def main():
# gpu configuration
toolkits.initialize_GPU(args)
import model
# ==================================
# Get Train/Val.
# ==================================
print('==> calculating test({}) data lists...'.format(args.test_type))
trnlist, trnlb = toolkits.get_voxceleb2_datalist(
args, path='../meta/voxlb2_train.txt')
vallist, vallb = toolkits.get_voxceleb2_datalist(
args, path='../meta/voxlb2_val.txt')
if args.test_type == 'normal':
verify_list = np.loadtxt('../meta/voxceleb1_veri_test.txt', str)
elif args.test_type == 'hard':
verify_list = np.loadtxt('../meta/voxceleb1_veri_test_hard.txt', str)
elif args.test_type == 'extend':
verify_list = np.loadtxt('../meta/voxceleb1_veri_test_extended.txt',
str)
else:
raise IOError('==> unknown test type.')
verify_lb = np.array([int(i[0]) for i in verify_list])
list1 = np.array([os.path.join(args.data_path, i[1]) for i in verify_list])
list2 = np.array([os.path.join(args.data_path, i[2]) for i in verify_list])
total_list = np.concatenate((list1, list2))
unique_list = np.unique(total_list)
# ==================================
# Get Model
# ==================================
# construct the data generator.
params = {
# 'dim': (257, None, 1),
'dim': (args.frame_num, args.hdim, args.wdim, 3),
'nfft': 512,
'spec_len': 250,
'win_length': 400,
'hop_length': 160,
'n_classes': 5994,
'sampling_rate': 16000,
'normalize': True,
}
Dim = (args.frame_num, args.hdim, args.wdim, 3)
network_eval = model.vggvox_resnet2d_icassp(input_dim=Dim,
num_class=params['n_classes'],
mode='eval',
args=args)
# ==> load pre-trained model ???
if args.resume:
# ==> get real_model from arguments input,
# load the model if the imag_model == real_model.
if os.path.isfile(args.resume):
network_eval.load_weights(os.path.join(args.resume), by_name=True)
result_path = set_result_path(args)
print('==> successfully loading model {}.'.format(args.resume))
else:
raise IOError("==> no checkpoint found at '{}'".format(
args.resume))
else:
raise IOError('==> please type in the model to load')
print('==> start testing.')
# The feature extraction process has to be done sample-by-sample,
# because each sample is of different lengths.
total_length = len(unique_list)
feats, scores, labels = [], [], []
for c, ID in enumerate(unique_list):
if c % 50 == 0:
print('Finish extracting features for {}/{}th wav.'.format(
c, total_length))
specs = ut.load_data(ID,
dim_w=args.wdim,
dim_h=args.hdim,
win_length=params['win_length'],
sr=params['sampling_rate'],
hop_length=params['hop_length'],
n_fft=params['nfft'],
spec_len=args.frame_num,
mode='eval')
# specs = np.expand_dims(np.expand_dims(specs, 0), -1)
specs = np.expand_dims(specs, 0)
v = network_eval.predict(specs)
feats += [v]
feats = np.array(feats)
# ==> compute the pair-wise similarity.
for c, (p1, p2) in enumerate(zip(list1, list2)):
ind1 = np.where(unique_list == p1)[0][0]
ind2 = np.where(unique_list == p2)[0][0]
v1 = feats[ind1, 0]
v2 = feats[ind2, 0]
scores += [np.sum(v1 * v2)]
labels += [verify_lb[c]]
print('scores : {}, gt : {}'.format(scores[-1], verify_lb[c]))
scores = np.array(scores)
labels = np.array(labels)
np.save(os.path.join(result_path, 'prediction_scores.npy'), scores)
np.save(os.path.join(result_path, 'groundtruth_labels.npy'), labels)
eer, thresh = toolkits.calculate_eer(labels, scores)
print('==> model : {}, EER: {}'.format(args.resume, eer))
def main():
# gpu configuration
toolkits.initialize_GPU(args)
