bms/predict.py at master · nofreewill42/bms · 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
129
130
131
132
133
134
135
136
import torch
from torch.utils.data import DataLoader
from data_loaders.test_loader import TestDS

import pickle
import pandas as pd
from tqdm import tqdm
from pathlib import Path
import sentencepiece as sp

from model_architecture.model import Model
from model_architecture.beam_search import BeamSearcher


if __name__ == '__main__':
    device = 'cuda:0' if torch.cuda.is_available() else 'cpu'
    bpe_num = 4096
    max_len = 256
    bs = 8
    bw = 2#0
    torch.backends.cudnn.benchmark = True

    type = 'valid'#'test'#

    names = ['160_224_210', '160_224_210_192_192_32_192_192_36',
             '256_544_128', '256_544_128_224_640_24', '256_544_128_224_640_24_320_480_16',
             '288_512_112', '288_512_112_320_480_20', '288_512_112_320_480_20_480_320_4',
             '288_512_160', '288_512_160_256_576_20', '288_512_160_256_576_20_320_448_24',
             '384_384_120', '384_384_120_352_416_20', '384_384_120_352_416_20_192_768_4']
    img_sizes = [(160,224), (192,192),
                 (256,544), (224,640), (320,480),
                 (288,512), (320,480), (480,320),
                 (288,512), (256,576), (320,448),
                 (384,384), (352,416), (192,768),]
    raw_lds = [1.5233, 1.5275,
               1.0955, 1.1850, 1.0699,
               1.1706, 1.0629, 1.3790,
               1.0664, 1.1234, 1.0832,
               1.4073, 1.2095, 2.1064,]
    norm_lds = [1.3661, 1.3879,
               0.9438, 1.0244, 0.9248,
               1.0205, 0.9280, 1.2254,
               0.9415, 0.9923, 0.9400,
               1.2318, 1.0646, 1.9285,]

    weights = [1/(raw_ld*norm_ld**3) for raw_ld, norm_ld in zip(raw_lds,norm_lds)]

    ds_path = Path("/home/nofreewill/Documents/kaggle/bms/bms-data/")
    imgs_path = ds_path / 'images' / ('train' if type=='valid' else 'test')
    test_stems = pickle.load((ds_path/'samples'/f'{type}_img_stems_sorted.pkl').open('rb'))[::-1]
    if (ds_path/'predictions'/type/'same_norm.csv').exists():
        same_df = pd.read_csv(ds_path/'predictions'/type/'same_norm.csv')
        same_stems = set(same_df.iloc[:,0])
        test_stems = [stem for stem in test_stems if stem not in same_stems]
    print(len(test_stems)//bs)

    sp.SentencePieceProcessor()
    subwords_path = ds_path/'subwords'/f'bpe_{bpe_num}.model'
    swp = sp.SentencePieceProcessor(str(subwords_path))

    models = []
    test_dls = []
    ws = []
    for name, img_size in zip(names, img_sizes):
        print(name, img_size)

        if bw == 0:
            w_path = (ds_path/'predictions'/type/f'raw/submission_{name}.csv')
            if w_path.exists():
                print('\tSKIP')
                continue
            ws.append(w_path)

        test_ds = TestDS(imgs_path, test_stems, img_size)
        test_dl = DataLoader(test_ds, batch_size=bs, pin_memory=True, num_workers=4)

        # model
        nhead = 8 if name in ('160_224_210', '160_224_210_192_192_32_192_192_36') else 16
        N, n, ff, first_k, first_s, last_s = 32, 128, 128, 3,2,1
        enc_d_model, enc_nhead, enc_dim_feedforward, enc_num_layers =  512, nhead, 2048, 6
        dec_d_model, dec_nhead, dec_dim_feedforward, dec_num_layers =  512, nhead, 2048, 6
        model = Model(bpe_num, N, n, ff, first_k, first_s, last_s,
                      enc_d_model, enc_nhead, enc_dim_feedforward, enc_num_layers,
                      dec_d_model, dec_nhead, dec_dim_feedforward, dec_num_layers,
                      max_trn_len=max_len).to(device)
        model.load_state_dict(torch.load(f'model_weights/done/model_{name}.pth', map_location=device))
        model.eval()

        models.append(model)
        test_dls.append(test_dl)

    if bw == 0:
        for k in range(len(models)):
            w = ws[k].open('w', buffering=1)
            m = models[k]
            test_dl = test_dls[k]
            w.write('image_id,InChI\n')

            for j, batch in enumerate(tqdm(test_dl)):
                imgs_tensor, img_stems, _ = batch
                imgs_tensor = imgs_tensor.to(device)

                with torch.no_grad():
                    lbl_ids, lens = m.predict([imgs_tensor])
                    lbl_ids = torch.stack(lbl_ids).T if not isinstance(lbl_ids, torch.Tensor) else lbl_ids

                for i in range(len(lbl_ids)):
                    pred = lbl_ids[i].tolist()[:lens[i]]
                    pred_text = swp.decode(pred)

                    if w is not None:
                        w.write(f'{img_stems[i]},"InChI=1S/C{pred_text}"\n')
            w.close()

    else:
        w = (ds_path/'predictions'/type/f'submission_{bw}.csv').open('w', buffering=1)
        w.write('image_id,InChI\n')
        m = BeamSearcher(models, weights, bpe_num, bw)

        for j, batch in enumerate(tqdm(zip(*test_dls))):
            imgs_tensors = [b[0].to(device) for b in batch]
            img_stems = batch[0][1]
            ratios_tensors = torch.stack([b[2] for b in batch]).float().to(device)
            ratios_tensors = ratios_tensors.reshape(-1,1).repeat(1,bw).reshape(-1,bs*bw)

            with torch.no_grad():
                lbl_ids, lens = m.predict(imgs_tensors, ratios_tensors, max_pred_len=256)
                lbl_ids = torch.stack(lbl_ids).T if not isinstance(lbl_ids, torch.Tensor) else lbl_ids
            if (lens == test_dls[0].dataset.max_len).sum() == len(lens):
                print('ERR')

            for i in range(len(lbl_ids)):
                pred = lbl_ids[i].tolist()[:lens[i]]
                pred_text = swp.decode(pred)

                w.write(f'{img_stems[i]},"InChI=1S/C{pred_text}"\n')