Add DeeBERT (entropy-based early exiting for *BERT) (#5477)

* Add deebert code

* Add readme of deebert

* Add test for deebert

Update test for Deebert

* Update DeeBert (README, class names, function refactoring); remove requirements.txt

* Format update

* Update test

* Update readme and model init methods

examples/deebert/README.md

@@ -0,0 +1,54 @@
+# DeeBERT: Early Exiting for *BERT
+
+This is the code base for the paper [DeeBERT: Dynamic Early Exiting for Accelerating BERT Inference](https://www.aclweb.org/anthology/2020.acl-main.204/), modified from its [original code base](https://github.com/castorini/deebert).
+
+The original code base also has information for downloading sample models that we have trained in advance.
+
+## Usage
+
+There are three scripts in the folder which can be run directly.
+
+In each script, there are several things to modify before running:
+
+* `PATH_TO_DATA`: path to the GLUE dataset.
+* `--output_dir`: path for saving fine-tuned models. Default: `./saved_models`.
+* `--plot_data_dir`: path for saving evaluation results. Default: `./results`. Results are printed to stdout and also saved to `npy` files in this directory to facilitate plotting figures and further analyses.
+* `MODEL_TYPE`: bert or roberta
+* `MODEL_SIZE`: base or large
+* `DATASET`: SST-2, MRPC, RTE, QNLI, QQP, or MNLI
+
+#### train_deebert.sh
+
+This is for fine-tuning DeeBERT models.
+
+#### eval_deebert.sh
+
+This is for evaluating each exit layer for fine-tuned DeeBERT models.
+
+#### entropy_eval.sh
+
+This is for evaluating fine-tuned DeeBERT models, given a number of different early exit entropy thresholds.
+
+
+
+## Citation
+
+Please cite our paper if you find the resource useful:
+```
+@inproceedings{xin-etal-2020-deebert,
+    title = "{D}ee{BERT}: Dynamic Early Exiting for Accelerating {BERT} Inference",
+    author = "Xin, Ji  and
+      Tang, Raphael  and
+      Lee, Jaejun  and
+      Yu, Yaoliang  and
+      Lin, Jimmy",
+    booktitle = "Proceedings of the 58th Annual Meeting of the Association for Computational Linguistics",
+    month = jul,
+    year = "2020",
+    address = "Online",
+    publisher = "Association for Computational Linguistics",
+    url = "https://www.aclweb.org/anthology/2020.acl-main.204",
+    pages = "2246--2251",
+}
+```
+

examples/deebert/entropy_eval.sh

@@ -0,0 +1,33 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+
+PATH_TO_DATA=/h/xinji/projects/GLUE
+
+MODEL_TYPE=bert  # bert or roberta
+MODEL_SIZE=base  # base or large
+DATASET=MRPC  # SST-2, MRPC, RTE, QNLI, QQP, or MNLI
+
+MODEL_NAME=${MODEL_TYPE}-${MODEL_SIZE}
+if [ $MODEL_TYPE = 'bert' ]
+then
+  MODEL_NAME=${MODEL_NAME}-uncased
+fi
+
+ENTROPIES="0 0.1 0.2 0.3 0.4 0.5 0.6 0.7"
+
+for ENTROPY in $ENTROPIES; do
+  python -u run_glue_deebert.py \
+    --model_type $MODEL_TYPE \
+    --model_name_or_path ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
+    --task_name $DATASET \
+    --do_eval \
+    --do_lower_case \
+    --data_dir $PATH_TO_DATA/$DATASET \
+    --output_dir ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
+    --plot_data_dir ./results/ \
+    --max_seq_length 128 \
+    --early_exit_entropy $ENTROPY \
+    --eval_highway \
+    --overwrite_cache \
+    --per_gpu_eval_batch_size=1
+done

examples/deebert/eval_deebert.sh

@@ -0,0 +1,30 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+
+PATH_TO_DATA=/h/xinji/projects/GLUE
+
+MODEL_TYPE=bert  # bert or roberta
+MODEL_SIZE=base  # base or large
+DATASET=MRPC  # SST-2, MRPC, RTE, QNLI, QQP, or MNLI
+
+MODEL_NAME=${MODEL_TYPE}-${MODEL_SIZE}
+if [ $MODEL_TYPE = 'bert' ]
+then
+  MODEL_NAME=${MODEL_NAME}-uncased
+fi
+
+
+python -u run_glue_deebert.py  \
+  --model_type $MODEL_TYPE \
+  --model_name_or_path ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
+  --task_name $DATASET \
+  --do_eval \
+  --do_lower_case \
+  --data_dir $PATH_TO_DATA/$DATASET \
+  --output_dir ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
+  --plot_data_dir ./results/ \
+  --max_seq_length 128 \
+  --eval_each_highway \
+  --eval_highway \
+  --overwrite_cache \
+  --per_gpu_eval_batch_size=1

