Add BERT Loses Patience (Patience-based Early Exit) (#5078)

* Add BERT Loses Patience (Patience-based Early Exit)

* update model archive

* update format

* sort import

* flake8

* Add results

* full results

* align the table

* refactor to inherit

* default per gpu eval = 1

* Formatting

* Formatting

* isort

* modify readme

* Add check

* Fix format

* Fix format

* Doc strings

* ALBERT & BERT for sequence classification don't inherit from the original anymore

* Remove incorrect comments

* Remove incorrect comments

* Remove incorrect comments

* Sync up with new code

* Sync up with new code

* Add a test

* Add a test

* Add a test

* Add a test

* Add a test

* Add a test

* Finishing up!

examples/bert-loses-patience/README.md

@@ -0,0 +1,89 @@
+# Patience-based Early Exit
+
+Patience-based Early Exit (PABEE) is a plug-and-play inference method for pretrained language models.
+We have already implemented it on BERT and ALBERT. Basically, you can make your LM faster and more robust with PABEE. It can even improve the performance of ALBERT on GLUE. The only sacrifice is that the batch size can only be 1.
+Learn more in the paper ["BERT Loses Patience: Fast and Robust Inference with Early Exit"](https://arxiv.org/abs/2006.04152) and the official [GitHub repo](https://github.com/JetRunner/PABEE).
+
+![PABEE](https://github.com/JetRunner/PABEE/raw/master/bert-loses-patience.png)
+
+## Training
+
+You can fine-tune a pretrained language model (you can choose from BERT and ALBERT) and train the internal classifiers by:
+```bash
+export GLUE_DIR=/path/to/glue_data
+export TASK_NAME=MRPC
+
+python ./run_glue_with_pabee.py \
+  --model_type albert \
+  --model_name_or_path bert-base-uncased/albert-base-v2 \
+  --task_name $TASK_NAME \
+  --do_train \
+  --do_eval \
+  --do_lower_case \
+  --data_dir "$GLUE_DIR/$TASK_NAME" \
+  --max_seq_length 128 \
+  --per_gpu_train_batch_size 32 \
+  --per_gpu_eval_batch_size 32 \
+  --learning_rate 2e-5 \
+  --save_steps 50 \
+  --logging_steps 50 \
+  --num_train_epochs 5 \
+  --output_dir /path/to/save/ \
+  --evaluate_during_training
+```
+
+## Inference
+
+You can inference with different patience settings by:
+```bash
+export GLUE_DIR=/path/to/glue_data
+export TASK_NAME=MRPC
+
+python ./run_glue_with_pabee.py \
+  --model_type albert \
+  --model_name_or_path /path/to/save/ \
+  --task_name $TASK_NAME \
+  --do_eval \
+  --do_lower_case \
+  --data_dir "$GLUE_DIR/$TASK_NAME" \
+  --max_seq_length 128 \
+  --per_gpu_eval_batch_size 1 \
+  --learning_rate 2e-5 \
+  --logging_steps 50 \
+  --num_train_epochs 15 \
+  --output_dir /path/to/save/ \
+  --eval_all_checkpoints \
+  --patience 3,4,5,6,7,8
+```
+where `patience` can be a list of patience settings, separated by a comma. It will help determine which patience works best.
+
+When evaluating on a regression task (STS-B), you may add `--regression_threshold 0.1` to define the regression threshold.
+
+## Results
+On the GLUE dev set:
+
+| Model        | \#Param | Speed  | CoLA  | MNLI  | MRPC  | QNLI  | QQP   | RTE   | SST\-2 | STS\-B |
+|--------------|---------|--------|-------|-------|-------|-------|-------|-------|--------|--------|
+| ALBERT\-base | 12M     |        | 58\.9 | 84\.6 | 89\.5 | 91\.7 | 89\.6 | 78\.6 | 92\.8  | 89\.5  |
+| \+PABEE      | 12M     | 1\.57x | 61\.2 | 85\.1 | 90\.0 | 91\.8 | 89\.6 | 80\.1 | 93\.0  | 90\.1  |
+
+| Model         | \#Param | Speed\-up | MNLI  | SST\-2 | STS\-B |
+|---------------|---------|-----------|-------|--------|--------|
+| BERT\-base    | 108M    |           | 84\.5 | 92\.1  | 88\.9  |
+| \+PABEE       | 108M    | 1\.62x    | 83\.6 | 92\.0  | 88\.7  |
+| ALBERT\-large | 18M     |           |       |        |        |
+| \+PABEE       | 18M     | 2\.42x    | 86\.8 | 95\.2  | 90\.6  |
+
+
+## Citation
+If you find this resource useful, please consider citing the following paper:
+```bibtex
+@misc{zhou2020bert,
+    title={BERT Loses Patience: Fast and Robust Inference with Early Exit},
+    author={Wangchunshu Zhou and Canwen Xu and Tao Ge and Julian McAuley and Ke Xu and Furu Wei},
+    year={2020},
+    eprint={2006.04152},
+    archivePrefix={arXiv},
+    primaryClass={cs.CL}
+}
+```

