / 2 - np.arange(L)[None, :]) / tau
+        s = np.tril(s, 0) + np.triu(s - float("inf"), 1)
+        s = np.exp(s - s.max(1, keepdims=True))
+        return s / s.sum(1, keepdims=True)
+
+
+neg_scorer = NegativeDistanceScore()
+
+
+def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx, vocab_size, tau=None):
+    B = in_tokens.size(0)
+    T = in_tokens.size(1)
+    V = vocab_size
+
+    with torch.cuda.device_of(in_tokens):
+        in_tokens_list = [
+            [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist())
+        ]
+        out_tokens_list = [
+            [t for t in s if t != padding_idx]
+            for i, s in enumerate(out_tokens.tolist())
+        ]
+
+    full_labels = libnat.suggested_ed2_path(
+        in_tokens_list, out_tokens_list, padding_idx
+    )
+    insert_labels = [a[:-1] for a in full_labels]
+
+    # numericalize1
+    insert_label_tensors = in_tokens.new_zeros(B * (T - 1) * V).float()
+    insert_index, insert_labels = zip(
+        *[
+            (w + (j + i * (T - 1)) * V, neg_scorer(k, len(label), tau))
+            for i, labels in enumerate(insert_labels)
+            for j, label in enumerate(labels[1:-1])
+            for k, w in enumerate(label)
+        ]
+    )  # HACK 1:-1
+    insert_index, insert_labels = [
+        torch.tensor(list(a), device=in_tokens.device)
+        for a in [insert_index, insert_labels]
+    ]
+    insert_label_tensors.scatter_(0, insert_index.long(), insert_labels)
+    insert_label_tensors = insert_label_tensors.view(B, T - 1, V)
+
+    return insert_label_tensors
+
+
+def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, padding_idx):
+
+    padding_masks = in_tokens[:, 1:].eq(padding_idx)
+    word_ins_scores.masked_fill_(padding_masks, 0.0)
+    word_ins_pred.masked_fill_(padding_masks, padding_idx)
+
+    in_coords = torch.arange(in_tokens.size(1), device=in_tokens.device)
+    in_coords = in_coords.unsqueeze(0).repeat(in_tokens.size(0), 1).type_as(in_scores)
+
+    # shift all padding predictions to infinite
+    out_coords = (in_coords[:, 1:] - 0.5).masked_fill(
+        word_ins_pred.eq(padding_idx), float("inf")
+    )
+    out_coords = torch.cat([in_coords, out_coords], 1).sort(-1)[1]
+    out_tokens = torch.cat([in_tokens, word_ins_pred], 1).gather(1, out_coords)
+    out_scores = torch.cat([in_scores, word_ins_scores], 1).gather(1, out_coords)
+    return out_tokens, out_scores
+
+
+@register_model("insertion_transformer")
+class InsertionTransformerModel(LevenshteinTransformerModel):
+    def __init__(self, encoder, decoder):
+        super().__init__(encoder, decoder)
+
+    @staticmethod
+    def add_args(parser):
+        TransformerModel.add_args(parser)
+        parser.add_argument(
+            "--apply-bert-init",
+            action="store_true",
+            help="use custom param initialization for BERT",
+        )
+        parser.add_argument("--label-tau", default=None, type=float)
+
+    @classmethod
+    def build_decoder(cls, args, tgt_dict, embed_tokens):
+        decoder = InsertionTransformerDecoder(args, tgt_dict, embed_tokens)
+        if getattr(args, "apply_bert_init", False):
+            decoder.apply(init_bert_params)
+        return decoder
+
+    def forward(
+        self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs
+    ):
+
+        assert tgt_tokens is not None, "forward function only supports training."
+
+        # encoding
+        encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs)
+
+        # generate training labels for insertion
+        word_ins_out = self.decoder.forward_word_ins(
+            prev_output_tokens, encoder_out=encoder_out
+        )
+        word_ins_tgt = _get_ins_targets(
+            prev_output_tokens,
+            tgt_tokens,
+            self.pad,
+            self.unk,
+            len(self.tgt_dict),
+            tau=self.decoder.label_tau,
+        ).type_as(word_ins_out)
+        word_ins_masks = prev_output_tokens[:, 1:].ne(self.pad)
+
+        return {
+            "word_ins_out": word_ins_out,
+            "word_ins_tgt": word_ins_tgt,
+            "word_ins_mask": word_ins_masks,
+        }
+
+    def forward_decoder(
+        self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs
+    ):
+
+        output_tokens = decoder_out["output_tokens"]
+        output_scores = decoder_out["output_scores"]
+        # TODO: decoding for InsertionTransformer
+        word_ins_out = self.decoder.forward_word_ins(
+            output_tokens, encoder_out=encoder_out
+        )
+        word_ins_score = F.log_softmax(word_ins_out, 2)
+        if eos_penalty > 0.0:
+            word_ins_score[:, :, self.pad] -= eos_penalty
+        word_ins_score, word_ins_pred = word_ins_score.max(-1)
+        output_tokens, output_scores = _apply_ins_words(
+            output_tokens, output_scores, word_ins_pred, word_ins_score, self.pad
+        )
+
+        # delete some unnecessary paddings
+        cut_off = output_tokens.ne(self.pad).sum(1).max()
+        output_tokens = output_tokens[:, :cut_off]
+        output_scores = output_scores[:, :cut_off]
+        return {"output_tokens": output_tokens, "output_scores": output_scores}
+
+
+class InsertionTransformerDecoder(LevenshteinTransformerDecoder):
+    def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False):
+        # use the TransformerDecoder's __init__
+        super(LevenshteinTransformerDecoder, self).__init__(
+            args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn
+        )
+
+        self.dictionary = dictionary
+        self.bos = dictionary.bos()
+        self.unk = dictionary.unk()
+        self.eos = dictionary.eos()
+        self.pool_out = Linear(self.output_embed_dim * 2, self.output_embed_dim)
+
+        self.label_tau = getattr(args, "label_tau", None)
+
+    def forward_word_ins(self, prev_output_tokens, encoder_out=None):
+        features, _ = self.extract_features(prev_output_tokens, encoder_out=encoder_out)
+        features = self.pool_out(
+            torch.cat([features[:, :-1, :], features[:, 1:, :]], 2)
+        )
+        return self.output_layer(features)
+
+    def forward_mask_ins(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def forward_word_del(self, *args, **kwargs):
+        raise NotImplementedError
+
+    def forward_word_del_mask_ins(self, *args, **kwargs):
+        raise NotImplementedError
+
+
+@register_model_architecture("insertion_transformer", "insertion_transformer")
+def base_architecture(args):
+    args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
+    args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+    args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048)
+    args.encoder_layers = getattr(args, "encoder_layers", 6)
+    args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8)
+    args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+    args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False)
+    args.decoder_embed_path = getattr(args, "decoder_embed_path", None)
+    args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim)
+    args.decoder_ffn_embed_dim = getattr(
+        args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim
+    )
+    args.decoder_layers = getattr(args, "decoder_layers", 6)
+    args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8)
+    args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False)
+    args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False)
+    args.attention_dropout = getattr(args, "attention_dropout", 0.0)
+    args.activation_dropout = getattr(args, "activation_dropout", 0.0)
+    args.activation_fn = getattr(args, "activation_fn", "relu")
+    args.dropout = getattr(args, "dropout", 0.1)
+    args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None)
+    args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0)
+    args.share_decoder_input_output_embed = getattr(
+        args, "share_decoder_input_output_embed", False
+    )
+    args.share_all_embeddings = getattr(args, "share_all_embeddings", False)
+    args.no_token_positional_embeddings = getattr(
+        args, "no_token_positional_embeddings", False
+    )
+    args.adaptive_input = getattr(args, "adaptive_input", False)
+    args.apply_bert_init = getattr(args, "apply_bert_init", False)
+
+    args.decoder_output_dim = getattr(
+        args, "decoder_output_dim", args.decoder_embed_dim
+    )
+    args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim)
+
+    # special for insertion transformer
+    args.label_tau = getattr(args, "label_tau", None)

