RXF OS Implementation (facebookresearch#2455)

Summary:
## What does this PR do?
Implements R3F and R4F coming from Facebook Research: https://arxiv.org/abs/2008.03156

This code was used to generate all the results from the paper excluding probing results.

Pull Request resolved: facebookresearch#2455

Reviewed By: myleott

Differential Revision: D23444863

Pulled By: AkshatSh

fbshipit-source-id: b724a6d6cc9cebfdb4bd219828afbb5679f2259b

README.md

@@ -51,11 +51,14 @@ We provide reference implementations of various sequence modeling papers:
   - 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:
 
+- 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)
 - September 2020: [Added Linformer code](examples/linformer/README.md)

examples/rxf/README.md

@@ -0,0 +1,52 @@
+[Better Fine-Tuning by Reducing Representational Collapse](https://arxiv.org/abs/2008.03156)
+=====================
+This repo contains the code to replicate all experiments from the _Better Fine-Tuning by Reducing Representational Collapse_ paper excluding the probing results.
+
+The R3F sentence prediction criterion is registered as `sentence_prediction_r3f` while the label smoothing version of it is implemented as `label_smoothed_cross_entropy_r3f`. The R4F version of the sentence prediction criterion can be achieved by applying spectral norm to the classification head via the `--spectral-norm-classification-head` parameter.
+
+## Hyper-parameters
+Our methods introduce 3 new hyper-parameters; `--eps` which sets the standard deviation or range of the distribution we're sampling from, `--r3f-lambda` which controls the combining of logistic loss and noisy KL loss and `--noise-type` which controls which parametric distribution we use ('normal', 'uniform').
+
+For example to run R3F on RTE from GLUE
+
+```
+TOTAL_NUM_UPDATES=3120
+WARMUP_UPDATES=187
+LR=1e-05
+NUM_CLASSES=2
+MAX_SENTENCES=8        # Batch size.
+ROBERTA_PATH=/path/to/roberta/model.pt
+
+CUDA_VISIBLE_DEVICES=0 fairseq-train RTE-bin \
+    --restore-file $ROBERTA_PATH \
+    --max-positions 512 \
+    --max-sentences $MAX_SENTENCES \
+    --max-tokens 4400 \
+    --task sentence_prediction \
+    --reset-optimizer --reset-dataloader --reset-meters \
+    --required-batch-size-multiple 1 \
+    --init-token 0 --separator-token 2 \
+    --arch roberta_large \
+    --criterion sentence_prediction_r3f \
+    --num-classes $NUM_CLASSES \
+    --dropout 0.1 --attention-dropout 0.1 \
+    --weight-decay 0.1 --optimizer adam --adam-betas "(0.9, 0.98)" --adam-eps 1e-06 \
+    --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 \
+    --noise-type uniform --r3f-lambda 0.7 \
+    --user-dir examples/rxf;
+```
+
+## Citation
+```bibtex
+@article{aghajanyan2020better,
+  title={Better Fine-Tuning by Reducing Representational Collapse},
+  author={Aghajanyan, Armen and Shrivastava, Akshat and Gupta, Anchit and Goyal, Naman and Zettlemoyer, Luke and Gupta, Sonal},
+  journal={arXiv preprint arXiv:2008.03156},
+  year={2020}
+}
+```

examples/rxf/__init__.py

@@ -0,0 +1,6 @@
+# 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 . import src  # noqa

examples/rxf/src/__init__.py

@@ -0,0 +1,6 @@
+# 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 . import label_smoothed_cross_entropy_r3f, sentence_prediction_r3f  # noqa

