def eval_model(working_dir: str, model: DeepSpeakerModel):
enable_deterministic()
audio = Audio(working_dir)
batcher = LazyTripletBatcher(working_dir, NUM_FRAMES, model)
speakers_list = list(audio.speakers_to_utterances.keys())
num_negative_speakers = 99
num_speakers = len(speakers_list)
y_pred = np.zeros(shape=(num_speakers, num_negative_speakers +
1)) # negatives + positive
for i, positive_speaker in tqdm(enumerate(speakers_list),
desc='test',
total=num_speakers):
# convention id[0] is anchor speaker, id[1] is positive, id[2:] are negative.
input_data = batcher.get_speaker_verification_data(
positive_speaker, num_negative_speakers)
# batch size is not relevant. just making sure we don't push too much on the GPU.
predictions = model.m.predict(input_data, batch_size=BATCH_SIZE)
anchor_embedding = predictions[0]
for j, other_than_anchor_embedding in enumerate(
predictions[1:]): # positive + negatives
y_pred[i][j] = batch_cosine_similarity(
[anchor_embedding], [other_than_anchor_embedding])[0]
# y_pred[i] = softmax(y_pred[i])
# could apply softmax here.
y_true = np.zeros_like(y_pred) # positive is at index 0.
y_true[:, 0] = 1.0
print(np.matrix(y_true))
print(np.matrix(y_pred))
print(np.min(y_pred), np.max(y_pred))
fm, tpr, acc, eer = evaluate(y_pred, y_true)
return fm, tpr, acc, eer
def fit_model(dsm: DeepSpeakerModel,
working_dir: str,
max_length: int = NUM_FRAMES,
batch_size=BATCH_SIZE):
batcher = LazyTripletBatcher(working_dir, max_length, dsm)
# build small test set.
test_batches = []
for _ in tqdm(range(200), desc='Build test set'):
test_batches.append(batcher.get_batch_test(batch_size))
def test_generator():
while True:
for bb in test_batches:
yield bb
def train_generator():
while True:
yield batcher.get_random_batch(batch_size, is_test=False)
checkpoint_name = dsm.m.name + '_checkpoint'
checkpoint_filename = os.path.join(CHECKPOINTS_TRIPLET_DIR,
checkpoint_name + '_{epoch}.h5')
checkpoint = ModelCheckpoint(monitor='val_loss',
filepath=checkpoint_filename,
save_best_only=True)
dsm.m.fit(x=train_generator(),
y=None,
steps_per_epoch=2000,
shuffle=False,
epochs=1000,
validation_data=test_generator(),
validation_steps=len(test_batches),
callbacks=[checkpoint])
def main2():
batch_input_shape = [None, NUM_FRAMES, NUM_FBANKS, 1]
dsm = DeepSpeakerModel(batch_input_shape, include_softmax=False)
dsm.m.compile(optimizer='adam', loss=deep_speaker_loss)
dsm.m.load_weights('/Users/premy/deep-speaker/ResCNN_checkpoint_102.h5',
by_name=True)
dsm.m.summary()
batcher = LazyTripletBatcher(working_dir='/Users/premy/deep-speaker',
max_length=NUM_FRAMES,
model=dsm)
bs = 18
print(
np.mean([
dsm.m.evaluate(*batcher.get_batch_train(batch_size=bs),
batch_size=bs,
verbose=0) for _ in range(100)
]))
print(
np.mean([
dsm.m.evaluate(*batcher.get_batch_test(batch_size=bs),
batch_size=bs,
verbose=0) for _ in range(100)
]))
print(
np.mean([
dsm.m.evaluate(*batcher.get_random_batch(batch_size=bs,
is_test=False),
batch_size=bs,
verbose=0) for _ in range(100)
]))
print(
np.mean([
dsm.m.evaluate(*batcher.get_random_batch(batch_size=bs,
is_test=True),
batch_size=bs,
verbose=0) for _ in range(100)
]))
def fit_model(dsm: DeepSpeakerModel,
working_dir: str,
max_length: int = NUM_FRAMES,
batch_size=BATCH_SIZE,
epochs=1000,
classify=False,
initial_epoch=0):
batcher = LazyTripletBatcher(working_dir,
max_length,
dsm,
classify=classify)
# build small test set.
test_batches = []
for _ in tqdm(range(200), desc='Build test set'):
test_batches.append(
batcher.get_batch_test(batch_size, classify=classify))
def test_generator():
while True:
for bb in test_batches:
yield bb
def train_generator():
while True:
yield batcher.get_random_batch(batch_size,
is_test=False,
classify=classify)
checkpoint_name = dsm.m.name + '_checkpoint'
if classify:
checkpoint_filename = os.path.join(CHECKPOINTS_CLASSIFY_DIR,
checkpoint_name + '_{epoch}.h5')
else:
checkpoint_filename = os.path.join(CHECKPOINTS_TRIPLET_DIR,
checkpoint_name + '_{epoch}.h5')
checkpoint = ModelCheckpoint(monitor='val_loss',
filepath=checkpoint_filename,
save_best_only=True)
early_stopping = EarlyStopping(monitor='loss',
min_delta=0.001,
patience=20,
verbose=1)
dsm.m.fit_generator(train_generator(),
steps_per_epoch=2000,
shuffle=False,
epochs=epochs,
validation_data=test_generator(),
validation_steps=len(test_batches),
initial_epoch=initial_epoch,
callbacks=[checkpoint, early_stopping])
mfccs, y = load_data(os.path.join(working_dir, 'samples/train'))
features = dsm.m.predict(mfccs)
clf = svm.SVC()
clf.fit(features, y)
svm_pickle = open('svm.pkl', 'wb')
pickle.dump(clf, svm_pickle)
svm_pickle.close()