init

.gitattributes

@@ -1,35 +1,27 @@
 *.7z filter=lfs diff=lfs merge=lfs -text
 *.arrow filter=lfs diff=lfs merge=lfs -text
 *.bin filter=lfs diff=lfs merge=lfs -text
+*.bin.* filter=lfs diff=lfs merge=lfs -text
 *.bz2 filter=lfs diff=lfs merge=lfs -text
-*.ckpt filter=lfs diff=lfs merge=lfs -text
 *.ftz filter=lfs diff=lfs merge=lfs -text
 *.gz filter=lfs diff=lfs merge=lfs -text
 *.h5 filter=lfs diff=lfs merge=lfs -text
 *.joblib filter=lfs diff=lfs merge=lfs -text
 *.lfs.* filter=lfs diff=lfs merge=lfs -text
-*.mlmodel filter=lfs diff=lfs merge=lfs -text
 *.model filter=lfs diff=lfs merge=lfs -text
 *.msgpack filter=lfs diff=lfs merge=lfs -text
-*.npy filter=lfs diff=lfs merge=lfs -text
-*.npz filter=lfs diff=lfs merge=lfs -text
 *.onnx filter=lfs diff=lfs merge=lfs -text
 *.ot filter=lfs diff=lfs merge=lfs -text
 *.parquet filter=lfs diff=lfs merge=lfs -text
 *.pb filter=lfs diff=lfs merge=lfs -text
-*.pickle filter=lfs diff=lfs merge=lfs -text
-*.pkl filter=lfs diff=lfs merge=lfs -text
 *.pt filter=lfs diff=lfs merge=lfs -text
 *.pth filter=lfs diff=lfs merge=lfs -text
 *.rar filter=lfs diff=lfs merge=lfs -text
-*.safetensors filter=lfs diff=lfs merge=lfs -text
 saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.tar.* filter=lfs diff=lfs merge=lfs -text
-*.tar filter=lfs diff=lfs merge=lfs -text
 *.tflite filter=lfs diff=lfs merge=lfs -text
 *.tgz filter=lfs diff=lfs merge=lfs -text
-*.wasm filter=lfs diff=lfs merge=lfs -text
 *.xz filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
-*.zst filter=lfs diff=lfs merge=lfs -text
+*.zstandard filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1 @@
+.idea

LICENSE

@@ -0,0 +1,201 @@
+                                 Apache License
+                           Version 2.0, January 2004
+                        http://www.apache.org/licenses/
+
+   TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION
+
+   1. Definitions.
+
+      "License" shall mean the terms and conditions for use, reproduction,
+      and distribution as defined by Sections 1 through 9 of this document.
+
+      "Licensor" shall mean the copyright owner or entity authorized by
+      the copyright owner that is granting the License.
+
+      "Legal Entity" shall mean the union of the acting entity and all
+      other entities that control, are controlled by, or are under common
+      control with that entity. For the purposes of this definition,
+      "control" means (i) the power, direct or indirect, to cause the
+      direction or management of such entity, whether by contract or
+      otherwise, or (ii) ownership of fifty percent (50%) or more of the
+      outstanding shares, or (iii) beneficial ownership of such entity.
+
+      "You" (or "Your") shall mean an individual or Legal Entity
+      exercising permissions granted by this License.
+
+      "Source" form shall mean the preferred form for making modifications,
+      including but not limited to software source code, documentation
+      source, and configuration files.
+
+      "Object" form shall mean any form resulting from mechanical
+      transformation or translation of a Source form, including but
+      not limited to compiled object code, generated documentation,
+      and conversions to other media types.
+
+      "Work" shall mean the work of authorship, whether in Source or
+      Object form, made available under the License, as indicated by a
+      copyright notice that is included in or attached to the work
+      (an example is provided in the Appendix below).
+
+      "Derivative Works" shall mean any work, whether in Source or Object
+      form, that is based on (or derived from) the Work and for which the
+      editorial revisions, annotations, elaborations, or other modifications
+      represent, as a whole, an original work of authorship. For the purposes
+      of this License, Derivative Works shall not include works that remain
+      separable from, or merely link (or bind by name) to the interfaces of,
+      the Work and Derivative Works thereof.
+
+      "Contribution" shall mean any work of authorship, including
+      the original version of the Work and any modifications or additions
+      to that Work or Derivative Works thereof, that is intentionally
+      submitted to Licensor for inclusion in the Work by the copyright owner
+      or by an individual or Legal Entity authorized to submit on behalf of
+      the copyright owner. For the purposes of this definition, "submitted"
+      means any form of electronic, verbal, or written communication sent
+      to the Licensor or its representatives, including but not limited to
+      communication on electronic mailing lists, source code control systems,
+      and issue tracking systems that are managed by, or on behalf of, the
+      Licensor for the purpose of discussing and improving the Work, but
+      excluding communication that is conspicuously marked or otherwise
+      designated in writing by the copyright owner as "Not a Contribution."
+
+      "Contributor" shall mean Licensor and any individual or Legal Entity
+      on behalf of whom a Contribution has been received by Licensor and
+      subsequently incorporated within the Work.
+
+   2. Grant of Copyright License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      copyright license to reproduce, prepare Derivative Works of,
+      publicly display, publicly perform, sublicense, and distribute the
+      Work and such Derivative Works in Source or Object form.
+
+   3. Grant of Patent License. Subject to the terms and conditions of
+      this License, each Contributor hereby grants to You a perpetual,
+      worldwide, non-exclusive, no-charge, royalty-free, irrevocable
+      (except as stated in this section) patent license to make, have made,
+      use, offer to sell, sell, import, and otherwise transfer the Work,
+      where such license applies only to those patent claims licensable
+      by such Contributor that are necessarily infringed by their
+      Contribution(s) alone or by combination of their Contribution(s)
+      with the Work to which such Contribution(s) was submitted. If You
+      institute patent litigation against any entity (including a
+      cross-claim or counterclaim in a lawsuit) alleging that the Work
+      or a Contribution incorporated within the Work constitutes direct
+      or contributory patent infringement, then any patent licenses
+      granted to You under this License for that Work shall terminate
+      as of the date such litigation is filed.
+
+   4. Redistribution. You may reproduce and distribute copies of the
+      Work or Derivative Works thereof in any medium, with or without
+      modifications, and in Source or Object form, provided that You
+      meet the following conditions:
+
+      (a) You must give any other recipients of the Work or
+          Derivative Works a copy of this License; and
+
+      (b) You must cause any modified files to carry prominent notices
+          stating that You changed the files; and
+
+      (c) You must retain, in the Source form of any Derivative Works
+          that You distribute, all copyright, patent, trademark, and
+          attribution notices from the Source form of the Work,
+          excluding those notices that do not pertain to any part of
+          the Derivative Works; and
+
+      (d) If the Work includes a "NOTICE" text file as part of its
+          distribution, then any Derivative Works that You distribute must
+          include a readable copy of the attribution notices contained
+          within such NOTICE file, excluding those notices that do not
+          pertain to any part of the Derivative Works, in at least one
+          of the following places: within a NOTICE text file distributed
+          as part of the Derivative Works; within the Source form or
+          documentation, if provided along with the Derivative Works; or,
+          within a display generated by the Derivative Works, if and
+          wherever such third-party notices normally appear. The contents
+          of the NOTICE file are for informational purposes only and
+          do not modify the License. You may add Your own attribution
+          notices within Derivative Works that You distribute, alongside
+          or as an addendum to the NOTICE text from the Work, provided
+          that such additional attribution notices cannot be construed
+          as modifying the License.
+
+      You may add Your own copyright statement to Your modifications and
+      may provide additional or different license terms and conditions
+      for use, reproduction, or distribution of Your modifications, or
+      for any such Derivative Works as a whole, provided Your use,
+      reproduction, and distribution of the Work otherwise complies with
+      the conditions stated in this License.
+
+   5. Submission of Contributions. Unless You explicitly state otherwise,
+      any Contribution intentionally submitted for inclusion in the Work
+      by You to the Licensor shall be under the terms and conditions of
+      this License, without any additional terms or conditions.
+      Notwithstanding the above, nothing herein shall supersede or modify
+      the terms of any separate license agreement you may have executed
+      with Licensor regarding such Contributions.
+
+   6. Trademarks. This License does not grant permission to use the trade
+      names, trademarks, service marks, or product names of the Licensor,
+      except as required for reasonable and customary use in describing the
+      origin of the Work and reproducing the content of the NOTICE file.
+
+   7. Disclaimer of Warranty. Unless required by applicable law or
+      agreed to in writing, Licensor provides the Work (and each
+      Contributor provides its Contributions) on an "AS IS" BASIS,
+      WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or
+      implied, including, without limitation, any warranties or conditions
+      of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A
+      PARTICULAR PURPOSE. You are solely responsible for determining the
+      appropriateness of using or redistributing the Work and assume any
+      risks associated with Your exercise of permissions under this License.
+
+   8. Limitation of Liability. In no event and under no legal theory,
+      whether in tort (including negligence), contract, or otherwise,
+      unless required by applicable law (such as deliberate and grossly
+      negligent acts) or agreed to in writing, shall any Contributor be
+      liable to You for damages, including any direct, indirect, special,
+      incidental, or consequential damages of any character arising as a
+      result of this License or out of the use or inability to use the
+      Work (including but not limited to damages for loss of goodwill,
+      work stoppage, computer failure or malfunction, or any and all
+      other commercial damages or losses), even if such Contributor
+      has been advised of the possibility of such damages.
+
+   9. Accepting Warranty or Additional Liability. While redistributing
+      the Work or Derivative Works thereof, You may choose to offer,
+      and charge a fee for, acceptance of support, warranty, indemnity,
+      or other liability obligations and/or rights consistent with this
+      License. However, in accepting such obligations, You may act only
+      on Your own behalf and on Your sole responsibility, not on behalf
+      of any other Contributor, and only if You agree to indemnify,
+      defend, and hold each Contributor harmless for any liability
+      incurred by, or claims asserted against, such Contributor by reason
+      of your accepting any such warranty or additional liability.
+
+   END OF TERMS AND CONDITIONS
+
+   APPENDIX: How to apply the Apache License to your work.
+
+      To apply the Apache License to your work, attach the following
+      boilerplate notice, with the fields enclosed by brackets "[]"
+      replaced with your own identifying information. (Don't include
+      the brackets!)  The text should be enclosed in the appropriate
+      comment syntax for the file format. We also recommend that a
+      file or class name and description of purpose be included on the
+      same "printed page" as the copyright notice for easier
+      identification within third-party archives.
+
+   Copyright 1999-2022 Alibaba Group Holding Ltd.
+
+   Licensed under the Apache License, Version 2.0 (the "License");
+   you may not use this file except in compliance with the License.
+   You may obtain a copy of the License at
+
+       http://www.apache.org/licenses/LICENSE-2.0
+
+   Unless required by applicable law or agreed to in writing, software
+   distributed under the License is distributed on an "AS IS" BASIS,
+   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+   See the License for the specific language governing permissions and
+   limitations under the License.

README.md

@@ -1,12 +1,27 @@
 ---
-title: OFA
-emoji: ⚡
-colorFrom: green
-colorTo: blue
+title: OFA-Visual_Grounding
+emoji: 👀
+colorFrom: yellow
+colorTo: green
 sdk: gradio
-sdk_version: 3.39.0
 app_file: app.py
 pinned: false
+duplicated_from: OFA-Sys/OFA-Visual_Grounding
 ---
+# Configuration
+`title`: _string_
+OFA Image Caption
+`emoji`: _string_
+🖼
+`colorFrom`: _string_
+red
+`colorTo`: _string_
+indigo
+`sdk`: _string_
+gradio
+`app_file`: _string_
+app.py
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+
+`pinned`: _boolean_
+false

app.py

@@ -0,0 +1,155 @@
+import os
+
+os.system('cd fairseq;'
+          'pip install ./; cd ..')
+os.system('ls -l')
+
+import torch
+import numpy as np
+from fairseq import utils, tasks
+from fairseq import checkpoint_utils
+from utils.eval_utils import eval_step
+from tasks.mm_tasks.refcoco import RefcocoTask
+from models.ofa import OFAModel
+from PIL import Image
+from torchvision import transforms
+import cv2
+import gradio as gr
+
+# Register refcoco task
+tasks.register_task('refcoco', RefcocoTask)
+
+# turn on cuda if GPU is available
+use_cuda = torch.cuda.is_available()
+# use fp16 only when GPU is available
+use_fp16 = False
+
+os.system('wget https://ofa-silicon.oss-us-west-1.aliyuncs.com/checkpoints/refcocog_large_best.pt; '
+          'mkdir -p checkpoints; mv refcocog_large_best.pt checkpoints/refcocog.pt')
+
+# Load pretrained ckpt & config
+overrides = {"bpe_dir": "utils/BPE", "eval_cider": False, "beam": 5,
+             "max_len_b": 16, "no_repeat_ngram_size": 3, "seed": 7}
+models, cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+    utils.split_paths('checkpoints/refcocog.pt'),
+    arg_overrides=overrides
+)
+
+cfg.common.seed = 7
+cfg.generation.beam = 5
+cfg.generation.min_len = 4
+cfg.generation.max_len_a = 0
+cfg.generation.max_len_b = 4
+cfg.generation.no_repeat_ngram_size = 3
+
+# Fix seed for stochastic decoding
+if cfg.common.seed is not None and not cfg.generation.no_seed_provided:
+    np.random.seed(cfg.common.seed)
+    utils.set_torch_seed(cfg.common.seed)
+
+# Move models to GPU
+for model in models:
+    model.eval()
+    if use_fp16:
+        model.half()
+    if use_cuda and not cfg.distributed_training.pipeline_model_parallel:
+        model.cuda()
+    model.prepare_for_inference_(cfg)
+
+# Initialize generator
+generator = task.build_generator(models, cfg.generation)
+
+mean = [0.5, 0.5, 0.5]
+std = [0.5, 0.5, 0.5]
+
+patch_resize_transform = transforms.Compose([
+    lambda image: image.convert("RGB"),
+    transforms.Resize((cfg.task.patch_image_size, cfg.task.patch_image_size), interpolation=Image.BICUBIC),
+    transforms.ToTensor(),
+    transforms.Normalize(mean=mean, std=std),
+])
+
+# Text preprocess
+bos_item = torch.LongTensor([task.src_dict.bos()])
+eos_item = torch.LongTensor([task.src_dict.eos()])
+pad_idx = task.src_dict.pad()
+
+
+def encode_text(text, length=None, append_bos=False, append_eos=False):
+    s = task.tgt_dict.encode_line(
+        line=task.bpe.encode(text),
+        add_if_not_exist=False,
+        append_eos=False
+    ).long()
+    if length is not None:
+        s = s[:length]
+    if append_bos:
+        s = torch.cat([bos_item, s])
+    if append_eos:
+        s = torch.cat([s, eos_item])
+    return s
+
+
+patch_image_size = cfg.task.patch_image_size
+
+
+def construct_sample(image: Image, text: str):
+    w, h = image.size
+    w_resize_ratio = torch.tensor(patch_image_size / w).unsqueeze(0)
+    h_resize_ratio = torch.tensor(patch_image_size / h).unsqueeze(0)
+    patch_image = patch_resize_transform(image).unsqueeze(0)
+    patch_mask = torch.tensor([True])
+    src_text = encode_text(' which region does the text " {} " describe?'.format(text), append_bos=True,
+                           append_eos=True).unsqueeze(0)
+    src_length = torch.LongTensor([s.ne(pad_idx).long().sum() for s in src_text])
+    sample = {
+        "id": np.array(['42']),
+        "net_input": {
+            "src_tokens": src_text,
+            "src_lengths": src_length,
+            "patch_images": patch_image,
+            "patch_masks": patch_mask,
+        },
+        "w_resize_ratios": w_resize_ratio,
+        "h_resize_ratios": h_resize_ratio,
+        "region_coords": torch.randn(1, 4)
+    }
+    return sample
+
+
+# Function to turn FP32 to FP16
+def apply_half(t):
+    if t.dtype is torch.float32:
+        return t.to(dtype=torch.half)
+    return t
+
+
+# Function for visual grounding
+def visual_grounding(Image, Text):
+    sample = construct_sample(Image, Text.lower())
+    sample = utils.move_to_cuda(sample) if use_cuda else sample
+    sample = utils.apply_to_sample(apply_half, sample) if use_fp16 else sample
+    with torch.no_grad():
+        result, scores = eval_step(task, generator, models, sample)
+    img = np.asarray(Image)
+    cv2.rectangle(
+        img,
+        (int(result[0]["box"][0]), int(result[0]["box"][1])),
+        (int(result[0]["box"][2]), int(result[0]["box"][3])),
+        (0, 255, 0),
+        3
+    )
+    return img
+
+
+title = "OFA Visual Grounding"
+description = "Démonstration pour OFA Visual Grounding. Téléchargez votre image ou cliquez sur l'un des exemples, et rédigez une description concernant un objet spécifique."
+
+examples = [['test-1.jpeg', 'black chair'],
+            ['test-2.jpeg', 'orange door'],
+            ['test-3.jpeg', 'fire extinguisher']]
+io = gr.Interface(fn=visual_grounding, inputs=[gr.inputs.Image(type='pil'), "textbox"],
+                  outputs=gr.outputs.Image(type='numpy'),
+                  title=title, description=description, examples=examples,
+                  allow_flagging=False, allow_screenshot=False)
+io.launch()

criterions/__init__.py

@@ -0,0 +1,2 @@
+from .scst_loss import ScstRewardCriterion
+from .label_smoothed_cross_entropy import AjustLabelSmoothedCrossEntropyCriterion

criterions/label_smoothed_cross_entropy.py

@@ -0,0 +1,343 @@
+# 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
+from dataclasses import dataclass, field
+from typing import Optional
+
+import torch
+import torch.nn.functional as F
+import numpy as np
+from fairseq import metrics, utils
+from fairseq.criterions import FairseqCriterion, register_criterion
+from fairseq.dataclass import FairseqDataclass
+from omegaconf import II
+
+
+@dataclass
+class AjustLabelSmoothedCrossEntropyCriterionConfig(FairseqDataclass):
+    label_smoothing: float = field(
+        default=0.0,
+        metadata={"help": "epsilon for label smoothing, 0 means no label smoothing"},
+    )
+    report_accuracy: bool = field(
+        default=False,
+        metadata={"help": "report accuracy metric"},
+    )
+    ignore_prefix_size: int = field(
+        default=0,
+        metadata={"help": "Ignore first N tokens"},
+    )
+    ignore_eos: bool = field(
+        default=False,
+        metadata={"help": "Ignore eos token"},
+    )
+    sentence_avg: bool = II("optimization.sentence_avg")
+    drop_worst_ratio: float = field(
+        default=0.0,
+        metadata={"help": "ratio for discarding bad samples"},
+    )
+    drop_worst_after: int = field(
+        default=0,
+        metadata={"help": "steps for discarding bad samples"},
+    )
+    use_rdrop: bool = field(
+        default=False, metadata={"help": "use R-Drop"}
+    )
+    reg_alpha: float = field(
+        default=1.0, metadata={"help": "weight for R-Drop"}
+    )
+    sample_patch_num: int = field(
+        default=196, metadata={"help": "sample patchs for v1"}
+    )
+    constraint_range: Optional[str] = field(
+        default=None,
+        metadata={"help": "constraint range"}
+    )
+
+
+def construct_rdrop_sample(x):
+    if isinstance(x, dict):
+        for key in x:
+            x[key] = construct_rdrop_sample(x[key])
+        return x
+    elif isinstance(x, torch.Tensor):
+        return x.repeat(2, *([1] * (x.dim()-1)))
+    elif isinstance(x, int):
+        return x * 2
+    elif isinstance(x, np.ndarray):
+        return x.repeat(2)
+    else:
+        raise NotImplementedError
+
+
+def kl_loss(p, q):
+    p_loss = F.kl_div(p, torch.exp(q), reduction='sum')
+    q_loss = F.kl_div(q, torch.exp(p), reduction='sum')
+    loss = (p_loss + q_loss) / 2
+    return loss
+
+
+def label_smoothed_nll_loss(
+        lprobs, target, epsilon, update_num, reduce=True,
+        drop_worst_ratio=0.0, drop_worst_after=0, use_rdrop=False, reg_alpha=1.0,
+        constraint_masks=None, constraint_start=None, constraint_end=None
+):
+    if target.dim() == lprobs.dim() - 1:
+        target = target.unsqueeze(-1)
+    nll_loss = -lprobs.gather(dim=-1, index=target).squeeze(-1)
+    if constraint_masks is not None:
+        smooth_loss = -lprobs.masked_fill(~constraint_masks, 0).sum(dim=-1, keepdim=True).squeeze(-1)
+        eps_i = epsilon / (constraint_masks.sum(1) - 1 + 1e-6)
+    elif constraint_start is not None and constraint_end is not None:
+        constraint_range = [0, 1, 2, 3] + list(range(constraint_start, constraint_end))
+        smooth_loss = -lprobs[:, constraint_range].sum(dim=-1, keepdim=True).squeeze(-1)
+        eps_i = epsilon / (len(constraint_range) - 1 + 1e-6)
+    else:
+        smooth_loss = -lprobs.sum(dim=-1, keepdim=True).squeeze(-1)
+        eps_i = epsilon / (lprobs.size(-1) - 1)
+    loss = (1.0 - epsilon - eps_i) * nll_loss + eps_i * smooth_loss
+    if drop_worst_ratio > 0 and update_num > drop_worst_after:
+        if use_rdrop:
+            true_batch_size = loss.size(0) // 2
+            _, indices = torch.topk(loss[:true_batch_size], k=int(true_batch_size * (1 - drop_worst_ratio)), largest=False)
+            loss = torch.cat([loss[indices], loss[indices+true_batch_size]])
+            nll_loss = torch.cat([nll_loss[indices], nll_loss[indices+true_batch_size]])
+            lprobs = torch.cat([lprobs[indices], lprobs[indices+true_batch_size]])
+        else:
+            loss, indices = torch.topk(loss, k=int(loss.shape[0] * (1 - drop_worst_ratio)), largest=False)
+            nll_loss = nll_loss[indices]
+            lprobs = lprobs[indices]
+
+    ntokens = loss.numel()
+    nll_loss = nll_loss.sum()
+    loss = loss.sum()
+    if use_rdrop:
+        true_batch_size = lprobs.size(0) // 2
+        p = lprobs[:true_batch_size]
+        q = lprobs[true_batch_size:]
+        if constraint_start is not None and constraint_end is not None:
+            constraint_range = [0, 1, 2, 3] + list(range(constraint_start, constraint_end))
+            p = p[:, constraint_range]
+            q = q[:, constraint_range]
+        loss += kl_loss(p, q) * reg_alpha
+
+    return loss, nll_loss, ntokens
+
+
+@register_criterion(
+    "ajust_label_smoothed_cross_entropy", dataclass=AjustLabelSmoothedCrossEntropyCriterionConfig
+)
+class AjustLabelSmoothedCrossEntropyCriterion(FairseqCriterion):
+    def __init__(
+        self,
+        task,
+        sentence_avg,
+        label_smoothing,
+        ignore_prefix_size=0,
+        ignore_eos=False,
+        report_accuracy=False,
+        drop_worst_ratio=0,
+        drop_worst_after=0,
+        use_rdrop=False,
+        reg_alpha=1.0,
+        sample_patch_num=196,
+        constraint_range=None
+    ):
+        super().__init__(task)
+        self.sentence_avg = sentence_avg
+        self.eps = label_smoothing
+        self.ignore_prefix_size = ignore_prefix_size
+        self.ignore_eos = ignore_eos
+        self.report_accuracy = report_accuracy
+        self.drop_worst_ratio = drop_worst_ratio
+        self.drop_worst_after = drop_worst_after
+        self.use_rdrop = use_rdrop
+        self.reg_alpha = reg_alpha
+        self.sample_patch_num = sample_patch_num
+
+        self.constraint_start = None
+        self.constraint_end = None
+        if constraint_range is not None:
+            constraint_start, constraint_end = constraint_range.split(',')
+            self.constraint_start = int(constraint_start)
+            self.constraint_end = int(constraint_end)
+
+    def forward(self, model, sample, update_num=0, 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
+        """
+        if isinstance(sample, list):
+            if self.sample_patch_num > 0:
+                sample[0]['net_input']['sample_patch_num'] = self.sample_patch_num
+            loss_v1, sample_size_v1, logging_output_v1 = self.forward(model, sample[0], update_num, reduce)
+            loss_v2, sample_size_v2, logging_output_v2 = self.forward(model, sample[1], update_num, reduce)
+            loss = loss_v1 / sample_size_v1 + loss_v2 / sample_size_v2
+            sample_size = 1
+            logging_output = {
+                "loss": loss.data,
+                "loss_v1": loss_v1.data,
+                "loss_v2": loss_v2.data,
+                "nll_loss": logging_output_v1["nll_loss"].data / sample_size_v1 + logging_output_v2["nll_loss"].data / sample_size_v2,
+                "ntokens": logging_output_v1["ntokens"] + logging_output_v2["ntokens"],
+                "nsentences": logging_output_v1["nsentences"] + logging_output_v2["nsentences"],
+                "sample_size": 1,
+                "sample_size_v1": sample_size_v1,
+                "sample_size_v2": sample_size_v2,
+            }
+            return loss, sample_size, logging_output
+
+        if self.use_rdrop:
+            construct_rdrop_sample(sample)
+
+        net_output = model(**sample["net_input"])
+        loss, nll_loss, ntokens = self.compute_loss(model, net_output, sample, update_num, reduce=reduce)
+        sample_size = (
+            sample["target"].size(0) if self.sentence_avg else ntokens
+        )
+        logging_output = {
+            "loss": loss.data,
+            "nll_loss": nll_loss.data,
+            "ntokens": sample["ntokens"],
+            "nsentences": sample["nsentences"],
+            "sample_size": sample_size,
+        }
+        if self.report_accuracy:
+            n_correct, total = self.compute_accuracy(model, net_output, sample)
+            logging_output["n_correct"] = utils.item(n_correct.data)
+            logging_output["total"] = utils.item(total.data)
+        return loss, sample_size, logging_output
+
+    def get_lprobs_and_target(self, model, net_output, sample):
+        conf = sample['conf'][:, None, None] if 'conf' in sample and sample['conf'] is not None else 1
+        constraint_masks = None
+        if "constraint_masks" in sample and sample["constraint_masks"] is not None:
+            constraint_masks = sample["constraint_masks"]
+            net_output[0].masked_fill_(~constraint_masks, -math.inf)
+        if self.constraint_start is not None and self.constraint_end is not None:
+            net_output[0][:, :, 4:self.constraint_start] = -math.inf
+            net_output[0][:, :, self.constraint_end:] = -math.inf
+        lprobs = model.get_normalized_probs(net_output, log_probs=True) * conf
+        target = model.get_targets(sample, net_output)
+        if self.ignore_prefix_size > 0:
+            lprobs = lprobs[:, self.ignore_prefix_size :, :].contiguous()
+            target = target[:, self.ignore_prefix_size :].contiguous()
+            if constraint_masks is not None:
+                constraint_masks = constraint_masks[:, self.ignore_prefix_size :, :].contiguous()
+        if self.ignore_eos:
+            bsz, seq_len, embed_dim = lprobs.size()
+            eos_indices = target.eq(self.task.tgt_dict.eos())
+            lprobs = lprobs[~eos_indices].reshape(bsz, seq_len-1, embed_dim)
+            target = target[~eos_indices].reshape(bsz, seq_len-1)
+            if constraint_masks is not None:
+                constraint_masks = constraint_masks[~eos_indices].reshape(bsz, seq_len-1, embed_dim)
+        if constraint_masks is not None:
+            constraint_masks = constraint_masks.view(-1, constraint_masks.size(-1))
+        return lprobs.view(-1, lprobs.size(-1)), target.view(-1), constraint_masks
+
+    def compute_loss(self, model, net_output, sample, update_num, reduce=True):
+        lprobs, target, constraint_masks = self.get_lprobs_and_target(model, net_output, sample)
+        if constraint_masks is not None:
+            constraint_masks = constraint_masks[target != self.padding_idx]
+        lprobs = lprobs[target != self.padding_idx]
+        target = target[target != self.padding_idx]
+        loss, nll_loss, ntokens = label_smoothed_nll_loss(
+            lprobs,
+            target,
+            self.eps,
+            update_num,
+            reduce=reduce,
+            drop_worst_ratio=self.drop_worst_ratio,
+            drop_worst_after=self.drop_worst_after,
+            use_rdrop=self.use_rdrop,
+            reg_alpha=self.reg_alpha,
+            constraint_masks=constraint_masks,
+            constraint_start=self.constraint_start,
+            constraint_end=self.constraint_end
+        )
+        return loss, nll_loss, ntokens
+
+    def compute_accuracy(self, model, net_output, sample):
+        lprobs, target = self.get_lprobs_and_target(model, net_output, sample)
+        mask = target.ne(self.padding_idx)
+        n_correct = torch.sum(
+            lprobs.argmax(1).masked_select(mask).eq(target.masked_select(mask))
+        )
+        total = torch.sum(mask)
+        return n_correct, total
+
+    @classmethod
+    def reduce_metrics(cls, logging_outputs) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        loss_sum = sum(log.get("loss", 0) for log in logging_outputs)
+        loss_sum_v1 = sum(log.get("loss_v1", 0) for log in logging_outputs)
+        loss_sum_v2 = sum(log.get("loss_v2", 0) for log in logging_outputs)
+        nll_loss_sum = sum(log.get("nll_loss", 0) for log in logging_outputs)
+        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)
+        sample_size_v1 = sum(log.get("sample_size_v1", 0) for log in logging_outputs)
+        sample_size_v2 = sum(log.get("sample_size_v2", 0) for log in logging_outputs)
+
+        metrics.log_scalar(
+            "loss", loss_sum / sample_size, sample_size, round=3
+        )
+        metrics.log_scalar(
+            "loss_v1", loss_sum_v1 / max(sample_size_v1, 1), max(sample_size_v1, 1), round=3
+        )
+        metrics.log_scalar(
+            "loss_v2", loss_sum_v2 / max(sample_size_v2, 1), max(sample_size_v2, 1), round=3
+        )
+        metrics.log_scalar(
+            "nll_loss", nll_loss_sum / sample_size, ntokens, round=3
+        )
+        metrics.log_derived(
+            "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg)
+        )
+
+        metrics.log_scalar(
+            "ntokens", ntokens, 1, round=3
+        )
+        metrics.log_scalar(
+            "nsentences", nsentences, 1, round=3
+        )
+        metrics.log_scalar(
+            "sample_size", sample_size, 1, round=3
+        )
+        metrics.log_scalar(
+            "sample_size_v1", sample_size_v1, 1, round=3
+        )
+        metrics.log_scalar(
+            "sample_size_v2", sample_size_v2, 1, round=3
+        )
+
+        total = utils.item(sum(log.get("total", 0) for log in logging_outputs))
+        if total > 0:
+            metrics.log_scalar("total", total)
+            n_correct = utils.item(
+                sum(log.get("n_correct", 0) for log in logging_outputs)
+            )
+            metrics.log_scalar("n_correct", n_correct)
+            metrics.log_derived(
+                "accuracy",
+                lambda meters: round(
+                    meters["n_correct"].sum * 100.0 / meters["total"].sum, 3
+                )
+                if meters["total"].sum > 0
+                else float("nan"),
+            )
+
+    @staticmethod
+    def logging_outputs_can_be_summed() -> bool:
+        """
+        Whether the logging outputs returned by `forward` can be summed
+        across workers prior to calling `reduce_metrics`. Setting this
+        to True will improves distributed training speed.
+        """
+        return True

