File size: 8,548 Bytes
c39c017
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
#!/usr/bin/env python3
"""
GUI Task Evaluation Scripts for GUI-Shift.

Evaluates on:
- AndroidControl (Low / High)
- GUI Odyssey
- ScreenSpot-v2
- ScreenSpot-Pro

Metrics:
- TM: Type Match (correct action type)
- EM: Exact Match (correct type + all parameters)

From: GUI-Shift paper Section 4 (arXiv:2505.12493)
"""

import argparse
import json
import os
from pathlib import Path
from typing import Dict, Any, List, Tuple, Optional
from collections import defaultdict

import torch
from transformers import AutoModelForVision2Seq, AutoProcessor
from PIL import Image


def load_model(model_path: str, device: str = "cuda"):
    """Load trained GUI-Shift model."""
    processor = AutoProcessor.from_pretrained(model_path, trust_remote_code=True)
    model = AutoModelForVision2Seq.from_pretrained(
        model_path,
        torch_dtype=torch.bfloat16,
        device_map=device,
        trust_remote_code=True,
    )
    model.eval()
    return model, processor


def parse_predicted_action(text: str) -> Optional[Dict[str, Any]]:
    """Parse action from model output."""
    import re
    match = re.search(r'<answer>(.*?)</answer>', text, re.DOTALL)
    if not match:
        return None
    
    content = match.group(1).strip()
    try:
        return json.loads(content)
    except json.JSONDecodeError:
        # Fallback regex
        action_type_match = re.search(r'"action_type"\s*:\s*"([^"]+)"', content)
        if action_type_match:
            action = {"action_type": action_type_match.group(1)}
            for key in ["x", "y", "direction", "app_name", "text"]:
                match = re.search(rf'"{key}"\s*:\s*(?:"([^"]+)"|(\d+))', content)
                if match:
                    val = match.group(1) or int(match.group(2))
                    if key in ["x", "y"]:
                        val = int(val)
                    action[key] = val
            return action
    return None


def type_match(pred: Optional[Dict], gt: Dict) -> bool:
    """Check if predicted action type matches ground truth."""
    if not pred:
        return False
    return pred.get("action_type") == gt.get("action_type")


def exact_match(pred: Optional[Dict], gt: Dict) -> bool:
    """Check if predicted action exactly matches ground truth."""
    if not type_match(pred, gt):
        return False
    
    action_type = gt.get("action_type", "")
    
    if action_type in ["click", "long_press"]:
        bbox = gt.get("bbox", [0, 0, 0, 0])
        x = pred.get("x", 0)
        y = pred.get("y", 0)
        if bbox and len(bbox) >= 4:
            return bbox[0] <= x <= bbox[2] and bbox[1] <= y <= bbox[3]
        if "x" in gt and "y" in gt:
            tolerance = 20
            return abs(x - gt["x"]) <= tolerance and abs(y - gt["y"]) <= tolerance
        return False
    
    elif action_type == "scroll":
        return pred.get("direction") == gt.get("direction")
    
    elif action_type == "open_app":
        return pred.get("app_name") == gt.get("app_name")
    
    elif action_type == "input_text":
        return pred.get("text") == gt.get("text")
    
    elif action_type in ["navigate_back", "navigate_home", "wait"]:
        return True
    
    return False


def evaluate_sample(
    model,
    processor,
    sample: Dict[str, Any],
    device: str = "cuda",
) -> Tuple[bool, bool, str]:
    """Evaluate a single sample. Returns (type_match, exact_match, prediction_text)."""
    image_paths = sample.get("image_path", sample.get("image", []))
    problem = sample.get("problem", sample.get("instruction", ""))
    ground_truth = sample.get("ground_truth_action", sample.get("action", {}))
    
    # Load images
    images = []
    for img_path in image_paths:
        if isinstance(img_path, str) and os.path.exists(img_path):
            images.append(Image.open(img_path).convert("RGB"))
    
    # Build prompt
    messages = [
        {"role": "user", "content": [{"type": "image"} for _ in images] + [{"type": "text", "text": problem}]}
    ]
    
    text = processor.apply_chat_template(messages, add_generation_prompt=True)
    inputs = processor(text=text, images=images, return_tensors="pt", padding=True).to(model.device)
    
