|
|
from transformers import AutoModel
|
|
|
import torch.nn as nn
|
|
|
from tasks import SECONDARY_TASKS
|
|
|
from huggingface_hub import PyTorchModelHubMixin
|
|
|
|
|
|
class BertMultiTask(nn.Module, PyTorchModelHubMixin):
|
|
|
def __init__(
|
|
|
self, model_name, extra_layer_sizes=[], dropout_rate=0.1, finetune: bool = False
|
|
|
):
|
|
|
super(BertMultiTask, self).__init__()
|
|
|
|
|
|
self.model_name = model_name
|
|
|
self.extra_layer_sizes = extra_layer_sizes
|
|
|
self.dropout_rate = dropout_rate
|
|
|
self.finetune = finetune
|
|
|
|
|
|
self.bert = AutoModel.from_pretrained(model_name)
|
|
|
self.layers = nn.ModuleList()
|
|
|
if not finetune:
|
|
|
self.name = f"{model_name.split('/')[-1]}_all_tasks_{'_'.join(map(str, extra_layer_sizes))}"
|
|
|
for param in self.bert.parameters():
|
|
|
param.requires_grad = False
|
|
|
else:
|
|
|
self.name = f"{model_name.split('/')[-1]}_finetune_all_tasks_{'_'.join(map(str, extra_layer_sizes))}"
|
|
|
for param in self.bert.parameters():
|
|
|
param.requires_grad = True
|
|
|
|
|
|
prev_size = self.bert.config.hidden_size
|
|
|
for size in extra_layer_sizes:
|
|
|
self.layers.append(nn.Linear(prev_size, size))
|
|
|
self.layers.append(nn.ReLU())
|
|
|
self.layers.append(nn.Dropout(dropout_rate))
|
|
|
prev_size = size
|
|
|
|
|
|
self.reg_head = nn.Linear(prev_size, 1)
|
|
|
|
|
|
self.classification_heads = nn.ModuleDict()
|
|
|
for task_name, id_map in SECONDARY_TASKS.items():
|
|
|
self.classification_heads[task_name] = nn.Linear(prev_size, len(id_map))
|
|
|
|
|
|
def forward(self, input_ids, attention_mask):
|
|
|
outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
|
|
|
pooled_output = outputs.pooler_output
|
|
|
|
|
|
x = pooled_output
|
|
|
for layer in self.layers:
|
|
|
x = layer(x)
|
|
|
|
|
|
reg_output = self.reg_head(x).squeeze(-1)
|
|
|
classes_outputs = {}
|
|
|
for task, head in self.classification_heads.items():
|
|
|
classes_outputs[task] = head(x)
|
|
|
|
|
|
return reg_output, classes_outputs
|
|
|
|
|
|
def model_unique_name(self) -> str:
|
|
|
return self.name
|
|
|
|