examples/rxf/src/label_smoothed_cross_entropy_r3f.py

@@ -0,0 +1,157 @@
+# 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.functional as F
+from fairseq import metrics, utils
+from fairseq.criterions import FairseqCriterion, register_criterion
+from fairseq.criterions.label_smoothed_cross_entropy import label_smoothed_nll_loss
+
+
+@register_criterion("label_smoothed_cross_entropy_r3f")
+class LabelSmoothedCrossEntropyR3FCriterion(FairseqCriterion):
+    def __init__(
+        self, task, sentence_avg, label_smoothing, eps, r3f_lambda, noise_type
+    ):
+        super().__init__(task)
+        self.sentence_avg = sentence_avg
+        self.label_smoothing = label_smoothing
+        self.eps = eps
+        self.r3f_lambda = r3f_lambda
+        self.noise_type = noise_type
+        if self.noise_type in {"normal"}:
+            self.noise_sampler = torch.distributions.normal.Normal(
+                loc=0.0, scale=self.eps
+            )
+        elif self.noise_type == "uniform":
+            self.noise_sampler = torch.distributions.uniform.Uniform(
+                low=-self.eps, high=self.eps
+            )
+        else:
+            raise Exception(f"unrecognized noise type {self.noise_type}")
+
+    @staticmethod
+    def add_args(parser):
+        """Add criterion-specific arguments to the parser."""
+        # fmt: off
+        parser.add_argument('--label-smoothing', default=0., type=float, metavar='D',
+                            help='epsilon for label smoothing, 0 means no label smoothing')
+        parser.add_argument('--eps', type=float, default=1e-5,
+                            help='noise eps')
+        parser.add_argument('--r3f-lambda', type=float, default=1.0,
+                            help='lambda for combining logistic loss and noisy KL loss')
+        parser.add_argument('--noise-type', type=str, default='normal',
+                            choices=['normal', 'uniform'],
+                            help='type of noises')
+        # fmt: on
+
+    def _get_symm_kl(self, noised_logits, input_logits):
+        return (
+            F.kl_div(
+                F.log_softmax(noised_logits, dim=-1, dtype=torch.float32),
+                F.softmax(input_logits, dim=-1, dtype=torch.float32),
+                None,
+                None,
+                "sum",
+            )
+            + F.kl_div(
+                F.log_softmax(input_logits, dim=-1, dtype=torch.float32),
+                F.softmax(noised_logits, dim=-1, dtype=torch.float32),
+                None,
+                None,
+                "sum",
+            )
+        ) / noised_logits.size(0)
+
+    def forward(self, model, sample, reduce=True):
+        """Compute the loss for the given sample.
+
+        Returns a tuple with three elements:
+        1) the loss
+        2) the sample size, which is used as the denominator for the gradient
+        3) logging outputs to display while training
+        """
+        token_embeddings = model.encoder.embed_tokens(sample["net_input"]["src_tokens"])
+        input_logits, extra = model(**sample["net_input"])
+        loss, nll_loss = self.compute_loss(
+            model, (input_logits, extra), sample, reduce=reduce
+        )
+        sample_size = (
+            sample["target"].size(0) if self.sentence_avg else sample["ntokens"]
+        )
+
+        if model.training:
+            noise = self.noise_sampler.sample(sample_shape=token_embeddings.shape).to(
+                token_embeddings
+            )
+            noised_embeddings = token_embeddings.clone() + noise
+
+            noised_logits, _ = model(
+                **sample["net_input"], token_embeddings=noised_embeddings
+            )
+            symm_kl = self._get_symm_kl(noised_logits, input_logits)
+
+        if model.training:
+            symm_kl = symm_kl * sample_size
+            loss = loss + self.r3f_lambda * symm_kl
+
+        logging_output = {
+            "loss": loss.data,
+            "nll_loss": nll_loss.data,
+            "ntokens": sample["ntokens"],
+            "nsentences": sample["target"].size(0),
+            "sample_size": sample_size,
+        }
+
+        if model.training:
+            logging_output.update(
+                symm_kl=utils.item(symm_kl.data) if reduce else symm_kl.data
+            )
+
+        return loss, sample_size, logging_output
+
+    def compute_loss(self, model, net_output, sample, reduce=True):
+        lprobs = model.get_normalized_probs(net_output, log_probs=True)
+        lprobs = lprobs.view(-1, lprobs.size(-1))
+        target = model.get_targets(sample, net_output).view(-1, 1)
+        loss, nll_loss = label_smoothed_nll_loss(
+            lprobs,
+            target,
+            self.label_smoothing,
+            ignore_index=self.padding_idx,
+            reduce=reduce,
+        )
+        return loss, nll_loss
+
+    @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)
+        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)
+        sample_size = sum(log.get("sample_size", 0) for log in logging_outputs)
+        symm_kl_sum = sum(log.get("symm_kl", 0) for log in logging_outputs)
+
+        metrics.log_scalar("symm_kl", symm_kl_sum / sample_size, sample_size, round=3)
+        metrics.log_scalar(
+            "loss", loss_sum / sample_size / math.log(2), sample_size, round=3
+        )
+        metrics.log_scalar(
+            "nll_loss", nll_loss_sum / ntokens / math.log(2), ntokens, round=3
+        )
+        metrics.log_derived(
+            "ppl", lambda meters: utils.get_perplexity(meters["nll_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