File size: 6,392 Bytes
e96bc27
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
import os
import json
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer, AutoConfig

# ๋ชจ๋ธ๊ณผ ํ† ํฌ๋‚˜์ด์ €๋ฅผ ๋กœ๋“œํ•˜๋Š” ํ•จ์ˆ˜
def model_fn(model_dir):
    """

    SageMaker๊ฐ€ ๋ชจ๋ธ์„ ๋กœ๋“œํ•˜๊ธฐ ์œ„ํ•ด ํ˜ธ์ถœํ•˜๋Š” ํ•จ์ˆ˜

    

    Args:

        model_dir (str): ๋ชจ๋ธ ํŒŒ์ผ์ด ์ €์žฅ๋œ ๋””๋ ‰ํ† ๋ฆฌ ๊ฒฝ๋กœ

        

    Returns:

        dict: ๋ชจ๋ธ, ํ† ํฌ๋‚˜์ด์ €, ์„ค์ • ๋“ฑ์„ ํฌํ•จํ•œ ๋”•์…”๋„ˆ๋ฆฌ

    """
    # ํ™˜๊ฒฝ ๋ณ€์ˆ˜ ์„ค์ • (์„ ํƒ ์‚ฌํ•ญ)
    os.environ["TOKENIZERS_PARALLELISM"] = "false"
    
    # ์„ค์ • ํŒŒ์ผ ๋กœ๋“œ
    config_path = os.path.join(model_dir, "config.json")
    config = AutoConfig.from_pretrained(config_path)
    
    print(f"Loading model from {model_dir}")
    print(f"Device: {'cuda' if torch.cuda.is_available() else 'cpu'}")
    
    # ๋ ˆ์ด๋ธ” ๋งคํ•‘ ๋กœ๋“œ (์žˆ๋Š” ๊ฒฝ์šฐ)
    label_map = {}
    label_map_path = os.path.join(model_dir, "label_map.json")
    if os.path.exists(label_map_path):
        with open(label_map_path, 'r', encoding='utf-8') as f:
            label_map = json.load(f)
        print(f"Loaded label map from {label_map_path}")
    else:
        print("No label map found. Using numeric indices as labels.")
    
    # ๋ชจ๋ธ ๋กœ๋“œ
    model = AutoModelForSequenceClassification.from_pretrained(
        model_dir, 
        config=config,
        torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
    )
    
    # GPU ์‚ฌ์šฉ ๊ฐ€๋Šฅํ•œ ๊ฒฝ์šฐ ๋ชจ๋ธ์„ GPU๋กœ ์ด๋™
    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
    model = model.to(device)
    model.eval()
    
    # ํ† ํฌ๋‚˜์ด์ € ๋กœ๋“œ
    tokenizer = AutoTokenizer.from_pretrained(model_dir)
    
    return {
        "model": model,
        "tokenizer": tokenizer,
        "config": config,
        "device": device,
        "label_map": label_map
    }

# ์ž…๋ ฅ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ํ•จ์ˆ˜
def input_fn(request_body, request_content_type):
    """

    SageMaker๊ฐ€ ์š”์ฒญ ๋ฐ์ดํ„ฐ๋ฅผ ์ฒ˜๋ฆฌํ•˜๊ธฐ ์œ„ํ•ด ํ˜ธ์ถœํ•˜๋Š” ํ•จ์ˆ˜

    

    Args:

        request_body: ์š”์ฒญ ๋ณธ๋ฌธ ๋ฐ์ดํ„ฐ

        request_content_type (str): ์š”์ฒญ ์ฝ˜ํ…์ธ  ํƒ€์ž…

        

    Returns:

        dict: ์ฒ˜๋ฆฌ๋œ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ

    """
    if request_content_type == "application/json":
        input_data = json.loads(request_body)
        
        # ๋ฌธ์ž์—ด์ธ ๊ฒฝ์šฐ ํ…์ŠคํŠธ๋กœ ์ฒ˜๋ฆฌ
        if isinstance(input_data, str):
            return {"text": input_data}
        
        return input_data
    
    elif request_content_type == "text/plain":
        # ์ผ๋ฐ˜ ํ…์ŠคํŠธ ์ฒ˜๋ฆฌ
        return {"text": request_body.decode('utf-8')}
    
    else:
        raise ValueError(f"์ง€์›๋˜์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ  ํƒ€์ž…: {request_content_type}")

# ์˜ˆ์ธก ํ•จ์ˆ˜
def predict_fn(input_data, model_dict):
    """

    SageMaker๊ฐ€ ๋ชจ๋ธ ์˜ˆ์ธก์„ ์ˆ˜ํ–‰ํ•˜๊ธฐ ์œ„ํ•ด ํ˜ธ์ถœํ•˜๋Š” ํ•จ์ˆ˜

    

    Args:

        input_data (dict): ์ฒ˜๋ฆฌ๋œ ์ž…๋ ฅ ๋ฐ์ดํ„ฐ

        model_dict (dict): model_fn์—์„œ ๋ฐ˜ํ™˜ํ•œ ๋ชจ๋ธ ์ •๋ณด

        

