""" Head Tuning with Prefix / Adapter """ from typing import Optional, List, Union, Tuple import torch from torch._C import NoopLogger import torch.nn import torch.nn.functional as F from torch import Tensor from torch.nn import CrossEntropyLoss, MSELoss, BCEWithLogitsLoss from transformers import BertModel, BertPreTrainedModel from transformers import RobertaModel, RobertaPreTrainedModel from transformers.models.deberta.modeling_deberta import DebertaModel, DebertaPreTrainedModel, ContextPooler, StableDropout from transformers.models.gpt2.modeling_gpt2 import GPT2Model, GPT2PreTrainedModel from transformers.models.bart.modeling_bart import BartPretrainedModel, BartClassificationHead, BartModel from transformers.models.roberta.modeling_roberta import RobertaClassificationHead from transformers.models.bart.configuration_bart import BartConfig from transformers.modeling_outputs import SequenceClassifierOutput, Seq2SeqSequenceClassifierOutput, SequenceClassifierOutputWithPast from models.basic_modules.prefix_encoder import PrefixEncoder from models.basic_modules.adapter import BertAdaModel, RobertaAdaModel, init_adapter from tools.model_utils.parameter_freeze import ParameterFreeze from tools.runner_utils.log_util import logging logger = logging.getLogger(__name__) freezer = ParameterFreeze() ## ======== BERT ======== # Vanilla Fine-tuning For BERT class BertForSequenceClassification(BertPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.config = config self.bert = BertModel(config) if self.config.use_freezing: self.bert = freezer.freeze_lm(self.bert) self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) self.init_weights() def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.bert = freezer.freeze_lm(self.bert) else: self.bert = freezer.unfreeze_lm(self.bert) def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`): 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). """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict # print("input_ids.shape=", input_ids.shape) # e.g., [8, 128] # print("attention_mask.shape=", attention_mask.shape) # e.g., [8, 128] # print("token_type_ids.shape=", token_type_ids.shape) # e.g., [8, 128] outputs = self.bert( 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) pooled_output = outputs[1] pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_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,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # Prefix-tuning For BERT class BertPrefixForSequenceClassification(BertPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.config = config self.bert = BertModel(config) self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) # for param in self.bert.parameters(): # param.requires_grad = False if self.config.use_freezing: self.bert = freezer.freeze_lm(self.bert) self.pre_seq_len = config.pre_seq_len self.n_layer = config.num_hidden_layers self.n_head = config.num_attention_heads self.n_embd = config.hidden_size // config.num_attention_heads self.prefix_tokens = torch.arange(self.pre_seq_len).long() self.prefix_encoder = PrefixEncoder(config) bert_param = 0 for name, param in self.bert.named_parameters(): bert_param += param.numel() all_param = 0 for name, param in self.named_parameters(): all_param += param.numel() total_param = all_param - bert_param print("total param is {}".format(total_param)) # 9860105 def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.bert = freezer.freeze_lm(self.bert) else: self.bert = freezer.unfreeze_lm(self.bert) def get_prompt(self, batch_size): prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(self.bert.device) past_key_values = self.prefix_encoder(prefix_tokens) # bsz, seqlen, _ = past_key_values.shape past_key_values = past_key_values.view( batch_size, self.pre_seq_len, self.n_layer * 2, self.n_head, self.n_embd ) past_key_values = self.dropout(past_key_values) past_key_values = past_key_values.permute([2, 0, 3, 1, 4]).split(2) return past_key_values def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=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 # print("input_ids.shape=", input_ids.shape) # e.g., [8, 128] # print("attention_mask.shape=", attention_mask.shape) # e.g., [8, 128] # print("token_type_ids.shape=", token_type_ids.shape) # e.g., [8, 128] batch_size = input_ids.shape[0] past_key_values = self.get_prompt(batch_size=batch_size) prefix_attention_mask = torch.ones(batch_size, self.pre_seq_len).to(self.bert.device) attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=1) if position_ids is None: position_ids = torch.tensor([i for i in range(input_ids.shape[-1])]).expand(batch_size, -1).to(self.bert.device) outputs = self.bert( 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, past_key_values=past_key_values, ) pooled_output = outputs[1] pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_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,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # Prompt-tuning For BERT class BertPtuningForSequenceClassification(BertPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.bert = BertModel(config) self.embeddings = self.bert.embeddings self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) # for