baseline_dev.py

import sys
import tensorflow as tf
import numpy as np
sys.path.insert(0, './scripts')
from tqdm import tqdm
import feature_tools as ft

#init params
NN_MODEL = 'lang2vec'
INPUT_DIM = 40
is_batchnorm = False
SOFTMAX_NUM = 17
nn_model = __import__(NN_MODEL)
FEAT_TYPE='mfcc'
N_FFT=400
HOP=160
VAD=True
CMVN='m'

#load data
lines = open('data/language_id_initial').readlines()
lang2idx = {}
for line in lines:
    lang = line.rstrip().split()[0]
    idx = line.rstrip().split()[1]
    lang2idx[lang]=int(idx)

lines=open('data/dev/utt2lang').readlines()
utt2lang={}
for iter in range(len(lines)):
    utt=lines[iter].rstrip().split()[0]
    lang=lines[iter].rstrip().split()[1]
    idx = lang2idx[lang]
    utt2lang[utt]=idx

lines=open('data/dev/wav.scp').readlines()
utt2wav={}
for iter in range(len(lines)):
    utt=lines[iter].rstrip().split()[0]
    wav=lines[iter].rstrip().split()[1]
    utt2wav[utt]=wav

lines=open('data/dev/segments').readlines()
wav_list = []
seg_windows = []
utt_label =[]
devid = []
for iter in range(len(lines)):
    wav_list.append( utt2wav[lines[iter].rstrip().split()[1]] )
    devid.append( lines[iter].rstrip().split()[0] )
    utt_label.append( utt2lang[lines[iter].rstrip().split()[1]] )
    seg_windows.append([ np.float(lines[iter].rstrip().split()[2]), np.float(lines[iter].rstrip().split()[3])])

    
#feature extraction
feat, _, utt_shape, tffilename = ft.feat_extract(wav_list,FEAT_TYPE,N_FFT,HOP,VAD,CMVN,0,seg_windows)


#init placeholder
test_feat_batch = tf.placeholder(tf.float32, [None,None,np.int(INPUT_DIM)],name="test_feat_batch")
test_label_batch = tf.placeholder(tf.int32, [None],name="test_label_batch")
test_shape_batch = tf.placeholder(tf.int32, [None,2],name="test_shape_batch")

#init model
emnet_validation = nn_model.nn(test_feat_batch,test_label_batch,test_label_batch,test_shape_batch, SOFTMAX_NUM,False,INPUT_DIM,is_batchnorm);
tf.get_variable_scope().reuse_variables()
sess = tf.InteractiveSession()
saver = tf.train.Saver()
tf.initialize_all_variables().run()

#load pretrained model
RESUME_STARTPOINT = 7712000
saver.restore(sess,'data/pretrained_model/model'+str(RESUME_STARTPOINT)+'.ckpt-'+str(RESUME_STARTPOINT))

#extract output layer
output = []
for iter in tqdm(range(len(feat))):
    o1 = emnet_validation.o1.eval({test_feat_batch:[feat[iter]], test_shape_batch:[utt_shape[iter]]})
    output.extend([o1])
output = np.squeeze(output)

#write result file
fid = open('result_dev.csv','w')
for row in range(output.shape[0]):
    fid.write('%s'%devid[row])    
    for col in range(output.shape[1]):
        fid.write(',%f' % output[row,col])
    fid.write('\n')
fid.close()