File size: 11,978 Bytes
498ffec
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
from pathlib import Path
import multiprocessing
import logging
from PIL import Image
import io
import base64
import numpy as np
import gymnasium as gym
import os

from agent.checklist import generate_checklist
from agent.reward import get_ar_reward

from browser_agent import BrowserAgent


logger = logging.getLogger(__name__)
logger.setLevel('INFO')

templates_dir = Path(__file__).parent / "templates"
CSS_RM_CARDS: str = (templates_dir / "rm_cards.css").read_text()
CSS_TRAJECTORY: str = (templates_dir / "trajectory.css").read_text()
CARD_HTML_TEMPLATE: str = (templates_dir / "card.html").read_text()

RM_BASE_URL = os.environ['RM_BASE_URL']
RM_MODEL_NAME = os.environ['RM_MODEL_NAME']

def return_state(state, screenshot=None):
    return state, None, None, screenshot, None

def run_agent(instruction: str, model_name: str = "gpt-4o", start_url: str = "about:blank",
              use_html: bool = False, use_axtree: bool = True, use_screenshot: bool = False, max_steps: int = 20):
    logger.info(f"Starting agent with instruction: {instruction}")
    logger.info(f"Configuration: model={model_name}, start_url={start_url}")
    
    trajectory = []
    trajectory_str = ''
    agent = BrowserAgent(
        model_name=model_name,
        use_html=use_html,
        use_axtree=use_axtree,
        use_screenshot=use_screenshot
    )

    # Initialize BrowserGym environment
    logger.info("Initializing BrowserGym environment")
    yield return_state("## Initializing BrowserGym environment...", None)
    env = gym.make(
        "browsergym/openended",
        task_kwargs={
            "start_url": start_url,
            "goal": instruction,
        },
        wait_for_user_message=True
    )
    obs, info = env.reset()
    logger.info("Environment initialized")

    # Send user instruction to the environment
    logger.info("Sending user instruction to environment")
    obs, reward, terminated, truncated, info = env.step({
        "type": "send_msg_to_user",
        "message": instruction
    })
    processed_obs = agent.obs_preprocessor(obs)
    logger.info(f"Obs: {processed_obs.keys()}")
    logger.info(f"axtree_txt: {processed_obs['axtree_txt']}")

    yield return_state("## Generating checklist...", obs['som_screenshot'])
    checklist = generate_checklist(intent=instruction, start_url=start_url, text_observation=processed_obs['axtree_txt'])

    # yield initial state
    current_screenshot = obs['som_screenshot'].copy()
    yield "## Rollout actions from policy...", checklist, [], current_screenshot, trajectory.copy()

    try:
        step_count = 0
        while step_count < max_steps:
            logger.info(f"Step {step_count}: Getting next action")
            # Get next action from agent
            candidates, _ = agent.get_action(processed_obs)

            yield return_state(f"## Rewarding actions...", current_screenshot)

            total_rewards, total_thoughts = get_ar_reward(
                dataset=[
                    {
                        'text_observation': processed_obs['axtree_txt'],
                        'intent': instruction,
                        'trajectory': trajectory_str,
                        'current_url': processed_obs['open_pages_urls'][processed_obs['active_page_index'][0]],
                        'checklist': checklist,
                        'thought': cand['thought'],
                        'action': cand['action'],
                    } for cand in candidates
                ],
                base_url=RM_BASE_URL,
                model_name=RM_MODEL_NAME,
            )

            # process rewards
            diff_reward = abs(max(total_rewards) - total_rewards[0])  # reward difference between actions with the highest reward and the most frequent.
            if diff_reward <= 0.01:
                logger.info(f"diff_reward: {diff_reward} -> most frequent action")
                max_index = 0  # most frequent action
            else:
                logger.info(f"diff_reward: {diff_reward} -> highest reward")
                max_index = total_rewards.index(max(total_rewards))  # highest reward

            # sort by reward
            sorted_indices = sorted(list(enumerate(total_rewards)), key=lambda x: (-1 if x[0] == max_index else 0, -x[1]))
            new_order = [idx for idx, _ in sorted_indices]
            candidates = [candidates[idx] for idx in new_order]
            total_rewards = [total_rewards[idx] for idx in new_order]
            total_thoughts = [total_thoughts[idx] for idx in new_order]

            best_cand = candidates[0]

            agent.action_history.append(best_cand['response'])

            action = best_cand['action']

            # processing action
            step_info = {
                'thought': best_cand['thought'],
                'action': action
            }
            current_cards = [{'thought': cand['thought'], 'action': cand['action'], 'feedback': feedback, 'reward': round(reward, 2)} for idx, (cand, reward, feedback) in enumerate(zip(candidates, total_rewards, total_thoughts))]

            trajectory_str += f'THOUGHT {step_count+1}: {step_info["thought"]}\nACTION {step_count+1}: {step_info["action"]}\n\n'
            
            # Execute action
            logger.info(f"Step {step_count}: Executing action: {action}")
            yield f"## Executing action: {action}", checklist, current_cards, current_screenshot, trajectory.copy()
            if action.startswith('send_msg_to_user'):
                terminated = True
                truncated = False
            else:
                obs, reward, terminated, truncated, info = env.step(action)
            trajectory.append((processed_obs['som_screenshot'], [{'action': cand['action'], 'reward': round(reward, 2)} for cand, reward in zip(candidates, total_rewards)]))
            processed_obs = agent.obs_preprocessor(obs)
            current_screenshot = processed_obs['som_screenshot'].copy()

            while '\n\n' in step_info['thought']:
                step_info['thought'] = step_info['thought'].replace('\n\n', '\n')
            
