File size: 2,190 Bytes
28f1640 9015082 28f1640 9015082 28f1640 9015082 28f1640 9015082 28f1640 9015082 28f1640 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 | """
Custom HuggingFace-compatible wrapper for the DeBERTa-v3-small MCQ scorer
trained in the Kaggle notebook (DeBERTaMCQModel).
This makes the model loadable on the Hub via:
AutoConfig.from_pretrained(repo_id, trust_remote_code=True)
AutoModel.from_pretrained(repo_id, trust_remote_code=True)
"""
import torch
import torch.nn as nn
from transformers import PretrainedConfig, PreTrainedModel, AutoModel
class DebertaMCQConfig(PretrainedConfig):
model_type = "deberta_mcq"
def __init__(
self,
base_model_name="microsoft/deberta-v3-small",
hidden_dropout=0.3,
num_options=5,
max_len=64,
**kwargs,
):
self.base_model_name = base_model_name
self.hidden_dropout = hidden_dropout
self.num_options = num_options
self.max_len = max_len
super().__init__(**kwargs)
class DebertaMCQForMultipleChoice(PreTrainedModel):
config_class = DebertaMCQConfig
def __init__(self, config: DebertaMCQConfig):
super().__init__(config)
self.deberta = AutoModel.from_pretrained(
config.base_model_name,
torch_dtype=torch.float32,
low_cpu_mem_usage=False,
device_map=None,
)
hidden_size = self.deberta.config.hidden_size
self.classifier = nn.Sequential(
nn.Dropout(config.hidden_dropout),
nn.Linear(hidden_size, 1),
)
self.post_init()
def forward(self, input_ids, attention_mask=None, labels=None):
# input_ids / attention_mask shape: (batch, num_options, seq_len)
batch_size, num_options, seq_len = input_ids.shape
input_ids = input_ids.view(-1, seq_len)
attention_mask = attention_mask.view(-1, seq_len)
outputs = self.deberta(input_ids=input_ids, attention_mask=attention_mask)
cls_output = outputs.last_hidden_state[:, 0, :].float()
scores = self.classifier(cls_output).view(batch_size, num_options)
loss = None
if labels is not None:
loss = nn.CrossEntropyLoss()(scores, labels)
return {"loss": loss, "logits": scores} if loss is not None else {"logits": scores} |