criterions/scst_loss.py

@@ -0,0 +1,280 @@
+# 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 string
+from dataclasses import dataclass, field
+from collections import OrderedDict
+from typing import Optional
+
+import torch
+from fairseq import metrics, utils
+from fairseq.criterions import FairseqCriterion, register_criterion
+from fairseq.dataclass import FairseqDataclass
+from omegaconf import II
+
+from data import data_utils
+from utils.cider.pyciderevalcap.ciderD.ciderD import CiderD
+
+
+def scst_loss(lprobs, target, reward, ignore_index=None, reduce=True):
+    loss = -lprobs.gather(dim=-1, index=target.unsqueeze(-1)).squeeze() * reward.unsqueeze(-1)
+    if ignore_index is not None:
+        pad_mask = target.eq(ignore_index)
+        loss.masked_fill_(pad_mask, 0.0)
+        ntokens = (~pad_mask).sum()
+    else:
+        loss = loss.squeeze(-1)
+        ntokens = target.numel()
+    if reduce:
+        loss = loss.sum()
+    return loss, ntokens
+
+@dataclass
+class ScstRewardCriterionConfig(FairseqDataclass):
+    scst_cider_cached_tokens: str = field(
+        default="coco-train-words.p",
+        metadata={"help": "path to cached cPickle file used to calculate CIDEr scores"},
+    )
+    ignore_prefix_size: int = field(
+        default=0,
+        metadata={"help": "Ignore first N tokens"},
+    )
+    sentence_avg: bool = II("optimization.sentence_avg")
+    constraint_range: Optional[str] = field(
+        default=None,
+        metadata={"help": "constraint range"}
+    )
+
+
+@register_criterion(
+    "scst_reward_criterion", dataclass=ScstRewardCriterionConfig
+)
+class ScstRewardCriterion(FairseqCriterion):
+    CIDER_REWARD_WEIGHT = 1
+
+    def __init__(
+        self,
+        task,
+        scst_cider_cached_tokens,
+        sentence_avg,
+        ignore_prefix_size=0,
+        constraint_range=None
+    ):
+        super().__init__(task)
+        self.scst_cider_scorer = CiderD(df=scst_cider_cached_tokens)
+        self.sentence_avg = sentence_avg
+        self.ignore_prefix_size = ignore_prefix_size
+        self.transtab = str.maketrans({key: None for key in string.punctuation})
+
+        self.constraint_start = None
+        self.constraint_end = None
+        if constraint_range is not None:
+            constraint_start, constraint_end = constraint_range.split(',')
+            self.constraint_start = int(constraint_start)
+            self.constraint_end = int(constraint_end)
+
+    def forward(self, model, sample, update_num=0, 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
+        """
+        loss, score, ntokens, nsentences = self.compute_loss(model, sample, reduce=reduce)
+
+        sample_size = (
+            nsentences if self.sentence_avg else ntokens
+        )
+        logging_output = {
+            "loss": loss.data,
+            "score": score,
+            "ntokens": ntokens,
+            "nsentences": nsentences,
+            "sample_size": sample_size,
+        }
+        return loss, sample_size, logging_output
+
+    def _calculate_eval_scores(self, gen_res, gt_idx, gt_res):
+        '''
+        gen_res: generated captions, list of str
+        gt_idx: list of int, of the same length as gen_res
+        gt_res: ground truth captions, list of list of str.
+            gen_res[i] corresponds to gt_res[gt_idx[i]]
+            Each image can have multiple ground truth captions
+        '''
+        gen_res_size = len(gen_res)
+
+        res = OrderedDict()
+        for i in range(gen_res_size):
+            res[i] = [self._wrap_sentence(gen_res[i].strip().translate(self.transtab))]
+
+        gts = OrderedDict()
+        gt_res_ = [
+            [self._wrap_sentence(gt_res[i][j].strip().translate(self.transtab)) for j in range(len(gt_res[i]))]
+                for i in range(len(gt_res))
+        ]
+        for i in range(gen_res_size):
+            gts[i] = gt_res_[gt_idx[i]]
+
+        res_ = [{'image_id':i, 'caption': res[i]} for i in range(len(res))]
+        _, batch_cider_scores = self.scst_cider_scorer.compute_score(gts, res_)
+        scores = self.CIDER_REWARD_WEIGHT * batch_cider_scores
+        return scores
+
+    @classmethod
+    def _wrap_sentence(self, s):
+        # ensure the sentence ends with <eos> token
+        # in order to keep consisitent with cider_cached_tokens
+        r = s.strip()
+        if r.endswith('.'):
+            r = r[:-1]
+        r += ' <eos>'
+        return r
+
+    def get_generator_out(self, model, sample):
+        def decode(toks):
+            hypo = toks.int().cpu()
+            hypo_str = self.task.tgt_dict.string(hypo)
+            hypo_str = self.task.bpe.decode(hypo_str).strip()
+            return hypo, hypo_str
+
+        model.eval()
+        with torch.no_grad():
+            self.task.scst_generator.model.eval()
+            gen_out = self.task.scst_generator.generate([model], sample)
+
+        gen_target = []
+        gen_res = []
+        gt_res = []
+        for i in range(len(gen_out)):
+            for j in range(len(gen_out[i])):
+                hypo, hypo_str = decode(gen_out[i][j]["tokens"])
+                gen_target.append(hypo)
+                gen_res.append(hypo_str)
+            gt_res.append(
+                decode(utils.strip_pad(sample["target"][i], self.padding_idx))[1].split('&&')
+            )
+
+        return gen_target, gen_res, gt_res
+
+    def get_reward_and_scores(self, gen_res, gt_res, device):
+        batch_size = len(gt_res)
+        gen_res_size = len(gen_res)
+        seq_per_img = gen_res_size // batch_size
+
+        gt_idx = [i // seq_per_img for i in range(gen_res_size)]
+        scores = self._calculate_eval_scores(gen_res, gt_idx, gt_res)
+        sc_ = scores.reshape(batch_size, seq_per_img)
+        baseline = (sc_.sum(1, keepdims=True) - sc_) / (sc_.shape[1] - 1)
+        # sample - baseline
+        reward = scores.reshape(batch_size, seq_per_img)
+        reward = reward - baseline
+        reward = reward.reshape(gen_res_size)
+        reward = torch.as_tensor(reward, device=device, dtype=torch.float64)
+
+        return reward, scores
+
+    def get_net_output(self, model, sample, gen_target):
+        def merge(sample_list, eos=self.task.tgt_dict.eos(), move_eos_to_beginning=False):
+            return data_utils.collate_tokens(
+                sample_list,
+                pad_idx=self.padding_idx,
+                eos_idx=eos,
+                left_pad=False,
+                move_eos_to_beginning=move_eos_to_beginning,
+            )
+
+        batch_size = len(sample["target"])
+        gen_target_size = len(gen_target)
+        seq_per_img = gen_target_size // batch_size
+
+        model.train()
+        sample_src_tokens = torch.repeat_interleave(
+            sample['net_input']['src_tokens'], seq_per_img, dim=0
+        )
+        sample_src_lengths = torch.repeat_interleave(
+            sample['net_input']['src_lengths'], seq_per_img, dim=0
+        )
+        sample_patch_images = torch.repeat_interleave(
+            sample['net_input']['patch_images'], seq_per_img, dim=0
+        )
+        sample_patch_masks = torch.repeat_interleave(
+            sample['net_input']['patch_masks'], seq_per_img, dim=0
+        )
+        gen_prev_output_tokens = torch.as_tensor(
+            merge(gen_target, eos=self.task.tgt_dict.bos(), move_eos_to_beginning=True),
+            device=sample["target"].device, dtype=torch.int64
+        )
+        gen_target_tokens = torch.as_tensor(
+            merge(gen_target), device=sample["target"].device, dtype=torch.int64
+        )
+        net_output = model(
+            src_tokens=sample_src_tokens, src_lengths=sample_src_lengths,
+            patch_images=sample_patch_images, patch_masks=sample_patch_masks,
+            prev_output_tokens=gen_prev_output_tokens
+        )
+
+        return net_output, gen_target_tokens
+
+    def get_lprobs_and_target(self, model, net_output, gen_target):
+        if self.constraint_start is not None and self.constraint_end is not None:
+            net_output[0][:, :, 4:self.constraint_start] = -math.inf
+            net_output[0][:, :, self.constraint_end:] = -math.inf
+        lprobs = model.get_normalized_probs(net_output, log_probs=True)
+        if self.ignore_prefix_size > 0:
+            if getattr(lprobs, "batch_first", False):
+                lprobs = lprobs[:, self.ignore_prefix_size :, :].contiguous()
+                gen_target = gen_target[:, self.ignore_prefix_size :].contiguous()
+            else:
+                lprobs = lprobs[self.ignore_prefix_size :, :, :].contiguous()
+                gen_target = gen_target[self.ignore_prefix_size :, :].contiguous()
+        return lprobs, gen_target
+
+    def compute_loss(self, model, sample, reduce=True):
+        gen_target, gen_res, gt_res = self.get_generator_out(model, sample)
+        reward, scores = self.get_reward_and_scores(gen_res, gt_res, device=sample["target"].device)
+        net_output, gen_target_tokens = self.get_net_output(model, sample, gen_target)
+        gen_lprobs, gen_target_tokens = self.get_lprobs_and_target(model, net_output, gen_target_tokens)
+        loss, ntokens = scst_loss(gen_lprobs, gen_target_tokens, reward, ignore_index=self.padding_idx, reduce=reduce)
+        nsentences = gen_target_tokens.size(0)
+
+        return loss, scores.sum(), ntokens, nsentences
+
+    @classmethod
+    def reduce_metrics(cls, logging_outputs) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        loss_sum = sum(log.get("loss", 0) for log in logging_outputs)
+        score_sum = sum(log.get("score", 0) for log in logging_outputs)
+        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)
+
+        metrics.log_scalar(
+            "loss", loss_sum / sample_size, sample_size, round=3
+        )
+        metrics.log_scalar(
+            "score", score_sum / nsentences, nsentences, round=3
+        )
+
+        metrics.log_scalar(
+            "ntokens", ntokens, 1, round=3
+        )
+        metrics.log_scalar(
+            "nsentences", nsentences, 1, round=3
+        )
+        metrics.log_scalar(
+            "sample_size", sample_size, 1, round=3
+        )
+
+    @staticmethod
+    def logging_outputs_can_be_summed() -> bool:
+        """
+        Whether the logging outputs returned by `forward` can be summed
+        across workers prior to calling `reduce_metrics`. Setting this
+        to True will improves distributed training speed.
+        """
+        return True