            # trajectory에 numpy array 직접 저장
            logger.info(f"Step {step_count}: Saved screenshot and updated trajectory")
            step_count += 1

            # yield by each step
            yield "## Rollout actions from policy...", checklist, current_cards, current_screenshot, trajectory.copy()

            if terminated or truncated:
                logger.info(f"Episode ended: terminated={terminated}, truncated={truncated}")
                yield return_state("## Episode ended", current_screenshot)
                break

    finally:
        logger.info("Finished")


def run_agent_worker(instruction, model_name, start_url, use_html, use_axtree, use_screenshot, max_steps, return_queue):
    """Worker function that runs the agent in a separate process and puts results in a queue."""
    try:
        for result in run_agent(instruction, model_name, start_url, use_html, use_axtree, use_screenshot, max_steps):
            return_queue.put(result)
    except Exception as e:
        logger.error(f"Error in agent worker process: {e}")
        return_queue.put(("Error occurred in agent process", [], None, []))
        import traceback
        traceback.print_exc()
    finally:
        # Signal that the process is done
        return_queue.put(None)

def run_agent_wrapper(instruction, model_name="gpt-4o", start_url="about:blank",
                     use_html=False, use_axtree=True, use_screenshot=False, max_steps=20):
    """Wrapper function that runs the agent in a separate process and yields results."""
    return_queue = multiprocessing.Queue()
    
    # Start the agent in a separate process
    p = multiprocessing.Process(
        target=run_agent_worker, 
        args=(instruction, model_name, start_url, use_html, use_axtree, use_screenshot, max_steps, return_queue)
    )
    p.daemon = True  # Ensure process terminates when parent terminates
    p.start()
    
    # Get results from the queue and yield them
    while True:
        result = return_queue.get()
        if result is None:  # End signal
            break
        yield result
    
    # Clean up
    if p.is_alive():
        p.terminate()
    p.join()

def process_run(instruction, model_name, start_url):
    # Use the wrapper function instead of directly calling run_agent
    trajectory_generator = run_agent_wrapper(
        instruction, 
        model_name, 
        start_url,
        use_html=False,
        use_axtree=True,
        use_screenshot=False
    )

    all_trajectory = []
    last_checklist_view, last_trajectory_html = None, None

    for state, checklist_view, rm_cards, screenshot, trajectory in trajectory_generator:
        if checklist_view is None:
            yield state, screenshot, last_checklist_view, None, last_trajectory_html
            continue
        # Create HTML for reward model cards
        rm_cards_html = f"""
        <style>
            {CSS_RM_CARDS}
        </style>
        <div class="rm-cards-container">
        """
        
        for idx, card in enumerate(rm_cards):
            rm_cards_html += CARD_HTML_TEMPLATE.format(
                additional_class='top-candidate' if idx == 0 else '',
                k=idx+1,
                suffix='(best)' if idx == 0 else '',
                thought=card['thought'],
                action=card['action'],
                reward=card['reward'],
                feedback=card['feedback']
            )
        
        rm_cards_html += "</div>"
        all_trajectory = trajectory
        
        # Create HTML for trajectory display
        trajectory_html = f"""
        <style>
            {CSS_TRAJECTORY}
        </style>
        <div class="trajectory-container">
        """
        
        for idx, (after_img, cands) in enumerate(all_trajectory):
            # Convert image to base64 if needed
            img = all_trajectory[idx][0]
            if isinstance(img, np.ndarray):
                img = Image.fromarray(img)
            if isinstance(img, Image.Image):
                buffer = io.BytesIO()
                img.save(buffer, format="JPEG")
                img_str = base64.b64encode(buffer.getvalue()).decode()
                img_src = f"data:image/jpeg;base64,{img_str}"
            else:
                img_src = img
            
            trajectory_html += f"""
            <div class="step-container">
                <div class="step-header">Step {idx + 1}</div>
                <div class="step-content">
                    <div class="step-image">
                        <img src="{img_src}" alt="Browser state">
                    </div>
                    <div class="step-info">
                        <div class="box-title">Action Candidates:</div>
                        <div class="action-candidates">
            """
            
            # Display all candidates for this step
            for i, cand in enumerate(cands):
                action = cand['action']
                reward = cand['reward']
                
                trajectory_html += f"""
                <div class="candidate-box{' selected' if i == 0 else ''}">
                    <div class="box-title">
                        Action {i+1}{' (Selected)' if i == 0 else ''}
                        <span class="reward-text">Reward: {reward}</span>
                    </div>
                    <pre>{action}</pre>
                </div>
                """
            
            trajectory_html += """
                        </div>
                    </div>
                </div>
            </div>
            """
        
        trajectory_html += "</div>" 
        
        last_checklist_view, last_trajectory_html = checklist_view, trajectory_html
        yield state, screenshot, last_checklist_view, rm_cards_html, last_trajectory_html
    yield state, screenshot, last_checklist_view, rm_cards_html, last_trajectory_html