param in self.bert.parameters(): # param.requires_grad = False if self.config.use_freezing: self.bert = freezer.freeze_lm(self.bert) self.pre_seq_len = config.pre_seq_len self.n_layer = config.num_hidden_layers self.n_head = config.num_attention_heads self.n_embd = config.hidden_size // config.num_attention_heads self.prefix_tokens = torch.arange(self.pre_seq_len).long() self.prefix_encoder = torch.nn.Embedding(self.pre_seq_len, config.hidden_size) def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.bert = freezer.freeze_lm(self.bert) else: self.bert = freezer.unfreeze_lm(self.bert) def get_prompt(self, batch_size): prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(self.bert.device) prompts = self.prefix_encoder(prefix_tokens) return prompts def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=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 batch_size = input_ids.shape[0] raw_embedding = self.embeddings( input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, ) prompts = self.get_prompt(batch_size=batch_size) inputs_embeds = torch.cat((prompts, raw_embedding), dim=1) prefix_attention_mask = torch.ones(batch_size, self.pre_seq_len).to(self.bert.device) attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=1) outputs = self.bert( # 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, # past_key_values=past_key_values, ) # pooled_output = outputs[1] sequence_output = outputs[0] sequence_output = sequence_output[:, self.pre_seq_len:, :].contiguous() first_token_tensor = sequence_output[:, 0] pooled_output = self.bert.pooler.dense(first_token_tensor) pooled_output = self.bert.pooler.activation(pooled_output) pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_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,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # Adapter-tuning For BERT class BertAdapterForSequenceClassification(BertPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.bert = BertAdaModel(config) self.embeddings = self.bert.embeddings self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) # for param in self.bert.parameters(): # param.requires_grad = False if self.config.use_freezing: self.bert = freezer.freeze_lm_component(self.bert, "adapter") def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.bert = freezer.freeze_lm_component(self.bert, "adapter") else: self.bert = freezer.unfreeze_lm(self.bert) def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=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 batch_size = input_ids.shape[0] inputs_embeds = self.embeddings( input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, ) outputs = self.bert( # 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, # past_key_values=past_key_values, ) # pooled_output = outputs[1] sequence_output = outputs[0] # sequence_output = sequence_output[:, self.pre_seq_len:, :].contiguous() first_token_tensor = sequence_output[:, 0] pooled_output = self.bert.pooler.dense(first_token_tensor) pooled_output = self.bert.pooler.activation(pooled_output) pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_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,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # ========= RoBERTa ========= # Vanilla Fine-tuning For RoBERTa class RobertaForSequenceClassification(RobertaPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.config = config self.roberta = RobertaModel(config) if self.config.use_freezing: self.roberta = freezer.freeze_lm(self.roberta) self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) # self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) self.classifier = RobertaClassificationHead(config) self.init_weights() def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.roberta = freezer.freeze_lm(self.roberta) else: self.roberta = freezer.unfreeze_lm(self.roberta) def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=None, head_mask=None, inputs_embeds=None, labels=None, output_attentions=None, output_hidden_states=None, return_dict=None, ): r""" labels (:obj:`torch.LongTensor` of shape :obj:`(batch_size,)`, `optional`): 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). """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict outputs = self.roberta( 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) pooled_output = outputs[1] pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_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,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # Prefix-tuning For RoBERTa class RobertaPrefixForSequenceClassification(RobertaPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.config = config self.roberta = RobertaModel(config) self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) # self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) self.classifier = RobertaClassificationHead(config) self.init_weights() for param in self.roberta.parameters(): param.requires_grad = False self.pre_seq_len = config.pre_seq_len self.n_layer = config.num_hidden_layers self.n_head = config.num_attention_heads self.n_embd = config.hidden_size // config.num_attention_heads self.prefix_tokens = torch.arange(self.pre_seq_len).long() self.prefix_encoder = PrefixEncoder(config) bert_param = 0 for name, param in self.roberta.named_parameters(): bert_param += param.numel() all_param = 0 for name, param in self.named_parameters(): all_param += param.numel() total_param = all_param - bert_param print("total param is {}".format(total_param)) # 9860105 def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.roberta = freezer.freeze_lm(self.roberta) else: self.roberta = freezer.unfreeze_lm(self.roberta) def get_prompt(self, batch_size): prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(self.roberta.device) # print("prefix_tokens.shape=", prefix_tokens.shape) past_key_values = self.prefix_encoder(prefix_tokens) # print("past_key_values[0].shape=", past_key_values[0].shape) past_key_values = past_key_values.view( batch_size, self.pre_seq_len, self.n_layer * 2, self.n_head, self.n_embd ) # print("past_key_values[0].shape=", past_key_values[0].shape) past_key_values = self.dropout(past_key_values) past_key_values = past_key_values.permute([2, 0, 3, 1, 4]).split(2) # print("past_key_values[0].shape=", past_key_values[0].shape) return past_key_values def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=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 batch_size = input_ids.shape[0] past_key_values = self.get_prompt(batch_size=batch_size) prefix_attention_mask = torch.ones(batch_size, self.pre_seq_len).to(self.roberta.device) attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=1) if position_ids is None: position_ids = torch.tensor([i for i in range(input_ids.shape[-1])]).expand(batch_size, -1).to(self.roberta.device) outputs = self.roberta( 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, past_key_values=past_key_values, ) pooled_output = outputs[1] pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_output) loss = None if labels is not None: labels = (labels < 0).long().to(labels.device) + labels 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[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # Prompt-tuning For RoBERTa class RobertaPtuningForSequenceClassification(RobertaPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.roberta = RobertaModel(config) self.embeddings = self.roberta.embeddings self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) # self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) self.classifier = RobertaClassificationHead(config) # for param in self.roberta.parameters(): # param.requires_grad = False if self.config.use_freezing: self.roberta = freezer.freeze_lm(self.roberta) self.pre_seq_len = config.pre_seq_len self.n_layer = config.num_hidden_layers self.n_head = config.num_attention_heads self.n_embd = config.hidden_size // config.num_attention_heads self.prefix_tokens = torch.arange(self.pre_seq_len).long() self.prefix_encoder = torch.nn.Embedding(self.pre_seq_len, config.hidden_size) def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.roberta = freezer.freeze_lm(self.roberta) else: self.roberta = freezer.unfreeze_lm(self.roberta) def get_prompt(self, batch_size): prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(self.roberta.device) prompts = self.prefix_encoder(prefix_tokens) return prompts def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=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 batch_size = input_ids.shape[0] raw_embedding = self.embeddings( input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, ) prompts = self.get_prompt(batch_size=batch_size) inputs_embeds = torch.cat((prompts, raw_embedding), dim=1) # print(input_embeddings.shape) # exit() prefix_attention_mask = torch.ones(batch_size, self.pre_seq_len).to(self.roberta.device) attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=1) outputs = self.roberta( # 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, # past_key_values=past_key_values, ) # pooled_output = outputs[1] sequence_output = outputs[0] sequence_output = sequence_output[:, self.pre_seq_len:, :].contiguous() first_token_tensor = sequence_output[:, 0] pooled_output = self.roberta.pooler.dense(first_token_tensor) pooled_output = self.roberta.pooler.activation(pooled_output) pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_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,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # Adapter-tuning For RoBERTa class RobertaAdapterForSequenceClassification(RobertaPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.roberta = RobertaAdaModel(config) self.embeddings = self.roberta.embeddings self.dropout = torch.nn.Dropout(config.hidden_dropout_prob) # self.classifier = torch.nn.Linear(config.hidden_size, config.num_labels) self.classifier = RobertaClassificationHead(config) self.init_weights() # for param in self.roberta.parameters(): # param.requires_grad = False self.roberta = init_adapter(self.roberta) if self.config.use_freezing: self.roberta = freezer.freeze_lm_component(self.roberta, "adapter") def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.roberta = freezer.freeze_lm_component(self.roberta, "adapter") else: self.roberta = freezer.unfreeze_lm(self.roberta) def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=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 