data/data_utils.py

@@ -0,0 +1,601 @@
+# 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.
+
+try:
+    from collections.abc import Iterable
+except ImportError:
+    from collections import Iterable
+import contextlib
+import itertools
+import logging
+import re
+import warnings
+from typing import Optional, Tuple
+
+import numpy as np
+import torch
+
+from fairseq.file_io import PathManager
+from fairseq import utils
+import os
+
+logger = logging.getLogger(__name__)
+
+
+def infer_language_pair(path):
+    """Infer language pair from filename: <split>.<lang1>-<lang2>.(...).idx"""
+    src, dst = None, None
+    for filename in PathManager.ls(path):
+        parts = filename.split(".")
+        if len(parts) >= 3 and len(parts[1].split("-")) == 2:
+            return parts[1].split("-")
+    return src, dst
+
+
+def collate_tokens(
+    values,
+    pad_idx,
+    eos_idx=None,
+    left_pad=False,
+    move_eos_to_beginning=False,
+    pad_to_length=None,
+    pad_to_multiple=1,
+    pad_to_bsz=None,
+):
+    """Convert a list of 1d tensors into a padded 2d tensor."""
+    size = max(v.size(0) for v in values)
+    size = size if pad_to_length is None else max(size, pad_to_length)
+    if pad_to_multiple != 1 and size % pad_to_multiple != 0:
+        size = int(((size - 0.1) // pad_to_multiple + 1) * pad_to_multiple)
+
+    def copy_tensor(src, dst):
+        assert dst.numel() == src.numel()
+        if move_eos_to_beginning:
+            if eos_idx is None:
+                # if no eos_idx is specified, then use the last token in src
+                dst[0] = src[-1]
+            else:
+                dst[0] = eos_idx
+            dst[1:] = src[:-1]
+        else:
+            dst.copy_(src)
+
+    if values[0].dim() == 1:
+        res = values[0].new(len(values), size).fill_(pad_idx)
+    elif values[0].dim() == 2:
+        assert move_eos_to_beginning is False
+        res = values[0].new(len(values), size, values[0].size(1)).fill_(pad_idx)
+    else:
+        raise NotImplementedError
+
+    for i, v in enumerate(values):
+        copy_tensor(v, res[i][size - len(v) :] if left_pad else res[i][: len(v)])
+    return res
+
+
+def load_indexed_dataset(
+    path, dictionary=None, dataset_impl=None, combine=False, default="cached"
+):
+    """A helper function for loading indexed datasets.
+
+    Args:
+        path (str): path to indexed dataset (e.g., 'data-bin/train')
+        dictionary (~fairseq.data.Dictionary): data dictionary
+        dataset_impl (str, optional): which dataset implementation to use. If
+            not provided, it will be inferred automatically. For legacy indexed
+            data we use the 'cached' implementation by default.
+        combine (bool, optional): automatically load and combine multiple
+            datasets. For example, if *path* is 'data-bin/train', then we will
+            combine 'data-bin/train', 'data-bin/train1', ... and return a
+            single ConcatDataset instance.
+    """
+    import fairseq.data.indexed_dataset as indexed_dataset
+    from fairseq.data.concat_dataset import ConcatDataset
+
+    datasets = []
+    for k in itertools.count():
+        path_k = path + (str(k) if k > 0 else "")
+        try:
+            path_k = indexed_dataset.get_indexed_dataset_to_local(path_k)
+        except Exception as e:
+            if "StorageException: [404] Path not found" in str(e):
+                logger.warning(f"path_k: {e} not found")
+            else:
+                raise e
+
+        dataset_impl_k = dataset_impl
+        if dataset_impl_k is None:
+            dataset_impl_k = indexed_dataset.infer_dataset_impl(path_k)
+        dataset = indexed_dataset.make_dataset(
+            path_k,
+            impl=dataset_impl_k or default,
+            fix_lua_indexing=True,
+            dictionary=dictionary,
+        )
+        if dataset is None:
+            break
+        logger.info("loaded {:,} examples from: {}".format(len(dataset), path_k))
+        datasets.append(dataset)
+        if not combine:
+            break
+    if len(datasets) == 0:
+        return None
+    elif len(datasets) == 1:
+        return datasets[0]
+    else:
+        return ConcatDataset(datasets)
+
+
+@contextlib.contextmanager
+def numpy_seed(seed, *addl_seeds):
+    """Context manager which seeds the NumPy PRNG with the specified seed and
+    restores the state afterward"""
+    if seed is None:
+        yield
+        return
+    if len(addl_seeds) > 0:
+        seed = int(hash((seed, *addl_seeds)) % 1e6)
+    state = np.random.get_state()
+    np.random.seed(seed)
+    try:
+        yield
+    finally:
+        np.random.set_state(state)
+
+
+def collect_filtered(function, iterable, filtered):
+    """
+    Similar to :func:`filter` but collects filtered elements in ``filtered``.
+
+    Args:
+        function (callable): function that returns ``False`` for elements that
+            should be filtered
+        iterable (iterable): iterable to filter
+        filtered (list): list to store filtered elements
+    """
+    for el in iterable:
+        if function(el):
+            yield el
+        else:
+            filtered.append(el)
+
+
+def _filter_by_size_dynamic(indices, size_fn, max_positions, raise_exception=False):
+    def compare_leq(a, b):
+        return a <= b if not isinstance(a, tuple) else max(a) <= b
+
+    def check_size(idx):
+        if isinstance(max_positions, float) or isinstance(max_positions, int):
+            return size_fn(idx) <= max_positions
+        elif isinstance(max_positions, dict):
+            idx_size = size_fn(idx)
+            assert isinstance(idx_size, dict)
+            intersect_keys = set(max_positions.keys()) & set(idx_size.keys())
+            return all(
+                all(
+                    a is None or b is None or a <= b
+                    for a, b in zip(idx_size[key], max_positions[key])
+                )
+                for key in intersect_keys
+            )
+        else:
+            # For MultiCorpusSampledDataset, will generalize it later
+            if not isinstance(size_fn(idx), Iterable):
+                return all(size_fn(idx) <= b for b in max_positions)
+            return all(
+                a is None or b is None or a <= b
+                for a, b in zip(size_fn(idx), max_positions)
+            )
+
+    ignored = []
+    itr = collect_filtered(check_size, indices, ignored)
+    indices = np.fromiter(itr, dtype=np.int64, count=-1)
+    return indices, ignored
+
+
+def filter_by_size(indices, dataset, max_positions, raise_exception=False):
+    """
+    [deprecated] Filter indices based on their size.
+    Use `FairseqDataset::filter_indices_by_size` instead.
+
+    Args:
+        indices (List[int]): ordered list of dataset indices
+        dataset (FairseqDataset): fairseq dataset instance
+        max_positions (tuple): filter elements larger than this size.
+            Comparisons are done component-wise.
+        raise_exception (bool, optional): if ``True``, raise an exception if
+            any elements are filtered (default: False).
+    """
+    warnings.warn(
+        "data_utils.filter_by_size is deprecated. "
+        "Use `FairseqDataset::filter_indices_by_size` instead.",
+        stacklevel=2,
+    )
+    if isinstance(max_positions, float) or isinstance(max_positions, int):
+        if hasattr(dataset, "sizes") and isinstance(dataset.sizes, np.ndarray):
+            ignored = indices[dataset.sizes[indices] > max_positions].tolist()
+            indices = indices[dataset.sizes[indices] <= max_positions]
+        elif (
+            hasattr(dataset, "sizes")
+            and isinstance(dataset.sizes, list)
+            and len(dataset.sizes) == 1
+        ):
+            ignored = indices[dataset.sizes[0][indices] > max_positions].tolist()
+            indices = indices[dataset.sizes[0][indices] <= max_positions]
+        else:
+            indices, ignored = _filter_by_size_dynamic(
+                indices, dataset.size, max_positions
+            )
+    else:
+        indices, ignored = _filter_by_size_dynamic(indices, dataset.size, max_positions)
+
+    if len(ignored) > 0 and raise_exception:
+        raise Exception(
+            (
+                "Size of sample #{} is invalid (={}) since max_positions={}, "
+                "skip this example with --skip-invalid-size-inputs-valid-test"
+            ).format(ignored[0], dataset.size(ignored[0]), max_positions)
+        )
+    if len(ignored) > 0:
+        logger.warning(
+            (
+                "{} samples have invalid sizes and will be skipped, "
+                "max_positions={}, first few sample ids={}"
+            ).format(len(ignored), max_positions, ignored[:10])
+        )
+    return indices
+
+
+def filter_paired_dataset_indices_by_size(src_sizes, tgt_sizes, indices, max_sizes):
+    """Filter a list of sample indices. Remove those that are longer
+        than specified in max_sizes.
+
+    Args:
+        indices (np.array): original array of sample indices
+        max_sizes (int or list[int] or tuple[int]): max sample size,
+            can be defined separately for src and tgt (then list or tuple)
+
+    Returns:
+        np.array: filtered sample array
+        list: list of removed indices
+    """
+    if max_sizes is None:
+        return indices, []
+    if type(max_sizes) in (int, float):
+        max_src_size, max_tgt_size = max_sizes, max_sizes
+    else:
+        max_src_size, max_tgt_size = max_sizes
+    if tgt_sizes is None:
+        ignored = indices[src_sizes[indices] > max_src_size]
+    else:
+        ignored = indices[
+            (src_sizes[indices] > max_src_size) | (tgt_sizes[indices] > max_tgt_size)
+        ]
+    if len(ignored) > 0:
+        if tgt_sizes is None:
+            indices = indices[src_sizes[indices] <= max_src_size]
+        else:
+            indices = indices[
+                (src_sizes[indices] <= max_src_size)
+                & (tgt_sizes[indices] <= max_tgt_size)
+            ]
+    return indices, ignored.tolist()
+
+
+def batch_by_size(
+    indices,
+    num_tokens_fn,
+    num_tokens_vec=None,
+    max_tokens=None,
+    max_sentences=None,
+    required_batch_size_multiple=1,
+    fixed_shapes=None,
+):
+    """
+    Yield mini-batches of indices bucketed by size. Batches may contain
+    sequences of different lengths.
+
+    Args:
+        indices (List[int]): ordered list of dataset indices
+        num_tokens_fn (callable): function that returns the number of tokens at
+            a given index
+        num_tokens_vec (List[int], optional): precomputed vector of the number
+            of tokens for each index in indices (to enable faster batch generation)
+        max_tokens (int, optional): max number of tokens in each batch
+            (default: None).
+        max_sentences (int, optional): max number of sentences in each
+            batch (default: None).
+        required_batch_size_multiple (int, optional): require batch size to
+            be less than N or a multiple of N (default: 1).
+        fixed_shapes (List[Tuple[int, int]], optional): if given, batches will
+            only be created with the given shapes. *max_sentences* and
+            *required_batch_size_multiple* will be ignored (default: None).
+    """
+    try:
+        from fairseq.data.data_utils_fast import (
+            batch_by_size_fn,
+            batch_by_size_vec,
+            batch_fixed_shapes_fast,
+        )
+    except ImportError:
+        raise ImportError(
+            "Please build Cython components with: "
+            "`python setup.py build_ext --inplace`"
+        )
+    except ValueError:
+        raise ValueError(
+            "Please build (or rebuild) Cython components with `python setup.py build_ext --inplace`."
+        )
+
+    # added int() to avoid TypeError: an integer is required
+    max_tokens = (
+        int(max_tokens) if max_tokens is not None else -1
+    )
+    max_sentences = max_sentences if max_sentences is not None else -1
+    bsz_mult = required_batch_size_multiple
+
+    if not isinstance(indices, np.ndarray):
+        indices = np.fromiter(indices, dtype=np.int64, count=-1)
+
+    if num_tokens_vec is not None and not isinstance(num_tokens_vec, np.ndarray):
+        num_tokens_vec = np.fromiter(num_tokens_vec, dtype=np.int64, count=-1)
+
+    if fixed_shapes is None:
+        if num_tokens_vec is None:
+            return batch_by_size_fn(
+                indices,
+                num_tokens_fn,
+                max_tokens,
+                max_sentences,
+                bsz_mult,
+            )
+        else:
+            return batch_by_size_vec(
+                indices,
+                num_tokens_vec,
+                max_tokens,
+                max_sentences,
+                bsz_mult,
+            )
+
+    else:
+        fixed_shapes = np.array(fixed_shapes, dtype=np.int64)
+        sort_order = np.lexsort(
+            [
+                fixed_shapes[:, 1].argsort(),  # length
+                fixed_shapes[:, 0].argsort(),  # bsz
+            ]
+        )
+        fixed_shapes_sorted = fixed_shapes[sort_order]
+        return batch_fixed_shapes_fast(indices, num_tokens_fn, fixed_shapes_sorted)
+
+
+def post_process(sentence: str, symbol: str):
+    if symbol == "sentencepiece":
+        sentence = sentence.replace(" ", "").replace("\u2581", " ").strip()
+    elif symbol == "wordpiece":
+        sentence = sentence.replace(" ", "").replace("_", " ").strip()
+    elif symbol == "letter":
+        sentence = sentence.replace(" ", "").replace("|", " ").strip()
+    elif symbol == "silence":
+        import re
+        sentence = sentence.replace("<SIL>", "")
+        sentence = re.sub(' +', ' ', sentence).strip()
+    elif symbol == "_EOW":
+        sentence = sentence.replace(" ", "").replace("_EOW", " ").strip()
+    elif symbol in {"subword_nmt", "@@ ", "@@"}:
+        if symbol == "subword_nmt":
+            symbol = "@@ "
+        sentence = (sentence + " ").replace(symbol, "").rstrip()
+    elif symbol == "none":
+        pass
+    elif symbol is not None:
+        raise NotImplementedError(f"Unknown post_process option: {symbol}")
+    return sentence
+
+
+def compute_mask_indices(
+    shape: Tuple[int, int],
+    padding_mask: Optional[torch.Tensor],
+    mask_prob: float,
+    mask_length: int,
+    mask_type: str = "static",
+    mask_other: float = 0.0,
+    min_masks: int = 0,
+    no_overlap: bool = False,
+    min_space: int = 0,
+) -> np.ndarray:
+    """
+    Computes random mask spans for a given shape
+
+    Args:
+        shape: the the shape for which to compute masks.
+            should be of size 2 where first element is batch size and 2nd is timesteps
+        padding_mask: optional padding mask of the same size as shape, which will prevent masking padded elements
+        mask_prob: probability for each token to be chosen as start of the span to be masked. this will be multiplied by
+            number of timesteps divided by length of mask span to mask approximately this percentage of all elements.
+            however due to overlaps, the actual number will be smaller (unless no_overlap is True)
+        mask_type: how to compute mask lengths
+            static = fixed size
+            uniform = sample from uniform distribution [mask_other, mask_length*2]
+            normal = sample from normal distribution with mean mask_length and stdev mask_other. mask is min 1 element
+            poisson = sample from possion distribution with lambda = mask length
+        min_masks: minimum number of masked spans
+        no_overlap: if false, will switch to an alternative recursive algorithm that prevents spans from overlapping
+        min_space: only used if no_overlap is True, this is how many elements to keep unmasked between spans
+    """
+
+    bsz, all_sz = shape
+    mask = np.full((bsz, all_sz), False)
+
+    all_num_mask = int(
+        # add a random number for probabilistic rounding
+        mask_prob * all_sz / float(mask_length)
+        + np.random.rand()
+    )
+
+    all_num_mask = max(min_masks, all_num_mask)
+
+    mask_idcs = []
+    for i in range(bsz):
+        if padding_mask is not None:
+            sz = all_sz - padding_mask[i].long().sum().item()
+            num_mask = int(
+                # add a random number for probabilistic rounding
+                mask_prob * sz / float(mask_length)
+                + np.random.rand()
+            )
+            num_mask = max(min_masks, num_mask)
+        else:
+            sz = all_sz
+            num_mask = all_num_mask
+
+        if mask_type == "static":
+            lengths = np.full(num_mask, mask_length)
+        elif mask_type == "uniform":
+            lengths = np.random.randint(mask_other, mask_length * 2 + 1, size=num_mask)
+        elif mask_type == "normal":
+            lengths = np.random.normal(mask_length, mask_other, size=num_mask)
+            lengths = [max(1, int(round(x))) for x in lengths]
+        elif mask_type == "poisson":
+            lengths = np.random.poisson(mask_length, size=num_mask)
+            lengths = [int(round(x)) for x in lengths]
+        else:
+            raise Exception("unknown mask selection " + mask_type)
+
+        if sum(lengths) == 0:
+            lengths[0] = min(mask_length, sz - 1)
+
+        if no_overlap:
+            mask_idc = []
+
+            def arrange(s, e, length, keep_length):
+                span_start = np.random.randint(s, e - length)
+                mask_idc.extend(span_start + i for i in range(length))
+
+                new_parts = []
+                if span_start - s - min_space >= keep_length:
+                    new_parts.append((s, span_start - min_space + 1))
+                if e - span_start - keep_length - min_space > keep_length:
+                    new_parts.append((span_start + length + min_space, e))
+                return new_parts
+
+            parts = [(0, sz)]
+            min_length = min(lengths)
+            for length in sorted(lengths, reverse=True):
+                lens = np.fromiter(
+                    (e - s if e - s >= length + min_space else 0 for s, e in parts),
+                    np.int,
+                )
+                l_sum = np.sum(lens)
+                if l_sum == 0:
+                    break
+                probs = lens / np.sum(lens)
+                c = np.random.choice(len(parts), p=probs)
+                s, e = parts.pop(c)
+                parts.extend(arrange(s, e, length, min_length))
+            mask_idc = np.asarray(mask_idc)
+        else:
+            min_len = min(lengths)
+            if sz - min_len <= num_mask:
+                min_len = sz - num_mask - 1
+
+            mask_idc = np.random.choice(sz - min_len, num_mask, replace=False)
+
+            mask_idc = np.asarray(
+                [
+                    mask_idc[j] + offset
+                    for j in range(len(mask_idc))
+                    for offset in range(lengths[j])
+                ]
+            )
+
+        mask_idcs.append(np.unique(mask_idc[mask_idc < sz]))
+
+    min_len = min([len(m) for m in mask_idcs])
+    for i, mask_idc in enumerate(mask_idcs):
+        if len(mask_idc) > min_len:
+            mask_idc = np.random.choice(mask_idc, min_len, replace=False)
+        mask[i, mask_idc] = True
+
+    return mask
+
+
+def get_mem_usage():
+    try:
+        import psutil
+
+        mb = 1024 * 1024
+        return f"used={psutil.virtual_memory().used / mb}Mb; avail={psutil.virtual_memory().available / mb}Mb"
+    except ImportError:
+        return "N/A"
+
+
+# lens: torch.LongTensor
+# returns: torch.BoolTensor
+def lengths_to_padding_mask(lens):
+    bsz, max_lens = lens.size(0), torch.max(lens).item()
+    mask = torch.arange(max_lens).to(lens.device).view(1, max_lens)
+    mask = mask.expand(bsz, -1) >= lens.view(bsz, 1).expand(-1, max_lens)
+    return mask
+
+
+# lens: torch.LongTensor
+# returns: torch.BoolTensor
+def lengths_to_mask(lens):
+    return ~lengths_to_padding_mask(lens)
+
+
+def get_buckets(sizes, num_buckets):
+    buckets = np.unique(
+        np.percentile(
+            sizes,
+            np.linspace(0, 100, num_buckets + 1),
+            interpolation='lower',
+        )[1:]
+    )
+    return buckets
+
+
+def get_bucketed_sizes(orig_sizes, buckets):
+    sizes = np.copy(orig_sizes)
+    assert np.min(sizes) >= 0
+    start_val = -1
+    for end_val in buckets:
+        mask = (sizes > start_val) & (sizes <= end_val)
+        sizes[mask] = end_val
+        start_val = end_val
+    return sizes
+
+
+
+def _find_extra_valid_paths(dataset_path: str) -> set:
+    paths = utils.split_paths(dataset_path)
+    all_valid_paths = set()
+    for sub_dir in paths:
+        contents = PathManager.ls(sub_dir)
+        valid_paths = [c for c in contents if re.match("valid*[0-9].*", c) is not None]
+        all_valid_paths |= {os.path.basename(p) for p in valid_paths}
+    # Remove .bin, .idx etc
+    roots = {os.path.splitext(p)[0] for p in all_valid_paths}
+    return roots
+
+
+def raise_if_valid_subsets_unintentionally_ignored(train_cfg) -> None:
+    """Raises if there are paths matching 'valid*[0-9].*' which are not combined or ignored."""
+    if (
+        train_cfg.dataset.ignore_unused_valid_subsets
+        or train_cfg.dataset.combine_valid_subsets
+        or train_cfg.dataset.disable_validation
+        or not hasattr(train_cfg.task, "data")
+    ):
+        return
+    other_paths = _find_extra_valid_paths(train_cfg.task.data)
+    specified_subsets = train_cfg.dataset.valid_subset.split(",")
+    ignored_paths = [p for p in other_paths if p not in specified_subsets]
+    if ignored_paths:
+        advice = "Set --combine-val to combine them or --ignore-unused-valid-subsets to ignore them."
+        msg = f"Valid paths {ignored_paths} will be ignored. {advice}"
+        raise ValueError(msg)

data/file_dataset.py

@@ -0,0 +1,102 @@
+import os
+import torch
+import pickle
+
+
+class FileDataset:
+    def __init__(self, file_path, selected_col_ids=None, dtypes=None, separator="\t", cached_index=False):
+        self.file_path = file_path
+        assert os.path.exists(self.file_path), "Error: The local datafile {} not exists!".format(self.file_path)
+
+        self.separator = separator
+        if selected_col_ids is None:
+            # default to all fields
+            self.selected_col_ids = list(
+                range(len(open(self.file_path).readline().rstrip("\n").split(self.separator))))
+        else:
+            self.selected_col_ids = [int(col_id) for col_id in selected_col_ids.split(",")]
+        if dtypes is None:
+            # default to str
+            self.dtypes = [str for col_id in self.selected_col_ids]
+        else:
+            self.dtypes = [eval(col_dtype) for col_dtype in dtypes.split(",")]
+            assert len(self.dtypes) == len(self.selected_col_ids)
+
+        self.data_cnt = 0
+        try:
+            self.slice_id = torch.distributed.get_rank()
+            self.slice_count = torch.distributed.get_world_size()
+        except Exception:
+            self.slice_id = 0
+            self.slice_count = 1
+        self.cached_index = cached_index
+        self._init_seek_index()
+        self._reader = self._get_reader()
+        print("file {} slice_id {} row count {} total row count {}".format(
+            self.file_path, self.slice_id, self.row_count, self.total_row_count)
+        )
+
+    def _init_seek_index(self):
+        if self.cached_index:
+            cache_path = "{}.index".format(self.file_path)
+            assert os.path.exists(cache_path), "cache file {} not exists!".format(cache_path)
+            self.total_row_count, self.lineid_to_offset = pickle.load(open(cache_path, "rb"))
+            print("local datafile {} slice_id {} use cached row_count and line_idx-to-offset mapping".format(
+                self.file_path, self.slice_id))
+        else:
+            # make an iteration over the file to get row_count and line_idx-to-offset mapping
+            fp = open(self.file_path, "r")
+            print("local datafile {} slice_id {} begin to initialize row_count and line_idx-to-offset mapping".format(
+                self.file_path, self.slice_id))
+            self.total_row_count = 0
+            offset = 0
+            self.lineid_to_offset = []
+            for line in fp:
+                self.lineid_to_offset.append(offset)
+                self.total_row_count += 1
+                offset += len(line.encode('utf-8'))
+        self._compute_start_pos_and_row_count()
+        print("local datafile {} slice_id {} finished initializing row_count and line_idx-to-offset mapping".format(
+            self.file_path, self.slice_id))
+
+    def _compute_start_pos_and_row_count(self):
+        self.row_count = self.total_row_count // self.slice_count
+        if self.slice_id < self.total_row_count - self.row_count * self.slice_count:
+            self.row_count += 1
+            self.start_pos = self.row_count * self.slice_id
+        else:
+            self.start_pos = self.row_count * self.slice_id + (self.total_row_count - self.row_count * self.slice_count)
+
+    def _get_reader(self):
+        fp = open(self.file_path, "r")
+        fp.seek(self.lineid_to_offset[self.start_pos])
+        return fp
+
+    def _seek(self, offset=0):
+        try:
+            print("slice_id {} seek offset {}".format(self.slice_id, self.start_pos + offset))
+            self._reader.seek(self.lineid_to_offset[self.start_pos + offset])
+            self.data_cnt = offset
+        except Exception:
+            print("slice_id {} seek offset {}".format(self.slice_id, offset))
+            self._reader.seek(self.lineid_to_offset[offset])
+            self.data_cnt = offset
+
+    def __del__(self):
+        self._reader.close()
+
+    def __len__(self):
+        return self.row_count
+
+    def get_total_row_count(self):
+        return self.total_row_count
+
+    def __getitem__(self, index):
+        if self.data_cnt == self.row_count:
+            print("reach the end of datafile, start a new reader")
+            self.data_cnt = 0
+            self._reader = self._get_reader()
+        column_l = self._reader.readline().rstrip("\n").split(self.separator)
+        self.data_cnt += 1
+        column_l = [dtype(column_l[col_id]) for col_id, dtype in zip(self.selected_col_ids, self.dtypes)]
+        return column_l

data/mm_data/caption_dataset.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.
+from io import BytesIO
+
+import logging
+import warnings
+import string
+
+import numpy as np
+import torch
+import base64
+from torchvision import transforms
+
+from PIL import Image, ImageFile
+
+from data import data_utils
+from data.ofa_dataset import OFADataset
+
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+ImageFile.MAX_IMAGE_PIXELS = None
+Image.MAX_IMAGE_PIXELS = None
+
+logger = logging.getLogger(__name__)
+warnings.filterwarnings("ignore", "(Possibly )?corrupt EXIF data", UserWarning)
+
+IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
+
+
+def collate(samples, pad_idx, eos_idx):
+    if len(samples) == 0:
+        return {}
+
+    def merge(key):
+        return data_utils.collate_tokens(
+            [s[key] for s in samples],
+            pad_idx,
+            eos_idx=eos_idx,
+        )
+
+    id = np.array([s["id"] for s in samples])
+    src_tokens = merge("source")
+    src_lengths = torch.LongTensor([s["source"].ne(pad_idx).long().sum() for s in samples])
+
+    patch_images = torch.stack([sample['patch_image'] for sample in samples], dim=0)
+    patch_masks = torch.cat([sample['patch_mask'] for sample in samples])
+
+    prev_output_tokens = None
+    target = None
+    if samples[0].get("target", None) is not None:
+        target = merge("target")
+        tgt_lengths = torch.LongTensor([s["target"].ne(pad_idx).long().sum() for s in samples])
+        ntokens = tgt_lengths.sum().item()
+
+        if samples[0].get("prev_output_tokens", None) is not None:
+            prev_output_tokens = merge("prev_output_tokens")
+    else:
+        ntokens = src_lengths.sum().item()
+
+    batch = {
+        "id": id,
+        "nsentences": len(samples),
+        "ntokens": ntokens,
+        "net_input": {
+            "src_tokens": src_tokens,
+            "src_lengths": src_lengths,
+            "patch_images": patch_images,
+            "patch_masks": patch_masks,
+            "prev_output_tokens": prev_output_tokens
+        },
+        "target": target,
+    }
+
+    return batch
+
+
+class CaptionDataset(OFADataset):
+    def __init__(
+        self,
+        split,
+        dataset,
+        bpe,
+        src_dict,
+        tgt_dict=None,
+        max_src_length=128,
+        max_tgt_length=30,
+        patch_image_size=224,
+        imagenet_default_mean_and_std=False,
+        scst=False
+    ):
+        super().__init__(split, dataset, bpe, src_dict, tgt_dict)
+        self.max_src_length = max_src_length
+        self.max_tgt_length = max_tgt_length
+        self.patch_image_size = patch_image_size
+        self.scst = scst
+
+        self.transtab = str.maketrans({key: None for key in string.punctuation})
+
+        if imagenet_default_mean_and_std:
+            mean = IMAGENET_DEFAULT_MEAN
+            std = IMAGENET_DEFAULT_STD
+        else:
+            mean = [0.5, 0.5, 0.5]
+            std = [0.5, 0.5, 0.5]
+
+        self.patch_resize_transform = transforms.Compose([
+            lambda image: image.convert("RGB"),
+            transforms.Resize((patch_image_size, patch_image_size), interpolation=Image.BICUBIC),
+            transforms.ToTensor(),
+            transforms.Normalize(mean=mean, std=std),
+        ])
+
+    def __getitem__(self, index):
+        uniq_id, image, caption = self.dataset[index]
+
+        image = Image.open(BytesIO(base64.urlsafe_b64decode(image)))
+        patch_image = self.patch_resize_transform(image)
+        patch_mask = torch.tensor([True])
+
+        if self.split == 'train' and not self.scst:
+            caption = caption.translate(self.transtab).strip()
+            caption_token_list = caption.strip().split()
+            tgt_caption = ' '.join(caption_token_list[:self.max_tgt_length])
+        else:
+            caption = ' '.join(caption.strip().split())
+            caption_list = [cap.translate(self.transtab).strip() for cap in caption.strip().split('&&')]
+            tgt_caption = '&&'.join(caption_list)
+        src_item = self.encode_text(" what does the image describe?")
+        tgt_item = self.encode_text(" {}".format(tgt_caption))
+
+        src_item = torch.cat([self.bos_item, src_item, self.eos_item])
+        target_item = torch.cat([tgt_item, self.eos_item])
+        prev_output_item = torch.cat([self.bos_item, tgt_item])
+
+        example = {
+            "id": uniq_id,
+            "source": src_item,
+            "patch_image": patch_image,
+            "patch_mask": patch_mask,
+            "target": target_item,
+            "prev_output_tokens": prev_output_item
+        }
+        return example
+
+    def collater(self, samples, pad_to_length=None):
+        """Merge a list of samples to form a mini-batch.
+        Args:
+            samples (List[dict]): samples to collate
+        Returns:
+            dict: a mini-batch with the following keys:
+        """
+        return collate(samples, pad_idx=self.pad, eos_idx=self.eos)

data/mm_data/refcoco_dataset.py

@@ -0,0 +1,168 @@
+# 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.
+from io import BytesIO
+
+import logging
+import warnings
+
+import numpy as np
+import torch
+import base64
+import utils.transforms as T
+
+from PIL import Image, ImageFile
+
+from data import data_utils
+from data.ofa_dataset import OFADataset
+
+ImageFile.LOAD_TRUNCATED_IMAGES = True
+ImageFile.MAX_IMAGE_PIXELS = None
+Image.MAX_IMAGE_PIXELS = None
+
+logger = logging.getLogger(__name__)
+warnings.filterwarnings("ignore", "(Possibly )?corrupt EXIF data", UserWarning)
+
+IMAGENET_DEFAULT_MEAN = (0.485, 0.456, 0.406)
+IMAGENET_DEFAULT_STD = (0.229, 0.224, 0.225)
+
+
+def collate(samples, pad_idx, eos_idx):
+    if len(samples) == 0:
+        return {}
+
+    def merge(key):
+        return data_utils.collate_tokens(
+            [s[key] for s in samples],
+            pad_idx,
+            eos_idx=eos_idx,
+        )
+
+    id = np.array([s["id"] for s in samples])
+    src_tokens = merge("source")
+    src_lengths = torch.LongTensor([s["source"].ne(pad_idx).long().sum() for s in samples])
+
+    patch_images = torch.stack([sample['patch_image'] for sample in samples], dim=0)
+    patch_masks = torch.cat([sample['patch_mask'] for sample in samples])
+
+    w_resize_ratios = torch.stack([s["w_resize_ratio"] for s in samples], dim=0)
+    h_resize_ratios = torch.stack([s["h_resize_ratio"] for s in samples], dim=0)
+    region_coords = torch.stack([s['region_coord'] for s in samples], dim=0)
+
+    prev_output_tokens = None
+    target = None
+    if samples[0].get("target", None) is not None:
+        target = merge("target")
+        tgt_lengths = torch.LongTensor([s["target"].ne(pad_idx).long().sum() for s in samples])
+        ntokens = tgt_lengths.sum().item()
+
+        if samples[0].get("prev_output_tokens", None) is not None:
+            prev_output_tokens = merge("prev_output_tokens")
+    else:
+        ntokens = src_lengths.sum().item()
+
+    batch = {
+        "id": id,
+        "nsentences": len(samples),
+        "ntokens": ntokens,
+        "net_input": {
+            "src_tokens": src_tokens,
+            "src_lengths": src_lengths,
+            "patch_images": patch_images,
+            "patch_masks": patch_masks,
+            "prev_output_tokens": prev_output_tokens
+        },
+        "target": target,
+        "w_resize_ratios": w_resize_ratios,
+        "h_resize_ratios": h_resize_ratios,
+        "region_coords": region_coords
+    }
+
+    return batch
+
+
+class RefcocoDataset(OFADataset):
+    def __init__(
+        self,
+        split,
+        dataset,
+        bpe,
+        src_dict,
+        tgt_dict=None,
+        max_src_length=80,
+        max_tgt_length=30,
+        patch_image_size=512,
+        imagenet_default_mean_and_std=False,
+        num_bins=1000,
+        max_image_size=512
+    ):
+        super().__init__(split, dataset, bpe, src_dict, tgt_dict)
+        self.max_src_length = max_src_length
+        self.max_tgt_length = max_tgt_length
+        self.patch_image_size = patch_image_size
+        self.num_bins = num_bins
+
+        if imagenet_default_mean_and_std:
+            mean = IMAGENET_DEFAULT_MEAN
+            std = IMAGENET_DEFAULT_STD
+        else:
+            mean = [0.5, 0.5, 0.5]
+            std = [0.5, 0.5, 0.5]
+
+        # for positioning
+        self.positioning_transform = T.Compose([
+            T.RandomResize([patch_image_size], max_size=patch_image_size),
+            T.ToTensor(),
+            T.Normalize(mean=mean, std=std, max_image_size=max_image_size)
+        ])
+
+    def __getitem__(self, index):
+        uniq_id, base64_str, text, region_coord = self.dataset[index]
+
+        image = Image.open(BytesIO(base64.urlsafe_b64decode(base64_str))).convert("RGB")
+        w, h = image.size
+        boxes_target = {"boxes": [], "labels": [], "area": [], "size": torch.tensor([h, w])}
+        x0, y0, x1, y1 = region_coord.strip().split(',')
+        region = torch.tensor([float(x0), float(y0), float(x1), float(y1)])
+        boxes_target["boxes"] = torch.tensor([[float(x0), float(y0), float(x1), float(y1)]])
+        boxes_target["labels"] = np.array([0])
+        boxes_target["area"] = torch.tensor([(float(x1) - float(x0)) * (float(y1) - float(y0))])
+
+        patch_image, patch_boxes = self.positioning_transform(image, boxes_target)
+        resize_h, resize_w = patch_boxes["size"][0], patch_boxes["size"][1]
+        patch_mask = torch.tensor([True])
+        quant_x0 = "<bin_{}>".format(int((patch_boxes["boxes"][0][0] * (self.num_bins - 1)).round()))
+        quant_y0 = "<bin_{}>".format(int((patch_boxes["boxes"][0][1] * (self.num_bins - 1)).round()))
+        quant_x1 = "<bin_{}>".format(int((patch_boxes["boxes"][0][2] * (self.num_bins - 1)).round()))
+        quant_y1 = "<bin_{}>".format(int((patch_boxes["boxes"][0][3] * (self.num_bins - 1)).round()))
+        region_coord = "{} {} {} {}".format(quant_x0, quant_y0, quant_x1, quant_y1)
+        src_caption = self.pre_caption(text, self.max_src_length)
+        src_item = self.encode_text(' which region does the text " {} " describe?'.format(src_caption))
+        tgt_item = self.encode_text(region_coord, use_bpe=False)
+
+        src_item = torch.cat([self.bos_item, src_item, self.eos_item])
+        target_item = torch.cat([tgt_item, self.eos_item])
+        prev_output_item = torch.cat([self.bos_item, tgt_item])
+
+        example = {
+            "id": uniq_id,
+            "source": src_item,
+            "patch_image": patch_image,
+            "patch_mask": patch_mask,
+            "target": target_item,
+            "prev_output_tokens": prev_output_item,
+            "w_resize_ratio": resize_w / w,
+            "h_resize_ratio": resize_h / h,
+            "region_coord": region
+        }
+        return example
+
+    def collater(self, samples, pad_to_length=None):
+        """Merge a list of samples to form a mini-batch.
+        Args:
+            samples (List[dict]): samples to collate
+        Returns:
+            dict: a mini-batch with the following keys:
+        """
+        return collate(samples, pad_idx=self.pad, eos_idx=self.eos)