examples/deebert/run_glue_deebert.py

@@ -0,0 +1,720 @@
+from __future__ import absolute_import, division, print_function
+
+import argparse
+import glob
+import logging
+import os
+import random
+import time
+
+import numpy as np
+import torch
+from torch.utils.data import DataLoader, RandomSampler, SequentialSampler, TensorDataset
+from torch.utils.data.distributed import DistributedSampler
+from tqdm import tqdm, trange
+
+from src.modeling_highway_bert import DeeBertForSequenceClassification
+from src.modeling_highway_roberta import DeeRobertaForSequenceClassification
+from transformers import (
+    WEIGHTS_NAME,
+    AdamW,
+    BertConfig,
+    BertTokenizer,
+    RobertaConfig,
+    RobertaTokenizer,
+    get_linear_schedule_with_warmup,
+)
+from transformers import glue_compute_metrics as compute_metrics
+from transformers import glue_convert_examples_to_features as convert_examples_to_features
+from transformers import glue_output_modes as output_modes
+from transformers import glue_processors as processors
+
+
+try:
+    from torch.utils.tensorboard import SummaryWriter
+except ImportError:
+    from tensorboardX import SummaryWriter
+
+
+logger = logging.getLogger(__name__)
+
+
+MODEL_CLASSES = {
+    "bert": (BertConfig, DeeBertForSequenceClassification, BertTokenizer),
+    "roberta": (RobertaConfig, DeeRobertaForSequenceClassification, RobertaTokenizer),
+}
+
+
+def set_seed(args):
+    random.seed(args.seed)
+    np.random.seed(args.seed)
+    torch.manual_seed(args.seed)
+    if args.n_gpu > 0:
+        torch.cuda.manual_seed_all(args.seed)
+
+
+def get_wanted_result(result):
+    if "spearmanr" in result:
+        print_result = result["spearmanr"]
+    elif "f1" in result:
+        print_result = result["f1"]
+    elif "mcc" in result:
+        print_result = result["mcc"]
+    elif "acc" in result:
+        print_result = result["acc"]
+    else:
+        raise ValueError("Primary metric unclear in the results")
+    return print_result
+
+
+def train(args, train_dataset, model, tokenizer, train_highway=False):
+    """ Train the model """
+    if args.local_rank in [-1, 0]:
+        tb_writer = SummaryWriter()
+
+    args.train_batch_size = args.per_gpu_train_batch_size * max(1, args.n_gpu)
+    train_sampler = RandomSampler(train_dataset) if args.local_rank == -1 else DistributedSampler(train_dataset)
+    train_dataloader = DataLoader(train_dataset, sampler=train_sampler, batch_size=args.train_batch_size)
+
+    if args.max_steps > 0:
+        t_total = args.max_steps
+        args.num_train_epochs = args.max_steps // (len(train_dataloader) // args.gradient_accumulation_steps) + 1
+    else:
+        t_total = len(train_dataloader) // args.gradient_accumulation_steps * args.num_train_epochs
+
+    # Prepare optimizer and schedule (linear warmup and decay)
+    no_decay = ["bias", "LayerNorm.weight"]
+    if train_highway:
+        optimizer_grouped_parameters = [
+            {
+                "params": [
+                    p
+                    for n, p in model.named_parameters()
+                    if ("highway" in n) and (not any(nd in n for nd in no_decay))
+                ],
+                "weight_decay": args.weight_decay,
+            },
+            {
+                "params": [
+                    p for n, p in model.named_parameters() if ("highway" in n) and (any(nd in n for nd in no_decay))
+                ],
+                "weight_decay": 0.0,
+            },
+        ]
+    else:
+        optimizer_grouped_parameters = [
+            {
+                "params": [
+                    p
+                    for n, p in model.named_parameters()
+                    if ("highway" not in n) and (not any(nd in n for nd in no_decay))
+                ],
+                "weight_decay": args.weight_decay,
+            },
+            {
+                "params": [
+                    p
+                    for n, p in model.named_parameters()
+                    if ("highway" not in n) and (any(nd in n for nd in no_decay))
+                ],
+                "weight_decay": 0.0,
+            },
+        ]
+    optimizer = AdamW(optimizer_grouped_parameters, lr=args.learning_rate, eps=args.adam_epsilon)
+    scheduler = get_linear_schedule_with_warmup(
+        optimizer, num_warmup_steps=args.warmup_steps, num_training_steps=t_total
+    )
+    if args.fp16:
+        try:
+            from apex import amp
+        except ImportError:
+            raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.")
+        model, optimizer = amp.initialize(model, optimizer, opt_level=args.fp16_opt_level)
+
+    # multi-gpu training (should be after apex fp16 initialization)
+    if args.n_gpu > 1:
+        model = torch.nn.DataParallel(model)
+
+    # Distributed training (should be after apex fp16 initialization)
+    if args.local_rank != -1:
+        model = torch.nn.parallel.DistributedDataParallel(
+            model, device_ids=[args.local_rank], output_device=args.local_rank, find_unused_parameters=True
+        )
+
+    # Train!
+    logger.info("***** Running training *****")
+    logger.info("  Num examples = %d", len(train_dataset))
+    logger.info("  Num Epochs = %d", args.num_train_epochs)
+    logger.info("  Instantaneous batch size per GPU = %d", args.per_gpu_train_batch_size)
+    logger.info(
+        "  Total train batch size (w. parallel, distributed & accumulation) = %d",
+        args.train_batch_size
+        * args.gradient_accumulation_steps
+        * (torch.distributed.get_world_size() if args.local_rank != -1 else 1),
+    )
+    logger.info("  Gradient Accumulation steps = %d", args.gradient_accumulation_steps)
+    logger.info("  Total optimization steps = %d", t_total)
+
+    global_step = 0
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0])
+    set_seed(args)  # Added here for reproductibility (even between python 2 and 3)
+    for _ in train_iterator:
+        epoch_iterator = tqdm(train_dataloader, desc="Iteration", disable=args.local_rank not in [-1, 0])
+        for step, batch in enumerate(epoch_iterator):
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
+            if args.model_type != "distilbert":
+                inputs["token_type_ids"] = (
+                    batch[2] if args.model_type in ["bert", "xlnet"] else None
+                )  # XLM, DistilBERT and RoBERTa don't use segment_ids
+            inputs["train_highway"] = train_highway
+            outputs = model(**inputs)
+            loss = outputs[0]  # model outputs are always tuple in transformers (see doc)
+
+            if args.n_gpu > 1:
+                loss = loss.mean()  # mean() to average on multi-gpu parallel training
+            if args.gradient_accumulation_steps > 1:
+                loss = loss / args.gradient_accumulation_steps
+
+            if args.fp16:
+                with amp.scale_loss(loss, optimizer) as scaled_loss:
+                    scaled_loss.backward()
+            else:
+                loss.backward()
+
+            tr_loss += loss.item()
+            if (step + 1) % args.gradient_accumulation_steps == 0:
+                if args.fp16:
+                    torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm)
+                else:
+                    torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm)
+
+                optimizer.step()
+                scheduler.step()  # Update learning rate schedule
+                model.zero_grad()
+                global_step += 1
+
+                if args.local_rank in [-1, 0] and args.logging_steps > 0 and global_step % args.logging_steps == 0:
+                    # Log metrics
+                    if (
