# Copyright 2022 Cerebras Systems.
#
# 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.
import torch
from cerebras_pytorch.metrics import AccuracyMetric, FBetaScoreMetric
from modelzoo.transformers.pytorch.bert.bert_finetune_models import (
BertForSequenceClassification,
BertForSequenceClassificationLoss,
)
from modelzoo.transformers.pytorch.bert.utils import check_unused_model_params
[docs]class BertForSequenceClassificationModel(torch.nn.Module):
[docs] def __init__(self, params):
super().__init__()
model_params = params["model"].copy()
self.num_labels = model_params.pop("num_labels")
problem_type = model_params.pop("problem_type")
classifier_dropout = model_params.pop("task_dropout")
dropout_rate = model_params.pop("dropout_rate")
embedding_dropout_rate = model_params.pop(
"embedding_dropout_rate", dropout_rate
)
model_kwargs = {
"vocab_size": model_params.pop("vocab_size"),
"hidden_size": model_params.pop("hidden_size"),
"num_hidden_layers": model_params.pop("num_hidden_layers"),
"num_heads": model_params.pop("num_heads"),
"filter_size": model_params.pop("filter_size"),
"nonlinearity": model_params.pop("encoder_nonlinearity"),
"pooler_nonlinearity": model_params.pop(
"pooler_nonlinearity", None
),
"embedding_dropout_rate": embedding_dropout_rate,
"dropout_rate": dropout_rate,
"attention_dropout_rate": model_params.pop(
"attention_dropout_rate"
),
"max_position_embeddings": model_params.pop(
"max_position_embeddings"
),
"layer_norm_epsilon": float(model_params.pop("layer_norm_epsilon")),
}
self.model = BertForSequenceClassification(
self.num_labels, problem_type, classifier_dropout, **model_kwargs,
)
self.loss_fn = BertForSequenceClassificationLoss(
self.num_labels, problem_type
)
self.compute_eval_metrics = model_params.pop(
"compute_eval_metrics", False
)
# Below flag helps create two more accuracy objects for
# matched and mismatched partitions
self.is_mnli_dataset = model_params.pop("is_mnli_dataset", False)
check_unused_model_params(model_params)
if self.compute_eval_metrics:
self.accuracy_metric = AccuracyMetric(name="eval/accuracy")
if self.num_labels == 2:
self.f1_metric = FBetaScoreMetric(
num_classes=self.num_labels,
beta=1.0,
average_type="micro",
name="eval/f1_score",
)
if self.is_mnli_dataset:
self.matched_accuracy_metric = AccuracyMetric(
name="eval/accuracy_matched"
)
self.mismatched_accuracy_metric = AccuracyMetric(
name="eval/accuracy_mismatched"
)
def forward(self, data):
logits = self.model(
input_ids=data["input_ids"],
token_type_ids=data["token_type_ids"],
attention_mask=data["attention_mask"],
)
loss = self.loss_fn(data["labels"], logits)
if not self.model.training and self.compute_eval_metrics:
labels = data["labels"].clone()
predictions = logits.argmax(-1).int()
self.accuracy_metric(
labels=labels, predictions=predictions, dtype=logits.dtype
)
if self.num_labels == 2:
self.f1_metric(labels=labels, predictions=predictions)
if self.is_mnli_dataset:
self.matched_accuracy_metric(
labels=labels,
predictions=predictions,
weights=data["is_matched"],
dtype=logits.dtype,
)
self.mismatched_accuracy_metric(
labels=labels,
predictions=predictions,
weights=data["is_mismatched"],
dtype=logits.dtype,
)
return loss