data/ofa_dataset.py

@@ -0,0 +1,74 @@
+import logging
+import re
+import torch.utils.data
+from fairseq.data import FairseqDataset
+
+logger = logging.getLogger(__name__)
+
+
+class OFADataset(FairseqDataset):
+    def __init__(self, split, dataset, bpe, src_dict, tgt_dict):
+        self.split = split
+        self.dataset = dataset
+        self.bpe = bpe
+        self.src_dict = src_dict
+        self.tgt_dict = tgt_dict
+
+        self.bos = src_dict.bos()
+        self.eos = src_dict.eos()
+        self.pad = src_dict.pad()
+        self.bos_item = torch.LongTensor([self.bos])
+        self.eos_item = torch.LongTensor([self.eos])
+
+    def __len__(self):
+        return len(self.dataset)
+
+    def encode_text(self, text, length=None, append_bos=False, append_eos=False, use_bpe=True):
+        s = self.tgt_dict.encode_line(
+            line=self.bpe.encode(text) if use_bpe else text,
+            add_if_not_exist=False,
+            append_eos=False
+        ).long()
+        if length is not None:
+            s = s[:length]
+        if append_bos:
+            s = torch.cat([self.bos_item, s])
+        if append_eos:
+            s = torch.cat([s, self.eos_item])
+        return s
+
+    def pre_question(self, question, max_ques_words):
+        question = question.lower().lstrip(",.!?*#:;~").replace('-', ' ').replace('/', ' ')
+
+        question = re.sub(
+            r"\s{2,}",
+            ' ',
+            question,
+        )
+        question = question.rstrip('\n')
+        question = question.strip(' ')
+
+        # truncate question
+        question_words = question.split(' ')
+        if len(question_words) > max_ques_words:
+            question = ' '.join(question_words[:max_ques_words])
+
+        return question
+
+    def pre_caption(self, caption, max_words):
+        caption = caption.lower().lstrip(",.!?*#:;~").replace('-', ' ').replace('/', ' ').replace('<person>', 'person')
+
+        caption = re.sub(
+            r"\s{2,}",
+            ' ',
+            caption,
+        )
+        caption = caption.rstrip('\n')
+        caption = caption.strip(' ')
+
+        # truncate caption
+        caption_words = caption.split(' ')
+        if len(caption_words) > max_words:
+            caption = ' '.join(caption_words[:max_words])
+
+        return caption

evaluate.py

@@ -0,0 +1,152 @@
+#!/usr/bin/env python3 -u
+# 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 logging
+import os
+import sys
+import json
+from itertools import chain
+
+import numpy as np
+import torch
+import torch.distributed as dist
+from fairseq import distributed_utils, options, tasks, utils
+from fairseq.dataclass.utils import convert_namespace_to_omegaconf
+from fairseq.logging import progress_bar
+from fairseq.utils import reset_logging
+from omegaconf import DictConfig
+
+from utils import checkpoint_utils
+from utils.eval_utils import eval_step
+
+logging.basicConfig(
+    format="%(asctime)s | %(levelname)s | %(name)s | %(message)s",
+    datefmt="%Y-%m-%d %H:%M:%S",
+    level=os.environ.get("LOGLEVEL", "INFO").upper(),
+    stream=sys.stdout,
+)
+logger = logging.getLogger("ofa.evaluate")
+
+
+def apply_half(t):
+    if t.dtype is torch.float32:
+        return t.to(dtype=torch.half)
+    return t
+
+
+def main(cfg: DictConfig):
+    utils.import_user_module(cfg.common)
+
+    reset_logging()
+    logger.info(cfg)
+
+    assert (
+        cfg.dataset.max_tokens is not None or cfg.dataset.batch_size is not None
+    ), "Must specify batch size either with --max-tokens or --batch-size"
+
+    # Fix seed for stochastic decoding
+    if cfg.common.seed is not None and not cfg.generation.no_seed_provided:
+        np.random.seed(cfg.common.seed)
+        utils.set_torch_seed(cfg.common.seed)
+
+    use_fp16 = cfg.common.fp16
+    use_cuda = torch.cuda.is_available() and not cfg.common.cpu
+
+    if use_cuda:
+        torch.cuda.set_device(cfg.distributed_training.device_id)
+
+    # Load ensemble
+    overrides = eval(cfg.common_eval.model_overrides)
+    logger.info("loading model(s) from {}".format(cfg.common_eval.path))
+    models, saved_cfg, task = checkpoint_utils.load_model_ensemble_and_task(
+        utils.split_paths(cfg.common_eval.path),
+        arg_overrides=overrides,
+        suffix=cfg.checkpoint.checkpoint_suffix,
+        strict=(cfg.checkpoint.checkpoint_shard_count == 1),
+        num_shards=cfg.checkpoint.checkpoint_shard_count,
+    )
+
+    # loading the dataset should happen after the checkpoint has been loaded so we can give it the saved task config
+    task.load_dataset(cfg.dataset.gen_subset, task_cfg=saved_cfg.task)
+
+    # Move models to GPU
+    for model in models:
+        model.eval()
+        if use_fp16:
+            model.half()
+        if use_cuda and not cfg.distributed_training.pipeline_model_parallel:
+            model.cuda()
+        model.prepare_for_inference_(cfg)
+
+    # Load dataset (possibly sharded)
+    itr = task.get_batch_iterator(
+        dataset=task.dataset(cfg.dataset.gen_subset),
+        max_tokens=cfg.dataset.max_tokens,
+        max_sentences=cfg.dataset.batch_size,
+        max_positions=utils.resolve_max_positions(
+            task.max_positions(), *[m.max_positions() for m in models]
+        ),
+        ignore_invalid_inputs=cfg.dataset.skip_invalid_size_inputs_valid_test,
+        required_batch_size_multiple=cfg.dataset.required_batch_size_multiple,
+        seed=cfg.common.seed,
+        num_shards=cfg.distributed_training.distributed_world_size,
+        shard_id=cfg.distributed_training.distributed_rank,
+        num_workers=cfg.dataset.num_workers,
+        data_buffer_size=cfg.dataset.data_buffer_size,
+    ).next_epoch_itr(shuffle=False)
+    progress = progress_bar.progress_bar(
+        itr,
+        log_format=cfg.common.log_format,
+        log_interval=cfg.common.log_interval,
+        default_log_format=("tqdm" if not cfg.common.no_progress_bar else "simple"),
+    )
+
+    # Initialize generator
+    generator = task.build_generator(models, cfg.generation)
+
+    results = []
+    score_sum = torch.FloatTensor([0]).cuda()
+    score_cnt = torch.FloatTensor([0]).cuda()
+    for sample in progress:
+        if "net_input" not in sample:
+            continue
+        sample = utils.move_to_cuda(sample) if use_cuda else sample
+        sample = utils.apply_to_sample(apply_half, sample) if cfg.common.fp16 else sample
+        with torch.no_grad():
+            result, scores = eval_step(task, generator, models, sample)
+        results += result
+        score_sum += sum(scores) if scores is not None else 0
+        score_cnt += len(scores) if scores is not None else 0
+        progress.log({"sentences": sample["nsentences"]})
+
+    gather_results = None
+    if cfg.distributed_training.distributed_world_size > 1:
+        gather_results = [None for _ in range(dist.get_world_size())]
+        dist.all_gather_object(gather_results, results)
+        dist.all_reduce(score_sum.data)
+        dist.all_reduce(score_cnt.data)
+    if score_cnt.item() > 0:
+        logger.info("score_sum: {}, score_cnt: {}, score: {}".format(
+            score_sum, score_cnt, round(score_sum.item() / score_cnt.item(), 4)
+        ))
+
+    if cfg.distributed_training.distributed_world_size == 1 or dist.get_rank() == 0:
+        os.makedirs(cfg.common_eval.results_path, exist_ok=True)
+        output_path = os.path.join(cfg.common_eval.results_path, "{}_predict.json".format(cfg.dataset.gen_subset))
+        gather_results = list(chain(*gather_results)) if gather_results is not None else results
+        with open(output_path, 'w') as fw:
+            json.dump(gather_results, fw)
+
+
+def cli_main():
+    parser = options.get_generation_parser()
+    args = options.parse_args_and_arch(parser)
+    cfg = convert_namespace_to_omegaconf(args)
+    distributed_utils.call_main(cfg, main)
+
+
+if __name__ == "__main__":
+    cli_main()

fairseq/.github/ISSUE_TEMPLATE.md

@@ -0,0 +1,3 @@
+## 👉 [Please follow one of these issue templates](https://github.com/pytorch/fairseq/issues/new/choose) 👈
+
+Note: to keep the backlog clean and actionable, issues may be immediately closed if they do not follow one of the above issue templates.

fairseq/.github/ISSUE_TEMPLATE/bug_report.md

@@ -0,0 +1,43 @@
+---
+name: 🐛 Bug Report
+about: Submit a bug report to help us improve
+labels: 'bug, needs triage'
+---
+
+## 🐛 Bug
+
+<!-- A clear and concise description of what the bug is. -->
+
+### To Reproduce
+
+Steps to reproduce the behavior (**always include the command you ran**):
+
+1. Run cmd '....'
+2. See error
+
+<!-- If you have a code sample, error messages, stack traces, please provide it here as well -->
+
+
+#### Code sample
+<!-- Ideally attach a minimal code sample to reproduce the decried issue.
+Minimal means having the shortest code but still preserving the bug. -->
+
+### Expected behavior
+
+<!-- A clear and concise description of what you expected to happen. -->
+
+### Environment
+
+ - fairseq Version (e.g., 1.0 or main):
+ - PyTorch Version (e.g., 1.0)
+ - OS (e.g., Linux):
+ - How you installed fairseq (`pip`, source):
+ - Build command you used (if compiling from source):
+ - Python version:
+ - CUDA/cuDNN version:
+ - GPU models and configuration:
+ - Any other relevant information:
+
+### Additional context
+
+<!-- Add any other context about the problem here. -->

fairseq/.github/ISSUE_TEMPLATE/documentation.md

@@ -0,0 +1,15 @@
+---
+name: 📚 Documentation/Typos
+about: Report an issue related to documentation or a typo
+labels: 'documentation, needs triage'
+---
+
+## 📚 Documentation
+
+For typos and doc fixes, please go ahead and:
+
+1. Create an issue.
+2. Fix the typo.
+3. Submit a PR.
+
+Thanks!

fairseq/.github/ISSUE_TEMPLATE/feature_request.md

@@ -0,0 +1,24 @@
+---
+name: 🚀 Feature Request
+about: Submit a proposal/request for a new feature
+labels: 'enhancement, help wanted, needs triage'
+---
+
+## 🚀 Feature Request
+<!-- A clear and concise description of the feature proposal -->
+
+### Motivation
+
+<!-- Please outline the motivation for the proposal. Is your feature request related to a problem? e.g., I'm always frustrated when [...]. If this is related to another GitHub issue, please link here too -->
+
+### Pitch
+
+<!-- A clear and concise description of what you want to happen. -->
+
+### Alternatives
+
+<!-- A clear and concise description of any alternative solutions or features you've considered, if any. -->
+
+### Additional context
+
+<!-- Add any other context or screenshots about the feature request here. -->

fairseq/.github/ISSUE_TEMPLATE/how-to-question.md

@@ -0,0 +1,33 @@
+---
+name: ❓ Questions/Help
+about: If you have questions, please first search existing issues and docs
+labels: 'question, needs triage'
+---
+
+## ❓ Questions and Help
+
+### Before asking:
+1. search the issues.
+2. search the docs.
+
+<!-- If you still can't find what you need: -->
+
+#### What is your question?
+
+#### Code
+
+<!-- Please paste a code snippet if your question requires it! -->
+
+#### What have you tried?
+
+#### What's your environment?
+
+ - fairseq Version (e.g., 1.0 or main):
+ - PyTorch Version (e.g., 1.0)
+ - OS (e.g., Linux):
+ - How you installed fairseq (`pip`, source):
+ - Build command you used (if compiling from source):
+ - Python version:
+ - CUDA/cuDNN version:
+ - GPU models and configuration:
+ - Any other relevant information:

fairseq/.github/PULL_REQUEST_TEMPLATE.md

@@ -0,0 +1,16 @@
+# Before submitting
+
+- [ ] Was this discussed/approved via a Github issue? (no need for typos, doc improvements)
+- [ ] Did you read the [contributor guideline](https://github.com/pytorch/fairseq/blob/main/CONTRIBUTING.md)?
+- [ ] Did you make sure to update the docs?
+- [ ] Did you write any new necessary tests?
+
+## What does this PR do?
+Fixes # (issue).
+
+## PR review
+Anyone in the community is free to review the PR once the tests have passed.
+If we didn't discuss your PR in Github issues there's a high chance it will not be merged.
+
+## Did you have fun?
+Make sure you had fun coding 🙃

fairseq/.github/stale.yml

@@ -0,0 +1,30 @@
+# Configuration for probot-stale - https://github.com/probot/stale
+# Mostly copied from github.com/facebook/react/blob/master/.github/stale.yml
+# Number of days of inactivity before an issue becomes stale
+daysUntilStale: 90
+# Number of days of inactivity before a stale issue is closed
+daysUntilClose: 7
+# Issues with these labels will never be considered stale
+exemptLabels:
+  - bug
+# Label to use when marking an issue as stale
+staleLabel: stale
+issues:
+  # Comment to post when marking an issue as stale.
+  markComment: >
+    This issue has been automatically marked as stale.
+    **If this issue is still affecting you, please leave any comment** (for example, "bump"), and we'll keep it open.
+    We are sorry that we haven't been able to prioritize it yet. If you have any new additional information, please include it with your comment!
+  # Comment to post when closing a stale issue.
+  closeComment: >
+    Closing this issue after a prolonged period of inactivity. If this issue is still present in the latest release, please create a new issue with up-to-date information. Thank you!
+pulls:
+  # Comment to post when marking a pull request as stale.
+  markComment: >
+    This pull request has been automatically marked as stale.
+    **If this pull request is still relevant, please leave any comment** (for example, "bump"), and we'll keep it open.
+    We are sorry that we haven't been able to prioritize reviewing it yet. Your contribution is very much appreciated.
+  # Comment to post when closing a stale pull request.
+  closeComment: >
+    Closing this pull request after a prolonged period of inactivity. If this issue is still present in the latest release, please ask for this pull request to be reopened. Thank you!
+

fairseq/.github/workflows/build.yml

@@ -0,0 +1,55 @@
+name: build
+
+on:
+  # Trigger the workflow on push to main or any pull request
+  push:
+    branches:
+      - main
+  pull_request:
+
+jobs:
+  build:
+
+    strategy:
+      max-parallel: 4
+      matrix:
+        platform: [ubuntu-latest, macos-latest]
+        python-version: [3.6, 3.7]
+
+    runs-on: ${{ matrix.platform }}
+
+    steps:
+    - uses: actions/checkout@v2
+
+    - name: Set up Python ${{ matrix.python-version }}
+      uses: actions/setup-python@v2
+      with:
+        python-version: ${{ matrix.python-version }}
+
+    - name: Conditionally install pytorch
+      if: matrix.platform == 'windows-latest'
+      run: pip3 install torch -f https://download.pytorch.org/whl/torch_stable.html
+
+    - name: Install locally
+      run: |
+        python -m pip install --upgrade pip
+        git submodule update --init --recursive
+        python setup.py build_ext --inplace
+        python -m pip install --editable .
+
+    - name: Install optional test requirements
+      run: |
+        python -m pip install iopath transformers pyarrow
+        python -m pip install git+https://github.com/facebookresearch/fairscale.git@main
+
+    - name: Lint with flake8
+      run: |
+        pip install flake8
+        # stop the build if there are Python syntax errors or undefined names
+        flake8 . --count --select=E9,F63,F7,F82 --show-source --statistics --extend-exclude fairseq/model_parallel/megatron
+        # exit-zero treats all errors as warnings. The GitHub editor is 127 chars wide
+        flake8 . --count --exit-zero --max-complexity=10 --max-line-length=127 --statistics --extend-exclude fairseq/model_parallel/megatron
+
+    - name: Run tests
+      run: |
+          python setup.py test

fairseq/.github/workflows/build_wheels.yml

@@ -0,0 +1,41 @@
+name: build_wheels
+
+on:
+  push:
+    branches:
+      - v[0-9]+.[0-9]+.[x0-9]+
+    tags:
+      - v*
+
+jobs:
+  build_wheels:
+    name: Build wheels on ${{ matrix.os }}
+    runs-on: ${{ matrix.os }}
+    strategy:
+      matrix:
+        os: [ubuntu-latest, macos-latest]
+
+    steps:
+      - uses: actions/checkout@v2
+
+      - name: Install Python
+        uses: actions/setup-python@v2
+        with:
+          python-version: '3.7'
+
+      - name: Install cibuildwheel
+        run: |
+          python -m pip install cibuildwheel
+
+      - name: Build wheels for CPython
+        run: |
+          python -m cibuildwheel --output-dir dist
+        env:
+          CIBW_BUILD: "cp36-*64 cp37-*64 cp38-*64"
+          CIBW_MANYLINUX_X86_64_IMAGE: manylinux1
+          CIBW_BEFORE_BUILD: git submodule update --init --recursive && pip install .
+
+      - uses: actions/upload-artifact@v2
+        with:
+          name: wheels
+          path: ./dist/*.whl

fairseq/.gitmodules

@@ -0,0 +1,4 @@
+[submodule "fairseq/model_parallel/megatron"]
+    path = fairseq/model_parallel/megatron
+    url = https://github.com/ngoyal2707/Megatron-LM
+    branch = fairseq

fairseq/CODE_OF_CONDUCT.md

@@ -0,0 +1,77 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+  advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+  address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <conduct@pytorch.org>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq
+

fairseq/CONTRIBUTING.md

@@ -0,0 +1,28 @@
+# Contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq)
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Facebook's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+## License
+By contributing to Facebook AI Research Sequence-to-Sequence Toolkit (fairseq),
+you agree that your contributions will be licensed under the LICENSE file in
+the root directory of this source tree.

fairseq/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) Facebook, Inc. and its affiliates.
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

