File size: 6,050 Bytes
aefc9ef
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
# Copyright (c) Guangsheng Bao.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.

import numpy as np
import torch
import torch.nn.functional as F
import tqdm
import argparse
import json
from data_builder import load_data
from model import load_tokenizer, load_model
from metrics import get_roc_metrics, get_precision_recall_metrics

def get_likelihood(logits, labels):
    assert logits.shape[0] == 1
    assert labels.shape[0] == 1

    logits = logits.view(-1, logits.shape[-1])
    labels = labels.view(-1)
    log_probs = torch.nn.functional.log_softmax(logits, dim=-1)
    log_likelihood = log_probs.gather(dim=-1, index=labels.unsqueeze(-1)).squeeze(-1)
    return log_likelihood.mean().item()

def get_rank(logits, labels):
    assert logits.shape[0] == 1
    assert labels.shape[0] == 1

    # get rank of each label token in the model's likelihood ordering
    matches = (logits.argsort(-1, descending=True) == labels.unsqueeze(-1)).nonzero()
    assert matches.shape[1] == 3, f"Expected 3 dimensions in matches tensor, got {matches.shape}"

    ranks, timesteps = matches[:, -1], matches[:, -2]

    # make sure we got exactly one match for each timestep in the sequence
    assert (timesteps == torch.arange(len(timesteps)).to(timesteps.device)).all(), "Expected one match per timestep"

    ranks = ranks.float() + 1 # convert to 1-indexed rank
    return -ranks.mean().item()

def get_logrank(logits, labels):
    assert logits.shape[0] == 1
    assert labels.shape[0] == 1

    # get rank of each label token in the model's likelihood ordering
    matches = (logits.argsort(-1, descending=True) == labels.unsqueeze(-1)).nonzero()
    assert matches.shape[1] == 3, f"Expected 3 dimensions in matches tensor, got {matches.shape}"

    ranks, timesteps = matches[:, -1], matches[:, -2]

    # make sure we got exactly one match for each timestep in the sequence
    assert (timesteps == torch.arange(len(timesteps)).to(timesteps.device)).all(), "Expected one match per timestep"

    ranks = ranks.float() + 1  # convert to 1-indexed rank
    ranks = torch.log(ranks)
    return -ranks.mean().item()

def get_entropy(logits, labels):
    assert logits.shape[0] == 1
    assert labels.shape[0] == 1

    entropy = F.softmax(logits, dim=-1) * F.log_softmax(logits, dim=-1)
    entropy = -entropy.sum(-1)
    return entropy.mean().item()


def experiment(args):
    # load model
    scoring_tokenizer = load_tokenizer(args.scoring_model_name, args.dataset, args.cache_dir)
    scoring_model = load_model(args.scoring_model_name, args.device, args.cache_dir)
    scoring_model.eval()
    # load data
    data = load_data(args.dataset_file)
    n_samples = len(data["sampled"])
    # eval criterions
    criterion_fns = {'likelihood': get_likelihood,
                     'rank': get_rank,
                     'logrank': get_logrank,
                     'entropy': get_entropy}
    for name in criterion_fns:
        criterion_fn = criterion_fns[name]
        torch.manual_seed(args.seed)
        np.random.seed(args.seed)
        eval_results = []
        for idx in tqdm.tqdm(range(n_samples), desc=f"Computing {name} criterion"):
            original_text = data["original"][idx]
            sampled_text = data["sampled"][idx]
            # original text
            tokenized = scoring_tokenizer(original_text, return_tensors="pt", padding=True, return_token_type_ids=False).to(args.device)
            labels = tokenized.input_ids[:, 1:]
            with torch.no_grad():
                logits = scoring_model(**tokenized).logits[:, :-1]
                original_crit = criterion_fn(logits, labels)
            # sampled text
            tokenized = scoring_tokenizer(sampled_text, return_tensors="pt", padding=True, return_token_type_ids=False).to(args.device)
            labels = tokenized.input_ids[:, 1:]
            with torch.no_grad():
                logits = scoring_model(**tokenized).logits[:, :-1]
                sampled_crit = criterion_fn(logits, labels)
            # result
            eval_results.append({"original": original_text,
                            "original_crit": original_crit,
                            "sampled": sampled_text,
                            "sampled_crit": sampled_crit})

        # compute prediction scores for real/sampled passages
        predictions = {'real': [x["original_crit"] for x in eval_results],
                       'samples': [x["sampled_crit"] for x in eval_results]}
        fpr, tpr, roc_auc = get_roc_metrics(predictions['real'], predictions['samples'])
        p, r, pr_auc = get_precision_recall_metrics(predictions['real'], predictions['samples'])
        print(f"Criterion {name}_threshold ROC AUC: {roc_auc:.4f}, PR AUC: {pr_auc:.4f}")
        # log results
        results_file = f'{args.output_file}.{name}.json'
        results = { 'name': f'{name}_threshold',
                    'info': {'n_samples': n_samples},
                    'predictions': predictions,
                    'raw_results': eval_results,
                    'metrics': {'roc_auc': roc_auc, 'fpr': fpr, 'tpr': tpr},
                    'pr_metrics': {'pr_auc': pr_auc, 'precision': p, 'recall': r},
                    'loss': 1 - pr_auc}
        with open(results_file, 'w') as fout:
            json.dump(results, fout)
            print(f'Results written into {results_file}')

if __name__ == '__main__':
    parser = argparse.ArgumentParser()
    parser.add_argument('--output_file', type=str, default="./exp_test/results/xsum_gpt2")
    parser.add_argument('--dataset', type=str, default="xsum")
    parser.add_argument('--dataset_file', type=str, default="./exp_test/data/xsum_gpt2")
    parser.add_argument('--scoring_model_name', type=str, default="gpt2")
    parser.add_argument('--seed', type=int, default=0)
    parser.add_argument('--device', type=str, default="cuda")
    parser.add_argument('--cache_dir', type=str, default="../cache")
    args = parser.parse_args()

    experiment(args)