AttentiveMLP/InferenceWMT_valid.py at main · Shark-NLP/AttentiveMLP · 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
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
import sys
import torch
import os

modelname = sys.argv[1]
start = int(sys.argv[2])
end = int(sys.argv[3])
checkpointfolder = sys.argv[4]
bleufolder = sys.argv[5]
datafolder = sys.argv[6]


modelfolder = checkpointfolder + modelname + '/'
ensemblefolder = modelfolder + 'ensemblemodel/'
batch=100

if 'CMLM' in modelname:
    bleufolder = bleufolder + modelname + '/bestvalid_iter10_LEN3/'
else:
    bleufolder = bleufolder + modelname + '/bestvalid_beam4/'
is_s3_path = "s3://" in checkpointfolder
if not is_s3_path:

    os.system('mkdir -p {}'.format(bleufolder))
    os.system('mkdir -p {}'.format(ensemblefolder))

try:
    validbleu = torch.load(bleufolder+'validbleu.pt')
    bestepoch = max(validbleu, key=lambda k: validbleu[k])
    bestbleu = validbleu[bestepoch]
    print('best validation bleu = {} at epoch {}'.format(bestbleu, bestepoch))
except:
    validbleu = {}
    bestepoch = start
    bestbleu = 0

jlist = [j for j in range(start, end + 1)]

for j in jlist:
    if j not in validbleu.keys():
        cpname = modelfolder + 'checkpoint{}.pt'.format(j)
        bleu = bleufolder + 'checkpoint{}_valid.out'.format(j)
        print('evaluating {}'.format(bleu))
        if 'WMTdeen' in modelname:
            if 'CMLM' in modelname:
                if 'distill' in modelname:
                    command = 'python generate.py {} --gen-subset valid --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --remove-bpe --iter-decode-force-max-iter --iter-decode-with-beam 3 --quiet | tee {}'.format(datafolder,
                        cpname, batch, bleu)
                else:
                    command = 'python generate.py /nvme/jsy/data-bin/wmt14_deen_jointdict/ --gen-subset valid --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --remove-bpe --iter-decode-force-max-iter --iter-decode-with-beam 3 --quiet | tee {}'.format(
                        cpname, batch, bleu)
            else:
                if 'distill' in modelname:
                    command = 'python generate.py {} --gen-subset valid --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 --quiet | tee {}'.format(datafolder,
                        cpname, bleu)
                else:
                    command = 'python generate.py /nvme/jsy/data-bin/wmt14_deen_jointdict/ --gen-subset valid --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 --quiet | tee {}'.format(
                        cpname, bleu)
        elif 'WMTende' in modelname:
            if 'CMLM' in modelname:
                if 'distill' in modelname:
                    command = 'python generate.py {} --gen-subset valid --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --remove-bpe --iter-decode-force-max-iter --iter-decode-with-beam 3 --quiet | tee {}'.format(datafolder,
                        cpname, batch, bleu)
                else:
                    command = 'python generate.py /nvme/jsy/data-bin/wmt14_ende_jointdict/ --gen-subset valid --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --remove-bpe --iter-decode-force-max-iter --iter-decode-with-beam 3 --quiet | tee {}'.format(
                        cpname, batch, bleu)
            else:
                if 'distill' in modelname:
                    command = 'python generate.py {} --gen-subset valid --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 --quiet | tee {}'.format(datafolder,
                        cpname, bleu)
                else:
                    command = 'python generate.py /nvme/jsy/data-bin/wmt14_ende_jointdict/ --gen-subset valid --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 --quiet | tee {}'.format(
                        cpname, bleu)
        os.system(command)
        with open(bleu, 'r') as f:
            lines = f.read().splitlines()
            lastline = lines[-1].replace(',', '').split()
            validbleu[j] = float(lastline[6])
            if bestbleu < float(lastline[6]):
                bestbleu = float(lastline[6])
                bestepoch = j
            print('best validation bleu {} at epoch {}'.format(bestbleu, bestepoch))