fairseq/README.md

@@ -0,0 +1,229 @@
+<p align="center">
+  <img src="docs/fairseq_logo.png" width="150">
+  <br />
+  <br />
+  <a href="https://github.com/pytorch/fairseq/blob/main/LICENSE"><img alt="MIT License" src="https://img.shields.io/badge/license-MIT-blue.svg" /></a>
+  <a href="https://github.com/pytorch/fairseq/releases"><img alt="Latest Release" src="https://img.shields.io/github/release/pytorch/fairseq.svg" /></a>
+  <a href="https://github.com/pytorch/fairseq/actions?query=workflow:build"><img alt="Build Status" src="https://github.com/pytorch/fairseq/workflows/build/badge.svg" /></a>
+  <a href="https://fairseq.readthedocs.io/en/latest/?badge=latest"><img alt="Documentation Status" src="https://readthedocs.org/projects/fairseq/badge/?version=latest" /></a>
+</p>
+
+--------------------------------------------------------------------------------
+
+Fairseq(-py) is a sequence modeling toolkit that allows researchers and
+developers to train custom models for translation, summarization, language
+modeling and other text generation tasks.
+
+We provide reference implementations of various sequence modeling papers:
+
+<details><summary>List of implemented papers</summary><p>
+
+* **Convolutional Neural Networks (CNN)**
+  + [Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)](examples/language_model/conv_lm/README.md)
+  + [Convolutional Sequence to Sequence Learning (Gehring et al., 2017)](examples/conv_seq2seq/README.md)
+  + [Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)](https://github.com/pytorch/fairseq/tree/classic_seqlevel)
+  + [Hierarchical Neural Story Generation (Fan et al., 2018)](examples/stories/README.md)
+  + [wav2vec: Unsupervised Pre-training for Speech Recognition (Schneider et al., 2019)](examples/wav2vec/README.md)
+* **LightConv and DynamicConv models**
+  + [Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)](examples/pay_less_attention_paper/README.md)
+* **Long Short-Term Memory (LSTM) networks**
+  + Effective Approaches to Attention-based Neural Machine Translation (Luong et al., 2015)
+* **Transformer (self-attention) networks**
+  + Attention Is All You Need (Vaswani et al., 2017)
+  + [Scaling Neural Machine Translation (Ott et al., 2018)](examples/scaling_nmt/README.md)
+  + [Understanding Back-Translation at Scale (Edunov et al., 2018)](examples/backtranslation/README.md)
+  + [Adaptive Input Representations for Neural Language Modeling (Baevski and Auli, 2018)](examples/language_model/README.adaptive_inputs.md)
+  + [Lexically constrained decoding with dynamic beam allocation (Post & Vilar, 2018)](examples/constrained_decoding/README.md)
+  + [Transformer-XL: Attentive Language Models Beyond a Fixed-Length Context (Dai et al., 2019)](examples/truncated_bptt/README.md)
+  + [Adaptive Attention Span in Transformers (Sukhbaatar et al., 2019)](examples/adaptive_span/README.md)
+  + [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)
+  + [Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)](examples/joint_alignment_translation/README.md )
+  + [Multilingual Denoising Pre-training for Neural Machine Translation (Liu et at., 2020)](examples/mbart/README.md)
+  + [Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)](examples/byte_level_bpe/README.md)
+  + [Unsupervised Quality Estimation for Neural Machine Translation (Fomicheva et al., 2020)](examples/unsupervised_quality_estimation/README.md)
+  + [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)](examples/wav2vec/README.md)
+  + [Generating Medical Reports from Patient-Doctor Conversations Using Sequence-to-Sequence Models (Enarvi et al., 2020)](examples/pointer_generator/README.md)
+  + [Linformer: Self-Attention with Linear Complexity (Wang et al., 2020)](examples/linformer/README.md)
+  + [Cross-lingual Retrieval for Iterative Self-Supervised Training (Tran et al., 2020)](examples/criss/README.md)
+  + [Deep Transformers with Latent Depth (Li et al., 2020)](examples/latent_depth/README.md)
+  + [Unsupervised Cross-lingual Representation Learning for Speech Recognition (Conneau et al., 2020)](https://arxiv.org/abs/2006.13979)
+  + [Robust wav2vec 2.0: Analyzing Domain Shift in Self-Supervised Pre-Training (Hsu, et al., 2021)](https://arxiv.org/abs/2104.01027)
+  + [Unsupervised Speech Recognition (Baevski, et al., 2021)](https://arxiv.org/abs/2105.11084)
+* **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)
+* **Finetuning**
+  + [Better Fine-Tuning by Reducing Representational Collapse (Aghajanyan et al. 2020)](examples/rxf/README.md)
+
+</p></details>
+
+### What's New:
+
+* September 2021 [`master` branch renamed to `main`](https://github.com/github/renaming).
+* July 2021 [Released DrNMT code](examples/discriminative_reranking_nmt/README.md)
+* July 2021 [Released Robust wav2vec 2.0 model](examples/wav2vec/README.md)
+* June 2021 [Released XLMR-XL and XLMR-XXL models](examples/xlmr/README.md)
+* May 2021 [Released Unsupervised Speech Recognition code](examples/wav2vec/unsupervised/README.md)
+* March 2021 [Added full parameter and optimizer state sharding + CPU offloading](examples/fully_sharded_data_parallel/README.md)
+* February 2021 [Added LASER training code](examples/laser/README.md)
+* December 2020: [Added Adaptive Attention Span code](examples/adaptive_span/README.md)
+* December 2020: [GottBERT model and code released](examples/gottbert/README.md)
+* November 2020: Adopted the [Hydra](https://github.com/facebookresearch/hydra) configuration framework
+  * [see documentation explaining how to use it for new and existing projects](docs/hydra_integration.md)
+* November 2020: [fairseq 0.10.0 released](https://github.com/pytorch/fairseq/releases/tag/v0.10.0)
+* October 2020: [Added R3F/R4F (Better Fine-Tuning) code](examples/rxf/README.md)
+* October 2020: [Deep Transformer with Latent Depth code released](examples/latent_depth/README.md)
+* October 2020: [Added CRISS models and code](examples/criss/README.md)
+
+<details><summary>Previous updates</summary><p>
+
+* September 2020: [Added Linformer code](examples/linformer/README.md)
+* September 2020: [Added pointer-generator networks](examples/pointer_generator/README.md)
+* August 2020: [Added lexically constrained decoding](examples/constrained_decoding/README.md)
+* August 2020: [wav2vec2 models and code released](examples/wav2vec/README.md)
+* July 2020: [Unsupervised Quality Estimation code released](examples/unsupervised_quality_estimation/README.md)
+* May 2020: [Follow fairseq on Twitter](https://twitter.com/fairseq)
+* April 2020: [Monotonic Multihead Attention code released](examples/simultaneous_translation/README.md)
+* April 2020: [Quant-Noise code released](examples/quant_noise/README.md)
+* April 2020: [Initial model parallel support and 11B parameters unidirectional LM released](examples/megatron_11b/README.md)
+* March 2020: [Byte-level BPE code released](examples/byte_level_bpe/README.md)
+* February 2020: [mBART model and code released](examples/mbart/README.md)
+* February 2020: [Added tutorial for back-translation](https://github.com/pytorch/fairseq/tree/main/examples/backtranslation#training-your-own-model-wmt18-english-german)
+* December 2019: [fairseq 0.9.0 released](https://github.com/pytorch/fairseq/releases/tag/v0.9.0)
+* November 2019: [VizSeq released (a visual analysis toolkit for evaluating fairseq models)](https://facebookresearch.github.io/vizseq/docs/getting_started/fairseq_example)
+* November 2019: [CamemBERT model and code released](examples/camembert/README.md)
+* November 2019: [BART model and code released](examples/bart/README.md)
+* November 2019: [XLM-R models and code released](examples/xlmr/README.md)
+* 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)
+* June 2019: [wav2vec models and code released](examples/wav2vec/README.md)
+
+</p></details>
+
+### Features:
+
+* multi-GPU training on one machine or across multiple machines (data and model parallel)
+* fast generation on both CPU and GPU with multiple search algorithms implemented:
+  + beam search
+  + Diverse Beam Search ([Vijayakumar et al., 2016](https://arxiv.org/abs/1610.02424))
+  + sampling (unconstrained, top-k and top-p/nucleus)
+  + [lexically constrained decoding](examples/constrained_decoding/README.md) (Post & Vilar, 2018)
+* [gradient accumulation](https://fairseq.readthedocs.io/en/latest/getting_started.html#large-mini-batch-training-with-delayed-updates) enables training with large mini-batches even on a single GPU
+* [mixed precision training](https://fairseq.readthedocs.io/en/latest/getting_started.html#training-with-half-precision-floating-point-fp16) (trains faster with less GPU memory on [NVIDIA tensor cores](https://developer.nvidia.com/tensor-cores))
+* [extensible](https://fairseq.readthedocs.io/en/latest/overview.html): easily register new models, criterions, tasks, optimizers and learning rate schedulers
+* [flexible configuration](docs/hydra_integration.md) based on [Hydra](https://github.com/facebookresearch/hydra) allowing a combination of code, command-line and file based configuration
+* [full parameter and optimizer state sharding](examples/fully_sharded_data_parallel/README.md)
+* [offloading parameters to CPU](examples/fully_sharded_data_parallel/README.md)
+
+We also provide [pre-trained models for translation and language modeling](#pre-trained-models-and-examples)
+with a convenient `torch.hub` interface:
+
+``` python
+en2de = torch.hub.load('pytorch/fairseq', 'transformer.wmt19.en-de.single_model')
+en2de.translate('Hello world', beam=5)
+# 'Hallo Welt'
+```
+
+See the PyTorch Hub tutorials for [translation](https://pytorch.org/hub/pytorch_fairseq_translation/)
+and [RoBERTa](https://pytorch.org/hub/pytorch_fairseq_roberta/) for more examples.
+
+# Requirements and Installation
+
+* [PyTorch](http://pytorch.org/) version >= 1.5.0
+* Python version >= 3.6
+* For training new models, you'll also need an NVIDIA GPU and [NCCL](https://github.com/NVIDIA/nccl)
+* **To install fairseq** and develop locally:
+
+``` bash
+git clone https://github.com/pytorch/fairseq
+cd fairseq
+pip install --editable ./
+
+# on MacOS:
+# CFLAGS="-stdlib=libc++" pip install --editable ./
+
+# to install the latest stable release (0.10.x)
+# pip install fairseq
+```
+
+* **For faster training** install NVIDIA's [apex](https://github.com/NVIDIA/apex) library:
+
+``` bash
+git clone https://github.com/NVIDIA/apex
+cd apex
+pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" \
+  --global-option="--deprecated_fused_adam" --global-option="--xentropy" \
+  --global-option="--fast_multihead_attn" ./
+```
+
+* **For large datasets** install [PyArrow](https://arrow.apache.org/docs/python/install.html#using-pip): `pip install pyarrow`
+* If you use Docker make sure to increase the shared memory size either with `--ipc=host` or `--shm-size`
+ as command line options to `nvidia-docker run` .
+
+# Getting Started
+
+The [full documentation](https://fairseq.readthedocs.io/) contains instructions
+for getting started, training new models and extending fairseq with new model
+types and tasks.
+
+# Pre-trained models and examples
+
+We provide pre-trained models and pre-processed, binarized test sets for several tasks listed below,
+as well as example training and evaluation commands.
+
+* [Translation](examples/translation/README.md): convolutional and transformer models are available
+* [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:
+
+* [Cross-lingual Retrieval for Iterative Self-Supervised Training (Tran et al., 2020)](examples/criss/README.md)
+* [wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations (Baevski et al., 2020)](examples/wav2vec/README.md)
+* [Unsupervised Quality Estimation for Neural Machine Translation (Fomicheva et al., 2020)](examples/unsupervised_quality_estimation/README.md)
+* [Training with Quantization Noise for Extreme Model Compression ({Fan*, Stock*} et al., 2020)](examples/quant_noise/README.md)
+* [Neural Machine Translation with Byte-Level Subwords (Wang et al., 2020)](examples/byte_level_bpe/README.md)
+* [Multilingual Denoising Pre-training for Neural Machine Translation (Liu et at., 2020)](examples/mbart/README.md)
+* [Reducing Transformer Depth on Demand with Structured Dropout (Fan et al., 2019)](examples/layerdrop/README.md)
+* [Jointly Learning to Align and Translate with Transformer Models (Garg et al., 2019)](examples/joint_alignment_translation/README.md)
+* [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)
+* [Mixture Models for Diverse Machine Translation: Tricks of the Trade (Shen et al., 2019)](examples/translation_moe/README.md)
+* [Pay Less Attention with Lightweight and Dynamic Convolutions (Wu et al., 2019)](examples/pay_less_attention_paper/README.md)
+* [Understanding Back-Translation at Scale (Edunov et al., 2018)](examples/backtranslation/README.md)
+* [Classical Structured Prediction Losses for Sequence to Sequence Learning (Edunov et al., 2018)](https://github.com/pytorch/fairseq/tree/classic_seqlevel)
+* [Hierarchical Neural Story Generation (Fan et al., 2018)](examples/stories/README.md)
+* [Scaling Neural Machine Translation (Ott et al., 2018)](examples/scaling_nmt/README.md)
+* [Convolutional Sequence to Sequence Learning (Gehring et al., 2017)](examples/conv_seq2seq/README.md)
+* [Language Modeling with Gated Convolutional Networks (Dauphin et al., 2017)](examples/language_model/README.conv.md)
+
+# Join the fairseq community
+
+* Twitter: https://twitter.com/fairseq
+* Facebook page: https://www.facebook.com/groups/fairseq.users
+* Google group: https://groups.google.com/forum/#!forum/fairseq-users
+
+# License
+
+fairseq(-py) is MIT-licensed.
+The license applies to the pre-trained models as well.
+
+# Citation
+
+Please cite as:
+
+``` bibtex
+@inproceedings{ott2019fairseq,
+  title = {fairseq: A Fast, Extensible Toolkit for Sequence Modeling},
+  author = {Myle Ott and Sergey Edunov and Alexei Baevski and Angela Fan and Sam Gross and Nathan Ng and David Grangier and Michael Auli},
+  booktitle = {Proceedings of NAACL-HLT 2019: Demonstrations},
+  year = {2019},
+}
+```

fairseq/examples/__init__.py

@@ -0,0 +1,9 @@
+# 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.
+
+try:
+    from fairseq.version import __version__  # noqa
+except ImportError:
+    pass

fairseq/examples/adaptive_span/README.md

@@ -0,0 +1,90 @@
+# Adaptive Span
+
+Adaptive Span is a novel self-attention mechanism that can learn its optimal
+attention span. This allows us to extend significantly the maximum context size
+used in Transformer, while maintaining control over their memory footprint
+and computational time. It uses the Truncated BPTT technique for training,
+as in [transformerXL](https://github.com/pytorch/fairseq/blob/main/examples/truncated_bptt/README.md).
+
+Adaptive Span was introduced by paper:
+[Adaptive Attention Span in Transformers](https://arxiv.org/abs/1905.07799),
+which achieved state-of-the-art language modeling results at the time of publication.
+
+We manage to reproduce their result in fairseq and keep most of the
+[original implementation](https://github.com/facebookresearch/adaptive-span) untouched.
+You can refer to the their sweep file as well if any combination of hyperparameter is not clear.
+
+##### 0. Setup
+
+First you need to process the Enwik8 dataset, we use the pre-tokenized dataset
+from [adaptive span paper](https://github.com/facebookresearch/adaptive-span/blob/master/get_data.sh).
+You can download the dataset, and then run:
+```bash
+fairseq-preprocess --only-source --trainpref ~/data/enwik8/train.txt \
+    --validpref ~/data/enwik8/valid.txt --testpref ~/data/enwik8/test.txt \
+    --destdir ~/data/enwik8/data-bin/ --joined-dictionary --workers 20
+```
+
+##### 1. Train a Adaptive Span model on Enwik8
+
+We will train a 12-layer Adaptive Span model following the [hyperparameters
+used in the original
+paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh).
+
+The following command assumes 4 GPUs, so that the total batch size is 64
+sequences (4 x 16). Training should take 2-3 days on 4 V100 GPUs:
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \
+    --user-dir examples/adaptive_span \
+    --data  ~/data/enwik8/data-bin/ \
+    --fp16 --fp16-no-flatten-grads --max-update 600000 \
+    --task truncated_bptt_lm --tokens-per-sample 512 --arch adaptive_span \
+    --n-layer 12 --d-model 512 --n-head 8 --d-inner 2048 --dropout 0.3 \
+    --attn-span 8192 --optimizer adagrad_with_grad_clip --adagrad-clip 0.03 \
+    --validate-interval-updates 1000 \
+    --lr-scheduler fixed --warmup-updates 32000 --batch-size-valid 32 \
+    --lr 0.07 --criterion adaptive_span_loss --batch-size 16 --update-freq 1 \
+    --seed 2 --log-format json --log-interval 25 --aux-loss-scaler 5e-07
+```
+This should land around 1.05 on validation, 1.03 on test. You can lower the
+--aux-loss-scaler for better performance (longer span). It gives ~0.03 bpc
+improvement to the transformerXL baseline here.
+If training on a single GPU, set `--update-freq=4` to accumulate 4x gradients
+and simulate training on 4 GPUs.
+You can also reproduce the transformerXL result on enwik8 using this code base.
+It should land around 1.06 on test,matching the [original paper](https://github.com/kimiyoung/transformer-xl/blob/master/pytorch/run_enwik8_base.sh).
+You can try by
+```bash
+CUDA_VISIBLE_DEVICES=0,1,2,3 fairseq-train \
+    --user-dir examples/truncated_bptt \
+    ~/data/enwik8/data-bin/ \
+    --task truncated_bptt_lm  --fp16 --max-update 400000 \
+    --tokens-per-sample 512 --arch transformer_xl --n-layer 12 \
+    --d-model 512 --n-head 8 --d-head 64 --d-inner 2048 --dropout 0.1 \
+    --dropatt 0.0 --mem-len 512 --optimizer adam --clip-norm 0.25 \
+    --lr-scheduler cosine --warmup-updates 0 \
+    --lr 0.0 --lr 0.00025 --batch-size 15 \
+    --update-freq 1 --seed 2 --log-format json --log-interval 25 \
+    --fp16
+```
+
+##### 2. Evaluate
+For Adaptive Span:
+```bash
+fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \
+ --user-dir examples/adaptive_span \
+ --task truncated_bptt_lm --batch-size 8 --tokens-per-sample 512 --gen-subset test
+```
+For Transformer-XL evaluation:
+```bash
+fairseq-eval-lm ~/data/enwik8/data-bin/ --path model/checkpoint_best.pt \
+    --user-dir examples/truncated_bptt/ --task truncated_bptt_lm --batch-size 8 \
+    --tokens-per-sample 80 \
+    --model-overrides '{"mem_len":2100,"clamp_len":820,"same_length":True}' \
+    --gen-subset valid
+```
+
+*Note:* During training the model saw 512 tokens of context
+(``--tokens-per-sample=512``), with batch size 8. These settings match the evaluation
+settings from [the original
+paper](https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8.sh).

fairseq/examples/adaptive_span/__init__.py

@@ -0,0 +1,19 @@
+# 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 importlib
+import os
+
+# automatically import any Python files in the current directory
+cur_dir = os.path.dirname(__file__)
+for file in os.listdir(cur_dir):
+    path = os.path.join(cur_dir, file)
+    if (
+        not file.startswith("_")
+        and not file.startswith(".")
+        and (file.endswith(".py") or os.path.isdir(path))
+    ):
+        mod_name = file[: file.find(".py")] if file.endswith(".py") else file
+        module = importlib.import_module(__name__ + "." + mod_name)

fairseq/examples/adaptive_span/adagrad_with_grad_clip.py

@@ -0,0 +1,128 @@
+# 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.
+
+from torch.optim import Adagrad
+
+from fairseq.optim import LegacyFairseqOptimizer, register_optimizer
+
+
+@register_optimizer("adagrad_with_grad_clip")
+class FairseqAdagradWithGradClip(LegacyFairseqOptimizer):
+    def __init__(self, args, params):
+        super().__init__(args)
+        self._optimizer = AdagradWithGradClip(params, **self.optimizer_config)
+
+    @staticmethod
+    def add_args(parser):
+        """Add optimizer-specific arguments to the parser."""
+        # fmt: off
+        parser.add_argument('--weight-decay', '--wd', default=0.0, type=float, metavar='WD',
+                            help='weight decay')
+        parser.add_argument('--adagrad-clip', default=0.0, type=float, metavar='D',
+                            help='internal grad clip')
+        # fmt: on
+
+    @property
+    def optimizer_config(self):
+        """
+        Return a kwarg dictionary that will be used to override optimizer
+        args stored in checkpoints. This allows us to load a checkpoint and
+        resume training using a different set of optimizer args, e.g., with a
+        different learning rate.
+        """
+        return {
+            "lr": self.args.lr[0],
+            "weight_decay": self.args.weight_decay,
+            "grad_clip": self.args.adagrad_clip,
+        }
+
+    @property
+    def supports_flat_params(self):
+        return False
+
+
+def _clip_grad(clr, grad, group_grad_clip):
+    if group_grad_clip > 0:
+        norm = grad.norm(2).item()
+        if norm > group_grad_clip:
+            clr *= group_grad_clip / (norm + 1e-10)
+    return clr
+
+
+class AdagradWithGradClip(Adagrad):
+    """Adagrad algorithm with custom gradient clipping"""
+
+    def __init__(
+        self,
+        params,
+        lr=1e-2,
+        lr_decay=0,
+        weight_decay=0,
+        initial_accumulator_value=0,
+        grad_clip=0,
+    ):
+        Adagrad.__init__(
+            self,
+            params,
+            lr=lr,
+            lr_decay=lr_decay,
+            weight_decay=weight_decay,
+            initial_accumulator_value=initial_accumulator_value,
+        )
+        self.defaults["grad_clip"] = grad_clip
+        self.param_groups[0].setdefault("grad_clip", grad_clip)
+
+    def step(self, closure=None):
+        loss = None
+        if closure is not None:
+            loss = closure()
+
+        for group in self.param_groups:
+            for p in group["params"]:
+                if p.grad is None:
+                    continue
+
+                grad = p.grad.data
+                state = self.state[p]
+
+                state["step"] += 1
+
+                if group["weight_decay"] != 0:
+                    if p.grad.data.is_sparse:
+                        raise RuntimeError(
+                            "weight_decay option is "
+                            "not compatible with sparse "
+                            "gradients"
+                        )
+                    grad = grad.add(group["weight_decay"], p.data)
+
+                clr = group["lr"] / (1 + (state["step"] - 1) * group["lr_decay"])
+
+                # clip
+                clr = _clip_grad(clr=clr, grad=grad, group_grad_clip=group["grad_clip"])
+
+                if grad.is_sparse:
+                    # the update is non-linear so indices must be unique
+                    grad = grad.coalesce()
+                    grad_indices = grad._indices()
+                    grad_values = grad._values()
+                    size = grad.size()
+
+                    def make_sparse(values):
+                        constructor = grad.new
+                        if grad_indices.dim() == 0 or values.dim() == 0:
+                            return constructor().resize_as_(grad)
+                        return constructor(grad_indices, values, size)
+
+                    state["sum"].add_(make_sparse(grad_values.pow(2)))
+                    std = state["sum"]._sparse_mask(grad)
+                    std_values = std._values().sqrt_().add_(1e-10)
+                    p.data.add_(-clr, make_sparse(grad_values / std_values))
+                else:
+                    state["sum"].addcmul_(1, grad, grad)
+                    std = state["sum"].sqrt().add_(1e-10)
+                    p.data.addcdiv_(-clr, grad, std)
+
+        return loss

fairseq/examples/adaptive_span/adaptive_span_attention.py

@@ -0,0 +1,160 @@
+# 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
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class AdaptiveMask(nn.Module):
+    """Soft masking function for adaptive size.
+    It masks out the last K values of an input. The masking value
+    goes from 1 to 0 gradually, so K can be learned with
+    back-propagation.
+    Args:
+        max_size: maximum size (i.e. input dimension)
+        ramp_size: size of the ramp going from 0 to 1
+        init_val: initial size proportion not to be masked out
+        shape: learn multiple sizes independent of each other
+    """
+
+    def __init__(self, max_size, ramp_size, init_val=0, shape=(1,)):
+        nn.Module.__init__(self)
+        self._max_size = max_size
+        self._ramp_size = ramp_size
+        self.current_val = nn.Parameter(torch.zeros(*shape) + init_val)
+        mask_template = torch.linspace(1 - max_size, 0, steps=max_size)
+        self.register_buffer("mask_template", mask_template)
+
+    def forward(self, x):
+        mask = self.mask_template.float() + self.current_val.float() * self._max_size
+        mask = mask / self._ramp_size + 1
+        mask = mask.clamp(0, 1)
+        if x.size(-1) < self._max_size:
+            # the input could have been trimmed beforehand to save computation
+            mask = mask.narrow(-1, self._max_size - x.size(-1), x.size(-1))
+        x = (x * mask).type_as(x)
+        return x
+
+    def get_current_max_size(self, include_ramp=True):
+        current_size = math.ceil(self.current_val.max().item() * self._max_size)
+        if include_ramp:
+            current_size += self._ramp_size
+        current_size = max(0, min(self._max_size, current_size))
+        return current_size
+
+    def get_current_avg_size(self, include_ramp=True):
+        current_size = math.ceil(
+            self.current_val.float().mean().item() * self._max_size
+        )
+        if include_ramp:
+            current_size += self._ramp_size
+        current_size = max(0, min(self._max_size, current_size))
+        return current_size
+
+    def clamp_param(self):
+        """this need to be called after each update"""
+        self.current_val.data.clamp_(0, 1)
+
+
+class AdaptiveSpan(nn.Module):
+    """Adaptive attention span for Transformerself.
+    This module learns an attention span length from data for each
+    self-attention head.
+    Args:
+        attn_span: maximum attention span
+        adapt_span_loss: loss coefficient for the span length
+        adapt_span_ramp: length of the masking ramp
+        adapt_span_init: initial size ratio
+        adapt_span_cache: adapt cache size to reduce memory usage
+    """
+
+    def __init__(
+        self,
+        attn_span,
+        adapt_span_ramp,
+        adapt_span_init,
+        n_head,
+        adapt_span_layer,
+        **kargs
+    ):
+        nn.Module.__init__(self)
+        self._max_span = attn_span
+        self._n_head = n_head
+        self._adapt_span_layer = adapt_span_layer
+        if self._adapt_span_layer:
+            self._mask = AdaptiveMask(
+                max_size=self._max_span,
+                ramp_size=adapt_span_ramp,
+                init_val=adapt_span_init,
+            )
+        else:
+            self._mask = AdaptiveMask(
+                max_size=self._max_span,
+                ramp_size=adapt_span_ramp,
+                init_val=adapt_span_init,
+                shape=(n_head, 1, 1),
+            )
+
+    def forward(self, attn, normalize=True):
+        """mask attention with the right span"""
+        # batch and head dimensions are merged together, so separate them first
+        self.clamp_param()
+        if self._adapt_span_layer:
+            attn = self._mask(attn)
+        else:
+            B = attn.size(0)  # batch size
+            M = attn.size(1)  # block size
+            attn = attn.reshape(B // self._n_head, self._n_head, M, -1)
+            attn = self._mask(attn)
+            attn = attn.view(B, M, -1)
+        return attn
+
+    def get_trim_len(self):
+        """how much of memory can be trimmed to reduce computation"""
+        L = self._max_span
+        trim_len = min(L - 1, L - self._mask.get_current_max_size())
+        # too fine granularity might be bad for the memory management
+        trim_len = math.floor(trim_len / 64) * 64
+        return trim_len
+
+    def trim_memory(self, query, key, value, key_pe):
+        """trim out unnecessary memory beforehand to reduce computation"""
+        trim_len = self.get_trim_len()
+        cache_size = key.size(1) - query.size(1)
+        trim_len_cache = trim_len - (self._max_span - cache_size)
+        if trim_len_cache > 0:
+            key = key[:, trim_len_cache:, :]
+            value = value[:, trim_len_cache:, :]
+        elif trim_len_cache < 0:
+            # cache is too short! this happens when validation resumes
+            # after a lot of updates.
+            key = F.pad(key, [0, 0, -trim_len_cache, 0])
+            value = F.pad(value, [0, 0, -trim_len_cache, 0])
+        if trim_len > 0:
+            if key_pe is not None:
+                key_pe = key_pe[:, :, trim_len:]
+        return key, value, key_pe
+
+    def get_cache_size(self):
+        """determine how long the cache should be"""
+        trim_len = self.get_trim_len()
+        # give a buffer of 64 steps since a span might increase
+        # in future updates
+        return min(self._max_span, self._max_span - trim_len + 64)
+
+    def get_loss(self):
+        """a loss term for regularizing the span length"""
+        return self._max_span * self._mask.current_val.float().mean()
+
+    def get_current_max_span(self):
+        return self._mask.get_current_max_size()
+
+    def get_current_avg_span(self):
+        return self._mask.get_current_avg_size()
+
+    def clamp_param(self):
+        self._mask.clamp_param()

fairseq/examples/adaptive_span/adaptive_span_loss.py

@@ -0,0 +1,106 @@
+# 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
+from dataclasses import dataclass
+
+import torch.nn.functional as F
+from fairseq import metrics, utils
+from fairseq.criterions import register_criterion
+from fairseq.criterions.cross_entropy import CrossEntropyCriterion
+from fairseq.dataclass import FairseqDataclass
+from omegaconf import II
+
+
+@dataclass
+class AdaptiveSpanCriterionConfig(FairseqDataclass):
+    sentence_avg: bool = II("optimization.sentence_avg")
+
+
+@register_criterion("adaptive_span_loss", dataclass=AdaptiveSpanCriterionConfig)
+class AdaptiveSpanCriterion(CrossEntropyCriterion):
+    def __init__(self, task, sentence_avg):
+        super().__init__(task, sentence_avg)
+
+    def forward(self, model, sample, reduce=True):
+        """Compute the loss for the given sample.
+
+        Returns a tuple with three elements:
+        1) the loss here is summed, different from the adaptive span code
+        2) the sample size, which is used as the denominator for the gradient
+        3) logging outputs to display while training
+        """
+        net_output = model(**sample["net_input"])
+        loss, aux_loss, avg_span, max_span = self.compute_loss(
+            model, net_output, sample, reduce=reduce
+        )
+        sample_size = (
+            sample["target"].size(0) if self.sentence_avg else sample["ntokens"]
+        )
+        loss /= sample_size
+        total_loss = loss + aux_loss
+        sample_size = 1
+
+        logging_output = {
+            "loss": loss.data,
+            "ntokens": sample["ntokens"],
+            "nsentences": sample["target"].size(0),
+            "sample_size": sample_size,
+            "total_loss": total_loss.data,
+            "avg_span": avg_span * sample_size,
+            "max_span": max_span * sample_size,
+        }
+        return total_loss, sample_size, logging_output
+
+    def compute_loss(self, model, net_output, sample, reduce=True):
+        loss, _ = super().compute_loss(model, net_output, sample, reduce)
+        aux_loss = model.get_aux_loss()
+        avg_span = model.get_current_avg_span()
+        max_span = model.get_current_max_span()
+        return loss, aux_loss, avg_span, max_span
+
+    @staticmethod
+    def reduce_metrics(logging_outputs) -> None:
+        """Aggregate logging outputs from data parallel training."""
+        loss_sum = sum(log.get("loss", 0) for log in logging_outputs)
+        ntokens = sum(log.get("ntokens", 0) for log in logging_outputs)
+        sample_size = sum(log.get("sample_size", 0) for log in logging_outputs)
+        total_loss_sum = sum(log.get("total_loss", 0) for log in logging_outputs)
+        avg_span_sum = sum(log.get("avg_span", 0) for log in logging_outputs)
+        max_span_sum = sum(log.get("max_span", 0) for log in logging_outputs)
+
+        # we divide by log(2) to convert the loss from base e to base 2
+        metrics.log_scalar(
+            "loss", loss_sum / sample_size / math.log(2), sample_size, round=3
+        )
+        metrics.log_scalar("avg_span", avg_span_sum / sample_size, sample_size, round=3)
+        metrics.log_scalar("max_span", max_span_sum / sample_size, sample_size, round=3)
+        # total loss contains the L1 norm on adaptive-span
+        metrics.log_scalar(
+            "total_loss",
+            total_loss_sum / sample_size / math.log(2),
+            sample_size,
+            round=3,
+        )
+        if sample_size != ntokens:
+            metrics.log_scalar(
+                "nll_loss", loss_sum / ntokens / math.log(2), ntokens, round=3
+            )
+            metrics.log_derived(
+                "ppl", lambda meters: utils.get_perplexity(meters["nll_loss"].avg)
+            )
+        else:
+            metrics.log_derived(
+                "ppl", lambda meters: utils.get_perplexity(meters["loss"].avg)
+            )
+
+    @staticmethod
+    def logging_outputs_can_be_summed() -> bool:
+        """
+        Whether the logging outputs returned by `forward` can be summed
+        across workers prior to calling `reduce_metrics`. Setting this
+        to True will improves distributed training speed.
+        """
+        return True