examples/bert-loses-patience/pabee/modeling_pabee_albert.py

@@ -0,0 +1,310 @@
+# coding=utf-8
+# Copyright 2020 Google AI, Google Brain, the HuggingFace Inc. team and Microsoft Corporation.
+#
+# 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.
+"""PyTorch ALBERT model with Patience-based Early Exit. """
+
+import logging
+
+import torch
+import torch.nn as nn
+from torch.nn import CrossEntropyLoss, MSELoss
+
+from transformers.file_utils import add_start_docstrings, add_start_docstrings_to_callable
+from transformers.modeling_albert import (
+    ALBERT_INPUTS_DOCSTRING,
+    ALBERT_START_DOCSTRING,
+    AlbertModel,
+    AlbertPreTrainedModel,
+    AlbertTransformer,
+)
+
+
+logger = logging.getLogger(__name__)
+
+
+class AlbertTransformerWithPabee(AlbertTransformer):
+    def adaptive_forward(self, hidden_states, current_layer, attention_mask=None, head_mask=None):
+        if current_layer == 0:
+            hidden_states = self.embedding_hidden_mapping_in(hidden_states)
+        else:
+            hidden_states = hidden_states[0]
+
+        layers_per_group = int(self.config.num_hidden_layers / self.config.num_hidden_groups)
+
+        # Index of the hidden group
+        group_idx = int(current_layer / (self.config.num_hidden_layers / self.config.num_hidden_groups))
+
+        layer_group_output = self.albert_layer_groups[group_idx](
+            hidden_states,
+            attention_mask,
+            head_mask[group_idx * layers_per_group : (group_idx + 1) * layers_per_group],
+        )
+        hidden_states = layer_group_output[0]
+
+        return (hidden_states,)
+
+
+@add_start_docstrings(
+    "The bare ALBERT Model transformer with PABEE outputting raw hidden-states without any specific head on top.",
+    ALBERT_START_DOCSTRING,
+)
+class AlbertModelWithPabee(AlbertModel):
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.encoder = AlbertTransformerWithPabee(config)
+
+        self.init_weights()
+        self.patience = 0
+        self.inference_instances_num = 0
+        self.inference_layers_num = 0
+
+        self.regression_threshold = 0
+
+    def set_regression_threshold(self, threshold):
+        self.regression_threshold = threshold
+
+    def set_patience(self, patience):
+        self.patience = patience
+
+    def reset_stats(self):
+        self.inference_instances_num = 0
+        self.inference_layers_num = 0
+
+    def log_stats(self):
+        avg_inf_layers = self.inference_layers_num / self.inference_instances_num
+        message = f"*** Patience = {self.patience} Avg. Inference Layers = {avg_inf_layers:.2f} Speed Up = {1 - avg_inf_layers / self.config.num_hidden_layers:.2f} ***"
+        print(message)
+
+    @add_start_docstrings_to_callable(ALBERT_INPUTS_DOCSTRING)
+    def forward(
+        self,
+        input_ids=None,
+        attention_mask=None,
+        token_type_ids=None,
+        position_ids=None,
+        head_mask=None,
+        inputs_embeds=None,
+        output_dropout=None,
+        output_layers=None,
+        regression=False,
+    ):
+        r"""
+    Return:
+        :obj:`tuple(torch.FloatTensor)` comprising various elements depending on the configuration (:class:`~transformers.AlbertConfig`) 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 ``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 ``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.
+        """
+
+        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 token_type_ids is None:
+            token_type_ids = torch.zeros(input_shape, dtype=torch.long, device=device)
+
+        extended_attention_mask = attention_mask.unsqueeze(1).unsqueeze(2)
+        extended_attention_mask = extended_attention_mask.to(dtype=self.dtype)  # fp16 compatibility
+        extended_attention_mask = (1.0 - extended_attention_mask) * -10000.0
+        head_mask = self.get_head_mask(head_mask, self.config.num_hidden_layers)
+
+        embedding_output = self.embeddings(
+            input_ids, position_ids=position_ids, token_type_ids=token_type_ids, inputs_embeds=inputs_embeds
+        )
+        encoder_outputs = embedding_output
+
+        if self.training:
+            res = []
+            for i in range(self.config.num_hidden_layers):
+                encoder_outputs = self.encoder.adaptive_forward(
+                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask,
+                )
+
+                pooled_output = self.pooler_activation(self.pooler(encoder_outputs[0][:, 0]))
+                logits = output_layers[i](output_dropout(pooled_output))
+                res.append(logits)
+        elif self.patience == 0:  # Use all layers for inference
+            encoder_outputs = self.encoder(encoder_outputs, extended_attention_mask, head_mask=head_mask)
+            pooled_output = self.pooler_activation(self.pooler(encoder_outputs[0][:, 0]))