fairseq/models/iterative_nonautoregressive_transformer.py

@@ -0,0 +1,196 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+from fairseq.models import register_model, register_model_architecture
+from fairseq.models.nonautoregressive_transformer import NATransformerModel
+
+
+def _sequential_poisoning(s, V, beta=0.33, bos=2, eos=3, pad=1):
+    # s: input batch
+    # V: vocabulary size
+    rand_words = torch.randint(low=4, high=V, size=s.size(), device=s.device)
+    choices = torch.rand(size=s.size(), device=s.device)
+    choices.masked_fill_((s == pad) | (s == bos) | (s == eos), 1)
+
+    replace = choices < beta / 3
+    repeat = (choices >= beta / 3) & (choices < beta * 2 / 3)
+    swap = (choices >= beta * 2 / 3) & (choices < beta)
+    safe = choices >= beta
+
+    for i in range(s.size(1) - 1):
+        rand_word = rand_words[:, i]
+        next_word = s[:, i + 1]
+        self_word = s[:, i]
+
+        replace_i = replace[:, i]
+        swap_i = swap[:, i] & (next_word != 3)
+        repeat_i = repeat[:, i] & (next_word != 3)
+        safe_i = safe[:, i] | ((next_word == 3) & (~replace_i))
+
+        s[:, i] = (
+            self_word * (safe_i | repeat_i).long()
+            + next_word * swap_i.long()
+            + rand_word * replace_i.long()
+        )
+        s[:, i + 1] = (
+            next_word * (safe_i | replace_i).long()
+            + self_word * (swap_i | repeat_i).long()
+        )
+    return s
+
+
+def gumbel_noise(input, TINY=1e-8):
+    return input.new_zeros(*input.size()).uniform_().add_(
+        TINY).log_().neg_().add_(TINY).log_().neg_()
+
+
+@register_model("iterative_nonautoregressive_transformer")
+class IterNATransformerModel(NATransformerModel):
+    @staticmethod
+    def add_args(parser):
+        NATransformerModel.add_args(parser)
+        parser.add_argument("--train-step", type=int,
+                            help="number of refinement iterations during training")
+        parser.add_argument("--dae-ratio", type=float,
+                            help="the probability of switching to the denoising auto-encoder loss")
+        parser.add_argument("--stochastic-approx", action="store_true",
+                            help="sampling from the decoder as the inputs for next iteration")
+
+    @classmethod
+    def build_model(cls, args, task):
+        model = super().build_model(args, task)
+        model.train_step = getattr(args, "train_step", 4)
+        model.dae_ratio = getattr(args, "dae_ratio", 0.5)
+        model.stochastic_approx = getattr(args, "stochastic_approx", False)
+        return model
+
+    def forward(
+        self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs
+    ):
+
+        B, T = prev_output_tokens.size()
+
+        # encoding
+        encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs)
+        length_out, length_tgt = self.decoder.forward_length_prediction(
+            encoder_out, tgt_tokens
+        )
+        word_ins_outs, word_ins_tgts, word_ins_masks = [], [], []
+        for t in range(self.train_step):
+            word_ins_out, word_ins_tgt, word_ins_mask = self.decoder(
+                prev_output_tokens,
+                encoder_out=encoder_out,
+                tgt_tokens=tgt_tokens,
+                step=t,
+            )
+
+            word_ins_outs.append(word_ins_out)
+            word_ins_tgts.append(word_ins_tgt)
+            word_ins_masks.append(word_ins_mask)
+
+            if t < (self.train_step - 1):
+                # prediction for next iteration
+                if self.stochastic_approx:
+                    word_ins_prediction = (
+                        word_ins_out + gumbel_noise(word_ins_out)
+                    ).max(-1)[1]
+                else:
+                    word_ins_prediction = word_ins_out.max(-1)[1]
+
+                prev_output_tokens = prev_output_tokens.masked_scatter(
+                    word_ins_mask, word_ins_prediction[word_ins_mask]
+                )
+
+                if self.dae_ratio > 0:
+                    # we do not perform denoising for the first iteration
+                    corrputed = (
+                        torch.rand(size=(B,), device=prev_output_tokens.device)
+                        < self.dae_ratio
+                    )
+                    corrputed_tokens = _sequential_poisoning(
+                        tgt_tokens[corrputed],
+                        len(self.tgt_dict),
+                        0.33,
+                        self.bos,
+                        self.eos,
+                        self.pad,
+                    )
+                    prev_output_tokens[corrputed] = corrputed_tokens
+
+        # concat everything
+        word_ins_out = torch.cat(word_ins_outs, 0)
+        word_ins_tgt = torch.cat(word_ins_tgts, 0)
+        word_ins_mask = torch.cat(word_ins_masks, 0)
+
+        return {
+            "word_ins_out": word_ins_out,
+            "word_ins_tgt": word_ins_tgt,
+            "word_ins_mask": word_ins_mask,
+            "length_out": length_out,
+            "length_tgt": length_tgt,
+            "length_w": self.decoder.length_loss_factor,
+        }
+
+
+@register_model_architecture(
+    "iterative_nonautoregressive_transformer", "iterative_nonautoregressive_transformer"
+)
+def base_architecture(args):
+    args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
+    args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+    args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048)
+    args.encoder_layers = getattr(args, "encoder_layers", 6)
+    args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8)
+    args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+    args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False)
+    args.decoder_embed_path = getattr(args, "decoder_embed_path", None)
+    args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim)
+    args.decoder_ffn_embed_dim = getattr(
+        args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim
+    )
+    args.decoder_layers = getattr(args, "decoder_layers", 6)
+    args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8)
+    args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False)
+    args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False)
+    args.attention_dropout = getattr(args, "attention_dropout", 0.0)
+    args.activation_dropout = getattr(args, "activation_dropout", 0.0)
+    args.activation_fn = getattr(args, "activation_fn", "relu")
+    args.dropout = getattr(args, "dropout", 0.1)
+    args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None)
+    args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0)
+    args.share_decoder_input_output_embed = getattr(
+        args, "share_decoder_input_output_embed", False
+    )
+    args.share_all_embeddings = getattr(args, "share_all_embeddings", False)
+    args.no_token_positional_embeddings = getattr(
+        args, "no_token_positional_embeddings", False
+    )
+    args.adaptive_input = getattr(args, "adaptive_input", False)
+    args.apply_bert_init = getattr(args, "apply_bert_init", False)
+
+    args.decoder_output_dim = getattr(
+        args, "decoder_output_dim", args.decoder_embed_dim
+    )
+    args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim)
+
+    # --- special arguments ---
+    args.sg_length_pred = getattr(args, "sg_length_pred", False)
+    args.pred_length_offset = getattr(args, "pred_length_offset", False)
+    args.length_loss_factor = getattr(args, "length_loss_factor", 0.1)
+    args.ngram_predictor = getattr(args, "ngram_predictor", 1)
+    args.src_embedding_copy = getattr(args, "src_embedding_copy", False)
+
+    args.train_step = getattr(args, "train_step", 4)
+    args.dae_ratio = getattr(args, "dae_ratio", 0.5)
+    args.stochastic_approx = getattr(args, "stochastic_approx", False)
+
+
+@register_model_architecture(
+    "iterative_nonautoregressive_transformer",
+    "iterative_nonautoregressive_transformer_wmt_en_de",
+)
+def iter_nat_wmt_en_de(args):
+    base_architecture(args)