fairseq/examples/adaptive_span/adaptive_span_model.py

@@ -0,0 +1,263 @@
+# Copyright (c) Facebook, Inc. and its affiliates.
+# All rights reserved.
+#
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from fairseq.modules.layer_norm import LayerNorm
+
+from .adaptive_span_attention import AdaptiveSpan
+
+# Size notations:
+# B = batch_size, H = d_model, M = block_size, L = attn_span
+
+
+def _skew(X, pad_value):
+    """shift every row 1 step to right"""
+    # X = B x M x L
+    B, M, L = X.size()
+    X = F.pad(X, (0, M + 1), value=pad_value)  # B x M x (L+M+1)
+    X = X.view(B, -1)  # B x ML+MM+M
+    X = X[:, :-M]  # B x ML+MM
+    X = X.view(B, M, M + L)  # B x M x L+M
+    return X
+
+
+def _unskew(X):
+    """reverse _skew operation"""
+    # X = B x M x L+M
+    B, M, L = X.size()
+    L -= M
+    X = X.view(B, -1)  # B x ML+MM
+    X = F.pad(X, (0, M))  # B x ML+MM+M
+    X = X.view(B, M, M + L + 1)  # B x M x L+M+1
+    X = X[:, :, :L]  # B x M x L
+    return X
+
+
+class SeqAttention(nn.Module):
+    """Sequential self-attention layer.
+    Each token will attend to its previous fixed number of steps.
+    Note that attention doesn't include the current step itself.
+    """
+
+    def __init__(self, d_model, n_head, attn_span, dropout, adapt_span_layer, **kargs):
+        nn.Module.__init__(self)
+        self.dropout = nn.Dropout(dropout)
+        self.d_model = d_model  # size of a single head
+        self.attn_span = attn_span
+        self.adaptive_span = AdaptiveSpan(
+            attn_span=attn_span,
+            n_head=n_head,
+            adapt_span_layer=adapt_span_layer,
+            **kargs
+        )
+
+    def forward(self, query, key, value, key_pe):
+        # query size = B x M x H
+        # key, value sizes = B x (M+L) x H
+
+        key, value, key_pe = self.adaptive_span.trim_memory(query, key, value, key_pe)
+
+        # compute attention from context
+        # B x M (dest) x (M+L) (src)
+        attn_cont = torch.matmul(query, key.transpose(-1, -2))
+        attn_cont = _unskew(attn_cont)  # B x M x L
+
+        # compute the effect of position embedding
+        attn_pos = torch.matmul(query, key_pe)  # B x M x L_pos
+        attn = attn_cont + attn_pos
+
+        attn = attn / math.sqrt(self.d_model)  # B x M X L_pos
+
+        attn = F.softmax(attn.float(), dim=-1).type_as(attn)
+
+        # trim attention lengths according to the learned span
+        attn = self.adaptive_span(attn)
+
+        attn = self.dropout(attn)  # B x M X L_pos
+
+        attn_cont = _skew(attn, 0)  # B x M X (L+M)
+        out = torch.matmul(attn_cont, value)  # B x M x H
+        return out
+
+    def get_cache_size(self):
+        return self.adaptive_span.get_cache_size()
+
+
+class MultiHeadSeqAttention(nn.Module):
+    def __init__(self, d_model, n_head, **kargs):
+        nn.Module.__init__(self)
+        assert d_model % n_head == 0
+        self.n_head = n_head
+        self.head_dim = d_model // n_head
+        self.attn = SeqAttention(d_model=self.head_dim, n_head=n_head, **kargs)
+        self.proj_query = nn.Linear(d_model, d_model, bias=False)
+        nn.init.xavier_normal_(self.proj_query.weight)
+        self.proj_out = nn.Linear(d_model, d_model, bias=False)
+        nn.init.xavier_normal_(self.proj_out.weight)
+        self.proj_val = nn.Linear(d_model, d_model, bias=False)
+        nn.init.xavier_normal_(self.proj_val.weight)
+        self.proj_key = nn.Linear(d_model, d_model, bias=False)
+        nn.init.xavier_normal_(self.proj_key.weight)
+
+    def head_reshape(self, x):
+        K = self.n_head
+        D = self.head_dim
+        x = x.view(x.size()[:-1] + (K, D))  # B x (M+L) x K x D
+        x = x.transpose(1, 2).contiguous()  # B x K x (M+L) x D
+        x = x.view(-1, x.size(-2), x.size(-1))  # B_K x (M+L) x D
+        return x
+
+    def forward(self, query, key, value, key_pe):
+        B = query.size(0)
+        K = self.n_head
+        D = self.head_dim
+        M = query.size(1)
+
+        query = self.proj_query(query)
+        query = self.head_reshape(query)
+        value = self.proj_val(value)
+        value = self.head_reshape(value)
+        key = self.proj_key(key)
+        key = self.head_reshape(key)
+
+        out = self.attn(query, key, value, key_pe)  # B_K x M x D
+        out = out.view(B, K, M, D)  # B x K x M x D
+        out = out.transpose(1, 2).contiguous()  # B x M x K x D
+        out = out.view(B, M, -1)  # B x M x K_D
+        out = self.proj_out(out)
+        return out
+
+
+class FeedForwardLayer(nn.Module):
+    def __init__(self, d_model, d_inner, dropout, **kargs):
+        nn.Module.__init__(self)
+        self.fc1 = nn.Linear(d_model, d_inner)
+        self.fc2 = nn.Linear(d_inner, d_model)
+        nn.init.xavier_uniform_(self.fc1.weight)
+        nn.init.xavier_uniform_(self.fc2.weight)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, h):
+        h1 = F.relu(self.fc1(h))
+        h1 = self.dropout(h1)
+        h2 = self.fc2(h1)
+        return h2
+
+
+class TransformerSeqLayer(nn.Module):
+    def __init__(self, d_model, **kargs):
+        nn.Module.__init__(self)
+        self.attn = MultiHeadSeqAttention(d_model=d_model, **kargs)
+        self.norm1 = LayerNorm(d_model)
+        self.ff = FeedForwardLayer(d_model=d_model, **kargs)
+        self.norm2 = LayerNorm(d_model)
+
+    def forward(self, h, h_cache, key_pe):
+        # h = B x M x H
+        # h_cache = B x L x H
+        h_all = torch.cat([h_cache, h], dim=1)  # B x (M+L) x H
+        attn_out = self.attn(h, h_all, h_all, key_pe)
+        h = self.norm1(h + attn_out)  # B x M x H
+        if self.ff is not None:
+            ff_out = self.ff(h)
+            out = self.norm2(h + ff_out)  # B x M x H
+        else:
+            out = h
+        return out
+
+    def get_cache_size(self):
+        return self.attn.attn.get_cache_size()
+
+
+class TransformerSeq(nn.Module):
+    def __init__(
+        self,
+        vocab_size,
+        d_model,
+        n_head,
+        n_layer,
+        attn_span,
+        emb_dropout,
+        aux_loss_scaler,
+        adapt_span_layer,
+        **kargs
+    ):
+        nn.Module.__init__(self)
+        # token embeddings
+        self.in_emb = nn.Embedding(vocab_size, d_model)
+        nn.init.normal_(self.in_emb.weight, mean=0, std=d_model ** -0.5)
+        self.out_emb = nn.Linear(d_model, vocab_size)
+        self.aux_loss_scaler = aux_loss_scaler
+        if emb_dropout > 0:
+            self.emb_dropout = nn.Dropout(emb_dropout)
+        else:
+            self.emb_dropout = None
+        # position embeddings
+        self.key_pe = nn.Parameter(torch.randn(1, d_model // n_head, attn_span))
+
+        self.layers = nn.ModuleList()
+        self.layers.extend(
+            TransformerSeqLayer(
+                d_model=d_model,
+                n_head=n_head,
+                attn_span=attn_span,
+                adapt_span_layer=adapt_span_layer,
+                **kargs
+            )
+            for _ in range(n_layer)
+        )
+
+    def forward(self, x, h_cache, target=None):
+        # x size = B x M
+        block_size = x.size(1)
+        h = self.in_emb(x)  # B x M x H
+        if self.emb_dropout is not None:
+            h = self.emb_dropout(h)
+
+        h_cache_next = []
+        for l, layer in enumerate(self.layers):
+            cache_size = layer.attn.attn.get_cache_size()
+            if cache_size > block_size:
+                h_cache_next_l = torch.cat(
+                    [h_cache[l][:, -cache_size + block_size :, :], h], dim=1
+                ).detach()
+            else:
+                h_cache_next_l = h[:, -cache_size:, :].detach()
+            h_cache_next.append(h_cache_next_l)
+            h = layer(h, h_cache[l], self.key_pe)  # B x M x H
+
+        if self.emb_dropout is not None:
+            h = self.emb_dropout(h)
+
+        out = F.log_softmax(self.out_emb(h).float(), dim=-1).type_as(h)
+        dummy_loss = None
+
+        return out, h_cache_next, dummy_loss
+
+    def get_aux_loss(self):
+        loss = 0.0
+        for layer in self.layers:
+            loss += layer.attn.attn.adaptive_span.get_loss()
+        return self.aux_loss_scaler * loss
+
+    def get_current_max_span(self):
+        max_span = 0.0
+        for layer in self.layers:
+            max_span = max(
+                max_span, layer.attn.attn.adaptive_span.get_current_max_span()
+            )
+        return max_span
+
+    def get_current_avg_span(self):
+        avg_span = 0.0
+        for layer in self.layers:
+            avg_span += layer.attn.attn.adaptive_span.get_current_avg_span()
+        return avg_span / len(self.layers)

fairseq/examples/adaptive_span/adaptive_span_model_wrapper.py

@@ -0,0 +1,145 @@
+# 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 logging
+from dataclasses import dataclass
+from typing import Dict, List, Optional
+
+import torch
+from fairseq.dataclass import FairseqDataclass
+from fairseq.models import (
+    FairseqIncrementalDecoder,
+    FairseqLanguageModel,
+    register_model,
+)
+from .adaptive_span_model import TransformerSeq as AdaptiveSpanTransformerModel
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class AdaptiveSpanSmallConfig(FairseqDataclass):
+    # defaults come from https://github.com/facebookresearch/adaptive-span/blob/master/experiments/enwik8_small.sh
+    vocab_size: int = 50
+    d_model: int = 256
+    n_head: int = 4
+    d_inner: int = 1024
+    n_layer: int = 8
+    attn_span: int = 1024
+    dropout: float = 0.0
+    emb_dropout: float = 0.0
+    adapt_span_ramp: int = 32
+    adapt_span_init: float = 0.0
+    aux_loss_scaler: float = 0.000002
+    adapt_span_layer: bool = False
+
+
+@register_model("adaptive_span", dataclass=AdaptiveSpanSmallConfig)
+class AdaptiveSpanTransformer(FairseqLanguageModel):
+    @classmethod
+    def build_model(cls, cfg: AdaptiveSpanSmallConfig, task):
+        return cls(AdaptiveSpanDecoder(cfg, task))
+
+    def get_aux_loss(self):
+        return self.decoder.get_aux_loss()
+
+    def get_current_max_span(self):
+        return self.decoder.get_current_max_span()
+
+    def get_current_avg_span(self):
+        return self.decoder.get_current_avg_span()
+
+
+class AdaptiveSpanDecoder(FairseqIncrementalDecoder):
+    def __init__(self, cfg, task):
+
+        super().__init__(task.target_dictionary)
+
+        self.config = cfg
+        config = AdaptiveSpanSmallConfig(
+            vocab_size=len(task.target_dictionary),
+            d_model=cfg.d_model,
+            n_head=cfg.n_head,
+            d_inner=cfg.d_inner,
+            n_layer=cfg.n_layer,
+            attn_span=cfg.attn_span,
+            dropout=cfg.dropout,
+            emb_dropout=cfg.emb_dropout,
+            adapt_span_ramp=cfg.adapt_span_ramp,
+            adapt_span_init=cfg.adapt_span_init,
+            aux_loss_scaler=cfg.aux_loss_scaler,
+            adapt_span_layer=cfg.adapt_span_layer,
+        )
+        logger.info(config)
+        self.model = AdaptiveSpanTransformerModel(**config.__dict__)
+
+        self._mems = None
+
+    def forward(
+        self,
+        src_tokens,
+        incremental_state: Optional[Dict[str, List[torch.Tensor]]] = None,
+        encoder_out=None,
+    ):
+        bsz = src_tokens.size(0)
+        if incremental_state is not None:  # used during inference
+            mems = self.get_incremental_state("mems")
+            src_tokens = src_tokens[:, -1:]  # only keep the most recent token
+        else:
+            mems = self._mems
+
+        if mems is None:
+            # first time init
+            mems = self.init_hid_cache(bsz)
+        output = self.model(x=src_tokens, h_cache=mems,)
+        if incremental_state is not None:
+            self.set_incremental_state(incremental_state, "mems", output[1])
+        else:
+            self._mems = output[1]
+        return (output[0],)
+
+    def max_positions(self):
+        return self.config.attn_span
+
+    def init_hid_cache(self, batch_sz):
+        hid = []
+        for layer in self.model.layers:
+            param = next(self.model.parameters())
+            h = torch.zeros(
+                batch_sz,
+                layer.get_cache_size(),
+                self.config.d_model,
+                dtype=param.dtype,
+                device=param.device,
+            )
+            hid.append(h)
+        return hid
+
+    def get_aux_loss(self):
+        return self.model.get_aux_loss()
+
+    def get_current_max_span(self):
+        return self.model.get_current_max_span()
+
+    def get_current_avg_span(self):
+        return self.model.get_current_avg_span()
+
+    def reorder_incremental_state(
+        self,
+        incremental_state: Dict[str, Dict[str, Optional[torch.Tensor]]],
+        new_order: torch.Tensor,
+    ):
+        """Reorder incremental state.
+
+        This will be called when the order of the input has changed from the
+        previous time step. A typical use case is beam search, where the input
+        order changes between time steps based on the selection of beams.
+        """
+        raise NotImplementedError("This is required for generation/beam search")
+        # mems = self.get_incremental_state(incremental_state, "mems")
+        # if mems is not None:
+        #     new_mems = [mems_i.index_select(1, new_order) for mems_i in mems]
+        #     self.set_incremental_state(incremental_state, "mems", new_mems)

fairseq/examples/adaptive_span/truncated_bptt_lm_task.py

@@ -0,0 +1,281 @@
+# 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 logging
+import os
+from dataclasses import dataclass, field
+from typing import List, Optional, Tuple
+
+import torch
+from fairseq import utils
+from fairseq.data import (
+    Dictionary,
+    TokenBlockDataset,
+    data_utils,
+    iterators,
+)
+from fairseq.dataclass import FairseqDataclass
+from fairseq.distributed import utils as dist_utils
+from fairseq.tasks import FairseqTask, register_task
+from omegaconf import II
+
+
+logger = logging.getLogger(__name__)
+
+
+@dataclass
+class TruncatedBPTTLMConfig(FairseqDataclass):
+    data: str = field(default="???", metadata={"help": "path to data directory"})
+    tokens_per_sample: int = field(
+        default=1024,
+        metadata={"help": "max number of tokens per sequence"},
+    )
+    batch_size: int = II("dataset.batch_size")
+    # Some models use *max_target_positions* to know how many positional
+    # embeddings to learn. We use II(...) to make it default to
+    # *tokens_per_sample*, but in principle there could be more positional
+    # embeddings than tokens in a single batch. This may also be irrelevant for
+    # custom model implementations.
+    max_target_positions: int = II("task.tokens_per_sample")
+    # these will be populated automatically if not provided
+    data_parallel_rank: Optional[int] = None
+    data_parallel_size: Optional[int] = None
+
+
+@register_task("truncated_bptt_lm", dataclass=TruncatedBPTTLMConfig)
+class TruncatedBPTTLMTask(FairseqTask):
+    def __init__(self, cfg: TruncatedBPTTLMConfig):
+        super().__init__(cfg)
+
+        if cfg.data_parallel_rank is None or cfg.data_parallel_size is None:
+            if torch.distributed.is_initialized():
+                cfg.data_parallel_rank = dist_utils.get_data_parallel_rank()
+                cfg.data_parallel_size = dist_utils.get_data_parallel_world_size()
+            else:
+                cfg.data_parallel_rank = 0
+                cfg.data_parallel_size = 1
+
+        # load the dictionary
+        paths = utils.split_paths(cfg.data)
+        assert len(paths) > 0
+        self.dictionary = Dictionary.load(os.path.join(paths[0], "dict.txt"))
+        logger.info("dictionary: {} types".format(len(self.dictionary)))
+
+    def load_dataset(self, split, epoch=1, combine=False, **kwargs):
+        """Load a given dataset split (e.g., train, valid, test)"""
+
+        # support sharded datasets
+        paths = utils.split_paths(self.cfg.data)
+        assert len(paths) > 0
+        data_path = paths[(epoch - 1) % len(paths)]
+        split_path = os.path.join(data_path, split)
+
+        # each element of *data* will be a tensorized line from the original
+        # text dataset, similar to ``open(split_path).readlines()``
+        data = data_utils.load_indexed_dataset(
+            split_path, self.dictionary, combine=combine
+        )
+        if data is None:
+            raise FileNotFoundError(
+                "Dataset not found: {} ({})".format(split, split_path)
+            )
+
+        # this is similar to ``data.view(-1).split(tokens_per_sample)``
+        data = TokenBlockDataset(
+            data,
+            data.sizes,
+            block_size=self.cfg.tokens_per_sample,
+            pad=None,  # unused
+            eos=None,  # unused
+            break_mode="none",
+        )
+
+        self.datasets[split] = TruncatedBPTTDataset(
+            data=data,
+            bsz_per_shard=self.cfg.batch_size,
+            shard_id=self.cfg.data_parallel_rank,
+            num_shards=self.cfg.data_parallel_size,
+        )
+
+    def dataset(self, split):
+        return self.datasets[split]
+
+    def get_batch_iterator(
+        self, dataset, num_workers=0, epoch=1, data_buffer_size=0, **kwargs
+    ):
+        return iterators.EpochBatchIterator(
+            dataset=dataset,
+            collate_fn=self._collate_fn,
+            num_workers=num_workers,
+            epoch=epoch,
+            buffer_size=data_buffer_size,
+            # we don't use the batching functionality from EpochBatchIterator;
+            # instead every item in *dataset* is a whole batch
+            batch_sampler=[[i] for i in range(len(dataset))],
+            disable_shuffling=True,
+        )
+
+    def _collate_fn(self, items: List[List[torch.Tensor]]):
+        # we don't use fairseq's batching functionality, so we expect a single
+        # Tensor of type List[torch.Tensor]
+        assert len(items) == 1
+
+        # item will have shape B x T (the last batch may have length < T)
+        id, item = items[0]
+        item = data_utils.collate_tokens(item, pad_idx=self.source_dictionary.pad())
+        B, T = item.size()
+
+        # shift item one position over and append a padding token for the target
+        target = torch.nn.functional.pad(
+            item[:, 1:], (0, 1, 0, 0), value=self.target_dictionary.pad()
+        )
+
+        # fairseq expects batches to have the following structure
+        return {
+            "id": torch.tensor([id]*item.size(0)),
+            "net_input": {
+                "src_tokens": item,
+            },
+            "target": target,
+            "nsentences": item.size(0),
+            "ntokens": item.numel(),
+        }
+
+    def build_dataset_for_inference(
+        self, src_tokens: List[torch.Tensor], src_lengths: List[int], **kwargs
+    ) -> torch.utils.data.Dataset:
+        eos = self.source_dictionary.eos()
+        dataset = TokenBlockDataset(
+            src_tokens,
+            src_lengths,
+            block_size=None,  # ignored for "eos" break mode
+            pad=self.source_dictionary.pad(),
+            eos=eos,
+            break_mode="eos",
+        )
+
+        class Dataset(torch.utils.data.Dataset):
+            def __getitem__(self, i):
+                item = dataset[i]
+                if item[-1] == eos:
+                    # remove eos to support generating with a prefix
+                    item = item[:-1]
+                return (i, [item])
+
+            def __len__(self):
+                return len(dataset)
+
+        return Dataset()
+
+    def inference_step(
+        self, generator, models, sample, prefix_tokens=None, constraints=None
+    ):
+        with torch.no_grad():
+            if constraints is not None:
+                raise NotImplementedError
+
+            # SequenceGenerator doesn't use *src_tokens* directly, we need to
+            # pass the *prefix_tokens* argument instead.
+            if prefix_tokens is None and sample["net_input"]["src_tokens"].nelement():
+                prefix_tokens = sample["net_input"]["src_tokens"]
+
+            # begin generation with the end-of-sentence token
+            bos_token = self.source_dictionary.eos()
+
+            return generator.generate(
+                models, sample, prefix_tokens=prefix_tokens, bos_token=bos_token
+            )
+
+    def eval_lm_dataloader(
+        self,
+        dataset,
+        max_tokens: Optional[int] = 36000,
+        batch_size: Optional[int] = None,
+        max_positions: Optional[int] = None,
+        num_shards: int = 1,
+        shard_id: int = 0,
+        num_workers: int = 1,
+        data_buffer_size: int = 10,
+        context_window: int = 0,
+    ):
+        if context_window > 0:
+            raise NotImplementedError(
+                "Transformer-XL doesn't need --context-window, try "
+                "--model-overrides '{\"mem_len\":42}' instead "
+            )
+        return self.get_batch_iterator(
+            dataset=dataset,
+            max_tokens=max_tokens,
+            max_sentences=batch_size,
+            max_positions=max_positions,
+            ignore_invalid_inputs=True,
+            num_shards=num_shards,
+            shard_id=shard_id,
+            num_workers=num_workers,
+            data_buffer_size=data_buffer_size,
+        ).next_epoch_itr(shuffle=False)
+
+    @property
+    def source_dictionary(self):
+        return self.dictionary
+
+    @property
+    def target_dictionary(self):
+        return self.dictionary
+
+
+class TruncatedBPTTDataset(torch.utils.data.Dataset):
+    def __init__(
+        self,
+        data: List[torch.Tensor],  # ordered list of items
+        bsz_per_shard,  # number of items processed per GPUs per forward
+        shard_id,  # current GPU ID
+        num_shards,  # number of GPUs
+    ):
+        super().__init__()
+        self.data = data
+
+        def batchify(data, bsz):
+            # Work out how cleanly we can divide the dataset into bsz parts.
+            nbatch = data.size(0) // bsz
+            # Trim off any extra elements that wouldn't cleanly fit (remainders).
+            data = data.narrow(0, 0, nbatch * bsz)
+            # Evenly divide the data across the bsz batches.
+            data = data.view(bsz, -1).contiguous()
+            return data
+
+        # total number of sequences processed by all GPUs in each forward pass
+        global_batch_size = bsz_per_shard * num_shards
+
+        """
+        With a 16 item dataset, bsz_per_shard=2 and num_shards=3,
+        *indices* might look like:
+
+            indices = [[0, 1],
+                       [2, 3],
+                       [4, 5],
+                       [6, 7],
+                       [8, 9],
+                       [10, 11]]
+
+        The size of the TruncatedBPTTDataset instance will be 2,
+        and shard 1 will see items:
+
+            [(0, [data[4], data[6]]),
+             (1, [data[5], data[7]])]
+        """
+        indices = batchify(torch.arange(len(data)), global_batch_size)
+        assert indices.size(0) == global_batch_size
+
+        self.my_indices = indices[
+            shard_id * bsz_per_shard : (shard_id + 1) * bsz_per_shard
+        ]
+        assert self.my_indices.size(0) == bsz_per_shard
+
+    def __len__(self):
+        return self.my_indices.size(1)
+
+    def __getitem__(self, i) -> Tuple[int, List[torch.Tensor]]:
+        return (i, [self.data[idx] for idx in self.my_indices[:, i]])