+                        args.local_rank == -1 and args.evaluate_during_training
+                    ):  # Only evaluate when single GPU otherwise metrics may not average well
+                        results = evaluate(args, model, tokenizer)
+                        for key, value in results.items():
+                            tb_writer.add_scalar("eval_{}".format(key), value, global_step)
+                    tb_writer.add_scalar("lr", scheduler.get_lr()[0], global_step)
+                    tb_writer.add_scalar("loss", (tr_loss - logging_loss) / args.logging_steps, global_step)
+                    logging_loss = tr_loss
+
+                if args.local_rank in [-1, 0] and args.save_steps > 0 and global_step % args.save_steps == 0:
+                    # Save model checkpoint
+                    output_dir = os.path.join(args.output_dir, "checkpoint-{}".format(global_step))
+                    if not os.path.exists(output_dir):
+                        os.makedirs(output_dir)
+                    model_to_save = (
+                        model.module if hasattr(model, "module") else model
+                    )  # Take care of distributed/parallel training
+                    model_to_save.save_pretrained(output_dir)
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+            if args.max_steps > 0 and global_step > args.max_steps:
+                epoch_iterator.close()
+                break
+        if args.max_steps > 0 and global_step > args.max_steps:
+            train_iterator.close()
+            break
+
+    if args.local_rank in [-1, 0]:
+        tb_writer.close()
+
+    return global_step, tr_loss / global_step
+
+
+def evaluate(args, model, tokenizer, prefix="", output_layer=-1, eval_highway=False):
+    # Loop to handle MNLI double evaluation (matched, mis-matched)
+    eval_task_names = ("mnli", "mnli-mm") if args.task_name == "mnli" else (args.task_name,)
+    eval_outputs_dirs = (args.output_dir, args.output_dir + "-MM") if args.task_name == "mnli" else (args.output_dir,)
+
+    results = {}
+    for eval_task, eval_output_dir in zip(eval_task_names, eval_outputs_dirs):
+        eval_dataset = load_and_cache_examples(args, eval_task, tokenizer, evaluate=True)
+
+        if not os.path.exists(eval_output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(eval_output_dir)
+
+        args.eval_batch_size = args.per_gpu_eval_batch_size * max(1, args.n_gpu)
+        # Note that DistributedSampler samples randomly
+        eval_sampler = SequentialSampler(eval_dataset) if args.local_rank == -1 else DistributedSampler(eval_dataset)
+        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+        # multi-gpu eval
+        if args.n_gpu > 1:
+            model = torch.nn.DataParallel(model)
+
+        # Eval!
+        logger.info("***** Running evaluation {} *****".format(prefix))
+        logger.info("  Num examples = %d", len(eval_dataset))
+        logger.info("  Batch size = %d", args.eval_batch_size)
+        eval_loss = 0.0
+        nb_eval_steps = 0
+        preds = None
+        out_label_ids = None
+        exit_layer_counter = {(i + 1): 0 for i in range(model.num_layers)}
+        st = time.time()
+        for batch in tqdm(eval_dataloader, desc="Evaluating"):
+            model.eval()
+            batch = tuple(t.to(args.device) for t in batch)
+
+            with torch.no_grad():
+                inputs = {"input_ids": batch[0], "attention_mask": batch[1], "labels": batch[3]}
+                if args.model_type != "distilbert":
+                    inputs["token_type_ids"] = (
+                        batch[2] if args.model_type in ["bert", "xlnet"] else None
+                    )  # XLM, DistilBERT and RoBERTa don't use segment_ids
+                if output_layer >= 0:
+                    inputs["output_layer"] = output_layer
+                outputs = model(**inputs)
+                if eval_highway:
+                    exit_layer_counter[outputs[-1]] += 1
+                tmp_eval_loss, logits = outputs[:2]
+
+                eval_loss += tmp_eval_loss.mean().item()
+            nb_eval_steps += 1
+            if preds is None:
+                preds = logits.detach().cpu().numpy()
+                out_label_ids = inputs["labels"].detach().cpu().numpy()
+            else:
+                preds = np.append(preds, logits.detach().cpu().numpy(), axis=0)
+                out_label_ids = np.append(out_label_ids, inputs["labels"].detach().cpu().numpy(), axis=0)
+        eval_time = time.time() - st
+        logger.info("Eval time: {}".format(eval_time))
+
+        eval_loss = eval_loss / nb_eval_steps
+        if args.output_mode == "classification":
+            preds = np.argmax(preds, axis=1)
+        elif args.output_mode == "regression":
+            preds = np.squeeze(preds)
+        result = compute_metrics(eval_task, preds, out_label_ids)
+        results.update(result)
+
+        if eval_highway:
+            logger.info("Exit layer counter: {}".format(exit_layer_counter))
+            actual_cost = sum([l * c for l, c in exit_layer_counter.items()])
+            full_cost = len(eval_dataloader) * model.num_layers
+            logger.info("Expected saving: {}".format(actual_cost / full_cost))
+            if args.early_exit_entropy >= 0:
+                save_fname = (
+                    args.plot_data_dir
+                    + "/"
+                    + args.model_name_or_path[2:]
+                    + "/entropy_{}.npy".format(args.early_exit_entropy)
+                )
+                if not os.path.exists(os.path.dirname(save_fname)):
+                    os.makedirs(os.path.dirname(save_fname))
+                print_result = get_wanted_result(result)
+                np.save(save_fname, np.array([exit_layer_counter, eval_time, actual_cost / full_cost, print_result]))
+                logger.info("Entropy={}\tResult={:.2f}".format(args.early_exit_entropy, 100 * print_result))
+
+        output_eval_file = os.path.join(eval_output_dir, prefix, "eval_results.txt")
+        with open(output_eval_file, "w") as writer:
+            logger.info("***** Eval results {} *****".format(prefix))
+            for key in sorted(result.keys()):
+                logger.info("  %s = %s", key, str(result[key]))
+                writer.write("%s = %s\n" % (key, str(result[key])))
+
+    return results
+
+
+def load_and_cache_examples(args, task, tokenizer, evaluate=False):
+    if args.local_rank not in [-1, 0] and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    processor = processors[task]()
+    output_mode = output_modes[task]
+    # Load data features from cache or dataset file
+    cached_features_file = os.path.join(
+        args.data_dir,
+        "cached_{}_{}_{}_{}".format(
+            "dev" if evaluate else "train",
+            list(filter(None, args.model_name_or_path.split("/"))).pop(),
+            str(args.max_seq_length),
+            str(task),
+        ),
+    )
+    if os.path.exists(cached_features_file) and not args.overwrite_cache:
+        logger.info("Loading features from cached file %s", cached_features_file)
+        features = torch.load(cached_features_file)
+    else:
+        logger.info("Creating features from dataset file at %s", args.data_dir)
+        label_list = processor.get_labels()
+        if task in ["mnli", "mnli-mm"] and args.model_type in ["roberta"]:
+            # HACK(label indices are swapped in RoBERTa pretrained model)
+            label_list[1], label_list[2] = label_list[2], label_list[1]
+        examples = (
+            processor.get_dev_examples(args.data_dir) if evaluate else processor.get_train_examples(args.data_dir)
+        )
+        features = convert_examples_to_features(