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=256,
            do_sample=False,
        )
    
    generated_text = processor.batch_decode(
        outputs[:, inputs["input_ids"].shape[1]:],
        skip_special_tokens=True,
    )[0]
    
    pred_action = parse_predicted_action(generated_text)
    
    tm = type_match(pred_action, ground_truth)
    em = exact_match(pred_action, ground_truth)
    
    return tm, em, generated_text


def evaluate_dataset(
    model,
    processor,
    dataset_path: str,
    device: str = "cuda",
    output_path: Optional[str] = None,
) -> Dict[str, float]:
    """Evaluate a full dataset and compute metrics."""
    samples = []
    with open(dataset_path, "r") as f:
        for line in f:
            if line.strip():
                samples.append(json.loads(line))
    
    results = {
        "total": len(samples),
        "type_match": 0,
        "exact_match": 0,
        "by_type": defaultdict(lambda: {"total": 0, "tm": 0, "em": 0}),
        "details": [],
    }
    
    for i, sample in enumerate(samples):
        print(f"  Evaluating {i+1}/{len(samples)}...", end="\r")
        
        tm, em, pred_text = evaluate_sample(model, processor, sample, device)
        
        gt = sample.get("ground_truth_action", sample.get("action", {}))
        action_type = gt.get("action_type", "unknown")
        
        results["type_match"] += int(tm)
        results["exact_match"] += int(em)
        results["by_type"][action_type]["total"] += 1
        results["by_type"][action_type]["tm"] += int(tm)
        results["by_type"][action_type]["em"] += int(em)
        
        results["details"].append({
            "id": sample.get("id", i),
            "type_match": tm,
            "exact_match": em,
            "predicted": pred_text,
            "ground_truth": gt,
        })
    
    # Compute percentages
    results["type_match_pct"] = 100.0 * results["type_match"] / len(samples) if samples else 0
    results["exact_match_pct"] = 100.0 * results["exact_match"] / len(samples) if samples else 0
    
    for action_type, counts in results["by_type"].items():
        counts["tm_pct"] = 100.0 * counts["tm"] / counts["total"] if counts["total"] else 0
        counts["em_pct"] = 100.0 * counts["em"] / counts["total"] if counts["total"] else 0
    
    print(f"\n  TM: {results['type_match_pct']:.1f}% | EM: {results['exact_match_pct']:.1f}%")
    
    if output_path:
        with open(output_path, "w") as f:
            json.dump(results, f, indent=2)
        print(f"  Results saved to {output_path}")
    
    return results


def main():
    parser = argparse.ArgumentParser(description="Evaluate GUI-Shift model on GUI benchmarks")
    parser.add_argument("--model_path", type=str, required=True, help="Path to trained model")
    parser.add_argument("--dataset", type=str, required=True, help="Path to evaluation dataset (JSONL)")
    parser.add_argument("--output", type=str, default="evaluation_results.json", help="Output results file")
    parser.add_argument("--device", type=str, default="cuda", help="Device for inference")
    parser.add_argument("--benchmark", type=str, default="androidcontrol", 
                       choices=["androidcontrol_low", "androidcontrol_high", "gui_odyssey", 
                               "screenspot_v2", "screenspot_pro"],
                       help="Benchmark name")
    args = parser.parse_args()
    
    print(f"Loading model from {args.model_path}...")
    model, processor = load_model(args.model_path, args.device)
    
    print(f"Evaluating on {args.benchmark}...")
    results = evaluate_dataset(model, processor, args.dataset, args.device, args.output)
    
    print("\n=== Final Results ===")
    print(f"Benchmark: {args.benchmark}")
    print(f"Total samples: {results['total']}")
    print(f"Type Match (TM): {results['type_match']}/{results['total']} = {results['type_match_pct']:.2f}%")
    print(f"Exact Match (EM): {results['exact_match']}/{results['total']} = {results['exact_match_pct']:.2f}%")
    
    print("\nPer-action breakdown:")
    for action_type, counts in sorted(results["by_type"].items()):
        print(f"  {action_type:20s}: TM={counts['tm_pct']:.1f}% EM={counts['em_pct']:.1f}% ({counts['tm']}/{counts['total']})")


if __name__ == "__main__":
    main()