-
Notifications
You must be signed in to change notification settings - Fork 20
/
predict_noGT.py
103 lines (81 loc) · 3.37 KB
/
predict_noGT.py
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
# ==============================================================================
# Copyright (c) 2021, Yamagishi Laboratory, National Institute of Informatics
# Author: Erica Cooper
# All rights reserved.
# ==============================================================================
## run inference without requiring ground-truth answers
## or system info.
import os
import argparse
import torch
import torch.nn as nn
import fairseq
from torch.utils.data import DataLoader
from mos_fairseq import MosPredictor, MyDataset
import numpy as np
import scipy.stats
import datetime
import time
def unixnow():
return str(int(time.mktime(datetime.datetime.now().timetuple())))
def systemID(uttID):
return uttID.split('-')[0]
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--fairseq_base_model', type=str, required=True, help='Path to pretrained fairseq base model.')
parser.add_argument('--datadir', type=str, required=True, help='Path of your directory containing .wav files')
parser.add_argument('--finetuned_checkpoint', type=str, required=True, help='Path to finetuned MOS prediction checkpoint.')
parser.add_argument('--outfile', type=str, required=False, default='answer.txt', help='Output filename for your answer.txt file for submission to the CodaLab leaderboard.')
args = parser.parse_args()
cp_path = args.fairseq_base_model
my_checkpoint = args.finetuned_checkpoint
wavdir = args.datadir
outfile = args.outfile
model, cfg, task = fairseq.checkpoint_utils.load_model_ensemble_and_task([cp_path])
ssl_model = model[0]
ssl_model.remove_pretraining_modules()
print('Loading checkpoint')
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
ssl_model_type = cp_path.split('/')[-1]
if ssl_model_type == 'wav2vec_small.pt':
SSL_OUT_DIM = 768
elif ssl_model_type in ['w2v_large_lv_fsh_swbd_cv.pt', 'xlsr_53_56k.pt']:
SSL_OUT_DIM = 1024
else:
print('*** ERROR *** SSL model type ' + ssl_model_type + ' not supported.')
exit()
model = MosPredictor(ssl_model, SSL_OUT_DIM).to(device)
model.eval()
model.load_state_dict(torch.load(my_checkpoint))
wavfnames = [x for x in os.listdir(wavdir) if x.split('.')[-1] == 'wav']
wavlist = 'tmp_' + unixnow() + '.txt'
wavlistf = open(wavlist, 'w')
for w in wavfnames:
wavlistf.write(w + ',3.0\n')
wavlistf.close()
print('Loading data')
validset = MyDataset(wavdir, wavlist)
validloader = DataLoader(validset, batch_size=1, shuffle=True, num_workers=2, collate_fn=validset.collate_fn)
total_loss = 0.0
num_steps = 0.0
predictions = { } # filename : prediction
criterion = nn.L1Loss()
print('Starting prediction')
for i, data in enumerate(validloader, 0):
inputs, labels, filenames = data
inputs = inputs.to(device)
labels = labels.to(device)
outputs = model(inputs)
loss = criterion(outputs, labels)
total_loss += loss.item()
output = outputs.cpu().detach().numpy()[0]
predictions[filenames[0]] = output ## batch size = 1
## generate answer.txt for codalab
ans = open(outfile, 'w')
for k, v in predictions.items():
outl = k.split('.')[0] + ',' + str(v) + '\n'
ans.write(outl)
ans.close()
os.system('rm ' + wavlistf)
if __name__ == '__main__':
main()