for n in [5, 10]:
    bestcplist = sorted(validbleu, key=lambda key: validbleu[key], reverse=True)[:n]
    print('top {} checkpoints:'.format(n))
    for i in bestcplist:
        print('epoch {}: validation bleu {}'.format(i, validbleu[i]))

    if 'CMLM' in modelname:
        bestensemble = modelfolder + 'LEN3_bestmodel_bestvalid_ensemble{}_epoch{}_{}.pt'.format(n, min(bestcplist), max(bestcplist))
    else:
        bestensemble = modelfolder + 'bestmodel_bestvalid_ensemble{}_epoch{}_{}.pt'.format(n, min(bestcplist), max(bestcplist))

    cpname = ensemblefolder + 'checkpoint{}.pt'.format(bestcplist[0])
    try:
        model = torch.load(cpname)
    except:
        os.system('cp {} {}'.format(modelfolder + 'checkpoint{}.pt'.format(bestcplist[0]), cpname))
        model = torch.load(cpname)

    for i in range(1, len(bestcplist)):
        cpname2 = ensemblefolder + 'checkpoint{}.pt'.format(bestcplist[i])
        try:
            model2 = torch.load(cpname2)
        except:
            os.system('cp {} {}'.format(modelfolder + 'checkpoint{}.pt'.format(bestcplist[i]), cpname2))
            model2 = torch.load(cpname2)

        for param in model['model']:
            if 'decoder.embed_tokens.weight' in param:
                pass
            else:
                model['model'][param].add_(model2['model'][param])
        del model2
    for param in model['model']:
        if 'decoder.embed_tokens.weight' in param:
            pass
        else:
            model['model'][param].div_(float(n))
    torch.save(model, bestensemble)

    del model
    if 'CMLM' in modelname:
        bleu = bleufolder + 'LEN3_bestmodel_bestvalid_ensemble{}_epoch{}_{}.out'.format(n, min(bestcplist), max(bestcplist))
    else:
        bleu = bleufolder + 'bestmodel_bestvalid_ensemble{}_epoch{}_{}.out'.format(n, min(bestcplist), max(bestcplist))
    print('evaluating {}'.format(bleu))

    if 'WMTdeen' in modelname:
        if 'CMLM' in modelname:
            if 'distill' in modelname:
                command = 'python generate.py {} --gen-subset test --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --iter-decode-force-max-iter --iter-decode-with-beam 3 --remove-bpe | tee {}'.format(datafolder,
                    bestensemble, batch, bleu)
            else:
                command = 'python generate.py {} --gen-subset test --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --iter-decode-force-max-iter --iter-decode-with-beam 3 --remove-bpe | tee {}'.format(datafolder,
                    bestensemble, batch, bleu)
        else:
            if 'distill' in modelname:
                command = 'python generate.py {} --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 | tee {}'.format(datafolder,
                    bestensemble, bleu)
            else:
                command = 'python generate.py {} --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 | tee {}'.format(datafolder,
                    bestensemble, bleu)
    if 'WMTende' in modelname:
        if 'CMLM' in modelname:
            if 'distill' in modelname:
                command = 'python generate.py {} --gen-subset test --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --iter-decode-force-max-iter --iter-decode-with-beam 3 --remove-bpe | tee {}'.format(datafolder,
                    bestensemble, batch, bleu)
            else:
                command = 'python generate.py {} --gen-subset test --task translation_lev --path {} --batch-size {} --iter-decode-max-iter 10 --iter-decode-eos-penalty 0 --iter-decode-force-max-iter --iter-decode-with-beam 3 --remove-bpe | tee {}'.format(datafolder,
                    bestensemble, batch, bleu)
        else:
            if 'distill' in modelname:
                command = 'python generate.py {} --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 | tee {}'.format(datafolder,
                    bestensemble, bleu)
            else:
                command = 'python generate.py {} --path {} --beam 4 --batch-size 128 --remove-bpe --lenpen 0.3 | tee {}'.format(datafolder,
                    bestensemble, bleu)

    os.system(command)
    command = './compound_split_bleu.sh {}'.format(bleu)
    os.system(command)

torch.save(validbleu, bleufolder+'validbleu.pt')