Create handler.py
Browse files- handler.py +40 -0
handler.py
ADDED
@@ -0,0 +1,40 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from transformers import AutoTokenizer, AutoModelForSeq2SeqLM
|
2 |
+
import torch
|
3 |
+
|
4 |
+
class ModelHandler:
|
5 |
+
def __init__(self, model_dir):
|
6 |
+
# Cargar el modelo y el tokenizador desde el directorio del modelo
|
7 |
+
self.tokenizer = AutoTokenizer.from_pretrained(model_dir)
|
8 |
+
self.model = AutoModelForSeq2SeqLM.from_pretrained(model_dir)
|
9 |
+
self.model.eval() # Configurar el modelo en modo de evaluaci贸n
|
10 |
+
|
11 |
+
def preprocess(self, data):
|
12 |
+
# Preprocesamiento de la entrada
|
13 |
+
if isinstance(data, dict) and "input_text" in data:
|
14 |
+
input_text = data["input_text"]
|
15 |
+
else:
|
16 |
+
raise ValueError("Esperando un diccionario con la clave 'inputs'")
|
17 |
+
|
18 |
+
# Tokenizaci贸n de la entrada
|
19 |
+
tokens = self.tokenizer(input_text, return_tensors="pt", truncation=True, padding=True)
|
20 |
+
return tokens
|
21 |
+
|
22 |
+
def inference(self, tokens):
|
23 |
+
# Realizar la inferencia
|
24 |
+
with torch.no_grad():
|
25 |
+
outputs = self.model(**tokens)
|
26 |
+
# Obtener las predicciones y aplicar softmax para probabilidades
|
27 |
+
probabilities = torch.softmax(outputs.logits, dim=-1)
|
28 |
+
return probabilities
|
29 |
+
|
30 |
+
def postprocess(self, probabilities):
|
31 |
+
# Postprocesamiento para devolver la salida en formato JSON
|
32 |
+
predictions = torch.argmax(probabilities, dim=-1)
|
33 |
+
return {"predictions": predictions.tolist(), "probabilities": probabilities.tolist()}
|
34 |
+
|
35 |
+
def __call__(self, data):
|
36 |
+
# Llamada principal del handler para procesamiento completo
|
37 |
+
tokens = self.preprocess(data)
|
38 |
+
probabilities = self.inference(tokens)
|
39 |
+
result = self.postprocess(probabilities)
|
40 |
+
return result
|