Create handler.py
Browse files- handler.py +44 -0
handler.py
ADDED
@@ -0,0 +1,44 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from typing import Dict, List, Any
|
2 |
+
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer, pipeline
|
3 |
+
import torch
|
4 |
+
|
5 |
+
class EndpointHandler:
|
6 |
+
def __init__(self, path=""):
|
7 |
+
# Load model and processor from path
|
8 |
+
self.model = AutoModelForSeq2SeqLM.from_pretrained(path)
|
9 |
+
self.tokenizer = AutoTokenizer.from_pretrained(path)
|
10 |
+
|
11 |
+
def __call__(self, data: Dict[str, Any]) -> Dict[str, str]:
|
12 |
+
"""
|
13 |
+
Args:
|
14 |
+
data (:obj:):
|
15 |
+
Includes the deserialized image file as PIL.Image
|
16 |
+
"""
|
17 |
+
# Process input
|
18 |
+
inputs = data.pop("inputs", data)
|
19 |
+
parameters = data.pop("parameters", None)
|
20 |
+
|
21 |
+
# Preprocess
|
22 |
+
input_ids = self.tokenizer(inputs, return_tensors="pt").input_ids
|
23 |
+
|
24 |
+
# Modify parameters to increase max_length
|
25 |
+
if parameters is None:
|
26 |
+
parameters = {}
|
27 |
+
parameters['max_length'] = 512 # Set your desired max_length here
|
28 |
+
parameters['min_length'] = 100
|
29 |
+
parameters['length_penalty'] = 2.0
|
30 |
+
parameters['num_beams'] = 10
|
31 |
+
parameters['early_stopping'] = True
|
32 |
+
parameters['temperature'] = 0.0
|
33 |
+
parameters['top_k'] = 15
|
34 |
+
parameters['top_p'] = 0.8
|
35 |
+
|
36 |
+
|
37 |
+
|
38 |
+
# Generate output
|
39 |
+
outputs = self.model.generate(input_ids, **parameters)
|
40 |
+
|
41 |
+
# Postprocess the prediction
|
42 |
+
prediction = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
|
43 |
+
|
44 |
+
return [{"generated_text": prediction}]
|