batch_size = input_ids.shape[0] inputs_embeds = self.embeddings( input_ids=input_ids, position_ids=position_ids, token_type_ids=token_type_ids, ) outputs = self.roberta( # 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_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, # past_key_values=past_key_values, ) # pooled_output = outputs[1] sequence_output = outputs[0] # sequence_output = sequence_output[:, self.pre_seq_len:, :].contiguous() first_token_tensor = sequence_output[:, 0] pooled_output = self.roberta.pooler.dense(first_token_tensor) pooled_output = self.roberta.pooler.activation(pooled_output) pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_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,) + outputs[2:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # ========= DeBERTa ========= # Prefix-tuning For DeBERTa class DebertaPrefixForSequenceClassification(DebertaPreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.config = config self.deberta = DebertaModel(config) self.pooler = ContextPooler(config) output_dim = self.pooler.output_dim self.classifier = torch.nn.Linear(output_dim, self.num_labels) self.dropout = StableDropout(config.hidden_dropout_prob) self.init_weights() # for param in self.deberta.parameters(): # param.requires_grad = False if self.config.use_freezing: self.deberta = freezer.freeze_lm(self.deberta) self.pre_seq_len = config.pre_seq_len self.n_layer = config.num_hidden_layers self.n_head = config.num_attention_heads self.n_embd = config.hidden_size // config.num_attention_heads self.prefix_tokens = torch.arange(self.pre_seq_len).long() self.prefix_encoder = PrefixEncoder(config) deberta_param = 0 for name, param in self.deberta.named_parameters(): deberta_param += param.numel() all_param = 0 for name, param in self.named_parameters(): all_param += param.numel() total_param = all_param - deberta_param print("total param is {}".format(total_param)) # 9860105 def freeze_backbone(self, use_freezing: bool=True): if use_freezing: self.deberta = freezer.freeze_lm(self.deberta) else: self.deberta = freezer.unfreeze_lm(self.deberta) def get_prompt(self, batch_size): prefix_tokens = self.prefix_tokens.unsqueeze(0).expand(batch_size, -1).to(self.deberta.device) past_key_values = self.prefix_encoder(prefix_tokens) # bsz, seqlen, _ = past_key_values.shape past_key_values = past_key_values.view( batch_size, self.pre_seq_len, self.n_layer * 2, self.n_head, self.n_embd ) past_key_values = self.dropout(past_key_values) past_key_values = past_key_values.permute([2, 0, 3, 1, 4]).split(2) return past_key_values def forward( self, input_ids=None, attention_mask=None, token_type_ids=None, position_ids=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 batch_size = input_ids.shape[0] past_key_values = self.get_prompt(batch_size=batch_size) prefix_attention_mask = torch.ones(batch_size, self.pre_seq_len).to(self.deberta.device) attention_mask = torch.cat((prefix_attention_mask, attention_mask), dim=1) outputs = self.deberta( input_ids, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, inputs_embeds=inputs_embeds, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, past_key_values=past_key_values, ) encoder_layer = outputs[0] pooled_output = self.pooler(encoder_layer) pooled_output = self.dropout(pooled_output) logits = self.classifier(pooled_output) loss = None if labels is not None: if self.num_labels == 1: # regression task loss_fn = torch.nn.MSELoss() logits = logits.view(-1).to(labels.dtype) loss = loss_fn(logits, labels.view(-1)) elif labels.dim() == 1 or labels.size(-1) == 1: label_index = (labels >= 0).nonzero() labels = labels.long() if label_index.size(0) > 0: labeled_logits = torch.gather(logits, 0, label_index.expand(label_index.size(0), logits.size(1))) labels = torch.gather(labels, 0, label_index.view(-1)) loss_fct = CrossEntropyLoss() loss = loss_fct(labeled_logits.view(-1, self.num_labels).float(), labels.view(-1)) else: loss = torch.tensor(0).to(logits) else: log_softmax = torch.nn.LogSoftmax(-1) loss = -((log_softmax(logits) * labels).sum(-1)).mean() if not return_dict: output = (logits,) + outputs[1:] return ((loss,) + output) if loss is not None else output else: return SequenceClassifierOutput( loss=loss, logits=logits, hidden_states=outputs.hidden_states, attentions=outputs.attentions, ) # GPT2 for classification class GPT2ForSequenceClassification(GPT2PreTrainedModel): def __init__(self, config): super().