fairseq/models/levenshtein_transformer.py

@@ -0,0 +1,595 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn.functional as F
+from fairseq import libnat
+from fairseq.models import register_model, register_model_architecture
+from fairseq.models.model_utils import fill_tensors as _fill, skip_tensors as _skip
+from fairseq.models.transformer import (
+    Embedding,
+    TransformerDecoder,
+    TransformerEncoder,
+    TransformerModel,
+)
+from fairseq.modules.transformer_sentence_encoder import init_bert_params
+
+
+def _get_ins_targets(in_tokens, out_tokens, padding_idx, unk_idx):
+    in_seq_len, out_seq_len = in_tokens.size(1), out_tokens.size(1)
+
+    with torch.cuda.device_of(in_tokens):
+        in_tokens_list = [
+            [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist())
+        ]
+        out_tokens_list = [
+            [t for t in s if t != padding_idx]
+            for i, s in enumerate(out_tokens.tolist())
+        ]
+
+    full_labels = libnat.suggested_ed2_path(
+        in_tokens_list, out_tokens_list, padding_idx
+    )
+    mask_inputs = [
+        [len(c) if c[0] != padding_idx else 0 for c in a[:-1]] for a in full_labels
+    ]
+
+    # generate labels
+    masked_tgt_masks = []
+    for mask_input in mask_inputs:
+        mask_label = []
+        for beam_size in mask_input[1:-1]:  # HACK 1:-1
+            mask_label += [0] + [1 for _ in range(beam_size)]
+        masked_tgt_masks.append(
+            mask_label + [0 for _ in range(out_seq_len - len(mask_label))]
+        )
+    mask_ins_targets = [
+        mask_input[1:-1] + [0 for _ in range(in_seq_len - 1 - len(mask_input[1:-1]))]
+        for mask_input in mask_inputs
+    ]
+
+    # transform to tensor
+    masked_tgt_masks = torch.tensor(
+        masked_tgt_masks, device=out_tokens.device
+    ).bool()
+    mask_ins_targets = torch.tensor(mask_ins_targets, device=in_tokens.device)
+    masked_tgt_tokens = out_tokens.masked_fill(masked_tgt_masks, unk_idx)
+    return masked_tgt_masks, masked_tgt_tokens, mask_ins_targets
+
+
+def _get_del_targets(in_tokens, out_tokens, padding_idx):
+    out_seq_len = out_tokens.size(1)
+
+    with torch.cuda.device_of(in_tokens):
+        in_tokens_list = [
+            [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist())
+        ]
+        out_tokens_list = [
+            [t for t in s if t != padding_idx]
+            for i, s in enumerate(out_tokens.tolist())
+        ]
+
+    full_labels = libnat.suggested_ed2_path(
+        in_tokens_list, out_tokens_list, padding_idx
+    )
+    word_del_targets = [b[-1] for b in full_labels]
+    word_del_targets = [
+        labels + [0 for _ in range(out_seq_len - len(labels))]
+        for labels in word_del_targets
+    ]
+
+    # transform to tensor
+    word_del_targets = torch.tensor(word_del_targets, device=out_tokens.device)
+    return word_del_targets
+
+
+def _get_del_ins_targets(in_tokens, out_tokens, padding_idx):
+    in_seq_len, out_seq_len = in_tokens.size(1), out_tokens.size(1)
+
+    with torch.cuda.device_of(in_tokens):
+        in_tokens_list = [
+            [t for t in s if t != padding_idx] for i, s in enumerate(in_tokens.tolist())
+        ]
+        out_tokens_list = [
+            [t for t in s if t != padding_idx]
+            for i, s in enumerate(out_tokens.tolist())
+        ]
+
+    full_labels = libnat.suggested_ed2_path(
+        in_tokens_list, out_tokens_list, padding_idx
+    )
+
+    word_del_targets = [b[-1] for b in full_labels]
+    word_del_targets = [
+        labels + [0 for _ in range(out_seq_len - len(labels))]
+        for labels in word_del_targets
+    ]
+
+    mask_inputs = [
+        [len(c) if c[0] != padding_idx else 0 for c in a[:-1]] for a in full_labels
+    ]
+    mask_ins_targets = [
+        mask_input[1:-1] + [0 for _ in range(in_seq_len - 1 - len(mask_input[1:-1]))]
+        for mask_input in mask_inputs
+    ]
+
+    # transform to tensor
+    mask_ins_targets = torch.tensor(mask_ins_targets, device=in_tokens.device)
+    word_del_targets = torch.tensor(word_del_targets, device=out_tokens.device)
+    return word_del_targets, mask_ins_targets
+
+
+def _apply_ins_masks(
+    in_tokens, in_scores, mask_ins_pred, padding_idx, unk_idx, eos_idx
+):
+
+    in_masks = in_tokens.ne(padding_idx)
+    in_lengths = in_masks.sum(1)
+
+    # HACK: hacky way to shift all the paddings to eos first.
+    in_tokens.masked_fill_(~in_masks, eos_idx)
+    mask_ins_pred.masked_fill_(~in_masks[:, 1:], 0)
+
+    out_lengths = in_lengths + mask_ins_pred.sum(1)
+    out_max_len = out_lengths.max()
+    out_masks = (
+        torch.arange(out_max_len, device=out_lengths.device)[None, :]
+        < out_lengths[:, None]
+    )
+
+    reordering = (mask_ins_pred + in_masks[:, 1:].long()).cumsum(1)
+    out_tokens = (
+        in_tokens.new_zeros(in_tokens.size(0), out_max_len)
+        .fill_(padding_idx)
+        .masked_fill_(out_masks, unk_idx)
+    )
+    out_tokens[:, 0] = in_tokens[:, 0]
+    out_tokens.scatter_(1, reordering, in_tokens[:, 1:])
+
+    out_scores = None
+    if in_scores is not None:
+        in_scores.masked_fill_(~in_masks, 0)
+        out_scores = in_scores.new_zeros(*out_tokens.size())
+        out_scores[:, 0] = in_scores[:, 0]
+        out_scores.scatter_(1, reordering, in_scores[:, 1:])
+
+    return out_tokens, out_scores
+
+
+def _apply_ins_words(in_tokens, in_scores, word_ins_pred, word_ins_scores, unk_idx):
+    word_ins_masks = in_tokens.eq(unk_idx)
+    out_tokens = in_tokens.masked_scatter(word_ins_masks, word_ins_pred[word_ins_masks])
+
+    if in_scores is not None:
+        out_scores = in_scores.masked_scatter(
+            word_ins_masks, word_ins_scores[word_ins_masks]
+        )
+    else:
+        out_scores = None
+
+    return out_tokens, out_scores
+
+
+def _apply_del_words(
+    in_tokens, in_scores, in_attn, word_del_pred, padding_idx, bos_idx, eos_idx
+):
+    # apply deletion to a tensor
+    in_masks = in_tokens.ne(padding_idx)
+    bos_eos_masks = in_tokens.eq(bos_idx) | in_tokens.eq(eos_idx)
+
+    max_len = in_tokens.size(1)
+    word_del_pred.masked_fill_(~in_masks, 1)
+    word_del_pred.masked_fill_(bos_eos_masks, 0)
+
+    reordering = (
+        torch.arange(max_len, device=in_tokens.device)[None, :]
+        .expand_as(in_tokens)
+        .contiguous()
+        .masked_fill_(word_del_pred, max_len)
+        .sort(1)[1]
+    )
+
+    out_tokens = in_tokens.masked_fill(word_del_pred, padding_idx).gather(1, reordering)
+
+    out_scores = None
+    if in_scores is not None:
+        out_scores = in_scores.masked_fill(word_del_pred, 0).gather(1, reordering)
+
+    out_attn = None
+    if in_attn is not None:
+        _mask = word_del_pred[:, :, None].expand_as(in_attn)
+        _reordering = reordering[:, :, None].expand_as(in_attn)
+        out_attn = in_attn.masked_fill(_mask, 0.).gather(1, _reordering)
+
+    return out_tokens, out_scores, out_attn
+
+
+@register_model("levenshtein_transformer")
+class LevenshteinTransformerModel(TransformerModel):
+    def __init__(self, encoder, decoder):
+        super().__init__(encoder, decoder)
+        self.tgt_dict = decoder.dictionary
+        self.bos = decoder.dictionary.bos()
+        self.eos = decoder.dictionary.eos()
+        self.pad = decoder.dictionary.pad()
+        self.unk = decoder.dictionary.unk()
+
+    @staticmethod
+    def add_args(parser):
+        TransformerModel.add_args(parser)
+        parser.add_argument(
+            "--apply-bert-init",
+            action="store_true",
+            help="use custom param initialization for BERT",
+        )
+        parser.add_argument(
+            "--early-exit",
+            default="6,6,6",
+            type=str,
+            help="number of decoder layers before mask_ins, word_ins and word_del heads",
+        )
+
+    @classmethod
+    def build_decoder(cls, args, tgt_dict, embed_tokens):
+        decoder = LevenshteinTransformerDecoder(args, tgt_dict, embed_tokens)
+        if getattr(args, "apply_bert_init", False):
+            decoder.apply(init_bert_params)
+        return decoder
+
+    @classmethod
+    def build_encoder(cls, args, src_dict, embed_tokens):
+        encoder = TransformerEncoder(args, src_dict, embed_tokens)
+        if getattr(args, "apply_bert_init", False):
+            encoder.apply(init_bert_params)
+        return encoder
+
+    def forward(
+        self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs
+    ):
+
+        assert tgt_tokens is not None, "forward function only supports training."
+
+        # encoding
+        encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs)
+
+        # generate training labels for insertion
+        masked_tgt_masks, masked_tgt_tokens, mask_ins_targets = _get_ins_targets(
+            prev_output_tokens, tgt_tokens, self.pad, self.unk
+        )
+        mask_ins_targets = mask_ins_targets.clamp(min=0, max=255)  # for safe prediction
+        mask_ins_masks = prev_output_tokens[:, 1:].ne(self.pad)
+
+        mask_ins_out, _ = self.decoder.forward_mask_ins(
+            prev_output_tokens, encoder_out=encoder_out
+        )
+        word_ins_out, _ = self.decoder.forward_word_ins(
+            masked_tgt_tokens, encoder_out=encoder_out
+        )
+
+        # make online prediction
+        word_predictions = F.log_softmax(word_ins_out, dim=-1).max(2)[1]
+        word_predictions.masked_scatter_(
+            ~masked_tgt_masks, tgt_tokens[~masked_tgt_masks]
+        )
+
+        # generate training labels for deletion
+        word_del_targets = _get_del_targets(word_predictions, tgt_tokens, self.pad)
+        word_del_out, _ = self.decoder.forward_word_del(
+            word_predictions, encoder_out)
+
+        return {
+            "mask_ins_out": mask_ins_out,
+            "mask_ins_tgt": mask_ins_targets,
+            "mask_ins_mask": mask_ins_masks,
+            "word_ins_out": word_ins_out,
+            "word_ins_tgt": tgt_tokens,
+            "word_ins_mask": masked_tgt_masks,
+            "word_del_out": word_del_out,
+            "word_del_tgt": word_del_targets,
+            "word_del_mask": word_predictions.ne(self.pad),
+        }
+
+    def forward_encoder(self, encoder_inputs):
+        return self.encoder(*encoder_inputs)
+
+    def forward_decoder(
+        self, decoder_out, encoder_out, eos_penalty=0.0, max_ratio=None, **kwargs
+    ):
+
+        output_tokens = decoder_out["output_tokens"]
+        output_scores = decoder_out["output_scores"]
+        attn = decoder_out["attn"]
+
+        if max_ratio is None:
+            max_lens = output_tokens.new(output_tokens.size(0)).fill_(255)
+        else:
+            max_lens = (
+                (~encoder_out["encoder_padding_mask"]).sum(1) * max_ratio
+            ).clamp(min=10)
+
+        # delete words
+        # do not delete tokens if it is <s> </s>
+        can_del_word = output_tokens.ne(self.pad).sum(1) > 2
+        if can_del_word.sum() != 0:  # we cannot delete, skip
+            word_del_out, word_del_attn = self.decoder.forward_word_del(
+                _skip(output_tokens, can_del_word), _skip(encoder_out, can_del_word)
+            )
+            word_del_score = F.log_softmax(word_del_out, 2)
+            word_del_pred = word_del_score.max(-1)[1].bool()
+
+            _tokens, _scores, _attn = _apply_del_words(
+                output_tokens[can_del_word],
+                output_scores[can_del_word],
+                word_del_attn,
+                word_del_pred,
+                self.pad,
+                self.bos,
+                self.eos,
+            )
+            output_tokens = _fill(output_tokens, can_del_word, _tokens, self.pad)
+            output_scores = _fill(output_scores, can_del_word, _scores, 0)
+            attn = _fill(attn, can_del_word, _attn, 0.)
+
+        # insert placeholders
+        can_ins_mask = output_tokens.ne(self.pad).sum(1) < max_lens
+        if can_ins_mask.sum() != 0:
+            mask_ins_out, _ = self.decoder.forward_mask_ins(
+                _skip(output_tokens, can_ins_mask), _skip(encoder_out, can_ins_mask)
+            )
+            mask_ins_score = F.log_softmax(mask_ins_out, 2)
+            if eos_penalty > 0.0:
+                mask_ins_score[:, :, 0] -= eos_penalty
+            mask_ins_pred = mask_ins_score.max(-1)[1]
+            mask_ins_pred = torch.min(
+                mask_ins_pred, max_lens[:, None].expand_as(mask_ins_pred)
+            )
+
+            _tokens, _scores = _apply_ins_masks(
+                output_tokens[can_ins_mask],
+                output_scores[can_ins_mask],
+                mask_ins_pred,
+                self.pad,
+                self.unk,
+                self.eos,
+            )
+            output_tokens = _fill(output_tokens, can_ins_mask, _tokens, self.pad)
+            output_scores = _fill(output_scores, can_ins_mask, _scores, 0)
+
+        # insert words
+        can_ins_word = output_tokens.eq(self.unk).sum(1) > 0
+        if can_ins_word.sum() != 0:
+            word_ins_out, word_ins_attn = self.decoder.forward_word_ins(
+                _skip(output_tokens, can_ins_word), _skip(encoder_out, can_ins_word)
+            )
+            word_ins_score = F.log_softmax(word_ins_out, 2)
+            word_ins_pred = word_ins_score.max(-1)[1]
+
+            _tokens, _scores = _apply_ins_words(
+                output_tokens[can_ins_word],
+                output_scores[can_ins_word],
+                word_ins_pred,
+                word_ins_score,
+                self.unk,
+            )
+
+            output_tokens = _fill(output_tokens, can_ins_word, _tokens, self.pad)
+            output_scores = _fill(output_scores, can_ins_word, _scores, 0)
+            attn = _fill(attn, can_ins_word, word_ins_attn, 0.)
+
+        # delete some unnecessary paddings
+        cut_off = output_tokens.ne(self.pad).sum(1).max()
+        output_tokens = output_tokens[:, :cut_off]
+        output_scores = output_scores[:, :cut_off]
+        attn = None if attn is None else attn[:, :cut_off, :]
+        return {
+            "output_tokens": output_tokens,
+            "output_scores": output_scores,
+            "attn": attn,
+        }
+
+    def initialize_output_tokens(self, encoder_out, src_tokens):
+        initial_output_tokens = src_tokens.new_zeros(src_tokens.size(0), 2)
+        initial_output_tokens[:, 0] = self.bos
+        initial_output_tokens[:, 1] = self.eos
+
+        initial_output_scores = initial_output_tokens.new_zeros(
+            *initial_output_tokens.size()
+        ).type_as(encoder_out["encoder_out"])
+        return {
+            "output_tokens": initial_output_tokens,
+            "output_scores": initial_output_scores,
+            "attn": None,
+        }
+
+
+class LevenshteinTransformerDecoder(TransformerDecoder):
+    def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False):
+        super().__init__(
+            args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn
+        )
+        self.dictionary = dictionary
+        self.bos = dictionary.bos()
+        self.unk = dictionary.unk()
+        self.eos = dictionary.eos()
+
+        self.embed_mask_ins = Embedding(256, self.output_embed_dim * 2, None)
+        self.embed_word_del = Embedding(2, self.output_embed_dim, None)
+        # del_word, ins_mask, ins_word
+        self.early_exit = [int(i) for i in args.early_exit.split(',')]
+        assert len(self.early_exit) == 3
+
+    def extract_features(
+        self, prev_output_tokens, encoder_out=None, early_exit=None, **unused
+    ):
+        """
+        Similar to *forward* but only return features.
+
+        Inputs:
+            prev_output_tokens: Tensor(B, T)
+            encoder_out: a dictionary of hidden states and masks
+
+        Returns:
+            tuple:
+                - the decoder's features of shape `(batch, tgt_len, embed_dim)`
+                - a dictionary with any model-specific outputs
+            the LevenshteinTransformer decoder has full-attention to all generated tokens
+        """
+        # embed positions
+        positions = (
+            self.embed_positions(prev_output_tokens)
+            if self.embed_positions is not None
+            else None
+        )
+
+        # embed tokens and positions
+        x = self.embed_scale * self.embed_tokens(prev_output_tokens)
+        if self.project_in_dim is not None:
+            x = self.project_in_dim(x)
+
+        if positions is not None:
+            x += positions
+        x = F.dropout(x, p=self.dropout, training=self.training)
+
+        # B x T x C -> T x B x C
+        x = x.transpose(0, 1)
+        attn = None
+        inner_states = [x]
+
+        # decoder layers
+        decoder_padding_mask = prev_output_tokens.eq(self.padding_idx)
+        for i, layer in enumerate(self.layers):
+
+            # early exit from the decoder.
+            if (early_exit is not None) and (i >= early_exit):
+                break
+
+            x, attn = layer(
+                x,
+                encoder_out["encoder_out"] if encoder_out is not None else None,
+                encoder_out["encoder_padding_mask"]
+                if encoder_out is not None
+                else None,
+                self_attn_mask=None,
+                self_attn_padding_mask=decoder_padding_mask,
+            )
+            inner_states.append(x)
+
+        if self.layer_norm:
+            x = self.layer_norm(x)
+
+        # T x B x C -> B x T x C
+        x = x.transpose(0, 1)
+
+        if self.project_out_dim is not None:
+            x = self.project_out_dim(x)
+
+        return x, {"attn": attn, "inner_states": inner_states}
+
+    def forward_mask_ins(self, prev_output_tokens, encoder_out=None):
+        features, extra = self.extract_features(
+            prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[1]
+        )
+        features_cat = torch.cat([features[:, :-1, :], features[:, 1:, :]], 2)
+        return F.linear(features_cat, self.embed_mask_ins.weight), extra['attn']
+
+    def forward_word_ins(self, prev_output_tokens, encoder_out=None):
+        features, extra = self.extract_features(
+            prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[2]
+        )
+        return self.output_layer(features), extra['attn']
+
+    def forward_word_del(self, prev_output_tokens, encoder_out=None):
+        features, extra = self.extract_features(
+            prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[0]
+        )
+        return F.linear(features, self.embed_word_del.weight), extra['attn']
+
+    def forward_word_del_mask_ins(self, prev_output_tokens, encoder_out=None):
+        # merge the word-deletion and mask insertion into one operation,
+        assert self.early_exit[0] == self.early_exit[1], "must the same depth."
+        features, extra = self.extract_features(
+            prev_output_tokens, encoder_out=encoder_out, early_exit=self.early_exit[2]
+        )
+        features_cat = torch.cat([features[:, :-1, :], features[:, 1:, :]], 2)
+        f_word_del = F.linear(features, self.embed_word_del.weight)
+        f_mask_ins = F.linear(features_cat, self.embed_mask_ins.weight)
+        return f_word_del, f_mask_ins, extra['attn']
+
+
+@register_model_architecture("levenshtein_transformer", "levenshtein_transformer")
+def base_architecture(args):
+    args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
+    args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+    args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048)
+    args.encoder_layers = getattr(args, "encoder_layers", 6)
+    args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8)
+    args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+    args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False)
+    args.decoder_embed_path = getattr(args, "decoder_embed_path", None)
+    args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim)
+    args.decoder_ffn_embed_dim = getattr(
+        args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim
+    )
+    args.decoder_layers = getattr(args, "decoder_layers", 6)
+    args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8)
+    args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False)
+    args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False)
+    args.attention_dropout = getattr(args, "attention_dropout", 0.0)
+    args.activation_dropout = getattr(args, "activation_dropout", 0.0)
+    args.activation_fn = getattr(args, "activation_fn", "relu")
+    args.dropout = getattr(args, "dropout", 0.1)
+    args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None)
+    args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0)
+    args.share_decoder_input_output_embed = getattr(
+        args, "share_decoder_input_output_embed", False
+    )
+    args.share_all_embeddings = getattr(args, "share_all_embeddings", False)
+    args.no_token_positional_embeddings = getattr(
+        args, "no_token_positional_embeddings", False
+    )
+    args.adaptive_input = getattr(args, "adaptive_input", False)
+    args.apply_bert_init = getattr(args, "apply_bert_init", False)
+
+    args.decoder_output_dim = getattr(
+        args, "decoder_output_dim", args.decoder_embed_dim
+    )
+    args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim)
+
+    args.early_exit = getattr(args, "early_exit", "(6, 6, 6)")
+
+
+@register_model_architecture(
+    "levenshtein_transformer", "levenshtein_transformer_wmt_en_de"
+)
+def levenshtein_transformer_wmt_en_de(args):
+    base_architecture(args)
+
+
+# similar parameters used in the "Attention Is All You Need" paper (Vaswani et al., 2017)
+@register_model_architecture(
+    "levenshtein_transformer", "levenshtein_transformer_vaswani_wmt_en_de_big"
+)
+def levenshtein_transformer_vaswani_wmt_en_de_big(args):
+    args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 1024)
+    args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 4096)
+    args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 16)
+    args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+    args.decoder_embed_dim = getattr(args, "decoder_embed_dim", 1024)
+    args.decoder_ffn_embed_dim = getattr(args, "decoder_ffn_embed_dim", 4096)
+    args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 16)
+    args.dropout = getattr(args, "dropout", 0.3)
+    base_architecture(args)
+
+
+# default parameters used in tensor2tensor implementation
+@register_model_architecture(
+    "levenshtein_transformer", "levenshtein_transformer_wmt_en_de_big"
+)
+def levenshtein_transformer_wmt_en_de_big_t2t(args):
+    args.encoder_normalize_before = getattr(args, "encoder_normalize_before", True)
+    args.decoder_normalize_before = getattr(args, "decoder_normalize_before", True)
+    args.attention_dropout = getattr(args, "attention_dropout", 0.1)
+    args.activation_dropout = getattr(args, "activation_dropout", 0.1)
+    levenshtein_transformer_vaswani_wmt_en_de_big(args)