import model
import generator
import keras
# ==================================
# Get Train/Val.
# ==================================
trnlist, trnlb = toolkits.get_voxceleb2_datalist(
args, path='../meta/vox2_train_wav.txt')
vallist, vallb = toolkits.get_voxceleb2_datalist(
args, path='../meta/vox2_val_wav.txt')
# construct the data generator.
params = {
'dim': (257, 250, 1),
'mp_pooler': toolkits.set_mp(processes=args.multiprocess),
'nfft': 512,
'spec_len': 250,
'win_length': 400,
'hop_length': 160,
'n_classes': 5994,
'sampling_rate': 16000,
'batch_size': args.batch_size,
'shuffle': True,
'normalize': True,
}
# Datasets
partition = {'train': trnlist.flatten(), 'val': vallist.flatten()}
labels = {'train': trnlb.flatten(), 'val': vallb.flatten()}
# Generators
trn_gen = generator.DataGenerator(partition['train'], labels['train'],
**params)
network = model.vggvox_resnet2d_icassp(input_dim=params['dim'],
num_class=params['n_classes'],
mode='train',
args=args)
# ==> load pre-trained model ???
mgpu = len(keras.backend.tensorflow_backend._get_available_gpus())
if args.resume:
if os.path.isfile(args.resume):
if mgpu == 1: network.load_weights(os.path.join(args.resume))
else:
network.layers[mgpu + 1].load_weights(os.path.join(
args.resume))
print('==> successfully loading model {}.'.format(args.resume))
else:
print("==> no checkpoint found at '{}'".format(args.resume))
print(network.summary())
print('==> gpu {} is, training {} images, classes: 0-{} '
'loss: {}, aggregation: {}, ohemlevel: {}'.format(
args.gpu, len(partition['train']), np.max(labels['train']),
args.loss, args.aggregation_mode, args.ohem_level))
model_path, log_path = set_path(args)
normal_lr = keras.callbacks.LearningRateScheduler(step_decay)
tbcallbacks = keras.callbacks.TensorBoard(log_dir=log_path,
histogram_freq=0,
write_graph=True,
write_images=False,
update_freq=args.batch_size * 16)
callbacks = [
keras.callbacks.ModelCheckpoint(os.path.join(
model_path, 'weights-{epoch:02d}-{acc:.3f}.h5'),
monitor='loss',
mode='min',
save_best_only=True), normal_lr,
tbcallbacks
]
if args.ohem_level > 1: # online hard negative mining will be used
candidate_steps = int(len(partition['train']) // args.batch_size)
iters_per_epoch = int(
len(partition['train']) // (args.ohem_level * args.batch_size))
ohem_generator = generator.OHEM_generator(
network, trn_gen, candidate_steps, args.ohem_level,
args.batch_size, params['dim'], params['n_classes'])
A = ohem_generator.next(
) # for some reason, I need to warm up the generator
network.fit_generator(generator.OHEM_generator(
network, trn_gen, iters_per_epoch, args.ohem_level,
args.batch_size, params['dim'], params['n_classes']),
steps_per_epoch=iters_per_epoch,
epochs=args.epochs,
max_queue_size=10,
callbacks=callbacks,
use_multiprocessing=False,
workers=1,
verbose=1)
else:
network.fit_generator(trn_gen,
steps_per_epoch=int(
len(partition['train']) // args.batch_size),
epochs=args.epochs,
max_queue_size=10,
callbacks=callbacks,
use_multiprocessing=False,
workers=1,
verbose=1)
def main():
# gpu configuration
toolkits.initialize_GPU(args)
