def train(args):
workspace = args.workspace
# Load data.
t1 = time.time()
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram",
"data.h5")
tr_y = pp_data.load_hdf5(tr_hdf5_path)
tr_y = np.array(tr_y)[1]
print("data shape is {}".format(tr_y.shape))
print("Load data time: %s s" % (time.time() - t1, ))
# scaler data
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram",
"scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_y = pp_data.scale_on_2d(tr_y, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
# batch
batch_size = 500
print("%d iterations / epoch" % (tr_y.shape[0] / batch_size))
# data shape and input shape
(n_segs, n_freq) = tr_y.shape
def detect(args):
import log
val_loss = log.history["val_loss"]
val_acc = log.history["val_Acc"]
choose = []
for i in range(1):
'''
min_loss = np.argmin(val_loss)
choose.append(min_loss)
val_loss[min_loss] = np.inf
'''
max_acc = np.argmax(val_acc)
choose.append(max_acc)
val_acc[max_acc] = 0
(te_x, te_y, te_na_list) = load_hdf5(args.te_hdf5_path, verbose=1)
x = te_x
y = te_y
na_list = te_na_list
x = do_scale(x, args.scaler_path, verbose=1)
fusion_at_list = []
fusion_sed_list = []
#choose=[48]
for epoch in choose:
t1 = time.time()
[model_path
] = glob.glob(os.path.join(args.model_dir, "*-%04d-*hdf5" % epoch))
model = load_model(model_path,
custom_objects={
'focal_loss_fixed': focal_loss(),
'Acc': myacc()
})
print("load the model: %s" % model_path)
# Audio tagging
pred = model.predict(x)
fusion_at_list.append(pred)
print("Prediction time: %s" % (time.time() - t1, ))
# Write out AT probabilities
fusion_at = np.mean(np.array(fusion_at_list), axis=0)
print("AT shape: %s" % (fusion_at.shape, ))
if os.path.exists("result"):
shutil.rmtree("result")
for audio_ind in range(fusion_at.shape[0]):
#if na_list[audio_ind]=="mixture_babycry_0.0_0016_65021d74d0fb56db84b63896e2ff5ec9.wav":
# picture_path=os.path.join("result","picture",na_list[audio_ind].replace("wav","jpg"))
# pdb.set_trace()
# my_plot(fusion_at[audio_ind,...], y[audio_ind,...], picture_path)
picture_path = os.path.join("result", "picture",
na_list[audio_ind].replace("wav", "jpg"))
#my_plot(fusion_at[audio_ind,...], y[audio_ind,...], picture_path)
my_plot(fusion_at[audio_ind, ...],
y[audio_ind, ...],
picture_path,
threshold=cfg.threshold)
print("Prediction finished!")
def train(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
print(args)
workspace = args.workspace
tr_snr = args.tr_snr
te_snr = args.te_snr
lr = args.lr
iteration = args.iter
# Load data.
t1 = time.time()
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", "%ddb" % int(tr_snr), "data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram",
"test", "%ddb" % int(te_snr), "data.h5")
(tr_x1, tr_x2, tr_y1, tr_y2) = pp_data.load_hdf5(tr_hdf5_path)
(te_x1, te_x2, te_y1, te_y2) = pp_data.load_hdf5(te_hdf5_path)
print(tr_x1.shape, tr_y1.shape, tr_x2.shape, tr_y2.shape)
print(te_x1.shape, te_y1.shape, te_x2.shape, te_y2.shape)
print("Load data time: %s s" % (time.time() - t1, ))
batch_size = 500
print("%d iterations / epoch" % int(tr_x1.shape[0] / batch_size))
# Scale data.
if not True:
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", "%ddb" % int(tr_snr), "scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x1 = pp_data.scale_on_3d(tr_x1, scaler)
tr_y1 = pp_data.scale_on_2d(tr_y1, scaler)
te_x1 = pp_data.scale_on_3d(te_x1, scaler)
te_y1 = pp_data.scale_on_2d(te_y1, scaler)
tr_x2 = pp_data.scale_on_2d(tr_x2, scaler)
tr_y2 = pp_data.scale_on_2d(tr_y2, scaler)
te_x2 = pp_data.scale_on_2d(te_x2, scaler)
te_y2 = pp_data.scale_on_2d(te_y2, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