fairseq/models/model_utils.py

@@ -0,0 +1,62 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+
+def skip_tensors(x, mask):
+    """
+    Getting sliced (dim=0) tensor by mask. Supporting tensor and list/dict of tensors.
+    """
+    if isinstance(x, int):
+        return x
+
+    if x is None:
+        return None
+
+    if isinstance(x, torch.Tensor):
+        if x.size(0) == mask.size(0):
+            return x[mask]
+        elif x.size(1) == mask.size(0):
+            return x[:, mask]
+
+    if isinstance(x, list):
+        return [skip_tensors(x_i, mask) for x_i in x]
+
+    if isinstance(x, dict):
+        return {k: skip_tensors(v, mask) for k, v in x.items()}
+
+    raise NotImplementedError
+
+
+def fill_tensors(x, mask, y, padding_idx):
+    """
+    Filling tensor x with y at masked positions (dim=0).
+    """
+    if x is None:
+        return y
+    assert x.dim() == y.dim() and mask.size(0) == x.size(0)
+    assert x.dim() == 2 or (x.dim() == 3 and x.size(2) == y.size(2))
+    n_selected = mask.sum()
+    assert n_selected == y.size(0)
+
+    if n_selected == x.size(0):
+        return y
+
+    if x.size(1) < y.size(1):
+        dims = [x.size(0), y.size(1) - x.size(1)]
+        if x.dim() == 3:
+            dims.append(x.size(2))
+        x = torch.cat([x, x.new_zeros(*dims).fill_(padding_idx)], 1)
+        x[mask] = y
+    elif x.size(1) > y.size(1):
+        x[mask] = padding_idx
+        if x.dim() == 2:
+            x[mask, :y.size(1)] = y
+        else:
+            x[mask, :y.size(1), :] = y
+    else:
+        x[mask] = y
+    return x