+            examples, tokenizer, label_list=label_list, max_length=args.max_seq_length, output_mode=output_mode,
+        )
+        if args.local_rank in [-1, 0]:
+            logger.info("Saving features into cached file %s", cached_features_file)
+            torch.save(features, cached_features_file)
+
+    if args.local_rank == 0 and not evaluate:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training process the dataset, and the others will use the cache
+
+    # Convert to Tensors and build dataset
+    all_input_ids = torch.tensor([f.input_ids for f in features], dtype=torch.long)
+    all_attention_mask = torch.tensor([f.attention_mask for f in features], dtype=torch.long)
+
+    if features[0].token_type_ids is None:
+        # For RoBERTa (a potential bug!)
+        all_token_type_ids = torch.tensor([[0] * args.max_seq_length for f in features], dtype=torch.long)
+    else:
+        all_token_type_ids = torch.tensor([f.token_type_ids for f in features], dtype=torch.long)
+    if output_mode == "classification":
+        all_labels = torch.tensor([f.label for f in features], dtype=torch.long)
+    elif output_mode == "regression":
+        all_labels = torch.tensor([f.label for f in features], dtype=torch.float)
+
+    dataset = TensorDataset(all_input_ids, all_attention_mask, all_token_type_ids, all_labels)
+    return dataset
+
+
+def main():
+    parser = argparse.ArgumentParser()
+
+    # Required parameters
+    parser.add_argument(
+        "--data_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The input data dir. Should contain the .tsv files (or other data files) for the task.",
+    )
+    parser.add_argument(
+        "--model_type",
+        default=None,
+        type=str,
+        required=True,
+        help="Model type selected in the list: " + ", ".join(MODEL_CLASSES.keys()),
+    )
+    parser.add_argument(
+        "--model_name_or_path",
+        default=None,
+        type=str,
+        required=True,
+        help="Path to pre-trained model or shortcut name.",
+    )
+    parser.add_argument(
+        "--task_name",
+        default=None,
+        type=str,
+        required=True,
+        help="The name of the task to train selected in the list: " + ", ".join(processors.keys()),
+    )
+    parser.add_argument(
+        "--output_dir",
+        default=None,
+        type=str,
+        required=True,
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument(
+        "--plot_data_dir",
+        default="./plotting/",
+        type=str,
+        required=False,
+        help="The directory to store data for plotting figures.",
+    )
+
+    # Other parameters
+    parser.add_argument(
+        "--config_name", default="", type=str, help="Pretrained config name or path if not the same as model_name"
+    )
+    parser.add_argument(
+        "--tokenizer_name",
+        default="",
+        type=str,
+        help="Pretrained tokenizer name or path if not the same as model_name",
+    )
+    parser.add_argument(
+        "--cache_dir",
+        default="",
+        type=str,
+        help="Where do you want to store the pre-trained models downloaded from s3",
+    )
+    parser.add_argument(
+        "--max_seq_length",
+        default=128,
+        type=int,
+        help="The maximum total input sequence length after tokenization. Sequences longer "
+        "than this will be truncated, sequences shorter will be padded.",
+    )
+    parser.add_argument("--do_train", action="store_true", help="Whether to run training.")
+    parser.add_argument("--do_eval", action="store_true", help="Whether to run eval on the dev set.")
+    parser.add_argument(
+        "--evaluate_during_training", action="store_true", help="Rul evaluation during training at each logging step."
+    )
+    parser.add_argument(
+        "--do_lower_case", action="store_true", help="Set this flag if you are using an uncased model."
+    )
+    parser.add_argument("--eval_each_highway", action="store_true", help="Set this flag to evaluate each highway.")
+    parser.add_argument(
+        "--eval_after_first_stage",
+        action="store_true",
+        help="Set this flag to evaluate after training only bert (not highway).",
+    )
+    parser.add_argument("--eval_highway", action="store_true", help="Set this flag if it's evaluating highway models")
+
+    parser.add_argument("--per_gpu_train_batch_size", default=8, type=int, help="Batch size per GPU/CPU for training.")
+    parser.add_argument(
+        "--per_gpu_eval_batch_size", default=8, type=int, help="Batch size per GPU/CPU for evaluation."
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument("--learning_rate", default=5e-5, type=float, help="The initial learning rate for Adam.")
+    parser.add_argument("--weight_decay", default=0.0, type=float, help="Weight deay if we apply some.")
+    parser.add_argument("--adam_epsilon", default=1e-8, type=float, help="Epsilon for Adam optimizer.")
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument(
+        "--num_train_epochs", default=3.0, type=float, help="Total number of training epochs to perform."
+    )
+    parser.add_argument(
+        "--max_steps",
+        default=-1,
+        type=int,
+        help="If > 0: set total number of training steps to perform. Override num_train_epochs.",
+    )
+    parser.add_argument("--warmup_steps", default=0, type=int, help="Linear warmup over warmup_steps.")
+    parser.add_argument("--early_exit_entropy", default=-1, type=float, help="Entropy threshold for early exit.")
+
+    parser.add_argument("--logging_steps", type=int, default=50, help="Log every X updates steps.")
+    parser.add_argument("--save_steps", type=int, default=50, help="Save checkpoint every X updates steps.")
+    parser.add_argument(
+        "--eval_all_checkpoints",
+        action="store_true",
+        help="Evaluate all checkpoints starting with the same prefix as model_name ending and ending with step number",
+    )
+    parser.add_argument("--no_cuda", action="store_true", help="Avoid using CUDA when available")
+    parser.add_argument(
+        "--overwrite_output_dir", action="store_true", help="Overwrite the content of the output directory"
+    )
+    parser.add_argument(
+        "--overwrite_cache", action="store_true", help="Overwrite the cached training and evaluation sets"
+    )
+    parser.add_argument("--seed", type=int, default=42, help="random seed for initialization")
+
+    parser.add_argument(
+        "--fp16",
+        action="store_true",
+        help="Whether to use 16-bit (mixed) precision (through NVIDIA apex) instead of 32-bit",
+    )
+    parser.add_argument(
+        "--fp16_opt_level",
+        type=str,
+        default="O1",
+        help="For fp16: Apex AMP optimization level selected in ['O0', 'O1', 'O2', and 'O3']."
+        "See details at https://nvidia.github.io/apex/amp.html",
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument("--server_ip", type=str, default="", help="For distant debugging.")
+    parser.add_argument("--server_port", type=str, default="", help="For distant debugging.")
+    args = parser.parse_args()
+
+    if (
+        os.path.exists(args.output_dir)
+        and os.listdir(args.output_dir)
+        and args.do_train
+        and not args.overwrite_output_dir
+    ):
+        raise ValueError(
+            "Output directory ({}) already exists and is not empty. Use --overwrite_output_dir to overcome.".format(
+                args.output_dir
+            )
+        )
+
+    # Setup distant debugging if needed
+    if args.server_ip and args.server_port:
+        # Distant debugging - see https://code.visualstudio.com/docs/python/debugging#_attach-to-a-local-script