+            res = [output_layers[self.config.num_hidden_layers - 1](pooled_output)]
+        else:
+            patient_counter = 0
+            patient_result = None
+            calculated_layer_num = 0
+            for i in range(self.config.num_hidden_layers):
+                calculated_layer_num += 1
+                encoder_outputs = self.encoder.adaptive_forward(
+                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask,
+                )
+
+                pooled_output = self.pooler_activation(self.pooler(encoder_outputs[0][:, 0]))
+                logits = output_layers[i](pooled_output)
+                if regression:
+                    labels = logits.detach()
+                    if patient_result is not None:
+                        patient_labels = patient_result.detach()
+                    if (patient_result is not None) and torch.abs(patient_result - labels) < self.regression_threshold:
+                        patient_counter += 1
+                    else:
+                        patient_counter = 0
+                else:
+                    labels = logits.detach().argmax(dim=1)
+                    if patient_result is not None:
+                        patient_labels = patient_result.detach().argmax(dim=1)
+                    if (patient_result is not None) and torch.all(labels.eq(patient_labels)):
+                        patient_counter += 1
+                    else:
+                        patient_counter = 0
+
+                patient_result = logits
+                if patient_counter == self.patience:
+                    break
+            res = [patient_result]
+            self.inference_layers_num += calculated_layer_num
+            self.inference_instances_num += 1
+
+        return res
+
+
+@add_start_docstrings(
+    """Albert Model transformer with PABEE and a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    ALBERT_START_DOCSTRING,
+)
+class AlbertForSequenceClassificationWithPabee(AlbertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.albert = AlbertModelWithPabee(config)
+        self.dropout = nn.Dropout(config.classifier_dropout_prob)
+        self.classifiers = nn.ModuleList(
+            [nn.Linear(config.hidden_size, self.config.num_labels) for _ in range(config.num_hidden_layers)]
+        )
+
+        self.init_weights()
+
+    @add_start_docstrings_to_callable(ALBERT_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,
+    ):
+        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 ``[0, ..., config.num_labels - 1]``.
+            If ``config.num_labels == 1`` a regression loss is computed (Mean-Square loss),
+            If ``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.AlbertConfig`) and inputs:
+        loss: (`optional`, returned when ``labels`` is provided) ``torch.FloatTensor`` of shape ``(1,)``:
+            Classification (or regression if config.num_labels==1) loss.
+        logits ``torch.FloatTensor`` of shape ``(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 ``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 ``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.
+
+        Examples::
+
+            from transformers import AlbertTokenizer
+            from pabee import AlbertForSequenceClassificationWithPabee
+            import torch
+
+            tokenizer = AlbertTokenizer.from_pretrained('albert-base-v2')
+            model = AlbertForSequenceClassificationWithPabee.from_pretrained('albert-base-v2')
+            input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute")).unsqueeze(0)  # Batch size 1
+            labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+            outputs = model(input_ids, labels=labels)
+            loss, logits = outputs[:2]
+
+        """
+
+        logits = self.albert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_dropout=self.dropout,
+            output_layers=self.classifiers,
+            regression=self.num_labels == 1,
+        )
+
+        outputs = (logits[-1],)
+
+        if labels is not None:
+            total_loss = None
+            total_weights = 0
+            for ix, logits_item in enumerate(logits):
+                if self.num_labels == 1:
+                    #  We are doing regression
+                    loss_fct = MSELoss()
+                    loss = loss_fct(logits_item.view(-1), labels.view(-1))
+                else:
+                    loss_fct = CrossEntropyLoss()
+                    loss = loss_fct(logits_item.view(-1, self.num_labels), labels.view(-1))
+                if total_loss is None:
+                    total_loss = loss
+                else:
+                    total_loss += loss * (ix + 1)
+                total_weights += ix + 1
+            outputs = (total_loss / total_weights,) + outputs
+
+        return outputs