fairseq/models/nonautoregressive_transformer.py

@@ -0,0 +1,640 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn.functional as F
+from fairseq import utils
+from fairseq.models import register_model, register_model_architecture
+from fairseq.models.transformer import (
+    Embedding,
+    TransformerDecoder,
+    TransformerDecoderLayer,
+    TransformerEncoder,
+    TransformerModel,
+)
+from fairseq.modules import MultiheadAttention
+from fairseq.modules.transformer_sentence_encoder import init_bert_params
+
+
+def _mean_pooling(enc_feats, src_masks):
+    # enc_feats: T x B x C
+    # src_masks: B x T or None
+    if src_masks is None:
+        enc_feats = enc_feats.mean(0)
+    else:
+        src_masks = (~src_masks).transpose(0, 1).type_as(enc_feats)
+        enc_feats = (
+            (enc_feats / src_masks.sum(0)[None, :, None]) * src_masks[:, :, None]
+        ).sum(0)
+    return enc_feats
+
+
+def _argmax(x, dim):
+    return (x == x.max(dim, keepdim=True)[0]).type_as(x)
+
+
+def _dynamic_programming(tokens, scores):
+    N, B, T = tokens.size()
+    cum_scores = scores[:, :, 0].clone()  # N x B
+    cum_choice = tokens.new_zeros(B, T)
+
+    # forward
+    for t in range(T - 1):
+        score, choice = cum_scores.max(0)
+        cum_choice[:, t] = choice
+        cum_scores[0] = score + scores[0, :, t + 1]
+        cum_scores[1:] = cum_scores[:-1] + scores[1:, :, t + 1]
+
+    # back-tracking
+    end_score, end_choice = cum_scores.max(0)
+    cum_choice[:, T - 1] = end_choice
+    for t in range(T - 2, -1, -1):
+        is_start = (cum_choice[:, t + 1] == 0).type_as(cum_choice)
+        cum_choice[:, t] = (cum_choice[:, t + 1] - 1) * ~is_start + cum_choice[
+            :, t
+        ] * is_start
+
+    # finalize the prediction
+    tokens = tokens.gather(0, cum_choice.unsqueeze(0)).squeeze(0)
+    scores = scores.gather(0, cum_choice.unsqueeze(0)).squeeze(0)
+    return scores, tokens
+
+
+def _beam_search(tokens, scores, W=None):
+    N, B, T = tokens.size()
+
+    if (W is None) or (W > N):
+        W = N
+
+
+def _uniform_assignment(src_lens, trg_lens):
+    max_trg_len = trg_lens.max()
+    steps = (src_lens.float() - 1) / (trg_lens.float() - 1)  # step-size
+    # max_trg_len
+    index_t = torch.arange(max_trg_len, device=trg_lens.device).float()
+    index_t = steps[:, None] * index_t[None, :]  # batch_size X max_trg_len
+    index_t = torch.round(index_t).long().detach()
+    return index_t
+
+
+@register_model("nonautoregressive_transformer")
+class NATransformerModel(TransformerModel):
+    def __init__(self, encoder, decoder):
+        super().__init__(encoder, decoder)
+        self.tgt_dict = decoder.dictionary
+        self.bos = decoder.dictionary.bos()
+        self.eos = decoder.dictionary.eos()
+        self.pad = decoder.dictionary.pad()
+        self.unk = decoder.dictionary.unk()
+
+    @staticmethod
+    def add_args(parser):
+        TransformerModel.add_args(parser)
+        parser.add_argument(
+            "--apply-bert-init",
+            action="store_true",
+            help="use custom param initialization for BERT",
+        )
+
+        # length prediction
+        parser.add_argument("--src-embedding-copy", action="store_true",
+                            help="copy encoder word embeddings as the initial input of the decoder")
+        parser.add_argument("--pred-length-offset", action="store_true",
+                            help="predicting the length difference between the target and source sentences")
+        parser.add_argument("--sg-length-pred", action="store_true",
+                            help="stop the gradients back-propagated from the length predictor")
+        parser.add_argument("--length-loss-factor", type=float,
+                            help="weights on the length prediction loss")
+
+        # n-gram predictor
+        parser.add_argument(
+            "--ngram-predictor",
+            nargs="?",
+            const=4,
+            default=1,
+            type=int,
+            help="adding an additional n-gram predictor.",
+        )
+
+    @classmethod
+    def build_decoder(cls, args, tgt_dict, embed_tokens):
+        decoder = NATransformerDecoder(args, tgt_dict, embed_tokens)
+        if getattr(args, "apply_bert_init", False):
+            decoder.apply(init_bert_params)
+        return decoder
+
+    @classmethod
+    def build_encoder(cls, args, src_dict, embed_tokens):
+        encoder = TransformerEncoder(args, src_dict, embed_tokens)
+        if getattr(args, "apply_bert_init", False):
+            encoder.apply(init_bert_params)
+        return encoder
+
+    def forward(
+        self, src_tokens, src_lengths, prev_output_tokens, tgt_tokens, **kwargs
+    ):
+        # encoding
+        encoder_out = self.encoder(src_tokens, src_lengths=src_lengths, **kwargs)
+        length_out, length_tgt = self.decoder.forward_length_prediction(
+            encoder_out, tgt_tokens
+        )
+
+        word_ins_out, word_ins_tgt, word_ins_mask = self.decoder(
+            prev_output_tokens, encoder_out=encoder_out, tgt_tokens=tgt_tokens
+        )
+
+        return {
+            "word_ins_out": word_ins_out,
+            "word_ins_tgt": word_ins_tgt,
+            "word_ins_mask": word_ins_mask,
+            "length_out": length_out,
+            "length_tgt": length_tgt,
+            "length_w": self.decoder.length_loss_factor,
+        }
+
+    def forward_encoder(self, encoder_inputs):
+        return self.encoder(*encoder_inputs)
+
+    def forward_decoder(self, decoder_out, encoder_out, decoding_format=None, **kwargs):
+        step = decoder_out["step"]
+        output_tokens = decoder_out["output_tokens"]
+        output_scores = decoder_out["output_scores"]
+
+        # execute the decoder
+        output_masks = output_tokens.ne(self.pad)
+        _scores, _tokens = self.decoder(
+            output_tokens,
+            encoder_out=encoder_out,
+            decoding_format=decoding_format,
+            step=step,
+        )
+        output_tokens.masked_scatter_(output_masks, _tokens[output_masks])
+        output_scores.masked_scatter_(output_masks, _scores[output_masks])
+
+        return {"output_tokens": output_tokens, "output_scores": output_scores}
+
+    def initialize_output_tokens(self, encoder_out, src_tokens):
+        # length prediction
+        _, length_tgt = self.decoder.forward_length_prediction(encoder_out)
+        max_length = length_tgt.max()
+        idx_length = torch.arange(max_length, device=src_tokens.device)
+
+        initial_output_tokens = src_tokens.new_zeros(
+            src_tokens.size(0), max_length
+        ).fill_(self.pad)
+        initial_output_tokens.masked_fill_(
+            idx_length[None, :] < length_tgt[:, None], self.unk
+        )
+        initial_output_tokens[:, 0] = self.bos
+        initial_output_tokens.scatter_(1, length_tgt[:, None] - 1, self.eos)
+
+        initial_output_scores = initial_output_tokens.new_zeros(
+            *initial_output_tokens.size()
+        ).type_as(encoder_out["encoder_out"])
+
+        return {
+            "output_tokens": initial_output_tokens,
+            "output_scores": initial_output_scores,
+        }
+
+
+class NATransformerDecoder(TransformerDecoder):
+    def __init__(self, args, dictionary, embed_tokens, no_encoder_attn=False):
+        super().__init__(
+            args, dictionary, embed_tokens, no_encoder_attn=no_encoder_attn
+        )
+
+        self.dictionary = dictionary
+        self.bos = dictionary.bos()
+        self.unk = dictionary.unk()
+        self.eos = dictionary.eos()
+
+        self.encoder_embed_dim = args.encoder_embed_dim
+        self.sg_length_pred = getattr(args, "sg_length_pred", False)
+        self.pred_length_offset = getattr(args, "pred_length_offset", False)
+        self.length_loss_factor = getattr(args, "length_loss_factor", 0.1)
+        self.src_embedding_copy = getattr(args, "src_embedding_copy", False)
+        self.embed_length = Embedding(256, self.encoder_embed_dim, None)
+
+        self.ngram_predictor = getattr(args, "ngram_predictor", 1)
+        self.ngram_layer = (
+            None if (self.ngram_predictor == 1) else NgramDecoderLayer(args, True)
+        )
+
+    def forward(
+        self,
+        prev_output_tokens,
+        encoder_out=None,
+        tgt_tokens=None,
+        decoding_format=None,
+        step=0,
+        **kwargs
+    ):
+
+        features, _ = self.extract_features(
+            prev_output_tokens,
+            encoder_out=encoder_out,
+            embedding_copy=(step == 0) & self.src_embedding_copy,
+        )
+
+        if tgt_tokens is not None:
+            if self.ngram_layer is None:
+                word_ins_mask = tgt_tokens.ne(self.padding_idx)
+                word_ins_tgt = tgt_tokens
+            else:
+                context_embeds, context_masks = self.forward_ngram_context(tgt_tokens)
+                features = self.ngram_layer(features, context_embeds=context_embeds)
+                word_ins_tgt = tgt_tokens[:, :, None].repeat(1, 1, self.ngram_predictor)
+                word_ins_mask = word_ins_tgt.ne(self.padding_idx) & context_masks
+
+            return self.output_layer(features), word_ins_tgt, word_ins_mask
+
+        else:
+            if self.ngram_layer is None:
+                return F.log_softmax(self.output_layer(features), -1).max(-1)
+            else:
+                # inner iterations
+                return self.forward_ngram_decoding(
+                    features, prev_output_tokens.eq(self.padding_idx), decoding_format
+                )
+
+    def extract_features(
+        self,
+        prev_output_tokens,
+        encoder_out=None,
+        early_exit=None,
+        embedding_copy=False,
+        **unused
+    ):
+        """
+        Similar to *forward* but only return features.
+
+        Inputs:
+            prev_output_tokens: Tensor(B, T)
+            encoder_out: a dictionary of hidden states and masks
+
+        Returns:
+            tuple:
+                - the decoder's features of shape `(batch, tgt_len, embed_dim)`
+                - a dictionary with any model-specific outputs
+            the LevenshteinTransformer decoder has full-attention to all generated tokens
+        """
+        # embedding
+        if embedding_copy:
+            src_embd = encoder_out["encoder_embedding"]
+            src_mask = encoder_out["encoder_padding_mask"]
+            src_mask = (
+                ~src_mask
+                if src_mask is not None
+                else prev_output_tokens.new_ones(*src_embd.size()[:2]).bool()
+            )
+
+            x, decoder_padding_mask = self.forward_embedding(
+                prev_output_tokens,
+                self.forward_copying_source(
+                    src_embd, src_mask, prev_output_tokens.ne(self.padding_idx)
+                ),
+            )
+
+        else:
+
+            x, decoder_padding_mask = self.forward_embedding(prev_output_tokens)
+
+        # B x T x C -> T x B x C
+        x = x.transpose(0, 1)
+        attn = None
+        inner_states = [x]
+
+        # decoder layers
+        for i, layer in enumerate(self.layers):
+
+            # early exit from the decoder.
+            if (early_exit is not None) and (i >= early_exit):
+                break
+
+            x, attn = layer(
+                x,
+                encoder_out["encoder_out"] if encoder_out is not None else None,
+                encoder_out["encoder_padding_mask"]
+                if encoder_out is not None
+                else None,
+                self_attn_mask=None,
+                self_attn_padding_mask=decoder_padding_mask,
+            )
+            inner_states.append(x)
+
+        if self.layer_norm:
+            x = self.layer_norm(x)
+
+        # T x B x C -> B x T x C
+        x = x.transpose(0, 1)
+
+        if self.project_out_dim is not None:
+            x = self.project_out_dim(x)
+
+        return x, {"attn": attn, "inner_states": inner_states}
+
+    def forward_ngram_context(self, tgt_tokens):
+        tgt_embeds = self.forward_embedding(tgt_tokens)
+        n_contexts = self.ngram_predictor - 1
+
+        # shifting the embeddings
