|
import torch |
|
from transformers import AutoTokenizer, AutoModelForSequenceClassification |
|
from src.models import ModernBertForSentiment |
|
from transformers import ModernBertConfig |
|
from typing import Dict, Any |
|
import yaml |
|
import os |
|
|
|
|
|
class SentimentInference: |
|
def __init__(self, config_path: str = "config.yaml"): |
|
"""Load configuration and initialize model and tokenizer.""" |
|
with open(config_path, 'r') as f: |
|
config = yaml.safe_load(f) |
|
|
|
model_cfg = config.get('model', {}) |
|
inference_cfg = config.get('inference', {}) |
|
|
|
|
|
model_weights_path = inference_cfg.get('model_path', |
|
os.path.join(model_cfg.get('output_dir', 'checkpoints'), 'best_model.pt')) |
|
|
|
|
|
|
|
base_model_name = model_cfg.get('name', 'answerdotai/ModernBERT-base') |
|
|
|
self.max_length = inference_cfg.get('max_length', model_cfg.get('max_length', 256)) |
|
|
|
|
|
print(f"Loading tokenizer from: {base_model_name}") |
|
self.tokenizer = AutoTokenizer.from_pretrained(base_model_name) |
|
|
|
|
|
print(f"Loading ModernBertConfig from: {base_model_name}") |
|
bert_config = ModernBertConfig.from_pretrained(base_model_name) |
|
|
|
|
|
|
|
|
|
|
|
bert_config.classifier_dropout = model_cfg.get('dropout', bert_config.classifier_dropout) |
|
|
|
|
|
|
|
|
|
|
|
|
|
bert_config.pooling_strategy = model_cfg.get('pooling_strategy', 'cls') |
|
bert_config.num_weighted_layers = model_cfg.get('num_weighted_layers', 4) |
|
bert_config.loss_function = model_cfg.get('loss_function', {'name': 'SentimentWeightedLoss', 'params': {}}) |
|
|
|
bert_config.num_labels = 1 |
|
|
|
print("Instantiating ModernBertForSentiment model structure...") |
|
self.model = ModernBertForSentiment(bert_config) |
|
|
|
print(f"Loading model weights from local checkpoint: {model_weights_path}") |
|
|
|
checkpoint = torch.load(model_weights_path, map_location=torch.device('cpu')) |
|
|
|
|
|
|
|
if 'model_state_dict' in checkpoint: |
|
model_state_to_load = checkpoint['model_state_dict'] |
|
else: |
|
|
|
model_state_to_load = checkpoint |
|
|
|
self.model.load_state_dict(model_state_to_load) |
|
self.model.eval() |
|
print("Model loaded successfully.") |
|
|
|
def predict(self, text: str) -> Dict[str, Any]: |
|
inputs = self.tokenizer(text, return_tensors="pt", truncation=True, max_length=self.max_length) |
|
with torch.no_grad(): |
|
outputs = self.model(input_ids=inputs['input_ids'], attention_mask=inputs['attention_mask']) |
|
logits = outputs["logits"] |
|
prob = torch.sigmoid(logits).item() |
|
return {"sentiment": "positive" if prob > 0.5 else "negative", "confidence": prob} |