examples/bert-loses-patience/pabee/modeling_pabee_bert.py

@@ -0,0 +1,342 @@
+# coding=utf-8
+# Copyright 2020 The Google AI Language Team Authors, The HuggingFace Inc. team and Microsoft Corporation.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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.
+"""PyTorch BERT model with Patience-based Early Exit. """
+
+
+import logging
+
+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,
+    BertEncoder,
+    BertModel,
+    BertPreTrainedModel,
+)
+
+
+logger = logging.getLogger(__name__)
+
+
+class BertEncoderWithPabee(BertEncoder):
+    def adaptive_forward(self, hidden_states, current_layer, attention_mask=None, head_mask=None):
+        layer_outputs = self.layer[current_layer](hidden_states, attention_mask, head_mask[current_layer])
+
+        hidden_states = layer_outputs[0]
+
+        return hidden_states
+
+
+@add_start_docstrings(
+    "The bare Bert Model transformer with PABEE outputting raw hidden-states without any specific head on top.",
+    BERT_START_DOCSTRING,
+)
+class BertModelWithPabee(BertModel):
+    """
+
+    The model can behave as an encoder (with only self-attention) as well
+    as a decoder, in which case a layer of cross-attention is added between
+    the self-attention layers, following the architecture described in `Attention is all you need`_ by Ashish Vaswani,
+    Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser and Illia Polosukhin.
+
+    To behave as an decoder the model needs to be initialized with the
+    :obj:`is_decoder` argument of the configuration set to :obj:`True`; an
+    :obj:`encoder_hidden_states` is expected as an input to the forward pass.
+
+    .. _`Attention is all you need`:
+        https://arxiv.org/abs/1706.03762
+
+    """
+
+    def __init__(self, config):
+        super().__init__(config)
+
+        self.encoder = BertEncoderWithPabee(config)
+
+        self.init_weights()
+        self.patience = 0
+        self.inference_instances_num = 0
+        self.inference_layers_num = 0
+
+        self.regression_threshold = 0
+
+    def set_regression_threshold(self, threshold):
+        self.regression_threshold = threshold
+
+    def set_patience(self, patience):
+        self.patience = patience
+
+    def reset_stats(self):
+        self.inference_instances_num = 0
+        self.inference_layers_num = 0
+
+    def log_stats(self):
+        avg_inf_layers = self.inference_layers_num / self.inference_instances_num
+        message = f"*** Patience = {self.patience} Avg. Inference Layers = {avg_inf_layers:.2f} Speed Up = {1 - avg_inf_layers / self.config.num_hidden_layers:.2f} ***"
+        print(message)
+
+    @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,
+        output_dropout=None,
+        output_layers=None,
+        regression=False,
+    ):
+        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 ``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 ``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.
+        """
+
+        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 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 self.config.is_decoder and encoder_hidden_states is not None:
+            encoder_batch_size, encoder_sequence_length, _ = encoder_hidden_states.size()
+            encoder_hidden_shape = (encoder_batch_size, encoder_sequence_length)
+            if encoder_attention_mask is None:
+                encoder_attention_mask = torch.ones(encoder_hidden_shape, device=device)
+            encoder_extended_attention_mask = self.invert_attention_mask(encoder_attention_mask)
+        else:
+            encoder_extended_attention_mask = None
+
+        # 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 = embedding_output
+
+        if self.training:
+            res = []
+            for i in range(self.config.num_hidden_layers):
+                encoder_outputs = self.encoder.adaptive_forward(
+                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask
+                )
+
+                pooled_output = self.pooler(encoder_outputs)
+                logits = output_layers[i](output_dropout(pooled_output))
+                res.append(logits)
+        elif self.patience == 0:  # Use all layers for inference
+            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,
+            )
+            pooled_output = self.pooler(encoder_outputs[0])
+            res = [output_layers[self.config.num_hidden_layers - 1](pooled_output)]
+        else:
+            patient_counter = 0
+            patient_result = None
+            calculated_layer_num = 0
+            for i in range(self.config.num_hidden_layers):
+                calculated_layer_num += 1
+                encoder_outputs = self.encoder.adaptive_forward(
+                    encoder_outputs, current_layer=i, attention_mask=extended_attention_mask, head_mask=head_mask
+                )
+
+                pooled_output = self.pooler(encoder_outputs)
+                logits = output_layers[i](pooled_output)
+                if regression:
+                    labels = logits.detach()
+                    if patient_result is not None:
+                        patient_labels = patient_result.detach()
+                    if (patient_result is not None) and torch.abs(patient_result - labels) < self.regression_threshold:
+                        patient_counter += 1
+                    else:
+                        patient_counter = 0
+                else:
+                    labels = logits.detach().argmax(dim=1)
+                    if patient_result is not None:
+                        patient_labels = patient_result.detach().argmax(dim=1)
+                    if (patient_result is not None) and torch.all(labels.eq(patient_labels)):
+                        patient_counter += 1
+                    else:
+                        patient_counter = 0
+
+                patient_result = logits
+                if patient_counter == self.patience:
+                    break
+            res = [patient_result]
+            self.inference_layers_num += calculated_layer_num
+            self.inference_instances_num += 1
+
+        return res
+
+
+@add_start_docstrings(
+    """Bert Model transformer with PABEE and a sequence classification/regression head on top (a linear layer on top of
+    the pooled output) e.g. for GLUE tasks. """,
+    BERT_START_DOCSTRING,
+)
+class BertForSequenceClassificationWithPabee(BertPreTrainedModel):
+    def __init__(self, config):
+        super().__init__(config)
+        self.num_labels = config.num_labels
+
+        self.bert = BertModelWithPabee(config)
+        self.dropout = nn.Dropout(config.hidden_dropout_prob)
+        self.classifiers = nn.ModuleList(
+            [nn.Linear(config.hidden_size, self.config.num_labels) for _ in range(config.num_hidden_layers)]
+        )
+
+        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,
+    ):
+        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 ``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 ``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.
+
+    Examples::
+
+        from transformers import BertTokenizer, BertForSequenceClassification
+        from pabee import BertForSequenceClassificationWithPabee
+        import torch
+
+        tokenizer = BertTokenizer.from_pretrained('bert-base-uncased')
+        model = BertForSequenceClassificationWithPabee.from_pretrained('bert-base-uncased')
+
+        input_ids = torch.tensor(tokenizer.encode("Hello, my dog is cute", add_special_tokens=True)).unsqueeze(0)  # Batch size 1
+        labels = torch.tensor([1]).unsqueeze(0)  # Batch size 1
+        outputs = model(input_ids, labels=labels)
+
+        loss, logits = outputs[:2]
+
+        """
+
+        logits = self.bert(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            token_type_ids=token_type_ids,
+            position_ids=position_ids,
+            head_mask=head_mask,
+            inputs_embeds=inputs_embeds,
+            output_dropout=self.dropout,
+            output_layers=self.classifiers,
+            regression=self.num_labels == 1,
+        )
+
+        outputs = (logits[-1],)
+
+        if labels is not None:
+            total_loss = None
+            total_weights = 0
+            for ix, logits_item in enumerate(logits):
+                if self.num_labels == 1:
+                    #  We are doing regression
+                    loss_fct = MSELoss()
+                    loss = loss_fct(logits_item.view(-1), labels.view(-1))
+                else:
+                    loss_fct = CrossEntropyLoss()
+                    loss = loss_fct(logits_item.view(-1, self.num_labels), labels.view(-1))
+                if total_loss is None:
+                    total_loss = loss
+                else:
+                    total_loss += loss * (ix + 1)
+                total_weights += ix + 1
+            outputs = (total_loss / total_weights,) + outputs
+
+        return outputs