+        # context_embeds: N x B x T x C
+        # context_masks:  B x T x N
+        context_embeds = tgt_embeds.new_zeros(n_contexts, *tgt_embeds.size())
+        context_masks = tgt_embeds.new_ones(
+            *tgt_embeds.size()[:2], self.ngram_predictor
+        ).bool()
+
+        for k in range(n_contexts):
+            context_embeds[k, :, k + 1:] = tgt_embeds[:, : -k - 1]
+            context_masks[:, : k + 1, k + 1] = 0
+
+        return context_embeds, context_masks
+
+    def forward_ngram_decoding(self, features, padding_mask=None, decoding_format=None):
+        context_embeds = None
+        scores, tokens = [], []
+        ensemble_score = None
+        ensemble_index = None
+
+        if decoding_format is None:
+            decoding_format = "ensemble"
+
+        for k in range(self.ngram_predictor):
+            ngram_out = self.ngram_layer(
+                features, context_embeds=context_embeds, incremental=True
+            )
+            ngram_scores = F.log_softmax(self.output_layer(ngram_out), -1)
+            max_score, max_token = ngram_scores.max(-1)
+
+            if decoding_format == "vote":
+                ngram_scores = _argmax(ngram_scores, -1)
+
+            if ensemble_score is None:
+                ensemble_score = ngram_scores
+                ensemble_index = ensemble_score.new_ones(*ensemble_score.size()[:2])
+            else:
+                ensemble_index[:, k:] = ensemble_index[:, k:] + 1
+                ensemble_score = ensemble_score + ngram_scores.masked_fill_(
+                    (ensemble_index < k)
+                    .unsqueeze(2)
+                    .repeat(1, 1, ensemble_score.size(2)),
+                    0,
+                )
+                max_score[:, :k] = float("-inf")
+
+            if decoding_format == "unigram":
+                break
+
+            scores.append(max_score.masked_fill_(padding_mask, 0))
+            tokens.append(max_token.masked_fill_(padding_mask, self.padding_idx))
+
+            # context_embeds: N x B x T x C
+            if context_embeds is None:
+                context_embeds = self.forward_embedding(max_token).unsqueeze(0)
+
+            else:
+                context_embeds = torch.cat(
+                    [self.forward_embedding(max_token).unsqueeze(0), context_embeds], 0
+                )
+
+            context_embeds[:, :, 1:] = context_embeds[:, :, :-1]
+
+        if decoding_format != "dp":
+            ensemble_score = ensemble_score / ensemble_index.unsqueeze(2)
+            return ensemble_score.max(-1)
+
+        else:
+            tokens = torch.cat([t.unsqueeze(0) for t in tokens], 0)
+            scores = torch.cat([s.unsqueeze(0) for s in scores], 0)
+            return _dynamic_programming(tokens, scores)
+
+    def forward_embedding(self, prev_output_tokens, states=None):
+        # embed positions
+        positions = (
+            self.embed_positions(prev_output_tokens)
+            if self.embed_positions is not None
+            else None
+        )
+
+        # embed tokens and positions
+        if states is None:
+            x = self.embed_scale * self.embed_tokens(prev_output_tokens)
+            if self.project_in_dim is not None:
+                x = self.project_in_dim(x)
+        else:
+            x = states
+
+        if positions is not None:
+            x += positions
+        x = F.dropout(x, p=self.dropout, training=self.training)
+        decoder_padding_mask = prev_output_tokens.eq(self.padding_idx)
+        return x, decoder_padding_mask
+
+    def forward_copying_source(self, src_embeds, src_masks, tgt_masks):
+        length_sources = src_masks.sum(1)
+        length_targets = tgt_masks.sum(1)
+        mapped_inputs = _uniform_assignment(length_sources, length_targets).masked_fill(
+            ~tgt_masks, 0
+        )
+        copied_embedding = torch.gather(
+            src_embeds,
+            1,
+            mapped_inputs.unsqueeze(-1).expand(
+                *mapped_inputs.size(), src_embeds.size(-1)
+            ),
+        )
+        return copied_embedding
+
+    def forward_length_prediction(self, encoder_out, tgt_tokens=None):
+        enc_feats = encoder_out["encoder_out"]  # T x B x C
+        src_masks = encoder_out["encoder_padding_mask"]  # B x T or None
+
+        if self.pred_length_offset:
+            if src_masks is None:
+                src_lengs = enc_feats.new_ones(enc_feats.size(1)).fill_(
+                    enc_feats.size(0)
+                )
+            else:
+                src_lengs = (~src_masks).transpose(0, 1).type_as(enc_feats).sum(0)
+            src_lengs = src_lengs.long()
+
+        enc_feats = _mean_pooling(enc_feats, src_masks)
+        if self.sg_length_pred:
+            enc_feats = enc_feats.detach()
+
+        length_out = F.linear(enc_feats, self.embed_length.weight)
+
+        if tgt_tokens is not None:
+            # obtain the length target
+            tgt_lengs = tgt_tokens.ne(self.padding_idx).sum(1).long()
+            if self.pred_length_offset:
+                length_tgt = tgt_lengs - src_lengs + 128
+            else:
+                length_tgt = tgt_lengs
+            length_tgt = length_tgt.clamp(min=0, max=255)
+
+        else:
+            # predict the length target (greedy for now)
+            # TODO: implementing length-beam
+            pred_lengs = length_out.max(-1)[1]
+            if self.pred_length_offset:
+                length_tgt = pred_lengs - 128 + src_lengs
+            else:
+                length_tgt = pred_lengs
+
+        return length_out, length_tgt
+
+
+class NgramDecoderLayer(TransformerDecoderLayer):
+    """
+    N-gram Decoder Layer:
+
+    This module can be pluged in the last layer of any Non-autoregressive Model's
+    It provides an alternative way to capture local n-gram information by running the block multiple times.
+    """
+
+    def __init__(self, args, no_encoder_attn=False):
+        super(NgramDecoderLayer, self).__init__(args, no_encoder_attn=no_encoder_attn)
+        self.self_attn = MultiheadAttention(
+            embed_dim=self.embed_dim,
+            num_heads=1,  # maybe n-gram does not need too many heads.
+            dropout=args.attention_dropout,
+            self_attention=False,
+            encoder_decoder_attention=True,
+        )
+
+    def forward(
+        self,
+        x,
+        encoder_out=None,
+        encoder_padding_mask=None,
+        context_embeds=None,
+        incremental=False,
+    ):
+        # x: T x B x C
+        # context_embeds: N x T x B x C
+        T, B, C = x.size()
+
+        residual = x
+        x = self.maybe_layer_norm(self.self_attn_layer_norm, x, before=True)
+        x = x.contiguous().view(1, T * B, C).contiguous()
+
+        if context_embeds is not None:
+            N = context_embeds.size(0)
+            context_embeds = context_embeds.view(N, T * B, C).contiguous()
+
+        if not incremental:
+            assert context_embeds is not None, "we need context for training"
+            # attn_weights: (n_head x T x B) x 1 x N
+            # v: (n_head x T x B) x N x (dim / n_head)
+            # -- move the attention computation outside --
+            attn_weights, values = self.self_attn(
+                query=x, key=context_embeds, value=context_embeds, before_softmax=True
+            )
+
+            attn_weights = attn_weights.repeat(1, N, 1)
+            attn_masks = attn_weights.new_ones(N, N).triu_(1).bool()
+            attn_masks = attn_masks.unsqueeze(0).repeat(attn_weights.size(0), 1, 1)
+
+            attn_weights = attn_weights.masked_fill(attn_masks, float("-inf"))
+            attn_weights = utils.softmax(attn_weights, dim=-1).type_as(attn_weights)
+            attn_weights = F.dropout(
+                attn_weights, p=self.self_attn.dropout, training=self.training
+            )
+
+            # (n_head x T x B) x N x (dim / n_head)
+            attn = torch.bmm(attn_weights, values)
+            attn = attn.transpose(0, 1).contiguous()
+            attn = attn.view(N, T * B, C).contiguous()
+            attn = attn.transpose(1, 0).contiguous()
+            attn = attn.view(T, B, N, C)
+
+            residual = residual.unsqueeze(2)
+            x = self.self_attn.out_proj(attn)
+            x = F.dropout(x, p=self.dropout, training=self.training)
+            x = torch.cat([residual, residual + x], 2)
+
+        else:
+            if context_embeds is None:
+                x = residual
+
+            else:
+                x, _ = self.self_attn(query=x, key=context_embeds, value=context_embeds)
+                x = x.view(T, B, C)
+                x = F.dropout(x, p=self.dropout, training=self.training)
+                x = residual + x
+
+        x = self.maybe_layer_norm(self.self_attn_layer_norm, x, after=True)
+
+        if self.encoder_attn is not None:
+            raise NotImplementedError
+
+        residual = x
+        x = self.maybe_layer_norm(self.final_layer_norm, x, before=True)
+        x = self.activation_fn(self.fc1(x))
+        x = F.dropout(x, p=self.activation_dropout, training=self.training)
+        x = self.fc2(x)
+        x = F.dropout(x, p=self.dropout, training=self.training)
+        x = residual + x
+        x = self.maybe_layer_norm(self.final_layer_norm, x, after=True)
+        return x
+
+
+@register_model_architecture(
+    "nonautoregressive_transformer", "nonautoregressive_transformer"
+)
+def base_architecture(args):
+    args.encoder_embed_path = getattr(args, "encoder_embed_path", None)
+    args.encoder_embed_dim = getattr(args, "encoder_embed_dim", 512)
+    args.encoder_ffn_embed_dim = getattr(args, "encoder_ffn_embed_dim", 2048)
+    args.encoder_layers = getattr(args, "encoder_layers", 6)
+    args.encoder_attention_heads = getattr(args, "encoder_attention_heads", 8)
+    args.encoder_normalize_before = getattr(args, "encoder_normalize_before", False)
+    args.encoder_learned_pos = getattr(args, "encoder_learned_pos", False)
+    args.decoder_embed_path = getattr(args, "decoder_embed_path", None)
+    args.decoder_embed_dim = getattr(args, "decoder_embed_dim", args.encoder_embed_dim)
+    args.decoder_ffn_embed_dim = getattr(
+        args, "decoder_ffn_embed_dim", args.encoder_ffn_embed_dim
+    )
+    args.decoder_layers = getattr(args, "decoder_layers", 6)
+    args.decoder_attention_heads = getattr(args, "decoder_attention_heads", 8)
+    args.decoder_normalize_before = getattr(args, "decoder_normalize_before", False)
+    args.decoder_learned_pos = getattr(args, "decoder_learned_pos", False)
+    args.attention_dropout = getattr(args, "attention_dropout", 0.0)
+    args.activation_dropout = getattr(args, "activation_dropout", 0.0)
+    args.activation_fn = getattr(args, "activation_fn", "relu")
+    args.dropout = getattr(args, "dropout", 0.1)
+    args.adaptive_softmax_cutoff = getattr(args, "adaptive_softmax_cutoff", None)
+    args.adaptive_softmax_dropout = getattr(args, "adaptive_softmax_dropout", 0)
+    args.share_decoder_input_output_embed = getattr(
+        args, "share_decoder_input_output_embed", False
+    )
+    args.share_all_embeddings = getattr(args, "share_all_embeddings", False)
+    args.no_token_positional_embeddings = getattr(
+        args, "no_token_positional_embeddings", False
+    )
+    args.adaptive_input = getattr(args, "adaptive_input", False)
+    args.apply_bert_init = getattr(args, "apply_bert_init", False)
+
+    args.decoder_output_dim = getattr(
+        args, "decoder_output_dim", args.decoder_embed_dim
+    )
+    args.decoder_input_dim = getattr(args, "decoder_input_dim", args.decoder_embed_dim)
+
+    # --- special arguments ---
+    args.sg_length_pred = getattr(args, "sg_length_pred", False)
+    args.pred_length_offset = getattr(args, "pred_length_offset", False)
+    args.length_loss_factor = getattr(args, "length_loss_factor", 0.1)
+    args.src_embedding_copy = getattr(args, "src_embedding_copy", False)
+    args.ngram_predictor = getattr(args, "ngram_predictor", 1)
+
+
+@register_model_architecture(
+    "nonautoregressive_transformer", "nonautoregressive_transformer_wmt_en_de"
+)
+def nonautoregressive_transformer_wmt_en_de(args):
+    base_architecture(args)