import model
import generator
# ==================================
# Get Train/Val.
# ==================================
trnlist, trnlb = toolkits.get_voxceleb2_datalist(
args, path='../meta/voxlb2_train.txt')
vallist, vallb = toolkits.get_voxceleb2_datalist(
args, path='../meta/voxlb2_val.txt')
# construct the data generator.
params = {
'dim': (257, 250, 1),
'mp_pooler': toolkits.set_mp(processes=args.multiprocess),
'nfft': 512,
'spec_len': 250,
'win_length': 400,
'hop_length': 160,
'n_classes': 5994,
'sampling_rate': 16000,
'batch_size': args.batch_size,
'shuffle': True,
'normalize': True,
}
# Datasets
#The Flatten layer is a utility layer that flattens an input of shape n * c * h * w to a simple vector output of shape n * (c*h*w)
partition = {'train': trnlist.flatten(), 'val': vallist.flatten()}
# print("partition is: ",partition)
labels = {'train': trnlb.flatten(), 'val': vallb.flatten()}
# Generators
#make data
trn_gen = generator.DataGenerator(partition['train'], labels['train'],
**params)
# create model depend on args
network = model.vggvox_resnet2d_icassp(input_dim=params['dim'],
num_class=params['n_classes'],
mode='train',
args=args)
# ==> load pre-trained model ???
mgpu = len(keras.backend.tensorflow_backend._get_available_gpus())
if args.resume:
if os.path.isfile(args.resume):
if mgpu == 1: network.load_weights(os.path.join(args.resume))
else:
network.layers[mgpu + 1].load_weights(os.path.join(
args.resume))
print('==> successfully loading model {}.'.format(args.resume))
else:
print("==> no checkpoint found at '{}'".format(args.resume))
print(network.summary())
print('==> gpu {} is, training {} images, classes: 0-{} '
'loss: {}, aggregation: {}, ohemlevel: {}'.format(
args.gpu, len(partition['train']), np.max(labels['train']),
args.loss, args.aggregation_mode, args.ohem_level))
model_path, log_path = set_path(args)
normal_lr = keras.callbacks.LearningRateScheduler(step_decay)
# tbcallbacks = keras.callbacks.TensorBoard(log_dir=log_path, histogram_freq=0, write_graph=True, write_images=False,
# update_freq=args.batch_size * 16)
'''
Keras callbacks return information from a training algorithm while training is taking place. ...
A callback is a set of functions to be applied at given stages of the training procedure.
You can use callbacks to get a view on internal states and statistics of the model during trainin
'''
# callbacks = [keras.callbacks.ModelCheckpoint(os.path.join(model_path, 'weights-{epoch:02d}-{acc:.3f}.h5'),
# monitor='loss',
# mode='min',
# save_best_only=True),
# normal_lr, tbcallbacks]
if args.ohem_level > 1: # online hard negative mining will be used
candidate_steps = int(len(partition['train']) // args.batch_size)
iters_per_epoch = int(
len(partition['train']) // (args.ohem_level * args.batch_size))
ohem_generator = generator.OHEM_generator(
network, trn_gen, candidate_steps, args.ohem_level,
args.batch_size, params['dim'], params['n_classes'])
A = ohem_generator.next(
) # for some reason, I need to warm up the generator
network.fit_generator(generator.OHEM_generator(
network, trn_gen, iters_per_epoch, args.ohem_level,
args.batch_size, params['dim'], params['n_classes']),
steps_per_epoch=iters_per_epoch,
epochs=args.epochs,
max_queue_size=10,
callbacks=callbacks,
use_multiprocessing=False,
workers=1,
verbose=1)
else:
print("steps_per_epoch=", int(len(partition['train'])), " ",
args.batch_size)
print(trn_gen)
print(network)
print("epochs=", args.epochs, "--------")
print("***************", args.batch_size)
network.fit_generator(
trn_gen,
steps_per_epoch=2, #int(len(partition['train'])//args.batch_size),
epochs=args.epochs,
max_queue_size=2,
# callbacks=callbacks,
use_multiprocessing=True,
workers=1,
verbose=1) #should change to one?
print("end!")