examples/bert-loses-patience/run_glue_with_pabee.py

@@ -0,0 +1,708 @@
+# coding=utf-8
+# Copyright 2020 The Google AI Language Team Authors, The HuggingFace Inc. team and Microsoft Corporation.
+# Copyright (c) 2018, NVIDIA CORPORATION.  All rights reserved.
+#
+# 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.
+""" Training and inference using the library models for sequence classification on GLUE (Bert, Albert) with PABEE."""
+
+
+import argparse
+import glob
+import json
+import logging
+import os
+import random
+
+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 pabee.modeling_pabee_albert import AlbertForSequenceClassificationWithPabee
+from pabee.modeling_pabee_bert import BertForSequenceClassificationWithPabee
+from transformers import (
+    WEIGHTS_NAME,
+    AdamW,
+    AlbertConfig,
+    AlbertTokenizer,
+    BertConfig,
+    BertTokenizer,
+    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, BertForSequenceClassificationWithPabee, BertTokenizer),
+    "albert": (AlbertConfig, AlbertForSequenceClassificationWithPabee, AlbertTokenizer),
+}
+
+
+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 train(args, train_dataset, model, tokenizer):
+    """ 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"]
+    optimizer_grouped_parameters = [
+        {
+            "params": [p for n, p in model.named_parameters() if 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 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
+    )
+
+    # Check if saved optimizer or scheduler states exist
+    if os.path.isfile(os.path.join(args.model_name_or_path, "optimizer.pt")) and os.path.isfile(
+        os.path.join(args.model_name_or_path, "scheduler.pt")
+    ):
+        # Load in optimizer and scheduler states
+        optimizer.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "optimizer.pt")))
+        scheduler.load_state_dict(torch.load(os.path.join(args.model_name_or_path, "scheduler.pt")))
+
+    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
+    epochs_trained = 0
+    steps_trained_in_current_epoch = 0
+    # Check if continuing training from a checkpoint
+    if os.path.exists(args.model_name_or_path):
+        # set global_step to gobal_step of last saved checkpoint from model path
+        global_step = int(args.model_name_or_path.split("-")[-1].split("/")[0])
+        epochs_trained = global_step // (len(train_dataloader) // args.gradient_accumulation_steps)
+        steps_trained_in_current_epoch = global_step % (len(train_dataloader) // args.gradient_accumulation_steps)
+
+        logger.info("  Continuing training from checkpoint, will skip to saved global_step")
+        logger.info("  Continuing training from epoch %d", epochs_trained)
+        logger.info("  Continuing training from global step %d", global_step)
+        logger.info(
+            "  Will skip the first %d steps in the first epoch", steps_trained_in_current_epoch,
+        )
+
+    tr_loss, logging_loss = 0.0, 0.0
+    model.zero_grad()
+    train_iterator = trange(
+        epochs_trained, int(args.num_train_epochs), desc="Epoch", disable=args.local_rank not in [-1, 0],
+    )
+    set_seed(args)  # Added here for reproductibility
+    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):
+
+            # Skip past any already trained steps if resuming training
+            if steps_trained_in_current_epoch > 0:
+                steps_trained_in_current_epoch -= 1
+                continue
+
+            model.train()
+            batch = tuple(t.to(args.device) for t in batch)
+            inputs = {
+                "input_ids": batch[0],
+                "attention_mask": batch[1],
+                "labels": batch[3],
+            }
+            inputs["token_type_ids"] = batch[2]
+            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:
+                    logs = {}
+                    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():
+                            eval_key = "eval_{}".format(key)
+                            logs[eval_key] = value
+
+                    loss_scalar = (tr_loss - logging_loss) / args.logging_steps
+                    learning_rate_scalar = scheduler.get_lr()[0]
+                    logs["learning_rate"] = learning_rate_scalar
+                    logs["loss"] = loss_scalar
+                    logging_loss = tr_loss
+
+                    for key, value in logs.items():
+                        tb_writer.add_scalar(key, value, global_step)
+                    print(json.dumps({**logs, **{"step": global_step}}))
+
+                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)
+                    tokenizer.save_pretrained(output_dir)
+
+                    torch.save(args, os.path.join(output_dir, "training_args.bin"))
+                    logger.info("Saving model checkpoint to %s", output_dir)
+
+                    torch.save(optimizer.state_dict(), os.path.join(output_dir, "optimizer.pt"))
+                    torch.save(scheduler.state_dict(), os.path.join(output_dir, "scheduler.pt"))
+                    logger.info("Saving optimizer and scheduler states 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=""):
+
+    # PABEE STATS
+    if args.model_type == "albert":
+        model.albert.reset_stats()
+    elif args.model_type == "bert":
