import sys
import numpy as np
import triplet_loss
from batcher import KerasFormatConverter, TripletBatcherSelectHardNegatives, TripletBatcher
from constants import NUM_FBANKS, NUM_FRAMES, CHECKPOINTS_TRIPLET_DIR, CHECKPOINTS_SOFTMAX_DIR, BATCH_SIZE
from conv_models import DeepSpeakerModel
from triplet_loss import deep_speaker_loss
from utils import load_best_checkpoint
def predict(x):
y = np.tile(np.expand_dims(np.mean(x, axis=(1, 2, 3)), axis=1), (1, 512))
return y
# norm_y = np.linalg.norm(y, axis=1, ord=2, keepdims=True)
# return y / (norm_y + 1e-12)
def main2():
num_utterances_per_speaker = 50
num_speakers = 100
num_samples = num_speakers * num_utterances_per_speaker
kx_train = np.zeros(shape=(num_samples, 32, 64, 1))
ky_train = np.zeros(shape=(num_samples, num_speakers))
for i in range(num_samples):
speaker_id = i % num_speakers
ky_train[i, speaker_id] = 1
kx_train[i] = speaker_id
kx_test = np.array(kx_train)
ky_test = np.array(ky_train)
tpshn = TripletBatcherSelectHardNegatives(kx_train, ky_train, kx_test, ky_test, None)
tp = TripletBatcher(kx_train, ky_train, kx_test, ky_test)
avg = []
avg2 = []
while True:
bx, by = tp.get_batch(BATCH_SIZE, is_test=False)
avg.append(float(triplet_loss.deep_speaker_loss(predict(bx), predict(bx))))
bx, by = tpshn.get_batch(BATCH_SIZE, is_test=False, predict=predict)
avg2.append(float(triplet_loss.deep_speaker_loss(predict(bx), predict(bx))))
print(np.mean(avg), np.mean(avg2))
def main():
select = True
try:
sys.argv[1]
except:
select = False
print('select', select)
working_dir = '/media/philippe/8TB/deep-speaker'
# by construction this losses should be much higher than the normal losses.
# we select batches this way.
batch_input_shape = [None, NUM_FRAMES, NUM_FBANKS, 1]
print('Testing with the triplet losses.')
dsm = DeepSpeakerModel(batch_input_shape, include_softmax=False)
triplet_checkpoint = load_best_checkpoint(CHECKPOINTS_TRIPLET_DIR)
pre_training_checkpoint = load_best_checkpoint(CHECKPOINTS_SOFTMAX_DIR)
if triplet_checkpoint is not None:
print(f'Loading triplet checkpoint: {triplet_checkpoint}.')
dsm.m.load_weights(triplet_checkpoint)
elif pre_training_checkpoint is not None:
print(f'Loading pre-training checkpoint: {pre_training_checkpoint}.')
# If `by_name` is True, weights are loaded into layers only if they share the
# same name. This is useful for fine-tuning or transfer-learning models where
# some of the layers have changed.
dsm.m.load_weights(pre_training_checkpoint, by_name=True)
dsm.m.compile(optimizer='adam', loss=deep_speaker_loss)
kc = KerasFormatConverter(working_dir)
if select:
print('TripletBatcherSelectHardNegatives()')
batcher = TripletBatcherSelectHardNegatives(kc.kx_train, kc.ky_train, kc.kx_test, kc.ky_test, dsm)
else:
print('TripletBatcher()')
batcher = TripletBatcher(kc.kx_train, kc.ky_train, kc.kx_test, kc.ky_test)
batch_size = BATCH_SIZE
losses = []
while True:
_bx, _by = batcher.get_batch(batch_size, is_test=False)
losses.append(dsm.m.evaluate(_bx, _by, verbose=0, batch_size=BATCH_SIZE))
print(np.mean(losses))
if __name__ == '__main__':
main2()
