File size: 13,005 Bytes
0f43f8a
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import pandas as pd

from eval.evaluator import Eval
from dataset.base_dataset import DatasetBase
from utils.llm_chain import MetaChain
from estimator import give_estimator
from pathlib import Path
import pickle
import os
import json
import logging
import wandb


class OptimizationPipeline:
    """
    The main pipeline for optimization. The pipeline is composed of 4 main components:
    1. dataset - The dataset handle the data including the annotation and the prediction
    2. annotator - The annotator is responsible generate the GT
    3. predictor - The predictor is responsible to generate the prediction
    4. eval - The eval is responsible to calculate the score and the large errors
    """

    def __init__(self, config, task_description: str = None, initial_prompt: str = None, output_path: str = ''):
        """
        Initialize a new instance of the ClassName class.
        :param config: The configuration file (EasyDict)
        :param task_description: Describe the task that needed to be solved
        :param initial_prompt: Provide an initial prompt to solve the task
        :param output_path: The output dir to save dump, by default the dumps are not saved
        """

        if config.use_wandb:  # In case of using W&B
            wandb.login()
            self.wandb_run = wandb.init(
                project="AutoGPT",
                config=config,
            )
        if output_path == '':
            self.output_path = None
        else:
            if not os.path.isdir(output_path):
                os.makedirs(output_path)
            self.output_path = Path(output_path)
            logging.basicConfig(filename=self.output_path / 'info.log', level=logging.DEBUG,
                                format='%(asctime)s - %(levelname)s - %(message)s', force=True)

        self.dataset = None
        self.config = config
        self.meta_chain = MetaChain(config)
        self.initialize_dataset()

        self.task_description = task_description
        self.cur_prompt = initial_prompt

        self.predictor = give_estimator(config.predictor)
        self.annotator = give_estimator(config.annotator)
        self.eval = Eval(config.eval, self.meta_chain.error_analysis, self.dataset.label_schema)
        self.batch_id = 0
        self.patient = 0

    @staticmethod
    def log_and_print(message):
        print(message)
        logging.info(message)

    def initialize_dataset(self):
        """
        Initialize the dataset: Either empty dataset or loading an existing dataset
        """
        logging.info('Initialize dataset')
        self.dataset = DatasetBase(self.config.dataset)
        if 'initial_dataset' in self.config.dataset.keys():
            logging.info(f'Load initial dataset from {self.config.dataset.initial_dataset}')
            self.dataset.load_dataset(self.config.dataset.initial_dataset)

    def calc_usage(self):
        """
        Calculate the usage of the optimization process (either $ in case of openAI or #tokens the other cases)
        """
        total_usage = 0
        total_usage += self.meta_chain.calc_usage()
        total_usage += self.annotator.calc_usage()
        total_usage += self.predictor.calc_usage()
        return total_usage

    def extract_best_prompt(self):
        sorted_history = sorted(
            self.eval.history[min(self.config.meta_prompts.warmup - 1, len(self.eval.history) - 1):],
            key=lambda x: x['score'],
            reverse=False)
        return {'prompt': sorted_history[-1]['prompt'], 'score': sorted_history[-1]['score']}

    def run_step_prompt(self):
        """
        Run the meta-prompts and get new prompt suggestion, estimated prompt score and a set of challenging samples
        for the new prompts
        """
        step_num = len(self.eval.history)
        if (step_num < self.config.meta_prompts.warmup) or (step_num % 3) > 0:
            last_history = self.eval.history[-self.config.meta_prompts.history_length:]
        else:
            sorted_history = sorted(self.eval.history[self.config.meta_prompts.warmup - 1:], key=lambda x: x['score'],
                                    reverse=False)
            last_history = sorted_history[-self.config.meta_prompts.history_length:]
        history_prompt = '\n'.join([self.eval.sample_to_text(sample,
                                                        num_errors_per_label=self.config.meta_prompts.num_err_prompt,
                                                        is_score=True) for sample in last_history])
        prompt_input = {"history": history_prompt, "task_description": self.task_description,
                        'error_analysis': last_history[-1]['analysis']}
        if 'label_schema' in self.config.dataset.keys():
            prompt_input["labels"] = json.dumps(self.config.dataset.label_schema)
        prompt_suggestion = self.meta_chain.step_prompt_chain.invoke(prompt_input)
        self.log_and_print(f'Previous prompt score:\n{self.eval.mean_score}\n#########\n')
        self.log_and_print(f'Get new prompt:\n{prompt_suggestion["prompt"]}')
        self.batch_id += 1
        if len(self.dataset) < self.config.dataset.max_samples:
            batch_input = {"num_samples": self.config.meta_prompts.samples_generation_batch,
                           "task_description": self.task_description,
                           "prompt": prompt_suggestion['prompt']}
            batch_inputs = self.generate_samples_batch(batch_input, self.config.meta_prompts.num_generated_samples,
                                                       self.config.meta_prompts.samples_generation_batch)

            if sum([len(t['errors']) for t in last_history]) > 0:
                history_samples = '\n'.join([self.eval.sample_to_text(sample,
                                                                 num_errors_per_label=self.config.meta_prompts.num_err_samples,
                                                                 is_score=False) for sample in last_history])
                for batch in batch_inputs:
                    extra_samples = self.dataset.sample_records()
                    extra_samples_text = DatasetBase.samples_to_text(extra_samples)
                    batch['history'] = history_samples
                    batch['extra_samples'] = extra_samples_text
            else:
                for batch in batch_inputs:
                    extra_samples = self.dataset.sample_records()
                    extra_samples_text = DatasetBase.samples_to_text(extra_samples)
                    batch['history'] = 'No previous errors information'
                    batch['extra_samples'] = extra_samples_text