+        model.bert.reset_stats()
+    else:
+        raise NotImplementedError()
+
+    # 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)
+        eval_dataloader = DataLoader(eval_dataset, sampler=eval_sampler, batch_size=args.eval_batch_size)
+
+        # multi-gpu eval
+        if args.n_gpu > 1 and not isinstance(model, torch.nn.DataParallel):
+            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
+        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],
+                }
+                inputs["token_type_ids"] = batch[2]
+                outputs = model(**inputs)
+                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_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)
+
+        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]))
+                print("  %s = %s" % (key, str(result[key])))
+                writer.write("%s = %s\n" % (key, str(result[key])))
+
+    if args.eval_all_checkpoints:
+        if args.model_type == "albert":
+            model.albert.log_stats()
+        elif args.model_type == "bert":
+            model.bert.log_stats()
+        else:
+            raise NotImplementedError()
+
+    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", "xlmroberta"]:
+            # 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)
+    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(
+        "--patience", default="0", type=str, required=False,
+    )
+    parser.add_argument(
+        "--regression_threshold", default=0, type=float, required=False,
+    )
+
+    # 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="Run 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(
+        "--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=1, 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 decay 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("--logging_steps", type=int, default=500, help="Log every X updates steps.")
+    parser.add_argument(
+        "--save_steps", type=int, default=500, 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)
+
+    if args.patience != "0" and args.per_gpu_eval_batch_size != 1:
+        raise ValueError("The eval batch size must be 1 with PABEE inference on.")
+
+    # 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.local_rank == 0:
+        torch.distributed.barrier()  # Make sure only the first process in distributed training will download model & vocab
+
+    model.to(args.device)
+
+    print("Total Model Parameters:", sum(param.numel() for param in model.parameters()))
+    output_layers_param_num = sum(param.numel() for param in model.classifiers.parameters())
+    print("Output Layers Parameters:", output_layers_param_num)
+    single_output_layer_param_num = sum(param.numel() for param in model.classifiers[0].parameters())
+    print(
+        "Added Output Layers Parameters:", output_layers_param_num - single_output_layer_param_num,
+    )
+
+    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)
+
+    # 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]:
+        patience_list = [int(x) for x in args.patience.split(",")]
+        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)
+            model.to(args.device)
+
+            print(f"Evaluation for checkpoint {prefix}")
+            for patience in patience_list:
+                if args.model_type == "albert":
+                    model.albert.set_regression_threshold(args.regression_threshold)
+                    model.albert.set_patience(patience)
+                elif args.model_type == "bert":
+                    model.bert.set_regression_threshold(args.regression_threshold)
+                    model.bert.set_patience(patience)
+                else:
+                    raise NotImplementedError()
+                result = evaluate(args, model, tokenizer, prefix=prefix)
+                result = dict((k + "_{}".format(global_step), v) for k, v in result.items())
+                results.update(result)
+    return results
+
+
+if __name__ == "__main__":
+    main()

examples/bert-loses-patience/test_run_glue_with_pabee.py

@@ -0,0 +1,48 @@
+import argparse
+import logging
+import sys
+import unittest
+from unittest.mock import patch
+
+import run_glue_with_pabee
+
+
+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 PabeeTests(unittest.TestCase):
+    def test_run_glue(self):
+        stream_handler = logging.StreamHandler(sys.stdout)
+        logger.addHandler(stream_handler)
+
+        testargs = """
+            run_glue_with_pabee.py
+            --model_type albert
+            --model_name_or_path albert-base-v2
+            --data_dir ./tests/fixtures/tests_samples/MRPC/
+            --task_name mrpc
+            --do_train
+            --do_eval
+            --output_dir ./tests/fixtures/tests_samples/temp_dir
+            --per_gpu_train_batch_size=2
+            --per_gpu_eval_batch_size=1
+            --learning_rate=2e-5
+            --max_steps=50
+            --warmup_steps=2
+            --overwrite_output_dir
+            --seed=42
+            --max_seq_length=128
+            """.split()
+        with patch.object(sys, "argv", testargs):
+            result = run_glue_with_pabee.main()
+            for value in result.values():
+                self.assertGreaterEqual(value, 0.75)