fairseq/models/transformer.py

@@ -172,7 +172,7 @@ class TransformerModel(FairseqEncoderDecoderModel):
 
         encoder = cls.build_encoder(args, src_dict, encoder_embed_tokens)
         decoder = cls.build_decoder(args, tgt_dict, decoder_embed_tokens)
-        return TransformerModel(encoder, decoder)
+        return cls(encoder, decoder)
 
     @classmethod
     def build_encoder(cls, args, src_dict, embed_tokens):
@@ -222,7 +222,15 @@ class TransformerEncoder(FairseqEncoder):
         else:
             self.layer_norm = None
 
-    def forward(self, src_tokens, src_lengths):
+    def forward_embedding(self, src_tokens):
+        # embed tokens and positions
+        embed = self.embed_scale * self.embed_tokens(src_tokens)
+        if self.embed_positions is not None:
+            x = embed + self.embed_positions(src_tokens)
+        x = F.dropout(x, p=self.dropout, training=self.training)
+        return x, embed
+
+    def forward(self, src_tokens, src_lengths, cls_input=None):
         """
         Args:
             src_tokens (LongTensor): tokens in the source language of shape
@@ -237,11 +245,7 @@ class TransformerEncoder(FairseqEncoder):
                 - **encoder_padding_mask** (ByteTensor): the positions of
                   padding elements of shape `(batch, src_len)`
         """
-        # embed tokens and positions
-        x = self.embed_scale * self.embed_tokens(src_tokens)
-        if self.embed_positions is not None:
-            x += self.embed_positions(src_tokens)
-        x = F.dropout(x, p=self.dropout, training=self.training)
+        x, encoder_embedding = self.forward_embedding(src_tokens)
 
         # B x T x C -> T x B x C
         x = x.transpose(0, 1)
@@ -261,6 +265,7 @@ class TransformerEncoder(FairseqEncoder):
         return {
             'encoder_out': x,  # T x B x C
             'encoder_padding_mask': encoder_padding_mask,  # B x T
+            'encoder_embedding': encoder_embedding,  # B x T x C
         }
 
     def reorder_encoder_out(self, encoder_out, new_order):
@@ -332,7 +337,7 @@ class TransformerDecoder(FairseqIncrementalDecoder):
         embed_dim = args.decoder_embed_dim
         self.output_embed_dim = args.decoder_output_dim
 
-        padding_idx = embed_tokens.padding_idx
+        self.padding_idx = embed_tokens.padding_idx
         self.max_target_positions = args.max_target_positions
 
         self.embed_tokens = embed_tokens
@@ -341,7 +346,7 @@ class TransformerDecoder(FairseqIncrementalDecoder):
         self.project_in_dim = Linear(input_embed_dim, embed_dim, bias=False) if embed_dim != input_embed_dim else None
 
         self.embed_positions = PositionalEmbedding(
-            args.max_target_positions, embed_dim, padding_idx,
+            args.max_target_positions, embed_dim, self.padding_idx,
             learned=args.decoder_learned_pos,
         ) if not args.no_token_positional_embeddings else None
 

fairseq/modules/multihead_attention.py

@@ -91,7 +91,7 @@ class MultiheadAttention(nn.Module):
             nn.init.xavier_normal_(self.bias_v)
 
     def forward(self, query, key, value, key_padding_mask=None, incremental_state=None,
-                need_weights=True, static_kv=False, attn_mask=None):
+                need_weights=True, static_kv=False, attn_mask=None, before_softmax=False):
         """Input shape: Time x Batch x Channel
 
         Timesteps can be masked by supplying a T x T mask in the
@@ -239,6 +239,9 @@ class MultiheadAttention(nn.Module):
                 )
             attn_weights = attn_weights.view(bsz * self.num_heads, tgt_len, src_len)
 
+        if before_softmax:
+            return attn_weights, v
+
         attn_weights = utils.softmax(
             attn_weights, dim=-1, onnx_trace=self.onnx_trace,
         ).type_as(attn_weights)

fairseq/modules/transformer_layer.py

@@ -83,7 +83,7 @@ class TransformerEncoderLayer(nn.Module):
         residual = x
         x = self.maybe_layer_norm(self.self_attn_layer_norm, x, before=True)
         if attn_mask is not None:
-            attn_mask = attn_mask.masked_fill(attn_mask.byte(), -1e8)
+            attn_mask = attn_mask.masked_fill(attn_mask.bool(), -1e8)
         # anything in original attn_mask = 1, becomes -1e8
         # anything in original attn_mask = 0, becomes 0
         # Note that we cannot use -inf here, because at some edge cases,

fairseq/modules/transformer_sentence_encoder.py

@@ -36,7 +36,8 @@ def init_bert_params(module):
             module.bias.data.zero_()
     if isinstance(module, nn.Embedding):
         module.weight.data.normal_(mean=0.0, std=0.02)
-        module.weight.data[module.padding_idx].zero_()
+        if module.padding_idx is not None:
+            module.weight.data[module.padding_idx].zero_()
     if isinstance(module, MultiheadAttention):
         module.in_proj_weight.data.normal_(mean=0.0, std=0.02)
 

fairseq/options.py

@@ -280,6 +280,8 @@ def add_dataset_args(parser, train=False, gen=False):
                                 ' (train, valid, valid1, test, test1)')
         group.add_argument('--validate-interval', type=int, default=1, metavar='N',
                            help='validate every N epochs')
+        group.add_argument('--fixed-validation-seed', default=None, type=int, metavar='N',
+                           help='specified random seed for validation')
         group.add_argument('--disable-validation', action='store_true',
                            help='disable validation')
         group.add_argument('--max-tokens-valid', type=int, metavar='N',
@@ -493,6 +495,18 @@ def add_generation_args(parser):
                        help='strength of diversity penalty for Diverse Beam Search')
     group.add_argument('--print-alignment', action='store_true',
                        help='if set, uses attention feedback to compute and print alignment to source tokens')
+    group.add_argument('--print-step', action='store_true')
+
+    # arguments for iterative refinement generator
+    group.add_argument('---iter-decode-eos-penalty', default=0.0, type=float, metavar='N',
+                       help='if > 0.0, it penalized early-stopping in decoding.')
+    group.add_argument('--iter-decode-max-iter', default=10, type=int, metavar='N',
+                       help='maximum iterations for iterative refinement.')
+    group.add_argument('--iter-decode-force-max-iter', action='store_true',
+                       help='if set, run exact the maximum number of iterations without early stop')
+
+    # special decoding format for advanced decoding.
+    group.add_argument('--decoding-format', default=None, type=str, choices=['unigram', 'ensemble', 'vote', 'dp', 'bs'])
     # fmt: on
     return group
 

fairseq/tasks/translation.py

@@ -12,6 +12,7 @@ from fairseq.data import (
     data_utils,
     indexed_dataset,
     LanguagePairDataset,
+    PrependTokenDataset,
 )
 
 from . import FairseqTask, register_task
@@ -22,7 +23,8 @@ def load_langpair_dataset(
     src, src_dict,
     tgt, tgt_dict,
     combine, dataset_impl, upsample_primary,
-    left_pad_source, left_pad_target, max_source_positions, max_target_positions,
+    left_pad_source, left_pad_target, max_source_positions,
+    max_target_positions, prepend_bos=False,
 ):
     def split_exists(split, src, tgt, lang, data_path):
         filename = os.path.join(data_path, '{}.{}-{}.{}'.format(split, src, tgt, lang))
@@ -67,6 +69,11 @@ def load_langpair_dataset(
         src_dataset = ConcatDataset(src_datasets, sample_ratios)
         tgt_dataset = ConcatDataset(tgt_datasets, sample_ratios)
 