            samples_batches = self.meta_chain.step_samples.batch_invoke(batch_inputs,
                                                                         self.config.meta_prompts.num_workers)
            new_samples = [element for sublist in samples_batches for element in sublist['samples']]
            new_samples = self.dataset.remove_duplicates(new_samples)
            self.dataset.add(new_samples, self.batch_id)
            logging.info('Get new samples')
        self.cur_prompt = prompt_suggestion['prompt']

    def stop_criteria(self):
        """
        Check if the stop criteria holds. The conditions for stopping:
        1. Usage is above the threshold
        2. There was no improvement in the last > patient steps
        """
        if 0 < self.config.stop_criteria.max_usage < self.calc_usage():
            return True
        if len(self.eval.history) <= self.config.meta_prompts.warmup:
            self.patient = 0
            return False
        min_batch_id, max_score = self.eval.get_max_score(self.config.meta_prompts.warmup-1)
        if max_score - self.eval.history[-1]['score'] > -self.config.stop_criteria.min_delta:
            self.patient += 1
        else:
            self.patient = 0
        if self.patient > self.config.stop_criteria.patience:
            return True
        return False

    @staticmethod
    def generate_samples_batch(batch_input, num_samples, batch_size):
        """
        Generate samples in batch
        """
        batch_num = num_samples // batch_size
        all_batches = [batch_input.copy() for _ in range(batch_num)]
        reminder = num_samples - batch_num * batch_size
        if reminder > 0:
            all_batches.append(batch_input.copy())
            all_batches[-1]['num_samples'] = reminder
        return all_batches

    def generate_initial_samples(self):
        """
        In case the initial dataset is empty generate the initial samples
        """
        batch_input = {"num_samples": self.config.meta_prompts.samples_generation_batch,
                       "task_description": self.task_description,
                       "instruction": self.cur_prompt}
        batch_inputs = self.generate_samples_batch(batch_input, self.config.meta_prompts.num_initialize_samples,
                                                   self.config.meta_prompts.samples_generation_batch)

        samples_batches = self.meta_chain.initial_chain.batch_invoke(batch_inputs, self.config.meta_prompts.num_workers)
        samples_list = [element for sublist in samples_batches for element in sublist['samples']]
        samples_list = self.dataset.remove_duplicates(samples_list)
        self.dataset.add(samples_list, 0)

    def save_state(self):
        """
        Save the process state
        """
        if self.output_path is None:
            return
        logging.info('Save state')
        self.dataset.save_dataset(self.output_path / 'dataset.csv')
        state = {'history': self.eval.history, 'batch_id': self.batch_id,
                 'prompt': self.cur_prompt, 'task_description': self.task_description,
                 'patient': self.patient}
        pickle.dump(state, open(self.output_path / 'history.pkl', 'wb'))

    def load_state(self, path: str):
        """
        Load pretrain state
        """
        path = Path(path)
        if (path / 'dataset.csv').is_file():
            self.dataset.load_dataset(path / 'dataset.csv')
        if (path / 'history.pkl').is_file():
            state = pickle.load(open(path / 'history.pkl', 'rb'))
            self.eval.history = state['history']
            self.batch_id = state['batch_id']
            self.cur_prompt = state['prompt']
            self.task_description = state['task_description']
            self.patient = state['patient']

    def step(self, current_iter, total_iter):
        """
        This is the main optimization process step.
        """
        self.log_and_print(f'Starting step {self.batch_id}')
        if len(self.dataset.records) == 0:
            self.log_and_print('Dataset is empty generating initial samples')
            self.generate_initial_samples()
        if self.config.use_wandb:
            cur_batch = self.dataset.get_leq(self.batch_id)
            random_subset = cur_batch.sample(n=min(10, len(cur_batch)))[['text']]
            self.wandb_run.log(
                {"Prompt": wandb.Html(f"<p>{self.cur_prompt}</p>"), "Samples": wandb.Table(dataframe=random_subset)},
                step=self.batch_id)

        logging.info('Running annotator')
        records = self.annotator.apply(self.dataset, self.batch_id)
        self.dataset.update(records)

        self.predictor.cur_instruct = self.cur_prompt
        logging.info('Running Predictor')
        records = self.predictor.apply(self.dataset, self.batch_id, leq=True)
        self.dataset.update(records)

        self.eval.dataset = self.dataset.get_leq(self.batch_id)
        self.eval.eval_score()
        logging.info('Calculating Score')
        large_errors = self.eval.extract_errors()
        self.eval.add_history(self.cur_prompt, self.task_description)
        if self.config.use_wandb:
            large_errors = large_errors.sample(n=min(6, len(large_errors)))
            correct_samples = self.eval.extract_correct()
            correct_samples = correct_samples.sample(n=min(6, len(correct_samples)))
            vis_data = pd.concat([large_errors, correct_samples])
            self.wandb_run.log({"score": self.eval.history[-1]['score'],
                                "prediction_result": wandb.Table(dataframe=vis_data),
                                'Total usage': self.calc_usage()}, step=self.batch_id)
        if self.stop_criteria():
            self.log_and_print('Stop criteria reached')
            return True
        if current_iter != total_iter-1:
            self.run_step_prompt()
        self.save_state()
        return False

    def run_pipeline(self, num_steps: int):
        # Run the optimization pipeline for num_steps
        num_steps_remaining = num_steps - self.batch_id
        for i in range(num_steps_remaining):
            stop_criteria = self.step(i, num_steps_remaining)
            if stop_criteria:
                break
        final_result = self.extract_best_prompt()
        return final_result