fairseq/examples/backtranslation/README.md

@@ -0,0 +1,297 @@
+# Understanding Back-Translation at Scale (Edunov et al., 2018)
+
+This page includes pre-trained models from the paper [Understanding Back-Translation at Scale (Edunov et al., 2018)](https://arxiv.org/abs/1808.09381).
+
+## Pre-trained models
+
+Model | Description | Dataset | Download
+---|---|---|---
+`transformer.wmt18.en-de` | Transformer <br> ([Edunov et al., 2018](https://arxiv.org/abs/1808.09381)) <br> WMT'18 winner | [WMT'18 English-German](http://www.statmt.org/wmt18/translation-task.html) | [download (.tar.gz)](https://dl.fbaipublicfiles.com/fairseq/models/wmt18.en-de.ensemble.tar.gz) <br> See NOTE in the archive
+
+## Example usage (torch.hub)
+
+We require a few additional Python dependencies for preprocessing:
+```bash
+pip install subword_nmt sacremoses
+```
+
+Then to generate translations from the full model ensemble:
+```python
+import torch
+
+# List available models
+torch.hub.list('pytorch/fairseq')  # [..., 'transformer.wmt18.en-de', ... ]
+
+# Load the WMT'18 En-De ensemble
+en2de_ensemble = torch.hub.load(
+    'pytorch/fairseq', 'transformer.wmt18.en-de',
+    checkpoint_file='wmt18.model1.pt:wmt18.model2.pt:wmt18.model3.pt:wmt18.model4.pt:wmt18.model5.pt',
+    tokenizer='moses', bpe='subword_nmt')
+
+# The ensemble contains 5 models
+len(en2de_ensemble.models)
+# 5
+
+# Translate
+en2de_ensemble.translate('Hello world!')
+# 'Hallo Welt!'
+```
+
+## Training your own model (WMT'18 English-German)
+
+The following instructions can be adapted to reproduce the models from the paper.
+
+
+#### Step 1. Prepare parallel data and optionally train a baseline (English-German) model
+
+First download and preprocess the data:
+```bash
+# Download and prepare the data
+cd examples/backtranslation/
+bash prepare-wmt18en2de.sh
+cd ../..
+
+# Binarize the data
+TEXT=examples/backtranslation/wmt18_en_de
+fairseq-preprocess \
+    --joined-dictionary \
+    --source-lang en --target-lang de \
+    --trainpref $TEXT/train --validpref $TEXT/valid --testpref $TEXT/test \
+    --destdir data-bin/wmt18_en_de --thresholdtgt 0 --thresholdsrc 0 \
+    --workers 20
+
+# Copy the BPE code into the data-bin directory for future use
+cp examples/backtranslation/wmt18_en_de/code data-bin/wmt18_en_de/code
+```
+
+(Optionally) Train a baseline model (English-German) using just the parallel data:
+```bash
+CHECKPOINT_DIR=checkpoints_en_de_parallel
+fairseq-train --fp16 \
+    data-bin/wmt18_en_de \
+    --source-lang en --target-lang de \
+    --arch transformer_wmt_en_de_big --share-all-embeddings \
+    --dropout 0.3 --weight-decay 0.0 \
+    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
+    --lr 0.001 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+    --max-tokens 3584 --update-freq 16 \
+    --max-update 30000 \
+    --save-dir $CHECKPOINT_DIR
+# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a
+# different number of GPUs.
+```
+
+Average the last 10 checkpoints:
+```bash
+python scripts/average_checkpoints.py \
+    --inputs $CHECKPOINT_DIR \
+    --num-epoch-checkpoints 10 \
+    --output $CHECKPOINT_DIR/checkpoint.avg10.pt
+```
+
+Evaluate BLEU:
+```bash
+# tokenized BLEU on newstest2017:
+bash examples/backtranslation/tokenized_bleu.sh \
+    wmt17 \
+    en-de \
+    data-bin/wmt18_en_de \
+    data-bin/wmt18_en_de/code \
+    $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU4 = 29.57, 60.9/35.4/22.9/15.5 (BP=1.000, ratio=1.014, syslen=63049, reflen=62152)
+# compare to 29.46 in Table 1, which is also for tokenized BLEU
+
+# generally it's better to report (detokenized) sacrebleu though:
+bash examples/backtranslation/sacrebleu.sh \
+    wmt17 \
+    en-de \
+    data-bin/wmt18_en_de \
+    data-bin/wmt18_en_de/code \
+    $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU+case.mixed+lang.en-de+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 29.0 60.6/34.7/22.4/14.9 (BP = 1.000 ratio = 1.013 hyp_len = 62099 ref_len = 61287)
+```
+
+
+#### Step 2. Back-translate monolingual German data
+
+Train a reverse model (German-English) to do the back-translation:
+```bash
+CHECKPOINT_DIR=checkpoints_de_en_parallel
+fairseq-train --fp16 \
+    data-bin/wmt18_en_de \
+    --source-lang de --target-lang en \
+    --arch transformer_wmt_en_de_big --share-all-embeddings \
+    --dropout 0.3 --weight-decay 0.0 \
+    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
+    --lr 0.001 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+    --max-tokens 3584 --update-freq 16 \
+    --max-update 30000 \
+    --save-dir $CHECKPOINT_DIR
+# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a
+# different number of GPUs.
+```
+
+Let's evaluate the back-translation (BT) model to make sure it is well trained:
+```bash
+bash examples/backtranslation/sacrebleu.sh \
+    wmt17 \
+    de-en \
+    data-bin/wmt18_en_de \
+    data-bin/wmt18_en_de/code \
+    $CHECKPOINT_DIR/checkpoint_best.py
+# BLEU+case.mixed+lang.de-en+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 34.9 66.9/41.8/28.5/19.9 (BP = 0.983 ratio = 0.984 hyp_len = 63342 ref_len = 64399)
+# compare to the best system from WMT'17 which scored 35.1: http://matrix.statmt.org/matrix/systems_list/1868
+```
+
+Next prepare the monolingual data:
+```bash
+# Download and prepare the monolingual data
+# By default the script samples 25M monolingual sentences, which after
+# deduplication should be just over 24M sentences. These are split into 25
+# shards, each with 1M sentences (except for the last shard).
+cd examples/backtranslation/
+bash prepare-de-monolingual.sh
+cd ../..
+
+# Binarize each shard of the monolingual data
+TEXT=examples/backtranslation/wmt18_de_mono
+for SHARD in $(seq -f "%02g" 0 24); do \
+    fairseq-preprocess \
+        --only-source \
+        --source-lang de --target-lang en \
+        --joined-dictionary \
+        --srcdict data-bin/wmt18_en_de/dict.de.txt \
+        --testpref $TEXT/bpe.monolingual.dedup.${SHARD} \
+        --destdir data-bin/wmt18_de_mono/shard${SHARD} \
+        --workers 20; \
+    cp data-bin/wmt18_en_de/dict.en.txt data-bin/wmt18_de_mono/shard${SHARD}/; \
+done
+```
+
+Now we're ready to perform back-translation over the monolingual data. The
+following command generates via sampling, but it's possible to use greedy
+decoding (`--beam 1`), beam search (`--beam 5`),
+top-k sampling (`--sampling --beam 1 --sampling-topk 10`), etc.:
+```bash
+mkdir backtranslation_output
+for SHARD in $(seq -f "%02g" 0 24); do \
+    fairseq-generate --fp16 \
+        data-bin/wmt18_de_mono/shard${SHARD} \
+        --path $CHECKPOINT_DIR/checkpoint_best.pt \
+        --skip-invalid-size-inputs-valid-test \
+        --max-tokens 4096 \
+        --sampling --beam 1 \
+    > backtranslation_output/sampling.shard${SHARD}.out; \
+done
+```
+
+After BT, use the `extract_bt_data.py` script to re-combine the shards, extract
+the back-translations and apply length ratio filters:
+```bash
+python examples/backtranslation/extract_bt_data.py \
+    --minlen 1 --maxlen 250 --ratio 1.5 \
+    --output backtranslation_output/bt_data --srclang en --tgtlang de \
+    backtranslation_output/sampling.shard*.out
+
+# Ensure lengths are the same:
+# wc -l backtranslation_output/bt_data.{en,de}
+#   21795614 backtranslation_output/bt_data.en
+#   21795614 backtranslation_output/bt_data.de
+#   43591228 total
+```
+
+Binarize the filtered BT data and combine it with the parallel data:
+```bash
+TEXT=backtranslation_output
+fairseq-preprocess \
+    --source-lang en --target-lang de \
+    --joined-dictionary \
+    --srcdict data-bin/wmt18_en_de/dict.en.txt \
+    --trainpref $TEXT/bt_data \
+    --destdir data-bin/wmt18_en_de_bt \
+    --workers 20
+
+# We want to train on the combined data, so we'll symlink the parallel + BT data
+# in the wmt18_en_de_para_plus_bt directory. We link the parallel data as "train"
+# and the BT data as "train1", so that fairseq will combine them automatically
+# and so that we can use the `--upsample-primary` option to upsample the
+# parallel data (if desired).
+PARA_DATA=$(readlink -f data-bin/wmt18_en_de)
+BT_DATA=$(readlink -f data-bin/wmt18_en_de_bt)
+COMB_DATA=data-bin/wmt18_en_de_para_plus_bt
+mkdir -p $COMB_DATA
+for LANG in en de; do \
+    ln -s ${PARA_DATA}/dict.$LANG.txt ${COMB_DATA}/dict.$LANG.txt; \
+    for EXT in bin idx; do \
+        ln -s ${PARA_DATA}/train.en-de.$LANG.$EXT ${COMB_DATA}/train.en-de.$LANG.$EXT; \
+        ln -s ${BT_DATA}/train.en-de.$LANG.$EXT ${COMB_DATA}/train1.en-de.$LANG.$EXT; \
+        ln -s ${PARA_DATA}/valid.en-de.$LANG.$EXT ${COMB_DATA}/valid.en-de.$LANG.$EXT; \
+        ln -s ${PARA_DATA}/test.en-de.$LANG.$EXT ${COMB_DATA}/test.en-de.$LANG.$EXT; \
+    done; \
+done
+```
+
+
+#### 3. Train an English-German model over the combined parallel + BT data
+
+Finally we can train a model over the parallel + BT data:
+```bash
+CHECKPOINT_DIR=checkpoints_en_de_parallel_plus_bt
+fairseq-train --fp16 \
+    data-bin/wmt18_en_de_para_plus_bt \
+    --upsample-primary 16 \
+    --source-lang en --target-lang de \
+    --arch transformer_wmt_en_de_big --share-all-embeddings \
+    --dropout 0.3 --weight-decay 0.0 \
+    --criterion label_smoothed_cross_entropy --label-smoothing 0.1 \
+    --optimizer adam --adam-betas '(0.9, 0.98)' --clip-norm 0.0 \
+    --lr 0.0007 --lr-scheduler inverse_sqrt --warmup-updates 4000 \
+    --max-tokens 3584 --update-freq 16 \
+    --max-update 100000 \
+    --save-dir $CHECKPOINT_DIR
+# Note: the above command assumes 8 GPUs. Adjust `--update-freq` if you have a
+# different number of GPUs.
+```
+
+Average the last 10 checkpoints:
+```bash
+python scripts/average_checkpoints.py \
+    --inputs $CHECKPOINT_DIR \
+    --num-epoch-checkpoints 10 \
+    --output $CHECKPOINT_DIR/checkpoint.avg10.pt
+```
+
+Evaluate BLEU:
+```bash
+# tokenized BLEU on newstest2017:
+bash examples/backtranslation/tokenized_bleu.sh \
+    wmt17 \
+    en-de \
+    data-bin/wmt18_en_de \
+    data-bin/wmt18_en_de/code \
+    $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU4 = 32.35, 64.4/38.9/26.2/18.3 (BP=0.977, ratio=0.977, syslen=60729, reflen=62152)
+# compare to 32.35 in Table 1, which is also for tokenized BLEU
+
+# generally it's better to report (detokenized) sacrebleu:
+bash examples/backtranslation/sacrebleu.sh \
+    wmt17 \
+    en-de \
+    data-bin/wmt18_en_de \
+    data-bin/wmt18_en_de/code \
+    $CHECKPOINT_DIR/checkpoint.avg10.pt
+# BLEU+case.mixed+lang.en-de+numrefs.1+smooth.exp+test.wmt17+tok.13a+version.1.4.3 = 31.5 64.3/38.2/25.6/17.6 (BP = 0.971 ratio = 0.971 hyp_len = 59515 ref_len = 61287)
+```
+
+
+## Citation
+```bibtex
+@inproceedings{edunov2018backtranslation,
+  title = {Understanding Back-Translation at Scale},
+  author = {Edunov, Sergey and Ott, Myle and Auli, Michael and Grangier, David},
+  booktitle = {Conference of the Association for Computational Linguistics (ACL)},
+  year = 2018,
+}
+```

fairseq/examples/backtranslation/deduplicate_lines.py

@@ -0,0 +1,41 @@
+#!/usr/bin/python3
+# 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 argparse
+import fileinput
+import hashlib
+import sys
+from multiprocessing import Pool
+
+
+def get_hashes_and_lines(raw_line):
+    hash = hashlib.md5(raw_line).hexdigest()
+    return hash, raw_line
+
+
+def main():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("--workers", type=int, default=10)
+    parser.add_argument("files", nargs="*", help="input files")
+    args = parser.parse_args()
+
+    seen = set()
+    with fileinput.input(args.files, mode="rb") as h:
+        pool = Pool(args.workers)
+        results = pool.imap_unordered(get_hashes_and_lines, h, 1000)
+        for i, (hash, raw_line) in enumerate(results):
+            if hash not in seen:
+                seen.add(hash)
+                sys.stdout.buffer.write(raw_line)
+            if i % 1000000 == 0:
+                print(i, file=sys.stderr, end="", flush=True)
+            elif i % 100000 == 0:
+                print(".", file=sys.stderr, end="", flush=True)
+    print(file=sys.stderr, flush=True)
+
+
+if __name__ == "__main__":
+    main()

fairseq/examples/backtranslation/extract_bt_data.py

@@ -0,0 +1,72 @@
+#!/usr/bin/env python
+# 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 argparse
+import fileinput
+
+from tqdm import tqdm
+
+
+def main():
+    parser = argparse.ArgumentParser(
+        description=(
+            "Extract back-translations from the stdout of fairseq-generate. "
+            "If there are multiply hypotheses for a source, we only keep the first one. "
+        )
+    )
+    parser.add_argument("--output", required=True, help="output prefix")
+    parser.add_argument(
+        "--srclang", required=True, help="source language (extracted from H-* lines)"
+    )
+    parser.add_argument(
+        "--tgtlang", required=True, help="target language (extracted from S-* lines)"
+    )
+    parser.add_argument("--minlen", type=int, help="min length filter")
+    parser.add_argument("--maxlen", type=int, help="max length filter")
+    parser.add_argument("--ratio", type=float, help="ratio filter")
+    parser.add_argument("files", nargs="*", help="input files")
+    args = parser.parse_args()
+
+    def validate(src, tgt):
+        srclen = len(src.split(" ")) if src != "" else 0
+        tgtlen = len(tgt.split(" ")) if tgt != "" else 0
+        if (
+            (args.minlen is not None and (srclen < args.minlen or tgtlen < args.minlen))
+            or (
+                args.maxlen is not None
+                and (srclen > args.maxlen or tgtlen > args.maxlen)
+            )
+            or (
+                args.ratio is not None
+                and (max(srclen, tgtlen) / float(min(srclen, tgtlen)) > args.ratio)
+            )
+        ):
+            return False
+        return True
+
+    def safe_index(toks, index, default):
+        try:
+            return toks[index]
+        except IndexError:
+            return default
+
+    with open(args.output + "." + args.srclang, "w") as src_h, open(
+        args.output + "." + args.tgtlang, "w"
+    ) as tgt_h:
+        for line in tqdm(fileinput.input(args.files)):
+            if line.startswith("S-"):
+                tgt = safe_index(line.rstrip().split("\t"), 1, "")
+            elif line.startswith("H-"):
+                if tgt is not None:
+                    src = safe_index(line.rstrip().split("\t"), 2, "")
+                    if validate(src, tgt):
+                        print(src, file=src_h)
+                        print(tgt, file=tgt_h)
+                    tgt = None
+
+
+if __name__ == "__main__":
+    main()

fairseq/examples/backtranslation/prepare-de-monolingual.sh

@@ -0,0 +1,98 @@
+#!/bin/bash
+
+SCRIPTS=mosesdecoder/scripts
+TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl
+NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl
+REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl
+BPEROOT=subword-nmt/subword_nmt
+
+
+BPE_CODE=wmt18_en_de/code
+SUBSAMPLE_SIZE=25000000
+LANG=de
+
+
+OUTDIR=wmt18_${LANG}_mono
+orig=orig
+tmp=$OUTDIR/tmp
+mkdir -p $OUTDIR $tmp
+
+
+URLS=(
+    "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2007.de.shuffled.gz"
+    "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2008.de.shuffled.gz"
+    "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2009.de.shuffled.gz"
+    "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2010.de.shuffled.gz"
+    "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2011.de.shuffled.gz"
+    "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2012.de.shuffled.gz"
+    "http://www.statmt.org/wmt14/training-monolingual-news-crawl/news.2013.de.shuffled.gz"
+    "http://www.statmt.org/wmt15/training-monolingual-news-crawl-v2/news.2014.de.shuffled.v2.gz"
+    "http://data.statmt.org/wmt16/translation-task/news.2015.de.shuffled.gz"
+    "http://data.statmt.org/wmt17/translation-task/news.2016.de.shuffled.gz"
+    "http://data.statmt.org/wmt18/translation-task/news.2017.de.shuffled.deduped.gz"
+)
+FILES=(
+    "news.2007.de.shuffled.gz"
+    "news.2008.de.shuffled.gz"
+    "news.2009.de.shuffled.gz"
+    "news.2010.de.shuffled.gz"
+    "news.2011.de.shuffled.gz"
+    "news.2012.de.shuffled.gz"
+    "news.2013.de.shuffled.gz"
+    "news.2014.de.shuffled.v2.gz"
+    "news.2015.de.shuffled.gz"
+    "news.2016.de.shuffled.gz"
+    "news.2017.de.shuffled.deduped.gz"
+)
+
+
+cd $orig
+for ((i=0;i<${#URLS[@]};++i)); do
+    file=${FILES[i]}
+    if [ -f $file ]; then
+        echo "$file already exists, skipping download"
+    else
+        url=${URLS[i]}
+        wget "$url"
+    fi
+done
+cd ..
+
+
+if [ -f $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG} ]; then
+    echo "found monolingual sample, skipping shuffle/sample/tokenize"
+else
+    gzip -c -d -k $(for FILE in "${FILES[@]}"; do echo $orig/$FILE; done) \
+    | shuf -n $SUBSAMPLE_SIZE \
+    | perl $NORM_PUNC $LANG \
+    | perl $REM_NON_PRINT_CHAR \
+    | perl $TOKENIZER -threads 8 -a -l $LANG \
+    > $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG}
+fi
+
+
+if [ -f $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG} ]; then
+    echo "found BPE monolingual sample, skipping BPE step"
+else
+    python $BPEROOT/apply_bpe.py -c $BPE_CODE \
+        < $tmp/monolingual.${SUBSAMPLE_SIZE}.${LANG} \
+        > $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG}
+fi
+
+
+if [ -f $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG} ]; then
+    echo "found deduplicated monolingual sample, skipping deduplication step"
+else
+    python deduplicate_lines.py $tmp/bpe.monolingual.${SUBSAMPLE_SIZE}.${LANG} \
+    > $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG}
+fi
+
+
+if [ -f $OUTDIR/bpe.monolingual.dedup.00.de ]; then
+    echo "found sharded data, skipping sharding step"
+else
+    split --lines 1000000 --numeric-suffixes \
+        --additional-suffix .${LANG} \
+        $tmp/bpe.monolingual.dedup.${SUBSAMPLE_SIZE}.${LANG} \
+        $OUTDIR/bpe.monolingual.dedup.
+fi

fairseq/examples/backtranslation/prepare-wmt18en2de.sh

@@ -0,0 +1,135 @@
+#!/bin/bash
+# Adapted from https://github.com/facebookresearch/MIXER/blob/master/prepareData.sh
+
+echo 'Cloning Moses github repository (for tokenization scripts)...'
+git clone https://github.com/moses-smt/mosesdecoder.git
+
+echo 'Cloning Subword NMT repository (for BPE pre-processing)...'
+git clone https://github.com/rsennrich/subword-nmt.git
+
+SCRIPTS=mosesdecoder/scripts
+TOKENIZER=$SCRIPTS/tokenizer/tokenizer.perl
+CLEAN=$SCRIPTS/training/clean-corpus-n.perl
+NORM_PUNC=$SCRIPTS/tokenizer/normalize-punctuation.perl
+REM_NON_PRINT_CHAR=$SCRIPTS/tokenizer/remove-non-printing-char.perl
+BPEROOT=subword-nmt/subword_nmt
+BPE_TOKENS=32000
+
+URLS=(
+    "http://statmt.org/wmt13/training-parallel-europarl-v7.tgz"
+    "http://statmt.org/wmt13/training-parallel-commoncrawl.tgz"
+    "http://data.statmt.org/wmt18/translation-task/training-parallel-nc-v13.tgz"
+    "http://data.statmt.org/wmt18/translation-task/rapid2016.tgz"
+    "http://data.statmt.org/wmt17/translation-task/dev.tgz"
+    "http://statmt.org/wmt14/test-full.tgz"
+)
+FILES=(
+    "training-parallel-europarl-v7.tgz"
+    "training-parallel-commoncrawl.tgz"
+    "training-parallel-nc-v13.tgz"
+    "rapid2016.tgz"
+    "dev.tgz"
+    "test-full.tgz"
+)
+CORPORA=(
+    "training/europarl-v7.de-en"
+    "commoncrawl.de-en"
+    "training-parallel-nc-v13/news-commentary-v13.de-en"
+    "rapid2016.de-en"
+)
+
+if [ ! -d "$SCRIPTS" ]; then
+    echo "Please set SCRIPTS variable correctly to point to Moses scripts."
+    exit 1
+fi
+
+OUTDIR=wmt18_en_de
+
+src=en
+tgt=de
+lang=en-de
+prep=$OUTDIR
+tmp=$prep/tmp
+orig=orig
+
+mkdir -p $orig $tmp $prep
+
+cd $orig
+
+for ((i=0;i<${#URLS[@]};++i)); do
+    file=${FILES[i]}
+    if [ -f $file ]; then
+        echo "$file already exists, skipping download"
+    else
+        url=${URLS[i]}
+        wget "$url"
+        if [ -f $file ]; then
+            echo "$url successfully downloaded."
+        else
+            echo "$url not successfully downloaded."
+            exit 1
+        fi
+        if [ ${file: -4} == ".tgz" ]; then
+            tar zxvf $file
+        elif [ ${file: -4} == ".tar" ]; then
+            tar xvf $file
+        fi
+    fi
+done
+cd ..
+
+echo "pre-processing train data..."
+for l in $src $tgt; do
+    rm $tmp/train.tags.$lang.tok.$l
+    for f in "${CORPORA[@]}"; do
+        cat $orig/$f.$l | \
+            perl $NORM_PUNC $l | \
+            perl $REM_NON_PRINT_CHAR | \
+            perl $TOKENIZER -threads 8 -a -l $l >> $tmp/train.tags.$lang.tok.$l
+    done
+done
+
+echo "pre-processing test data..."
+for l in $src $tgt; do
+    if [ "$l" == "$src" ]; then
+        t="src"
+    else
+        t="ref"
+    fi
+    grep '<seg id' $orig/test-full/newstest2014-deen-$t.$l.sgm | \
+        sed -e 's/<seg id="[0-9]*">\s*//g' | \
+        sed -e 's/\s*<\/seg>\s*//g' | \
+        sed -e "s/\’/\'/g" | \
+    perl $TOKENIZER -threads 8 -a -l $l > $tmp/test.$l
+    echo ""
+done
+
+echo "splitting train and valid..."
+for l in $src $tgt; do
+    awk '{if (NR%100 == 0)  print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/valid.$l
+    awk '{if (NR%100 != 0)  print $0; }' $tmp/train.tags.$lang.tok.$l > $tmp/train.$l
+done
+
+TRAIN=$tmp/train.de-en
+BPE_CODE=$prep/code
+rm -f $TRAIN
+for l in $src $tgt; do
+    cat $tmp/train.$l >> $TRAIN
+done
+
+echo "learn_bpe.py on ${TRAIN}..."
+python $BPEROOT/learn_bpe.py -s $BPE_TOKENS < $TRAIN > $BPE_CODE
+
+for L in $src $tgt; do
+    for f in train.$L valid.$L test.$L; do
+        echo "apply_bpe.py to ${f}..."
+        python $BPEROOT/apply_bpe.py -c $BPE_CODE < $tmp/$f > $tmp/bpe.$f
+    done
+done
+
+perl $CLEAN -ratio 1.5 $tmp/bpe.train $src $tgt $prep/train 1 250
+perl $CLEAN -ratio 1.5 $tmp/bpe.valid $src $tgt $prep/valid 1 250
+
+for L in $src $tgt; do
+    cp $tmp/bpe.test.$L $prep/test.$L
+done

fairseq/examples/backtranslation/sacrebleu.sh

@@ -0,0 +1,37 @@
+#!/bin/bash
+
+if [ $# -ne 5 ]; then
+    echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]"
+    exit
+fi
+
+
+DATASET=$1
+LANGPAIR=$2
+DATABIN=$3
+BPECODE=$4
+MODEL=$5
+
+SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1)
+TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2)
+
+
+BPEROOT=examples/backtranslation/subword-nmt/subword_nmt
+if [ ! -e $BPEROOT ]; then
+    BPEROOT=subword-nmt/subword_nmt
+    if [ ! -e $BPEROOT ]; then
+        echo 'Cloning Subword NMT repository (for BPE pre-processing)...'
+        git clone https://github.com/rsennrich/subword-nmt.git
+    fi
+fi
+
+
+sacrebleu -t $DATASET -l $LANGPAIR --echo src \
+| sacremoses tokenize -a -l $SRCLANG -q \
+| python $BPEROOT/apply_bpe.py -c $BPECODE \
+| fairseq-interactive $DATABIN --path $MODEL \
+    -s $SRCLANG -t $TGTLANG \
+    --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \
+| grep ^H- | cut -f 3- \
+| sacremoses detokenize -l $TGTLANG -q \
+| sacrebleu -t $DATASET -l $LANGPAIR