+        import ptvsd
+
+        print("Waiting for debugger attach")
+        ptvsd.enable_attach(address=(args.server_ip, args.server_port), redirect_output=True)
+        ptvsd.wait_for_attach()
+
+    # Setup CUDA, GPU & distributed training
+    if args.local_rank == -1 or args.no_cuda:
+        device = torch.device("cuda" if torch.cuda.is_available() and not args.no_cuda else "cpu")
+        args.n_gpu = torch.cuda.device_count()
+    else:  # Initializes the distributed backend which will take care of sychronizing nodes/GPUs
+        torch.cuda.set_device(args.local_rank)
+        device = torch.device("cuda", args.local_rank)
+        torch.distributed.init_process_group(backend="nccl")
+        args.n_gpu = 1
+    args.device = device
+
+    # Setup logging
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s -   %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO if args.local_rank in [-1, 0] else logging.WARN,
+    )
+    logger.warning(
+        "Process rank: %s, device: %s, n_gpu: %s, distributed training: %s, 16-bits training: %s",
+        args.local_rank,
+        device,
+        args.n_gpu,
+        bool(args.local_rank != -1),
+        args.fp16,
+    )
+
+    # Set seed
+    set_seed(args)
+
+    # Prepare GLUE task
+    args.task_name = args.task_name.lower()
+    if args.task_name not in processors:
+        raise ValueError("Task not found: %s" % (args.task_name))
+    processor = processors[args.task_name]()
+    args.output_mode = output_modes[args.task_name]
+    label_list = processor.get_labels()
+    num_labels = len(label_list)
+
+    # Load pretrained model and tokenizer
+    if args.local_rank not in [-1, 0]:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    args.model_type = args.model_type.lower()
+    config_class, model_class, tokenizer_class = MODEL_CLASSES[args.model_type]
+    config = config_class.from_pretrained(
+        args.config_name if args.config_name else args.model_name_or_path,
+        num_labels=num_labels,
+        finetuning_task=args.task_name,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+    )
+    tokenizer = tokenizer_class.from_pretrained(
+        args.tokenizer_name if args.tokenizer_name else args.model_name_or_path,
+        do_lower_case=args.do_lower_case,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+    )
+    model = model_class.from_pretrained(
+        args.model_name_or_path,
+        from_tf=bool(".ckpt" in args.model_name_or_path),
+        config=config,
+        cache_dir=args.cache_dir if args.cache_dir else None,
+    )
+
+    if args.model_type == "bert":
+        model.bert.encoder.set_early_exit_entropy(args.early_exit_entropy)
+        model.bert.init_highway_pooler()
+    elif args.model_type == "roberta":
+        model.roberta.encoder.set_early_exit_entropy(args.early_exit_entropy)
+        model.roberta.init_highway_pooler()
+    else:
+        raise NotImplementedError()
+
+    if args.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    model.to(args.device)
+
+    logger.info("Training/evaluation parameters %s", args)
+
+    # Training
+    if args.do_train:
+        train_dataset = load_and_cache_examples(args, args.task_name, tokenizer, evaluate=False)
+        global_step, tr_loss = train(args, train_dataset, model, tokenizer)
+        logger.info(" global_step = %s, average loss = %s", global_step, tr_loss)
+
+        if args.eval_after_first_stage:
+            result = evaluate(args, model, tokenizer, prefix="")
+            print_result = get_wanted_result(result)
+
+        train(args, train_dataset, model, tokenizer, train_highway=True)
+
+    # Saving best-practices: if you use defaults names for the model, you can reload it using from_pretrained()
+    if args.do_train and (args.local_rank == -1 or torch.distributed.get_rank() == 0):
+        # Create output directory if needed
+        if not os.path.exists(args.output_dir) and args.local_rank in [-1, 0]:
+            os.makedirs(args.output_dir)
+
+        logger.info("Saving model checkpoint to %s", args.output_dir)
+        # Save a trained model, configuration and tokenizer using `save_pretrained()`.
+        # They can then be reloaded using `from_pretrained()`
+        model_to_save = (
+            model.module if hasattr(model, "module") else model
+        )  # Take care of distributed/parallel training
+        model_to_save.save_pretrained(args.output_dir)
+        tokenizer.save_pretrained(args.output_dir)
+
+        # Good practice: save your training arguments together with the trained model
+        torch.save(args, os.path.join(args.output_dir, "training_args.bin"))
+
+        # Load a trained model and vocabulary that you have fine-tuned
+        model = model_class.from_pretrained(args.output_dir)
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir)
+        model.to(args.device)
+
+    # Evaluation
+    results = {}
+    if args.do_eval and args.local_rank in [-1, 0]:
+        tokenizer = tokenizer_class.from_pretrained(args.output_dir, do_lower_case=args.do_lower_case)
+        checkpoints = [args.output_dir]
+        if args.eval_all_checkpoints:
+            checkpoints = list(
+                os.path.dirname(c) for c in sorted(glob.glob(args.output_dir + "/**/" + WEIGHTS_NAME, recursive=True))
+            )
+            logging.getLogger("transformers.modeling_utils").setLevel(logging.WARN)  # Reduce logging
+        logger.info("Evaluate the following checkpoints: %s", checkpoints)
+        for checkpoint in checkpoints:
+            global_step = checkpoint.split("-")[-1] if len(checkpoints) > 1 else ""
+            prefix = checkpoint.split("/")[-1] if checkpoint.find("checkpoint") != -1 else ""
+
+            model = model_class.from_pretrained(checkpoint)
+            if args.model_type == "bert":
+                model.bert.encoder.set_early_exit_entropy(args.early_exit_entropy)
+            elif args.model_type == "roberta":
+                model.roberta.encoder.set_early_exit_entropy(args.early_exit_entropy)
+            else:
+                raise NotImplementedError()
+
+            model.to(args.device)
+            result = evaluate(args, model, tokenizer, prefix=prefix, eval_highway=args.eval_highway)
+            print_result = get_wanted_result(result)
+            logger.info("Result: {}".format(print_result))
+            if args.eval_each_highway:
+                last_layer_results = print_result
+                each_layer_results = []
+                for i in range(model.num_layers):
+                    logger.info("\n")
+                    _result = evaluate(
+                        args, model, tokenizer, prefix=prefix, output_layer=i, eval_highway=args.eval_highway
+                    )
+                    if i + 1 < model.num_layers:
+                        each_layer_results.append(get_wanted_result(_result))
+                each_layer_results.append(last_layer_results)
+                save_fname = args.plot_data_dir + "/" + args.model_name_or_path[2:] + "/each_layer.npy"
+                if not os.path.exists(os.path.dirname(save_fname)):
+                    os.makedirs(os.path.dirname(save_fname))
+                np.save(save_fname, np.array(each_layer_results))
+                info_str = "Score of each layer:"
+                for i in range(model.num_layers):
+                    info_str += " {:.2f}".format(100 * each_layer_results[i])
+                logger.info(info_str)
+            result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
+            results.update(result)
+
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/deebert/src/modeling_highway_bert.py