    Returns:

        dict: ์˜ˆ์ธก ๊ฒฐ๊ณผ

    """
    model = model_dict["model"]
    tokenizer = model_dict["tokenizer"]
    device = model_dict["device"]
    label_map = model_dict["label_map"]
    
    # ์ž…๋ ฅ ํ…์ŠคํŠธ ๊ฐ€์ ธ์˜ค๊ธฐ
    if "text" in input_data:
        text = input_data["text"]
    else:
        raise ValueError("์ž…๋ ฅ ๋ฐ์ดํ„ฐ์— 'text' ํ•„๋“œ๊ฐ€ ์—†์Šต๋‹ˆ๋‹ค")
    
    # ํ† ํฐํ™” ์˜ต์…˜
    max_length = input_data.get("max_length", 512)
    padding = input_data.get("padding", "max_length")
    truncation = input_data.get("truncation", True)
    
    # ํ† ํฐํ™”
    inputs = tokenizer(
        text,
        return_tensors="pt",
        padding=padding,
        truncation=truncation,
        max_length=max_length
    )
    
    # ์ž…๋ ฅ ํ…์„œ๋ฅผ ๋””๋ฐ”์ด์Šค๋กœ ์ด๋™
    inputs = {name: tensor.to(device) for name, tensor in inputs.items()}
    
    # ๋ชจ๋ธ ์ถ”๋ก 
    with torch.no_grad():
        outputs = model(**inputs)
        logits = outputs.logits
        probabilities = torch.softmax(logits, dim=1)
    
    # ์ด์ง„ ๋ถ„๋ฅ˜ ๋ชจ๋ธ์ธ ๊ฒฝ์šฐ (ํด๋ž˜์Šค ์ˆ˜๊ฐ€ 2์ธ ๊ฒฝ์šฐ)
    if logits.shape[1] == 2:
        positive_prob = probabilities[0, 1].item()
        negative_prob = probabilities[0, 0].item()
        prediction = 1 if positive_prob > 0.5 else 0
        
        result = {
            "prediction": prediction,
            "positive_probability": positive_prob,
            "negative_probability": negative_prob
        }
        
        # ๋ ˆ์ด๋ธ” ๋งคํ•‘์ด ์žˆ๋Š” ๊ฒฝ์šฐ ๋ ˆ์ด๋ธ” ์ถ”๊ฐ€
        if label_map:
            pred_label = str(prediction)
            if pred_label in label_map:
                result["label"] = label_map[pred_label]
    
    # ๋‹ค์ค‘ ํด๋ž˜์Šค ๋ชจ๋ธ์ธ ๊ฒฝ์šฐ
    else:
        predictions = torch.argmax(probabilities, dim=1).cpu().numpy().tolist()
        probabilities = probabilities.cpu().numpy().tolist()[0]
        
        result = {
            "prediction": predictions[0],
            "probabilities": probabilities,
        }
        
        # ๋ ˆ์ด๋ธ” ๋งคํ•‘์ด ์žˆ๋Š” ๊ฒฝ์šฐ ๋ ˆ์ด๋ธ” ์ถ”๊ฐ€
        if label_map:
            pred_label = str(predictions[0])
            if pred_label in label_map:
                result["label"] = label_map[pred_label]
            
            # ๋ชจ๋“  ๋ ˆ์ด๋ธ”์— ๋Œ€ํ•œ ํ™•๋ฅ  ๋งคํ•‘ ์ถ”๊ฐ€
            result["label_probabilities"] = {
                label_map.get(str(idx), str(idx)): prob 
                for idx, prob in enumerate(probabilities)
            }
    
    return result

# ์ถœ๋ ฅ ๋ฐ์ดํ„ฐ ์ฒ˜๋ฆฌ ํ•จ์ˆ˜
def output_fn(prediction, response_content_type):
    """

    SageMaker๊ฐ€ ์˜ˆ์ธก ๊ฒฐ๊ณผ๋ฅผ ์‘๋‹ต ํ˜•์‹์œผ๋กœ ๋ณ€ํ™˜ํ•˜๊ธฐ ์œ„ํ•ด ํ˜ธ์ถœํ•˜๋Š” ํ•จ์ˆ˜

    

    Args:

        prediction: predict_fn์—์„œ ๋ฐ˜ํ™˜ํ•œ ์˜ˆ์ธก ๊ฒฐ๊ณผ

        response_content_type (str): ์›ํ•˜๋Š” ์‘๋‹ต ์ฝ˜ํ…์ธ  ํƒ€์ž…

        

    Returns:

        str: ์ง๋ ฌํ™”๋œ ์˜ˆ์ธก ๊ฒฐ๊ณผ

    """
    if response_content_type == "application/json":
        return json.dumps(prediction, ensure_ascii=False)
    else:
        raise ValueError(f"์ง€์›๋˜์ง€ ์•Š๋Š” ์ฝ˜ํ…์ธ  ํƒ€์ž…: {response_content_type}")