BayesAdapt/active_learn.py at main · SRI-CSL/BayesAdapt · GitHub

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
import sys
import math
import json
import torch
import numpy as np
from tqdm import tqdm
import os
import hydra
from hydra.utils import instantiate
from datasets import concatenate_datasets
from bayesadapt.utils import average_log_probs
from accelerate.utils import set_seed
import torch.nn.functional as F

def bald(logits):
    logprobs = F.log_softmax(logits, dim=-1) #B x N x C
    probs = F.softmax(logits, dim=-1) #B x N x C
    avg_probs = probs.mean(dim=1) #B x C
    log_avg_probs = average_log_probs(logits) #B x C
    entropy_avg = -torch.sum(avg_probs * log_avg_probs, dim=-1) #B
    sample_entropy = -torch.sum(probs * logprobs, dim=-1).mean(dim=1) #B
    bald_score = entropy_avg - sample_entropy
    return bald_score

@hydra.main(config_path="./conf", config_name="default", version_base=None)
def main(cfg):
    print(cfg)
    set_seed(cfg.seed)
    pool_dataset = instantiate(cfg.train_dataset, split='train')
    pool_ids = pool_dataset.data['question_id']
    test_dataset = instantiate(cfg.test_dataset, split='test')
    id_key = 'question_id'
    num_select = 10
    initial_train_ids = np.random.choice(
        pool_ids,
        size=num_select,
        replace=False
    ).tolist()
    train_dataset = instantiate(cfg.train_dataset, split='train')
    train_dataset.data = pool_dataset.data.filter(lambda x: x[id_key] in initial_train_ids)
    pool_dataset.data = pool_dataset.data.filter(lambda x: x[id_key] not in initial_train_ids)

    #randomly select 10 instances from pool to use as valset
    if cfg.use_val_set:
        val_ids = np.random.choice(
            pool_dataset.data[id_key],
            size=num_select,
            replace=False
        ).tolist()
        val_dataset = instantiate(cfg.train_dataset, split='train')
        val_dataset.data = pool_dataset.data.filter(lambda x: x[id_key] in val_ids)
        pool_dataset.data = pool_dataset.data.filter(lambda x: x[id_key] not in val_ids)


    history = []
    for i in range(20):
        print(f"Round {i+1} of active learning, training on {len(train_dataset)} samples")
        trainer = instantiate(cfg.trainer, cfg=cfg) #cold start each time
        evaldir = trainer.evaldir.replace('id', 'active_learn')

        json_path = os.path.join(evaldir, 'results.json')
        if os.path.exists(json_path):
            if not cfg.overwrite:
                sys.exit(f"Results already exist at {json_path}, experiment already complete")


        pool_ids = pool_dataset.data[id_key]
        if len(pool_ids) > 1000:
            sampled_pool_ids = np.random.choice(pool_ids, size=1000, replace=False).tolist()
            sampled_pool_dataset = instantiate(cfg.train_dataset, split='train')
            sampled_pool_dataset.data = pool_dataset.data.filter(lambda x: x[id_key] in sampled_pool_ids)
        else:
            sampled_pool_dataset = pool_dataset

        trainer.update_dataloaders(train_dataset=train_dataset, test_dataset=sampled_pool_dataset)
        if cfg.use_val_set:
            trainer.train(save=False, use_wandb=False, validation_dataset=val_dataset)
        else:
            trainer.train(save=False, use_wandb=False)
        pool_metrics, logits_dict = trainer.evaluate(save=False, verbose=False)

        logit_ids = list(logits_dict.keys())
        logits = torch.stack([logits_dict[qid] for qid in logit_ids]) #B x N x C
        logits = logits.to(torch.float64)
        B, num_samples, num_classes = logits.shape

        if num_samples == 1: #only one sample, probably MLE, so use predictive entropy instead of BALD
            logprobs = average_log_probs(logits) #B x C
            probs = torch.exp(logprobs)
            scores = -torch.sum(probs * logprobs, dim=1) / math.log(num_classes) #entropy normalized by log(num_classes)
        else:
            scores = bald(logits) #B

        top_values, top_indices = torch.topk(
            scores,
            k=num_select,
            largest=True
        )

        selected_qids = [logit_ids[idx] for idx in top_indices.tolist()]
        selected_hf = pool_dataset.data.filter(lambda x: x[id_key] in selected_qids)

        #test on true test set for plots later
        trainer.update_dataloaders(test_dataset=test_dataset)
        test_metrics, _ = trainer.evaluate(save=False)

        history.append({
            'selected_qids': selected_qids,
            'pool_metrics': pool_metrics,
            'test_metrics': test_metrics,
            'top_scores': top_values.cpu().tolist(),
        })

        train_dataset.data = concatenate_datasets([train_dataset.data, selected_hf])
        pool_dataset.data = pool_dataset.data.filter(lambda x: x[id_key] not in selected_qids)

    if not os.path.exists(evaldir):
        os.makedirs(evaldir, exist_ok=True)

    json_path = os.path.join(evaldir, 'results.json')
    with open(json_path, 'w') as f:
        json.dump(history, f)


if __name__ == "__main__":
    main()