@@ -0,0 +1,396 @@
+import torch
+from torch import nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.modeling_bert import (
+    BERT_INPUTS_DOCSTRING,
+    BERT_START_DOCSTRING,
+    BertEmbeddings,
+    BertLayer,
+    BertPooler,
+    BertPreTrainedModel,
+)
+
+
+def entropy(x):
+    """ Calculate entropy of a pre-softmax logit Tensor
+    """
+    exp_x = torch.exp(x)
+    A = torch.sum(exp_x, dim=1)  # sum of exp(x_i)
+    B = torch.sum(x * exp_x, dim=1)  # sum of x_i * exp(x_i)
+    return torch.log(A) - B / A
+
+
+class DeeBertEncoder(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.output_attentions = config.output_attentions
+        self.output_hidden_states = config.output_hidden_states
+        self.layer = nn.ModuleList([BertLayer(config) for _ in range(config.num_hidden_layers)])
+        self.highway = nn.ModuleList([BertHighway(config) for _ in range(config.num_hidden_layers)])
+
+        self.early_exit_entropy = [-1 for _ in range(config.num_hidden_layers)]
+
+    def set_early_exit_entropy(self, x):
+        if (type(x) is float) or (type(x) is int):
+            for i in range(len(self.early_exit_entropy)):
+                self.early_exit_entropy[i] = x
+        else:
+            self.early_exit_entropy = x
+
+    def init_highway_pooler(self, pooler):
+        loaded_model = pooler.state_dict()
+        for highway in self.highway:
+            for name, param in highway.pooler.state_dict().items():
+                param.copy_(loaded_model[name])
+
+    def forward(
+        self,
+        hidden_states,
+        attention_mask=None,
+        head_mask=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+    ):
+        all_hidden_states = ()
+        all_attentions = ()
+        all_highway_exits = ()
+        for i, layer_module in enumerate(self.layer):
+            if self.output_hidden_states:
+                all_hidden_states = all_hidden_states + (hidden_states,)
+
+            layer_outputs = layer_module(
+                hidden_states, attention_mask, head_mask[i], encoder_hidden_states, encoder_attention_mask
+            )
+            hidden_states = layer_outputs[0]
+
+            if self.output_attentions:
+                all_attentions = all_attentions + (layer_outputs[1],)
+
+            current_outputs = (hidden_states,)
+            if self.output_hidden_states:
+                current_outputs = current_outputs + (all_hidden_states,)
+            if self.output_attentions:
+                current_outputs = current_outputs + (all_attentions,)
+
+            highway_exit = self.highway[i](current_outputs)
+            # logits, pooled_output
+
+            if not self.training:
+                highway_logits = highway_exit[0]
+                highway_entropy = entropy(highway_logits)
+                highway_exit = highway_exit + (highway_entropy,)  # logits, hidden_states(?), entropy
+                all_highway_exits = all_highway_exits + (highway_exit,)
+
+                if highway_entropy < self.early_exit_entropy[i]:
+                    new_output = (highway_logits,) + current_outputs[1:] + (all_highway_exits,)
+                    raise HighwayException(new_output, i + 1)
+            else:
+                all_highway_exits = all_highway_exits + (highway_exit,)
+
+        # Add last layer
+        if self.output_hidden_states:
+            all_hidden_states = all_hidden_states + (hidden_states,)
+
+        outputs = (hidden_states,)
+        if self.output_hidden_states:
+            outputs = outputs + (all_hidden_states,)
+        if self.output_attentions:
+            outputs = outputs + (all_attentions,)
+
+        outputs = outputs + (all_highway_exits,)
+        return outputs  # last-layer hidden state, (all hidden states), (all attentions), all highway exits
+
+
+@add_start_docstrings(
+    "The Bert Model transformer with early exiting (DeeBERT). ", BERT_START_DOCSTRING,
+)
+class DeeBertModel(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.config = config
+
+        self.embeddings = BertEmbeddings(config)
+        self.encoder = DeeBertEncoder(config)
+        self.pooler = BertPooler(config)
+
+        self.init_weights()
+
+    def init_highway_pooler(self):
+        self.encoder.init_highway_pooler(self.pooler)
+
+    def get_input_embeddings(self):
+        return self.embeddings.word_embeddings
+
+    def set_input_embeddings(self, value):
+        self.embeddings.word_embeddings = value
+
+    def _prune_heads(self, heads_to_prune):
+        """ Prunes heads of the model.
+            heads_to_prune: dict of {layer_num: list of heads to prune in this layer}
+            See base class PreTrainedModel
+        """
+        for layer, heads in heads_to_prune.items():
+            self.encoder.layer[layer].attention.prune_heads(heads)
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        encoder_hidden_states=None,
+        encoder_attention_mask=None,
+    ):
+        r"""
+    Return:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        last_hidden_state (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, sequence_length, hidden_size)`):
+            Sequence of hidden-states at the output of the last layer of the model.
+        pooler_output (:obj:`torch.FloatTensor`: of shape :obj:`(batch_size, hidden_size)`):
+            Last layer hidden-state of the first token of the sequence (classification token)
+            further processed by a Linear layer and a Tanh activation function. The Linear
+            layer weights are trained from the next sentence prediction (classification)
+            objective during pre-training.
+
+            This output is usually *not* a good summary
+            of the semantic content of the input, you're often better with averaging or pooling
+            the sequence of hidden-states for the whole input sequence.
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        highway_exits (:obj:`tuple(tuple(torch.Tensor))`:
+            Tuple of each early exit's results (total length: number of layers)
+            Each tuple is again, a tuple of length 2 - the first entry is logits and the second entry is hidden states.
+        """
+        if input_ids is not None and inputs_embeds is not None:
+            raise ValueError("You cannot specify both input_ids and inputs_embeds at the same time")
+        elif input_ids is not None:
+            input_shape = input_ids.size()
+        elif inputs_embeds is not None:
+            input_shape = inputs_embeds.size()[:-1]
+        else:
+            raise ValueError("You have to specify either input_ids or inputs_embeds")
+
+        device = input_ids.device if input_ids is not None else inputs_embeds.device
+
+        if attention_mask is None:
+            attention_mask = torch.ones(input_shape, device=device)
+        if encoder_attention_mask is None:
+            encoder_attention_mask = torch.ones(input_shape, device=device)
+        if token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        # We can provide a self-attention mask of dimensions [batch_size, from_seq_length, to_seq_length]
+        # ourselves in which case we just need to make it broadcastable to all heads.
+        extended_attention_mask: torch.Tensor = self.get_extended_attention_mask(attention_mask, input_shape, device)
+
+        # If a 2D ou 3D attention mask is provided for the cross-attention
+        # we need to make broadcastabe to [batch_size, num_heads, seq_length, seq_length]
+        if encoder_attention_mask.dim() == 3:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, :, :]
+        if encoder_attention_mask.dim() == 2:
+            encoder_extended_attention_mask = encoder_attention_mask[:, None, None, :]
+
+        encoder_extended_attention_mask = encoder_extended_attention_mask.to(
+            dtype=next(self.parameters()).dtype
+        )  # fp16 compatibility
+        encoder_extended_attention_mask = (1.0 - encoder_extended_attention_mask) * -10000.0
+
+        # Prepare head mask if needed
+        # 1.0 in head_mask indicate we keep the head
+        # attention_probs has shape bsz x n_heads x N x N
+        # input head_mask has shape [num_heads] or [num_hidden_layers x num_heads]
+        # and head_mask is converted to shape [num_hidden_layers x batch x num_heads x seq_length x seq_length]
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+        )
+        encoder_outputs = self.encoder(
+            embedding_output,
+            attention_mask=extended_attention_mask,
+            head_mask=head_mask,
+            encoder_hidden_states=encoder_hidden_states,
+            encoder_attention_mask=encoder_extended_attention_mask,
+        )
+        sequence_output = encoder_outputs[0]
+        pooled_output = self.pooler(sequence_output)
+
+        outputs = (sequence_output, pooled_output,) + encoder_outputs[
+            1:
+        ]  # add hidden_states and attentions if they are here
+        return outputs  # sequence_output, pooled_output, (hidden_states), (attentions), highway exits
+
+
+class HighwayException(Exception):
+    def __init__(self, message, exit_layer):
+        self.message = message
+        self.exit_layer = exit_layer  # start from 1!
+
+
+class BertHighway(nn.Module):
+    """A module to provide a shortcut
+    from (the output of one non-final BertLayer in BertEncoder) to (cross-entropy computation in BertForSequenceClassification)
+    """
+
+    def __init__(self, config):
+        super().__init__()
+        self.pooler = BertPooler(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, config.num_labels)
+
+    def forward(self, encoder_outputs):
+        # Pooler
+        pooler_input = encoder_outputs[0]
+        pooler_output = self.pooler(pooler_input)
+        # "return" pooler_output
+
+        # BertModel
+        bmodel_output = (pooler_input, pooler_output) + encoder_outputs[1:]
+        # "return" bodel_output
+
+        # Dropout and classification
+        pooled_output = bmodel_output[1]
+
+        pooled_output = self.dropout(pooled_output)
+        logits = self.classifier(pooled_output)
+
+        return logits, pooled_output
+
+
+@add_start_docstrings(
+    """Bert Model (with early exiting - DeeBERT) with a classifier on top,
+    also takes care of multi-layer training. """,
+    BERT_START_DOCSTRING,
+)
+class DeeBertForSequenceClassification(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.num_layers = config.num_hidden_layers
+
+        self.bert = DeeBertModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(BERT_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_layer=-1,
+        train_highway=False,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.BertConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        highway_exits (:obj:`tuple(tuple(torch.Tensor))`:
+            Tuple of each early exit's results (total length: number of layers)
+            Each tuple is again, a tuple of length 2 - the first entry is logits and the second entry is hidden states.
+        """
+
+        exit_layer = self.num_layers
+        try:
+            outputs = self.bert(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+            )
+            # sequence_output, pooled_output, (hidden_states), (attentions), highway exits
+
+            pooled_output = outputs[1]
+
+            pooled_output = self.dropout(pooled_output)
+            logits = self.classifier(pooled_output)
+            outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        except HighwayException as e:
+            outputs = e.message
+            exit_layer = e.exit_layer
+            logits = outputs[0]
+
+        if not self.training:
+            original_entropy = entropy(logits)
+            highway_entropy = []
+            highway_logits_all = []
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+            # work with highway exits
+            highway_losses = []
+            for highway_exit in outputs[-1]:
+                highway_logits = highway_exit[0]
+                if not self.training:
+                    highway_logits_all.append(highway_logits)
+                    highway_entropy.append(highway_exit[2])
+                if self.num_labels == 1:
+                    #  We are doing regression
+                    loss_fct = MSELoss()
+                    highway_loss = loss_fct(highway_logits.view(-1), labels.view(-1))
+                else:
+                    loss_fct = CrossEntropyLoss()
+                    highway_loss = loss_fct(highway_logits.view(-1, self.num_labels), labels.view(-1))
+                highway_losses.append(highway_loss)
+
+            if train_highway:
+                outputs = (sum(highway_losses[:-1]),) + outputs
+                # exclude the final highway, of course
+            else:
+                outputs = (loss,) + outputs
+        if not self.training:
+            outputs = outputs + ((original_entropy, highway_entropy), exit_layer)
+            if output_layer >= 0:
+                outputs = (
+                    (outputs[0],) + (highway_logits_all[output_layer],) + outputs[2:]
+                )  # use the highway of the last layer
+
+        return outputs  # (loss), logits, (hidden_states), (attentions), (highway_exits)

