I found two implementations of T5Encoder for sequence classification in Github gists:
class T5EncoderClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
classifier_dropout = (
config.classifier_dropout if config.classifier_dropout is not None else config.hidden_dropout_prob
)
self.dropout = nn.Dropout(classifier_dropout)
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
def forward(self, hidden_states, **kwargs):
hidden_states = hidden_states[:, 0, :] # take <s> token (equiv. to [CLS])
hidden_states = self.dropout(hidden_states)
hidden_states = self.dense(hidden_states)
hidden_states = torch.tanh(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.out_proj(hidden_states)
return hidden_states
class T5EncoderForSequenceClassification:
"""
Use an in-memory T5Encoder to do sequence classification"""
def __init__(self, t5_encoder, config):
self.num_labels = config.num_labels
self.config = config
self.encoder = t5_encoder # already initialized model
self.classifier = T5EncoderClassificationHead(config)
def forward(
self,
input_ids=None,
attention_mask=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_hidden_states=None,
output_attentions=None,
return_dict=None,
):
r"""
labels (`torch.LongTensor` of shape `(batch_size,)`, *optional*):
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).
"""
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
encoder_outputs = self.encoder(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
sequence_output = encoder_outputs[0]
logits = self.classifier(sequence_output)
loss = None
if labels is not None:
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
loss_fct = MSELoss()
if self.num_labels == 1:
loss = loss_fct(logits.squeeze(), labels.squeeze())
else:
loss = loss_fct(logits, labels)
elif self.config.problem_type == "single_label_classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss_fct = BCEWithLogitsLoss()
loss = loss_fct(logits, labels)
if not return_dict:
output = (logits,) + encoder_outputs[2:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=encoder_outputs.hidden_states,
attentions=encoder_outputs.attentions,
)
The second being:
import copy
import torch
from torch import nn
from torch.nn import BCEWithLogitsLoss, CrossEntropyLoss, MSELoss
from transformers.modeling_outputs import SequenceClassifierOutput
from transformers.models.t5.modeling_t5 import T5Config, T5PreTrainedModel, T5Stack
from transformers.utils.model_parallel_utils import assert_device_map, get_device_map
class T5EncoderClassificationHead(nn.Module):
"""Head for sentence-level classification tasks."""
def __init__(self, config):
super().__init__()
self.dense = nn.Linear(config.hidden_size, config.hidden_size)
self.dropout = nn.Dropout(config.dropout_rate)
self.out_proj = nn.Linear(config.hidden_size, config.num_labels)
def forward(self, hidden_states):
hidden_states = hidden_states[:, 0, :] # take <s> token (equiv. to [CLS])
hidden_states = self.dropout(hidden_states)
hidden_states = self.dense(hidden_states)
hidden_states = torch.tanh(hidden_states)
hidden_states = self.dropout(hidden_states)
hidden_states = self.out_proj(hidden_states)
return hidden_states
class T5EncoderForSimpleSequenceClassification(T5PreTrainedModel):
# _keys_to_ignore_on_load_missing = [
# r"encoder\.embed_tokens\.weight",
# ]
def __init__(self, config: T5Config, dropout=0.1):
super().__init__(config)
self.num_labels = config.num_labels
self.config = config
self.shared = nn.Embedding(config.vocab_size, config.d_model)
encoder_config = copy.deepcopy(config)
encoder_config.use_cache = False
encoder_config.is_encoder_decoder = False
self.encoder = T5Stack(encoder_config, self.shared)
self.dropout = nn.Dropout(dropout)
self.classifier = T5EncoderClassificationHead(config)
# Initialize weights and apply final processing
self.post_init()
# Model parallel
self.model_parallel = False
self.device_map = None
def parallelize(self, device_map=None):
self.device_map = (
get_device_map(len(self.encoder.block), range(torch.cuda.device_count()))
if device_map is None
else device_map
)
assert_device_map(self.device_map, len(self.encoder.block))
self.encoder.parallelize(self.device_map)
self.classifier = self.classifier.to(self.encoder.first_device)
self.model_parallel = True
def deparallelize(self):
self.encoder.deparallelize()
self.encoder = self.encoder.to("cpu")
self.model_parallel = False
self.device_map = None
torch.cuda.empty_cache()
def get_input_embeddings(self):
return self.shared
def set_input_embeddings(self, new_embeddings):
self.shared = new_embeddings
self.encoder.set_input_embeddings(new_embeddings)
def get_encoder(self):
return self.encoder
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)
def forward(
self,
input_ids=None,
attention_mask=None,
head_mask=None,
inputs_embeds=None,
labels=None,
output_attentions=None,
output_hidden_states=None,
return_dict=None,
):
return_dict = return_dict if return_dict is not None else self.config.use_return_dict
outputs = self.encoder(
input_ids=input_ids,
attention_mask=attention_mask,
inputs_embeds=inputs_embeds,
head_mask=head_mask,
output_attentions=output_attentions,
output_hidden_states=output_hidden_states,
return_dict=return_dict,
)
hidden_states = outputs[0]
logits = self.classifier(hidden_states)
loss = None
if labels is not None:
if self.config.problem_type is None:
if self.num_labels == 1:
self.config.problem_type = "regression"
elif self.num_labels > 1 and (labels.dtype == torch.long or labels.dtype == torch.int):
self.config.problem_type = "single_label_classification"
else:
self.config.problem_type = "multi_label_classification"
if self.config.problem_type == "regression":
loss_fct = MSELoss()
if self.num_labels == 1:
loss = loss_fct(logits.squeeze(), labels.squeeze())
else:
loss = loss_fct(logits, labels)
elif self.config.problem_type == "single_label_classification":
loss_fct = CrossEntropyLoss()
loss = loss_fct(logits.view(-1, self.num_labels), labels.view(-1))
elif self.config.problem_type == "multi_label_classification":
loss_fct = BCEWithLogitsLoss()
loss = loss_fct(logits, labels)
if not return_dict:
output = (logits,) + outputs[1:]
return ((loss,) + output) if loss is not None else output
return SequenceClassifierOutput(
loss=loss,
logits=logits,
hidden_states=outputs.hidden_states,
attentions=outputs.attentions,
)
Both look similar other than that first defines encoder directly inside the T5EncoderForSequenceClassification class while the second seem to be a copy-paste of T5EncoderModel with classifer added.
I want to know which one is right or whether both are right? If I run the code, for first case, I get that embed_tokens and embed_weight are newly initialized (weird to me!) and I dont have that issue in second.
So, can someone tell me which version is correct? Though second version seem to use the pretrained embedding weights, it gives lower accuracy score than the first version. So, not sure which one is the correct definition for the classification model?