+    if prepend_bos:
+        assert hasattr(src_dict, "bos_index") and hasattr(tgt_dict, "bos_index")
+        src_dataset = PrependTokenDataset(src_dataset, src_dict.bos())
+        tgt_dataset = PrependTokenDataset(tgt_dataset, tgt_dict.bos())
+
     return LanguagePairDataset(
         src_dataset, src_dataset.sizes, src_dict,
         tgt_dataset, tgt_dataset.sizes, tgt_dict,

fairseq/tasks/translation_lev.py

@@ -0,0 +1,149 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+#
+# This source code is licensed under the MIT license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+
+from fairseq.tasks import register_task
+from fairseq.tasks.translation import TranslationTask, load_langpair_dataset
+
+
+@register_task('translation_lev')
+class TranslationLevenshteinTask(TranslationTask):
+    """
+    Translation (Sequence Generation) task for Levenshtein Transformer
+    See `"Levenshtein Transformer" <https://arxiv.org/abs/1905.11006>`_.
+    """
+
+    @staticmethod
+    def add_args(parser):
+        """Add task-specific arguments to the parser."""
+        # fmt: off
+        TranslationTask.add_args(parser)
+        parser.add_argument(
+            '--noise',
+            default='random_delete',
+            choices=['random_delete', 'random_mask', 'no_noise', 'full_mask'])
+
+    def load_dataset(self, split, epoch=0, combine=False, **kwargs):
+        """Load a given dataset split.
+
+        Args:
+            split (str): name of the split (e.g., train, valid, test)
+        """
+        paths = self.args.data.split(':')
+        assert len(paths) > 0
+        data_path = paths[epoch % len(paths)]
+
+        # infer langcode
+        src, tgt = self.args.source_lang, self.args.target_lang
+
+        self.datasets[split] = load_langpair_dataset(
+            data_path, split, src, self.src_dict, tgt, self.tgt_dict,
+            combine=combine, dataset_impl=self.args.dataset_impl,
+            upsample_primary=self.args.upsample_primary,
+            left_pad_source=self.args.left_pad_source,
+            left_pad_target=self.args.left_pad_target,
+            max_source_positions=self.args.max_source_positions,
+            max_target_positions=self.args.max_target_positions,
+            prepend_bos=True,
+        )
+
+    def inject_noise(self, target_tokens):
+        def _random_delete(target_tokens):
+            pad = self.tgt_dict.pad()
+            bos = self.tgt_dict.bos()
+            eos = self.tgt_dict.eos()
+
+            max_len = target_tokens.size(1)
+            target_mask = target_tokens.eq(pad)
+            target_score = target_tokens.clone().float().uniform_()
+            target_score.masked_fill_(
+                target_tokens.eq(bos) | target_tokens.eq(eos), 0.0)
+            target_score.masked_fill_(target_mask, 1)
+            target_score, target_rank = target_score.sort(1)
+            target_length = target_mask.size(1) - target_mask.float().sum(
+                1, keepdim=True)
+
+            # do not delete <bos> and <eos> (we assign 0 score for them)
+            target_cutoff = 2 + ((target_length - 2) * target_score.new_zeros(
+                target_score.size(0), 1).uniform_()).long()
+            target_cutoff = target_score.sort(1)[1] >= target_cutoff
+
+            prev_target_tokens = target_tokens.gather(
+                1, target_rank).masked_fill_(target_cutoff, pad).gather(
+                    1,
+                    target_rank.masked_fill_(target_cutoff,
+                                             max_len).sort(1)[1])
+            prev_target_tokens = prev_target_tokens[:, :prev_target_tokens.
+                                                    ne(pad).sum(1).max()]
+
+            return prev_target_tokens
+
+        def _random_mask(target_tokens):
+            pad = self.tgt_dict.pad()
+            bos = self.tgt_dict.bos()
+            eos = self.tgt_dict.eos()
+            unk = self.tgt_dict.unk()
+
+            target_mask = target_tokens.eq(bos) | target_tokens.eq(
+                eos) | target_tokens.eq(pad)
+            target_score = target_tokens.clone().float().uniform_()
+            target_score.masked_fill_(target_mask, 1.0)
+
+            prev_target_tokens = target_tokens.masked_fill(
+                target_score < target_score.new_zeros(target_score.size(0),
+                                                      1).uniform_(), unk)
+            return prev_target_tokens
+
+        def _full_mask(target_tokens):
+            pad = self.tgt_dict.pad()
+            bos = self.tgt_dict.bos()
+            eos = self.tgt_dict.eos()
+            unk = self.tgt_dict.unk()
+
+            target_mask = target_tokens.eq(bos) | target_tokens.eq(
+                eos) | target_tokens.eq(pad)
+            return target_tokens.masked_fill(~target_mask, unk)
+
+        if self.args.noise == 'random_delete':
+            return _random_delete(target_tokens)
+        elif self.args.noise == 'random_mask':
+            return _random_mask(target_tokens)
+        elif self.args.noise == 'full_mask':
+            return _full_mask(target_tokens)
+        elif self.args.noise == 'no_noise':
+            return target_tokens
+        else:
+            raise NotImplementedError
+
+    def build_generator(self, args):
+        from fairseq.iterative_refinement_generator import IterativeRefinementGenerator
+        return IterativeRefinementGenerator(
+            self.target_dictionary,
+            eos_penalty=getattr(args, 'iter_decode_eos_penalty', 0.0),
+            max_iter=getattr(args, 'iter_decode_max_iter', 10),
+            decoding_format=getattr(args, 'decoding_format', None),
+            adaptive=not getattr(args, 'iter_decode_force_max_iter', False))
+
+    def train_step(self,
+                   sample,
+                   model,
+                   criterion,
+                   optimizer,
+                   ignore_grad=False):
+        model.train()
+        sample['prev_target'] = self.inject_noise(sample['target'])
+        loss, sample_size, logging_output = criterion(model, sample)
+        if ignore_grad:
+            loss *= 0
+        optimizer.backward(loss)
+        return loss, sample_size, logging_output
+
+    def valid_step(self, sample, model, criterion):
+        model.eval()
+        with torch.no_grad():
+            sample['prev_target'] = self.inject_noise(sample['target'])
+            loss, sample_size, logging_output = criterion(model, sample)
+        return loss, sample_size, logging_output

fairseq/utils.py

@@ -359,3 +359,11 @@ def has_parameters(module):
         return True
     except StopIteration:
         return False
+
+
+def set_torch_seed(seed):
+    # Set seed based on args.seed and the update number so that we get
+    # reproducible results when resuming from checkpoints
+    assert isinstance(seed, int)
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)

generate.py

@@ -159,6 +159,9 @@ def main(args):
                                 ' '.join(map(lambda x: str(utils.item(x)), alignment))
                             ))
 
+                        if args.print_step:
+                            print('I-{}\t{}'.format(sample_id, hypo['steps']))
+
                     # Score only the top hypothesis
                     if has_target and j == 0:
                         if align_dict is not None or args.remove_bpe is not None:

setup.py

@@ -5,6 +5,7 @@
 # LICENSE file in the root directory of this source tree.
 
 from setuptools import setup, find_packages, Extension
+from torch.utils import cpp_extension
 import sys
 
 
@@ -60,6 +61,12 @@ extensions = [
         language='c++',
         extra_compile_args=extra_compile_args,
     ),
+    cpp_extension.CppExtension(
+        'fairseq.libnat',
+        sources=[
+            'fairseq/clib/libnat/edit_dist.cpp',
+        ],
+    )
 ]
 
 
@@ -106,5 +113,6 @@ setup(
             'fairseq-validate = fairseq_cli.validate:cli_main',
         ],
     },
+    cmdclass={'build_ext': cpp_extension.BuildExtension},
     zip_safe=False,
 )

tests/test_binaries.py

@@ -180,6 +180,52 @@ class TestTranslation(unittest.TestCase):
                 ])
                 generate_main(data_dir)
 
+    def test_levenshtein_transformer(self):
+        with contextlib.redirect_stdout(StringIO()):
+            with tempfile.TemporaryDirectory('test_levenshtein_transformer') as data_dir:
+                create_dummy_data(data_dir)
+                preprocess_translation_data(data_dir)
+                train_translation_model(data_dir, 'levenshtein_transformer', [
+                    '--apply-bert-init', '--early-exit', '6,6,6',
+                    '--criterion', 'nat_loss'
+                ], task='translation_lev')
+                generate_main(data_dir)
+
+    def test_nonautoregressive_transformer(self):
+        with contextlib.redirect_stdout(StringIO()):
+            with tempfile.TemporaryDirectory('test_nonautoregressive_transformer') as data_dir:
+                create_dummy_data(data_dir)
+                preprocess_translation_data(data_dir)
+                train_translation_model(data_dir, 'nonautoregressive_transformer', [
+                    '--apply-bert-init', '--src-embedding-copy', '--criterion',
+                    'nat_loss', '--noise', 'full_mask', '--pred-length-offset',
+                    '--length-loss-factor', '0.1'
+                ], task='translation_lev')
+                generate_main(data_dir)
+
+    def test_iterative_nonautoregressive_transformer(self):
+        with contextlib.redirect_stdout(StringIO()):
+            with tempfile.TemporaryDirectory('test_iterative_nonautoregressive_transformer') as data_dir:
+                create_dummy_data(data_dir)
+                preprocess_translation_data(data_dir)
+                train_translation_model(data_dir, 'iterative_nonautoregressive_transformer', [
+                    '--apply-bert-init', '--src-embedding-copy', '--criterion',
+                    'nat_loss', '--noise', 'full_mask', '--stochastic-approx',
+                    '--dae-ratio', '0.5', '--train-step', '3'
+                ], task='translation_lev')
+                generate_main(data_dir)
+
+    def test_insertion_transformer(self):
+        with contextlib.redirect_stdout(StringIO()):
+            with tempfile.TemporaryDirectory('test_insertion_transformer') as data_dir:
+                create_dummy_data(data_dir)
+                preprocess_translation_data(data_dir)
+                train_translation_model(data_dir, 'insertion_transformer', [
+                    '--apply-bert-init', '--criterion', 'nat_loss', '--noise',
+                    'random_mask'
+                ], task='translation_lev')
+                generate_main(data_dir)
+
     def test_mixture_of_experts(self):
         with contextlib.redirect_stdout(StringIO()):
             with tempfile.TemporaryDirectory('test_moe') as data_dir:

train.py

@@ -194,6 +194,11 @@ def get_training_stats(trainer):
 
 def validate(args, trainer, task, epoch_itr, subsets):
     """Evaluate the model on the validation set(s) and return the losses."""
+
+    if args.fixed_validation_seed is not None:
+        # set fixed seed for every validation
+        utils.set_torch_seed(args.fixed_validation_seed)
+
     valid_losses = []
     for subset in subsets:
         # Initialize data iterator