__init__(config) self.num_labels = config.num_labels self.transformer = GPT2Model(config) self.score = torch.nn.Linear(config.n_embd, self.num_labels, bias=False) # Model parallel self.model_parallel = False self.device_map = None # Initialize weights and apply final processing self.post_init() def forward( self, input_ids: Optional[torch.LongTensor] = None, past_key_values: Optional[Tuple[Tuple[torch.Tensor]]] = None, attention_mask: Optional[torch.FloatTensor] = None, token_type_ids: Optional[torch.LongTensor] = None, position_ids: Optional[torch.LongTensor] = None, head_mask: Optional[torch.FloatTensor] = None, inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, SequenceClassifierOutputWithPast]: 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 transformer_outputs = self.transformer( input_ids, past_key_values=past_key_values, attention_mask=attention_mask, token_type_ids=token_type_ids, position_ids=position_ids, head_mask=head_mask, inputs_embeds=inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = transformer_outputs[0] logits = self.score(hidden_states) if input_ids is not None: batch_size, sequence_length = input_ids.shape[:2] else: batch_size, sequence_length = inputs_embeds.shape[:2] assert ( self.config.pad_token_id is not None or batch_size == 1 ), "Cannot handle batch sizes > 1 if no padding token is defined." if self.config.pad_token_id is None: sequence_lengths = -1 else: if input_ids is not None: sequence_lengths = torch.ne(input_ids, self.config.pad_token_id).sum(-1) - 1 else: sequence_lengths = -1 logger.warning( f"{self.__class__.__name__} will not detect padding tokens in `inputs_embeds`. Results may be " "unexpected if using padding tokens in conjunction with `inputs_embeds.`" ) pooled_logits = logits[torch.arange(batch_size, device=logits.device), sequence_lengths] 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(pooled_logits.squeeze(), labels.squeeze()) else: loss = loss_fct(pooled_logits, labels) elif self.config.problem_type == "single_label_classification": loss_fct = CrossEntropyLoss() loss = loss_fct(pooled_logits.view(-1, self.num_labels), labels.view(-1)) elif self.config.problem_type == "multi_label_classification": loss_fct = BCEWithLogitsLoss() loss = loss_fct(pooled_logits, labels) if not return_dict: output = (pooled_logits,) + transformer_outputs[1:] return ((loss,) + output) if loss is not None else output return SequenceClassifierOutputWithPast( loss=loss, logits=pooled_logits, past_key_values=transformer_outputs.past_key_values, hidden_states=transformer_outputs.hidden_states, attentions=transformer_outputs.attentions, ) # Bart for classification class BartForSequenceClassification(BartPretrainedModel): def __init__(self, config: BartConfig, **kwargs): super().__init__(config, **kwargs) self.model = BartModel(config) self.classification_head = BartClassificationHead( config.d_model, config.d_model, config.num_labels, config.classifier_dropout, ) self.model._init_weights(self.classification_head.dense) self.model._init_weights(self.classification_head.out_proj) def forward( self, input_ids: torch.LongTensor = None, attention_mask: Optional[torch.Tensor] = None, decoder_input_ids: Optional[torch.LongTensor] = None, decoder_attention_mask: Optional[torch.LongTensor] = None, head_mask: Optional[torch.Tensor] = None, decoder_head_mask: Optional[torch.Tensor] = None, cross_attn_head_mask: Optional[torch.Tensor] = None, encoder_outputs: Optional[List[torch.FloatTensor]] = None, inputs_embeds: Optional[torch.FloatTensor] = None, decoder_inputs_embeds: Optional[torch.FloatTensor] = None, labels: Optional[torch.LongTensor] = None, use_cache: Optional[bool] = None, output_attentions: Optional[bool] = None, output_hidden_states: Optional[bool] = None, return_dict: Optional[bool] = None, ) -> Union[Tuple, Seq2SeqSequenceClassifierOutput]: 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 classification loss is computed (Cross-Entropy). """ return_dict = return_dict if return_dict is not None else self.config.use_return_dict if labels is not None: use_cache = False if input_ids is None and inputs_embeds is not None: raise NotImplementedError( f"Passing input embeddings is currently not supported for {self.__class__.__name__}" ) outputs = self.model( input_ids, attention_mask=attention_mask, decoder_input_ids=decoder_input_ids, decoder_attention_mask=decoder_attention_mask, head_mask=head_mask, decoder_head_mask=decoder_head_mask, cross_attn_head_mask=cross_attn_head_mask, encoder_outputs=encoder_outputs, inputs_embeds=inputs_embeds, decoder_inputs_embeds=decoder_inputs_embeds, use_cache=use_cache, output_attentions=output_attentions, output_hidden_states=output_hidden_states, return_dict=return_dict, ) hidden_states = outputs[0] # last hidden state # print("hidden_states.shape=", hidden_states.shape) # [bz, seq_len, dim] eos_mask = input_ids.eq(self.config.eos_token_id) if len(torch.unique_consecutive(eos_mask.sum(1))) > 1: raise ValueError("All examples must have the same number of tokens.") sentence_representation = hidden_states[eos_mask, :].view(hidden_states.size(0), -1, hidden_states.size(-1))[ :, -1, : ] logits = self.classification_head(sentence_representation) loss = None if labels is not None: if self.config.problem_type is None: if self.config.num_labels == 1: self.config.problem_type = "regression" elif self.config.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.config.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.config.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 Seq2SeqSequenceClassifierOutput( loss=loss, logits=logits, past_key_values=outputs.past_key_values, decoder_hidden_states=outputs.decoder_hidden_states, decoder_attentions=outputs.decoder_attentions, cross_attentions=outputs.cross_attentions, encoder_last_hidden_state=outputs.encoder_last_hidden_state, encoder_hidden_states=outputs.encoder_hidden_states, encoder_attentions=outputs.encoder_attentions, )