examples/deebert/src/modeling_highway_roberta.py

@@ -0,0 +1,151 @@
+from __future__ import absolute_import, division, print_function, unicode_literals
+
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from transformers.configuration_roberta import RobertaConfig
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.modeling_roberta import ROBERTA_INPUTS_DOCSTRING, ROBERTA_START_DOCSTRING, RobertaEmbeddings
+
+from .modeling_highway_bert import BertPreTrainedModel, DeeBertModel, HighwayException, entropy
+
+
+@add_start_docstrings(
+    "The RoBERTa Model transformer with early exiting (DeeRoBERTa). ", ROBERTA_START_DOCSTRING,
+)
+class DeeRobertaModel(DeeBertModel):
+
+    config_class = RobertaConfig
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.embeddings = RobertaEmbeddings(config)
+        self.init_weights()
+
+
+@add_start_docstrings(
+    """RoBERTa Model (with early exiting - DeeRoBERTa) with a classifier on top,
+    also takes care of multi-layer training. """,
+    ROBERTA_START_DOCSTRING,
+)
+class DeeRobertaForSequenceClassification(BertPreTrainedModel):
+
+    config_class = RobertaConfig
+    base_model_prefix = "roberta"
+
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+        self.num_layers = config.num_hidden_layers
+
+        self.roberta = DeeRobertaModel(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
+
+    @add_start_docstrings_to_callable(ROBERTA_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        labels=None,
+        output_layer=-1,
+        train_highway=False,
+    ):
+        r"""
+        labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`, defaults to :obj:`None`):
+            Labels for computing the sequence classification/regression loss.
+            Indices should be in :obj:`[0, ..., config.num_labels - 1]`.
+            If :obj:`config.num_labels == 1` a regression loss is computed (Mean-Square loss),
+            If :obj:`config.num_labels > 1` a classification loss is computed (Cross-Entropy).
+
+    Returns:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.RobertaConfig`) and inputs:
+        loss (:obj:`torch.FloatTensor` of shape :obj:`(1,)`, `optional`, returned when :obj:`label` is provided):
+            Classification (or regression if config.num_labels==1) loss.
+        logits (:obj:`torch.FloatTensor` of shape :obj:`(batch_size, config.num_labels)`):
+            Classification (or regression if config.num_labels==1) scores (before SoftMax).
+        hidden_states (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_hidden_states=True`` is passed or when ``config.output_hidden_states=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for the output of the embeddings + one for the output of each layer)
+            of shape :obj:`(batch_size, sequence_length, hidden_size)`.
+
+            Hidden-states of the model at the output of each layer plus the initial embedding outputs.
+        attentions (:obj:`tuple(torch.FloatTensor)`, `optional`, returned when ``output_attentions=True`` is passed or when ``config.output_attentions=True``):
+            Tuple of :obj:`torch.FloatTensor` (one for each layer) of shape
+            :obj:`(batch_size, num_heads, sequence_length, sequence_length)`.
+
+            Attentions weights after the attention softmax, used to compute the weighted average in the self-attention
+            heads.
+        highway_exits (:obj:`tuple(tuple(torch.Tensor))`:
+            Tuple of each early exit's results (total length: number of layers)
+            Each tuple is again, a tuple of length 2 - the first entry is logits and the second entry is hidden states.
+        """
+
+        exit_layer = self.num_layers
+        try:
+            outputs = self.roberta(
+                input_ids,
+                attention_mask=attention_mask,
+                token_type_ids=token_type_ids,
+                position_ids=position_ids,
+                head_mask=head_mask,
+                inputs_embeds=inputs_embeds,
+            )
+
+            pooled_output = outputs[1]
+
+            pooled_output = self.dropout(pooled_output)
+            logits = self.classifier(pooled_output)
+            outputs = (logits,) + outputs[2:]  # add hidden states and attention if they are here
+        except HighwayException as e:
+            outputs = e.message
+            exit_layer = e.exit_layer
+            logits = outputs[0]
+
+        if not self.training:
+            original_entropy = entropy(logits)
+            highway_entropy = []
+            highway_logits_all = []
+        if labels is not None:
+            if self.num_labels == 1:
+                #  We are doing regression
+                loss_fct = MSELoss()
+                loss = loss_fct(logits.view(-1), labels.view(-1))
+            else:
+                loss_fct = CrossEntropyLoss()
+                loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
+
+            # work with highway exits
+            highway_losses = []
+            for highway_exit in outputs[-1]:
+                highway_logits = highway_exit[0]
+                if not self.training:
+                    highway_logits_all.append(highway_logits)
+                    highway_entropy.append(highway_exit[2])
+                if self.num_labels == 1:
+                    #  We are doing regression
+                    loss_fct = MSELoss()
+                    highway_loss = loss_fct(highway_logits.view(-1), labels.view(-1))
+                else:
+                    loss_fct = CrossEntropyLoss()
+                    highway_loss = loss_fct(highway_logits.view(-1, self.num_labels), labels.view(-1))
+                highway_losses.append(highway_loss)
+
+            if train_highway:
+                outputs = (sum(highway_losses[:-1]),) + outputs
+                # exclude the final highway, of course
+            else:
+                outputs = (loss,) + outputs
+        if not self.training:
+            outputs = outputs + ((original_entropy, highway_entropy), exit_layer)
+            if output_layer >= 0:
+                outputs = (
+                    (outputs[0],) + (highway_logits_all[output_layer],) + outputs[2:]
+                )  # use the highway of the last layer
+
+        return outputs  # (loss), logits, (hidden_states), (attentions), entropy

