# 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.
# TODO: Move to a separate folder when more head models are added like MLM and NSP on bert, etc.
import torch
import torch.nn as nn
[docs]class GPTLMHeadModelLoss(nn.Module):
def __init__(
self,
vocab_size,
loss_scaling,
loss_weight,
):
super(GPTLMHeadModelLoss, self).__init__()
self.vocab_size = vocab_size
self.loss_weight = loss_weight
self.loss_scaling = loss_scaling
assert (
self.loss_scaling == "num_tokens"
or self.loss_scaling == "batch_size"
), f"Loss scaling can't be set to {self.loss_scaling}. \
Should be either 'num_tokens' or 'batch_size'"
if self.loss_scaling == "num_tokens":
assert (
self.loss_weight == 1.0
), f"Loss scaling with 'num_tokens' requires loss_weight == 1.0"
def forward(
self,
lm_logits,
labels,
attention_mask,
reduce_batch=True,
average_logps=False,
):
loss_fct = nn.CrossEntropyLoss(reduction='none')
lm_loss = loss_fct(
lm_logits.view(-1, self.vocab_size),
labels.view(-1).long(),
)
if reduce_batch:
assert (
not average_logps
), "average_logps can only be set to True, when reduce_batch=False"
lm_loss = lm_loss * attention_mask.to(dtype=lm_logits.dtype).view(
-1
)
if self.loss_scaling == "num_tokens":
lm_loss = torch.sum(lm_loss) / torch.sum(
attention_mask.to(dtype=lm_logits.dtype)
)
else:
lm_loss = (
torch.sum(lm_loss) / labels.shape[0]
) * self.loss_weight
loss = lm_loss
else:
loss = lm_loss.view(attention_mask.shape) * attention_mask.to(
dtype=lm_logits.dtype
)
loss = torch.sum(loss, dim=-1)
if average_logps:
loss = loss / torch.sum(
attention_mask.to(dtype=lm_logits.dtype), dim=-1
)
return loss