File size: 5,620 Bytes
10b912d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import argparse
import numpy as np
from pathlib import Path
import tqdm
from pprint import pprint
import torch
from torch.nn.utils.rnn import pad_sequence
from scrl.config import load_config
from scrl.training import setup_and_train
from scrl.model import labels_to_summary
from scrl.eval_metrics import compute_token_f1
import scrl.utils as utils
from nltk import word_tokenize


def evaluate_validation_reward(args, manager, model, tokenizer, reward_generator, dataset):
    device = args.device
    idx_range = list(range(len(dataset)))
    dataset_indices = list(utils.batchify(idx_range, args.batch_size))
    rewards = []
    for i, indices in enumerate(dataset_indices):
        if args.max_val_steps != None and i >= args.max_val_steps:
            break
        batch = dataset[indices]
        input_ids = batch["input_ids"]
        input_ids = pad_sequence(
            [torch.tensor(ids) for ids in input_ids], batch_first=True
        )
        logits = model(input_ids.to(device))
        probs = torch.softmax(logits, dim=2)
        argmax_labels = torch.argmax(logits, dim=2).to(device)
        argmax_summaries = labels_to_summary(input_ids, argmax_labels, tokenizer)
        argmax_rewards, _ = reward_generator(batch["document"], argmax_summaries)
        rewards += argmax_rewards
    avg_reward = np.mean(rewards)
    return avg_reward



def evaluate_validation_dataset(args, manager, model, tokenizer, reward_generator, dataset_path):
    f1_scores = []
    dataset = list(utils.read_jsonl(dataset_path))
    dump_data = []

    for item in tqdm.tqdm(dataset):
        src = item["text"]
        tgts = item["summaries"]

        input_ids = torch.tensor(tokenizer([src])["input_ids"]).to(args.device)
        logits = model.forward(input_ids)
        argmax_labels = torch.argmax(logits, dim=2)
        pred = labels_to_summary(input_ids, argmax_labels, tokenizer)[0]

        pred_tokens = word_tokenize(pred)
        src_tokens = word_tokenize(src)


        item_scores = []
        for tgt in tgts:
            tgt_tokens = word_tokenize(tgt)
            pred_tokens = [t.lower() for t in pred_tokens]
            tgt_tokens = [t.lower() for t in tgt_tokens]
            token_f1 = compute_token_f1(
                tgt_tokens, pred_tokens, use_counts=True
            )
            item_scores.append(token_f1)

        if args.dump:
            probs = torch.softmax(logits, dim=2)[0].detach().tolist()
            dump_item = {
                "probs": probs,
                "source": src,
                "target": tgts[0],
                "f1-score": item_scores[0],
                "pred_summary": pred,
                "pred_labels": argmax_labels[0].tolist(),
            }
            dump_data.append(dump_item)

        item_score = np.mean(item_scores)
        f1_scores.append(item_score)
    score = np.mean(f1_scores)


    if args.dump:
        dataset_name = dataset_path.name.split(".jsonl")[0]
        dump_dir = manager.dir / f"dump-{dataset_name}"
        dump_dir.mkdir(exist_ok=True)
        utils.write_jsonl(
            dump_data,
            dump_dir / f"step-{manager.step}.jsonl",
            "w"
        )
    return score


def evaluate(args, manager, model, tokenizer, reward_generator, holdout_data):
    step = manager.step
    val_reward = evaluate_validation_reward(args, manager, model, tokenizer, reward_generator, holdout_data)

    reward_path = manager.dir / "val_rewards.jsonl"
    if reward_path.exists():
        reward_results = list(utils.read_jsonl(reward_path))
        prev_max = max([x["score"] for x in reward_results])
    else:
        reward_results = []
        prev_max = 0
    if val_reward > prev_max:
        manager.save_model(model, step, "best_val_reward")
    reward_results.append({"step": step, "score": val_reward})
    utils.write_jsonl(reward_results, reward_path, "w")
    if args.verbose:
        print("Validation Rewards:")
        pprint(reward_results)
        print()

    # only used if a validation dataset is specified in config
    for val_data_path in args.validation_datasets:
        val_data_path = Path(val_data_path)
        dataset_name = val_data_path.name.split(".jsonl")[0]
        dataset_score = evaluate_validation_dataset(
            args, manager, model, tokenizer, reward_generator, val_data_path
        )
        result_path = Path(manager.dir / f"val_data_results.{dataset_name}.jsonl")
        if result_path.exists():
            dataset_results = list(utils.read_jsonl(result_path))
            prev_max = max([x["score"] for x in dataset_results])
        else:
            dataset_results = []
            prev_max = 0
        if dataset_score > prev_max:
            manager.save_model(model, step, f"best_on_{dataset_name}")
        dataset_results.append({"step": step, "score": dataset_score})
        utils.write_jsonl(dataset_results, result_path, "w")
        if args.verbose:
            print(f"Validation Dataset Results for {dataset_name}:")
            pprint(dataset_results)
            print()


def main(args):
    utils.set_random_seed(0)
    setup_and_train(args, eval_func=evaluate)


if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument("--config", help="path to JSON config file")
    parser.add_argument("--device", default="cuda")
    parser.add_argument("--dump", action="store_true")
    parser.add_argument("--verbose", action="store_true")
    parser.add_argument(
        "--fresh",
        action="store_true",
        help="delete model directory and start from scratch"
    )
    main(load_config(parser.parse_args()))