def build_criterion(vocab_size, padding_idx, smoothing):
if smoothing == 0.:
logging.info(f'Building CrossEntropyLoss')
criterion = nn.CrossEntropyLoss(ignore_index=padding_idx, size_average=False)
else:
logging.info(f'Building LabelSmoothingLoss (smoothing: {smoothing})')
criterion = LabelSmoothing(padding_idx, smoothing)
return criterion
def build_gnmt_criterion(vocab_size, padding_idx, smoothing):
if smoothing == 0.:
loss_weight = torch.ones(vocab_size)
loss_weight[padding_idx] = 0
criterion = CrossEntropyWrapper(weight=loss_weight, size_average=False)
else:
criterion = LabelSmoothing(padding_idx, smoothing)
return criterion
def build_criterion(vocab_size, padding_idx, smoothing, fusion):
if smoothing == 0.:
logging.info(f'Building CrossEntropyLoss')
loss_weight = torch.ones(vocab_size)
loss_weight[padding_idx] = 0
criterion = nn.CrossEntropyLoss(weight=loss_weight, size_average=False)
else:
logging.info(f'Building LabelSmoothingLoss (smoothing: {smoothing})')
criterion = LabelSmoothing(padding_idx, smoothing, fusion=fusion)
return criterion
def build_criterion(vocab_size, padding_idx, smoothing):
if smoothing == 0.:
logging.info(f'Building CrossEntropyLoss')
loss_weight = torch.ones(vocab_size)
loss_weight[padding_idx] = 0
criterion = nn.CrossEntropyLoss(weight=loss_weight, size_average=False)
gnmt_print(key=mlperf_log.MODEL_HP_LOSS_FN, value='Cross Entropy')
else:
logging.info(f'Building LabelSmoothingLoss (smoothing: {smoothing})')
criterion = LabelSmoothing(padding_idx, smoothing)
gnmt_print(key=mlperf_log.MODEL_HP_LOSS_FN,
value='Cross Entropy with label smoothing')
gnmt_print(key=mlperf_log.MODEL_HP_LOSS_SMOOTHING, value=smoothing)
return criterion