fairseq/examples/backtranslation/tokenized_bleu.sh

@@ -0,0 +1,46 @@
+#!/bin/bash
+
+if [ $# -ne 5 ]; then
+    echo "usage: $0 [dataset=wmt14/full] [langpair=en-de] [databin] [bpecode] [model]"
+    exit
+fi
+
+
+DATASET=$1
+LANGPAIR=$2
+DATABIN=$3
+BPECODE=$4
+MODEL=$5
+
+SRCLANG=$(echo $LANGPAIR | cut -d '-' -f 1)
+TGTLANG=$(echo $LANGPAIR | cut -d '-' -f 2)
+
+
+BPEROOT=examples/backtranslation/subword-nmt/subword_nmt
+if [ ! -e $BPEROOT ]; then
+    BPEROOT=subword-nmt/subword_nmt
+    if [ ! -e $BPEROOT ]; then
+        echo 'Cloning Subword NMT repository (for BPE pre-processing)...'
+        git clone https://github.com/rsennrich/subword-nmt.git
+    fi
+fi
+
+
+TMP_REF=$(mktemp)
+
+sacrebleu -t $DATASET -l $LANGPAIR --echo ref -q \
+| sacremoses normalize -l $TGTLANG -q \
+| sacremoses tokenize -a -l $TGTLANG -q \
+> $TMP_REF
+
+sacrebleu -t $DATASET -l $LANGPAIR --echo src -q \
+| sacremoses normalize -l $SRCLANG -q \
+| sacremoses tokenize -a -l $SRCLANG -q \
+| python $BPEROOT/apply_bpe.py -c $BPECODE \
+| fairseq-interactive $DATABIN --path $MODEL \
+    -s $SRCLANG -t $TGTLANG \
+    --beam 5 --remove-bpe --buffer-size 1024 --max-tokens 8000 \
+| grep ^H- | cut -f 3- \
+| fairseq-score --ref $TMP_REF
+
+rm -f $TMP_REF

fairseq/examples/bart/README.glue.md

@@ -0,0 +1,99 @@
+# Fine-tuning BART on GLUE tasks
+
+### 1) Download the data from GLUE website (https://gluebenchmark.com/tasks) using following commands:
+```bash
+wget https://gist.githubusercontent.com/W4ngatang/60c2bdb54d156a41194446737ce03e2e/raw/17b8dd0d724281ed7c3b2aeeda662b92809aadd5/download_glue_data.py
+python download_glue_data.py --data_dir glue_data --tasks all
+```
+
+### 2) Preprocess GLUE task data (same as RoBERTa):
+```bash
+./examples/roberta/preprocess_GLUE_tasks.sh glue_data <glue_task_name>
+```
+`glue_task_name` is one of the following:
+`{ALL, QQP, MNLI, QNLI, MRPC, RTE, STS-B, SST-2, CoLA}`
+Use `ALL` for preprocessing all the glue tasks.
+
+### 3) Fine-tuning on GLUE task:
+Example fine-tuning cmd for `RTE` task
+```bash
+TOTAL_NUM_UPDATES=2036  # 10 epochs through RTE for bsz 16
+WARMUP_UPDATES=61      # 6 percent of the number of updates
+LR=1e-05                # Peak LR for polynomial LR scheduler.
+NUM_CLASSES=2
+MAX_SENTENCES=16        # Batch size.
+BART_PATH=/path/to/bart/model.pt
+
+CUDA_VISIBLE_DEVICES=0,1 fairseq-train RTE-bin/ \
+    --restore-file $BART_PATH \
+    --batch-size $MAX_SENTENCES \
+    --max-tokens 4400 \
+    --task sentence_prediction \
+    --add-prev-output-tokens \
+    --layernorm-embedding \
+    --share-all-embeddings \
+    --share-decoder-input-output-embed \
+    --reset-optimizer --reset-dataloader --reset-meters \
+    --required-batch-size-multiple 1 \
+    --init-token 0 \
+    --arch bart_large \
+    --criterion sentence_prediction \
+    --num-classes $NUM_CLASSES \
+    --dropout 0.1 --attention-dropout 0.1 \
+    --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-08 \
+    --clip-norm 0.0 \
+    --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \
+    --fp16 --fp16-init-scale 4 --threshold-loss-scale 1 --fp16-scale-window 128 \
+    --max-epoch 10 \
+    --find-unused-parameters \
+    --best-checkpoint-metric accuracy --maximize-best-checkpoint-metric;
+```
+
+For each of the GLUE task, you will need to use following cmd-line arguments:
+
+Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B
+---|---|---|---|---|---|---|---|---
+`--num-classes` | 3 | 2 | 2 | 2 | 2 | 2 | 2 | 1
+`--lr` | 5e-6 | 1e-5 | 1e-5 | 1e-5 | 5e-6 | 2e-5 | 2e-5 | 2e-5
+`bsz` | 128 | 32 | 32 | 32 | 128 | 64 | 64 | 32
+`--total-num-update` | 30968 | 33112 | 113272 | 1018 | 5233 | 1148 | 1334 | 1799
+`--warmup-updates` | 1858 | 1986 | 6796 | 61 | 314 | 68 | 80 | 107
+
+For `STS-B` additionally add `--regression-target --best-checkpoint-metric loss` and remove `--maximize-best-checkpoint-metric`.
+
+**Note:**
+
+a) `--total-num-updates` is used by `--polynomial_decay` scheduler and is calculated for `--max-epoch=10` and `--batch-size=32/64/128` depending on the task.
+
+b) Above cmd-args and hyperparams are tested on Nvidia `V100` GPU with `32gb` of memory for each task. Depending on the GPU memory resources available to you, you can use increase `--update-freq` and reduce `--batch-size`.
+
+### Inference on GLUE task
+After training the model as mentioned in previous step, you can perform inference with checkpoints in `checkpoints/` directory using following python code snippet:
+
+```python
+from fairseq.models.bart import BARTModel
+
+bart = BARTModel.from_pretrained(
+    'checkpoints/',
+    checkpoint_file='checkpoint_best.pt',
+    data_name_or_path='RTE-bin'
+)
+
+label_fn = lambda label: bart.task.label_dictionary.string(
+    [label + bart.task.label_dictionary.nspecial]
+)   
+ncorrect, nsamples = 0, 0
+bart.cuda()
+bart.eval()
+with open('glue_data/RTE/dev.tsv') as fin:
+    fin.readline()
+    for index, line in enumerate(fin):
+        tokens = line.strip().split('\t')
+        sent1, sent2, target = tokens[1], tokens[2], tokens[3]
+        tokens = bart.encode(sent1, sent2)
+        prediction = bart.predict('sentence_classification_head', tokens).argmax().item()
+        prediction_label = label_fn(prediction)
+        ncorrect += int(prediction_label == target)
+        nsamples += 1
+print('| Accuracy: ', float(ncorrect)/float(nsamples))
+```

fairseq/examples/bart/README.md

@@ -0,0 +1,228 @@
+# BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension
+
+[https://arxiv.org/abs/1910.13461](https://arxiv.org/abs/1910.13461)
+
+## Introduction
+
+BART is sequence-to-sequence model trained with denoising as pretraining objective. We show that this pretraining objective is more generic and show that we can match [RoBERTa](../roberta) results on SQuAD and GLUE and gain state-of-the-art results on summarization (XSum, CNN dataset), long form generative question answering (ELI5) and dialog response genration (ConvAI2). See the associated paper for more details.
+
+## Pre-trained models
+
+Model | Description | # params | Download
+---|---|---|---
+`bart.base` | BART model with 6 encoder and decoder layers | 140M | [bart.base.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.base.tar.gz)
+`bart.large` | BART model with 12 encoder and decoder layers | 400M | [bart.large.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz)
+`bart.large.mnli` | `bart.large` finetuned on `MNLI` | 400M | [bart.large.mnli.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.mnli.tar.gz)
+`bart.large.cnn` | `bart.large` finetuned on `CNN-DM` | 400M | [bart.large.cnn.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.cnn.tar.gz)
+`bart.large.xsum` | `bart.large` finetuned on `Xsum` | 400M | [bart.large.xsum.tar.gz](https://dl.fbaipublicfiles.com/fairseq/models/bart.large.xsum.tar.gz)
+
+## Results
+
+**[GLUE (Wang et al., 2019)](https://gluebenchmark.com/)**
+_(dev set, single model, single-task finetuning)_
+
+Model | MNLI | QNLI | QQP | RTE | SST-2 | MRPC | CoLA | STS-B
+---|---|---|---|---|---|---|---|---
+`roberta.large` | 90.2 | 94.7 | 92.2 | 86.6 | 96.4 | 90.9 | 68.0 | 92.4
+`bart.large` | 89.9 | 94.9 | 92.5 | 87.0 | 96.6 | 90.4 | 62.8 | 91.2
+
+**[SQuAD (Rajpurkar et al., 2018)](https://rajpurkar.github.io/SQuAD-explorer/)**
+_(dev set, no additional data used)_
+
+Model | SQuAD 1.1 EM/F1 | SQuAD 2.0 EM/F1
+---|---|---
+`roberta.large` | 88.9/94.6 | 86.5/89.4
+`bart.large` | 88.8/94.6 | 86.1/89.2
+
+**[CNN/Daily Mail](http://nlpprogress.com/english/summarization.html)**
+_(test set, no additional data used)_
+
+Model | R1 | R2 | RL
+---|---|---|---
+`BERTSUMEXTABS` | 42.13 | 19.60 | 39.18
+`bart.large` | 44.16 | 21.28 | 40.90
+
+## Example usage
+
+##### Load BART from torch.hub (PyTorch >= 1.1):
+```python
+import torch
+bart = torch.hub.load('pytorch/fairseq', 'bart.large')
+bart.eval()  # disable dropout (or leave in train mode to finetune)
+```
+
+##### Load BART (for PyTorch 1.0 or custom models):
+```python
+# Download bart.large model
+wget https://dl.fbaipublicfiles.com/fairseq/models/bart.large.tar.gz
+tar -xzvf bart.large.tar.gz
+
+# Load the model in fairseq
+from fairseq.models.bart import BARTModel
+bart = BARTModel.from_pretrained('/path/to/bart.large', checkpoint_file='model.pt')
+bart.eval()  # disable dropout (or leave in train mode to finetune)
+```
+
+##### Apply Byte-Pair Encoding (BPE) to input text:
+```python
+tokens = bart.encode('Hello world!')
+assert tokens.tolist() == [0, 31414, 232, 328, 2]
+bart.decode(tokens)  # 'Hello world!'
+```
+
+##### Extract features from BART:
+```python
+# Extract the last layer's features
+last_layer_features = bart.extract_features(tokens)
+assert last_layer_features.size() == torch.Size([1, 5, 1024])
+
+# Extract all layer's features from decoder (layer 0 is the embedding layer)
+all_layers = bart.extract_features(tokens, return_all_hiddens=True)
+assert len(all_layers) == 13
+assert torch.all(all_layers[-1] == last_layer_features)
+```
+
+##### Use BART for sentence-pair classification tasks:
+```python
+# Download BART already finetuned for MNLI
+bart = torch.hub.load('pytorch/fairseq', 'bart.large.mnli')
+bart.eval()  # disable dropout for evaluation
+
+# Encode a pair of sentences and make a prediction
+tokens = bart.encode('BART is a seq2seq model.', 'BART is not sequence to sequence.')
+bart.predict('mnli', tokens).argmax()  # 0: contradiction
+
+# Encode another pair of sentences
+tokens = bart.encode('BART is denoising autoencoder.', 'BART is version of autoencoder.')
+bart.predict('mnli', tokens).argmax()  # 2: entailment
+```
+
+##### Register a new (randomly initialized) classification head:
+```python
+bart.register_classification_head('new_task', num_classes=3)
+logprobs = bart.predict('new_task', tokens)
+```
+
+##### Batched prediction:
+```python
+import torch
+from fairseq.data.data_utils import collate_tokens
+
+bart = torch.hub.load('pytorch/fairseq', 'bart.large.mnli')
+bart.eval()
+
+batch_of_pairs = [
+    ['BART is a seq2seq model.', 'BART is not sequence to sequence.'],
+    ['BART is denoising autoencoder.', 'BART is version of autoencoder.'],
+]
+
+batch = collate_tokens(
+    [bart.encode(pair[0], pair[1]) for pair in batch_of_pairs], pad_idx=1
+)
+
+logprobs = bart.predict('mnli', batch)
+print(logprobs.argmax(dim=1))
+# tensor([0, 2])
+```
+
+##### Using the GPU:
+```python
+bart.cuda()
+bart.predict('new_task', tokens)
+```
+
+#### Filling masks:
+
+BART can be used to fill multiple `<mask>` tokens in the input.
+```python
+bart = torch.hub.load('pytorch/fairseq', 'bart.base')
+bart.eval()
+bart.fill_mask(['The cat <mask> on the <mask>.'], topk=3, beam=10)
+# [[('The cat was on the ground.', tensor(-0.6183)), ('The cat was on the floor.', tensor(-0.6798)), ('The cat sleeps on the couch.', tensor(-0.6830))]]
+```
+
+Note that by default we enforce the output length to match the input length.
+This can be disabled by setting ``match_source_len=False``:
+```
+bart.fill_mask(['The cat <mask> on the <mask>.'], topk=3, beam=10, match_source_len=False)
+# [[('The cat was on the ground.', tensor(-0.6185)), ('The cat was asleep on the couch.', tensor(-0.6276)), ('The cat was on the floor.', tensor(-0.6800))]]
+```
+
+Example code to fill masks for a batch of sentences using GPU
+```
+bart.cuda()
+bart.fill_mask(['The cat <mask> on the <mask>.', 'The dog <mask> on the <mask>.'], topk=3, beam=10)
+# [[('The cat was on the ground.', tensor(-0.6183)), ('The cat was on the floor.', tensor(-0.6798)), ('The cat sleeps on the couch.', tensor(-0.6830))], [('The dog was on the ground.', tensor(-0.6190)), ('The dog lay on the ground.', tensor(-0.6711)),
+('The dog was asleep on the couch', tensor(-0.6796))]]
+```
+
+#### Evaluating the `bart.large.mnli` model:
+
+Example python code snippet to evaluate accuracy on the MNLI `dev_matched` set.
+```python
+label_map = {0: 'contradiction', 1: 'neutral', 2: 'entailment'}
+ncorrect, nsamples = 0, 0
+bart.cuda()
+bart.eval()
+with open('glue_data/MNLI/dev_matched.tsv') as fin:
+    fin.readline()
+    for index, line in enumerate(fin):
+        tokens = line.strip().split('\t')
+        sent1, sent2, target = tokens[8], tokens[9], tokens[-1]
+        tokens = bart.encode(sent1, sent2)
+        prediction = bart.predict('mnli', tokens).argmax().item()
+        prediction_label = label_map[prediction]
+        ncorrect += int(prediction_label == target)
+        nsamples += 1
+        print('| Accuracy: ', float(ncorrect)/float(nsamples))
+# Expected output: 0.9010
+```
+
+#### Evaluating the `bart.large.cnn` model:
+- Follow instructions [here](https://github.com/abisee/cnn-dailymail) to download and process into data-files such that `test.source` and `test.target` has one line for each non-tokenized sample.
+- For simpler preprocessing, you can also `wget https://cdn-datasets.huggingface.co/summarization/cnn_dm_v2.tgz`, although there is no guarantee of identical scores
+- `huggingface/transformers` has a simpler interface that supports [single-gpu](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/run_eval.py) and [multi-gpu](https://github.com/huggingface/transformers/blob/master/examples/legacy/seq2seq/run_distributed_eval.py) beam search.
+    In `huggingface/transformers`, the BART models' paths are `facebook/bart-large-cnn` and `facebook/bart-large-xsum`.
+
+In `fairseq`, summaries can be generated using:
+
+```bash
+cp data-bin/cnn_dm/dict.source.txt  checkpoints/
+python examples/bart/summarize.py \
+  --model-dir pytorch/fairseq \
+  --model-file bart.large.cnn \
+  --src cnn_dm/test.source \
+  --out cnn_dm/test.hypo
+```
+
+For calculating rouge, install `files2rouge` from [here](https://github.com/pltrdy/files2rouge).
+
+```bash
+export CLASSPATH=/path/to/stanford-corenlp-full-2016-10-31/stanford-corenlp-3.7.0.jar
+
+# Tokenize hypothesis and target files.
+cat test.hypo | java edu.stanford.nlp.process.PTBTokenizer -ioFileList -preserveLines > test.hypo.tokenized
+cat test.target | java edu.stanford.nlp.process.PTBTokenizer -ioFileList -preserveLines > test.hypo.target
+files2rouge test.hypo.tokenized test.hypo.target
+# Expected output: (ROUGE-2 Average_F: 0.21238)
+```
+
+
+## Finetuning
+
+- [Finetuning on GLUE](README.glue.md)
+- [Finetuning on CNN-DM](README.summarization.md)
+
+## Citation
+
+```bibtex
+@article{lewis2019bart,
+    title = {BART: Denoising Sequence-to-Sequence Pre-training for Natural
+Language Generation, Translation, and Comprehension},
+    author = {Mike Lewis and Yinhan Liu and Naman Goyal and Marjan Ghazvininejad and
+              Abdelrahman Mohamed and Omer Levy and Veselin Stoyanov
+              and Luke Zettlemoyer },
+    journal={arXiv preprint arXiv:1910.13461},
+    year = {2019},
+}
+```

fairseq/examples/bart/README.summarization.md

@@ -0,0 +1,102 @@
+# Fine-tuning BART on CNN-Dailymail summarization task
+
+### 1) Download the CNN and Daily Mail data and preprocess it into data files with non-tokenized cased samples.
+
+Follow the instructions [here](https://github.com/abisee/cnn-dailymail) to download the original CNN and Daily Mail datasets. To preprocess the data, refer to the pointers in [this issue](https://github.com/pytorch/fairseq/issues/1391) or check out the code [here](https://github.com/artmatsak/cnn-dailymail).
+
+Follow the instructions [here](https://github.com/EdinburghNLP/XSum) to download the original Extreme Summarization datasets, or check out the code [here](https://github.com/EdinburghNLP/XSum/tree/master/XSum-Dataset), Please keep the raw dataset and make sure no tokenization nor BPE on the dataset.
+
+### 2) BPE preprocess:
+
+```bash
+wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/encoder.json'
+wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/vocab.bpe'
+wget -N 'https://dl.fbaipublicfiles.com/fairseq/gpt2_bpe/dict.txt'
+
+TASK=cnn_dm
+for SPLIT in train val
+do
+  for LANG in source target
+  do
+    python -m examples.roberta.multiprocessing_bpe_encoder \
+    --encoder-json encoder.json \
+    --vocab-bpe vocab.bpe \
+    --inputs "$TASK/$SPLIT.$LANG" \
+    --outputs "$TASK/$SPLIT.bpe.$LANG" \
+    --workers 60 \
+    --keep-empty;
+  done
+done
+```
+
+### 3) Binarize dataset:
+```bash
+fairseq-preprocess \
+  --source-lang "source" \
+  --target-lang "target" \
+  --trainpref "${TASK}/train.bpe" \
+  --validpref "${TASK}/val.bpe" \
+  --destdir "${TASK}-bin/" \
+  --workers 60 \
+  --srcdict dict.txt \
+  --tgtdict dict.txt;
+```
+
+### 4) Fine-tuning on CNN-DM summarization task:
+Example fine-tuning CNN-DM
+```bash
+TOTAL_NUM_UPDATES=20000  
+WARMUP_UPDATES=500      
+LR=3e-05
+MAX_TOKENS=2048
+UPDATE_FREQ=4
+BART_PATH=/path/to/bart/model.pt
+
+CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7 fairseq-train cnn_dm-bin \
+    --restore-file $BART_PATH \
+    --max-tokens $MAX_TOKENS \
+    --task translation \
+    --source-lang source --target-lang target \
+    --truncate-source \
+    --layernorm-embedding \
+    --share-all-embeddings \
+    --share-decoder-input-output-embed \
+    --reset-optimizer --reset-dataloader --reset-meters \
+    --required-batch-size-multiple 1 \
+    --arch bart_large \
+    --criterion label_smoothed_cross_entropy \
+    --label-smoothing 0.1 \
+    --dropout 0.1 --attention-dropout 0.1 \
+    --weight-decay 0.01 --optimizer adam --adam-betas "(0.9, 0.999)" --adam-eps 1e-08 \
+    --clip-norm 0.1 \
+    --lr-scheduler polynomial_decay --lr $LR --total-num-update $TOTAL_NUM_UPDATES --warmup-updates $WARMUP_UPDATES \
+    --fp16 --update-freq $UPDATE_FREQ \
+    --skip-invalid-size-inputs-valid-test \
+    --find-unused-parameters;
+```
+Above is expected to run on `1` node with `8 32gb-V100`.
+Expected training time is about `5 hours`. Training time can be reduced with distributed training on `4` nodes and `--update-freq 1`.
+
+Use TOTAL_NUM_UPDATES=15000 UPDATE_FREQ=2 for Xsum task
+
+### Inference for CNN-DM test data using above trained checkpoint.
+After training the model as mentioned in previous step, you can perform inference with checkpoints in `checkpoints/` directory using `eval_cnn.py`, for example
+
+```bash
+cp data-bin/cnn_dm/dict.source.txt  checkpoints/
+python examples/bart/summarize.py \
+  --model-dir checkpoints \
+  --model-file checkpoint_best.pt \
+  --src cnn_dm/test.source \
+  --out cnn_dm/test.hypo
+```
+For XSUM, which uses beam=6, lenpen=1.0, max_len_b=60, min_len=10:
+```bash
+cp data-bin/cnn_dm/dict.source.txt  checkpoints/
+python examples/bart/summarize.py \
+  --model-dir checkpoints \
+  --model-file checkpoint_best.pt \
+  --src cnn_dm/test.source \
+  --out cnn_dm/test.hypo \
+  --xsum-kwargs
+```

fairseq/examples/bart/summarize.py

@@ -0,0 +1,100 @@
+# 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.bart import BARTModel
+import argparse
+
+XSUM_KWARGS = dict(beam=6, lenpen=1.0, max_len_b=60, min_len=10, no_repeat_ngram_size=3)
+CNN_KWARGS = dict(beam=4, lenpen=2.0, max_len_b=140, min_len=55, no_repeat_ngram_size=3)
+
+
+@torch.no_grad()
+def generate(bart, infile, outfile="bart_hypo.txt", bsz=32, n_obs=None, **eval_kwargs):
+    count = 1
+
+    # if n_obs is not None: bsz = min(bsz, n_obs)
+
+    with open(infile) as source, open(outfile, "w") as fout:
+        sline = source.readline().strip()
+        slines = [sline]
+        for sline in source:
+            if n_obs is not None and count > n_obs:
+                break
+            if count % bsz == 0:
+                hypotheses_batch = bart.sample(slines, **eval_kwargs)
+                for hypothesis in hypotheses_batch:
+                    fout.write(hypothesis + "\n")
+                    fout.flush()
+                slines = []
+
+            slines.append(sline.strip())
+            count += 1
+
+        if slines != []:
+            hypotheses_batch = bart.sample(slines, **eval_kwargs)
+            for hypothesis in hypotheses_batch:
+                fout.write(hypothesis + "\n")
+                fout.flush()
+
+
+def main():
+    """
+    Usage::
+
+         python examples/bart/summarize.py \
+            --model-dir $HOME/bart.large.cnn \
+            --model-file model.pt \
+            --src $HOME/data-bin/cnn_dm/test.source
+    """
+    parser = argparse.ArgumentParser()
+    parser.add_argument(
+        "--model-dir",
+        required=True,
+        type=str,
+        default="bart.large.cnn/",
+        help="path containing model file and src_dict.txt",
+    )
+    parser.add_argument(
+        "--model-file",
+        default="checkpoint_best.pt",
+        help="where in model_dir are weights saved",
+    )
+    parser.add_argument(
+        "--src", default="test.source", help="text to summarize", type=str
+    )
+    parser.add_argument(
+        "--out", default="test.hypo", help="where to save summaries", type=str
+    )
+    parser.add_argument("--bsz", default=32, help="where to save summaries", type=int)
+    parser.add_argument(
+        "--n", default=None, help="how many examples to summarize", type=int
+    )
+    parser.add_argument(
+        "--xsum-kwargs",
+        action="store_true",
+        default=False,
+        help="if true use XSUM_KWARGS else CNN_KWARGS",
+    )
+    args = parser.parse_args()
+    eval_kwargs = XSUM_KWARGS if args.xsum_kwargs else CNN_KWARGS
+    if args.model_dir == "pytorch/fairseq":
+        bart = torch.hub.load("pytorch/fairseq", args.model_file)
+    else:
+        bart = BARTModel.from_pretrained(
+            args.model_dir,
+            checkpoint_file=args.model_file,
+            data_name_or_path=args.model_dir,
+        )
+    bart = bart.eval()
+    if torch.cuda.is_available():
+        bart = bart.cuda().half()
+    generate(
+        bart, args.src, bsz=args.bsz, n_obs=args.n, outfile=args.out, **eval_kwargs
+    )
+
+
+if __name__ == "__main__":
+    main()