examples/deebert/test_glue_deebert.py

@@ -0,0 +1,97 @@
+import argparse
+import logging
+import sys
+import unittest
+from unittest.mock import patch
+
+import run_glue_deebert
+
+
+logging.basicConfig(level=logging.DEBUG)
+
+logger = logging.getLogger()
+
+
+def get_setup_file():
+    parser = argparse.ArgumentParser()
+    parser.add_argument("-f")
+    args = parser.parse_args()
+    return args.f
+
+
+class DeeBertTests(unittest.TestCase):
+    def test_glue_deebert(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        train_args = """
+            run_glue_deebert.py
+            --model_type roberta
+            --model_name_or_path roberta-base
+            --task_name MRPC
+            --do_train
+            --do_eval
+            --do_lower_case
+            --data_dir ./tests/fixtures/tests_samples/MRPC/
+            --max_seq_length 128
+            --per_gpu_eval_batch_size=1
+            --per_gpu_train_batch_size=8
+            --learning_rate 2e-4
+            --num_train_epochs 3
+            --overwrite_output_dir
+            --seed 42
+            --output_dir ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
+            --plot_data_dir ./examples/deebert/results/
+            --save_steps 0
+            --overwrite_cache
+            --eval_after_first_stage
+            """.split()
+
+        eval_args = """
+            run_glue_deebert.py
+            --model_type roberta
+            --model_name_or_path ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
+            --task_name MRPC
+            --do_eval
+            --do_lower_case
+            --data_dir ./tests/fixtures/tests_samples/MRPC/
+            --output_dir ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
+            --plot_data_dir ./examples/deebert/results/
+            --max_seq_length 128
+            --eval_each_highway
+            --eval_highway
+            --overwrite_cache
+            --per_gpu_eval_batch_size=1
+            """.split()
+
+        entropy_eval_args = """
+            run_glue_deebert.py
+            --model_type roberta
+            --model_name_or_path ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
+            --task_name MRPC
+            --do_eval
+            --do_lower_case
+            --data_dir ./tests/fixtures/tests_samples/MRPC/
+            --output_dir ./examples/deebert/saved_models/roberta-base/MRPC/two_stage
+            --plot_data_dir ./examples/deebert/results/
+            --max_seq_length 128
+            --early_exit_entropy 0.1
+            --eval_highway
+            --overwrite_cache
+            --per_gpu_eval_batch_size=1
+            """.split()
+
+        with patch.object(sys, "argv", train_args):
+            result = run_glue_deebert.main()
+            for value in result.values():
+                self.assertGreaterEqual(value, 0.75)
+
+        with patch.object(sys, "argv", eval_args):
+            result = run_glue_deebert.main()
+            for value in result.values():
+                self.assertGreaterEqual(value, 0.75)
+
+        with patch.object(sys, "argv", entropy_eval_args):
+            result = run_glue_deebert.main()
+            for value in result.values():
+                self.assertGreaterEqual(value, 0.75)

examples/deebert/train_deebert.sh

@@ -0,0 +1,38 @@
+#!/bin/bash
+export CUDA_VISIBLE_DEVICES=0
+
+PATH_TO_DATA=/h/xinji/projects/GLUE
+
+MODEL_TYPE=bert  # bert or roberta
+MODEL_SIZE=base  # base or large
+DATASET=MRPC  # SST-2, MRPC, RTE, QNLI, QQP, or MNLI
+
+MODEL_NAME=${MODEL_TYPE}-${MODEL_SIZE}
+EPOCHS=10
+if [ $MODEL_TYPE = 'bert' ]
+then
+  EPOCHS=3
+  MODEL_NAME=${MODEL_NAME}-uncased
+fi
+
+
+python -u run_glue_deebert.py \
+  --model_type $MODEL_TYPE \
+  --model_name_or_path $MODEL_NAME \
+  --task_name $DATASET \
+  --do_train \
+  --do_eval \
+  --do_lower_case \
+  --data_dir $PATH_TO_DATA/$DATASET \
+  --max_seq_length 128 \
+  --per_gpu_eval_batch_size=1 \
+  --per_gpu_train_batch_size=8 \
+  --learning_rate 2e-5 \
+  --num_train_epochs $EPOCHS \
+  --overwrite_output_dir \
+  --seed 42 \
+  --output_dir ./saved_models/${MODEL_TYPE}-${MODEL_SIZE}/$DATASET/two_stage \
+  --plot_data_dir ./results/ \
+  --save_steps 0 \
+  --overwrite_cache \
+  --eval_after_first_stage