# Debug plot.
if False:
plt.matshow(tr_x[0:1000, 0, :].T,
origin='lower',
aspect='auto',
cmap='jet')
plt.show()
pause
# Build model
(_, n_concat, n_freq) = tr_x1.shape
n_hid = 2048
input_dim1 = (257 + 40 + 30) * 2
input_dim2 = (257 + 40 + 30)
out_dim1 = (257 + 40 + 30) * 2
out_dim1_irm = 257 + 40 + 64
out_dim2 = (257 + 40 + 30)
out_dim2_irm = (257 + 40 + 64)
# model = Sequential()
# model.add(Flatten(input_shape=(n_concat, n_freq)))
# model.add(Dense(n_hid, activation='relu'))
# model.add(Dropout(0.2))
# model.add(Dense(n_hid, activation='relu'))
# model.add(Dropout(0.2))
# model.add(Dense(n_hid, activation='relu'))
# model.add(Dropout(0.2))
# model.add(Dense(n_freq, activation='linear'))
input1 = Input(shape=(n_concat, input_dim1), name='input1')
layer = Flatten(name='flatten')(input1)
layer = Dense(n_hid, activation='relu', name='dense1')(layer)
layer = Dropout(0.2)(layer)
layer = Dense(n_hid, activation='relu', name='dense2')(layer)
layer = Dropout(0.2)(layer)
partial_out1 = Dense(out_dim1, name='1_out_linear')(layer)
partial_out1_irm = Dense(out_dim1_irm,
name='1_out_irm',
activation='sigmoid')(layer)
out1 = concatenate([partial_out1, partial_out1_irm], name='out1')
input2 = Input(shape=(input_dim2, ), name='input2')
layer = concatenate([input2, out1], name='merge')
layer = Dense(n_hid, activation='relu', name='dense3')(layer)
layer = Dropout(0.2)(layer)
layer = Dense(n_hid, activation='relu', name='dense4')(layer)
layer = Dropout(0.2)(layer)
partial_out2 = Dense(out_dim2, name='2_out_linear')(layer)
partial_out2_irm = Dense(out_dim2_irm,
name='2_out_irm',
activation='sigmoid')(layer)
out2 = concatenate([partial_out2, partial_out2_irm], name='out2')
model = Model(inputs=[input1, input2], outputs=[out1, out2])
model.summary()
sys.stdout.flush()
model.compile(loss='mean_absolute_error',
optimizer=Adam(lr=lr, epsilon=1e-03))
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
eval_te_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
eval_tr_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models", "%ddb" % int(tr_snr))
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats", "%ddb" % int(tr_snr))
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x1, tr_x2, tr_y1, tr_y2)
te_loss = eval(model, eval_te_gen, te_x1, te_x2, te_y1, te_y2)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x1, tr_x2],
ys=[tr_y1, tr_y2]):
loss = model.train_on_batch(batch_x, batch_y)
iter += 1
# Validate and save training stats.
if iter % 100 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x1, tr_x2, tr_y1, tr_y2)
te_loss = eval(model, eval_te_gen, te_x1, te_x2, te_y1, te_y2)
print("Iteration: %d, tr_loss: %f, te_loss: %f" %
(iter, tr_loss, te_loss))
sys.stdout.flush()
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % (iteration / 20) == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == iteration + 1:
break
print("Training time: %s s" % (time.time() - t1, ))
def train(args):
if os.path.exists(args.out_model_dir):
shutil.rmtree(args.out_model_dir)
create_folder(args.out_model_dir)
num_classes = cfg.num_classes
# Load training & testing data
(tr_x, tr_y, tr_na_list) = load_hdf5(args.tr_hdf5_path, verbose=1)
(te_x, te_y, te_na_list) = load_hdf5(args.te_hdf5_path, verbose=1)
print("")
# Scale data
tr_x = do_scale(tr_x, args.scaler_path, verbose=1)
te_x = do_scale(te_x, args.scaler_path, verbose=1)
# Build model
(_, n_time, n_freq) = tr_x.shape
#pdb.set_trace()
input = Input(shape=(n_time, n_freq), name='input_layer')
input_ = Reshape((n_time, n_freq, 1))(input)
'''
block1 = Conv_BN(input_, 8, (3, 3), act="relu")
block1 = Conv_BN(block1, 32, (3, 3), act="relu")
block1 = Conv_BN(block1, 64, (3, 3), act="relu")
block1 = block_a(input_, 8)
block1 = block_a(block1, 32)
block1 = block_a(block1, 64)
'''
block1 = block_b(input_, 8)
block1 = block_b(block1, 32)
block1 = block_b(block1, 64)
block1 = MaxPooling2D(pool_size=(1, 2))(block1)
block2 = block_c(block1, 64)
block2 = MaxPooling2D(pool_size=(1, 2))(block2)
block3 = block_c(block2, 64)
block3 = MaxPooling2D(pool_size=(1, 2))(block3)
block4 = block_c(block3, 64)
block4 = MaxPooling2D(pool_size=(1, 2))(block4)
cnnout = Conv_BN(block4, 128, (1, 1), act="relu", bias=True)
cnnout = MaxPooling2D(pool_size=(1, 2))(cnnout)
cnnout = Reshape((240, 256))(cnnout)
rnn = Bidirectional(
GRU(128,
activation='relu',
return_sequences=True,
kernel_regularizer=regularizers.l2(0.01),
recurrent_regularizer=regularizers.l2(0.01)))(cnnout)
out = TimeDistributed(Dense(
num_classes,
activation='softmax',
kernel_regularizer=regularizers.l2(0.0),
),
name='output_layer')(rnn)
model = Model(input, out)
model.summary()
# Compile model
adam = optimizers.Adam(lr=0.001, beta_1=0.9, beta_2=0.999, decay=0.009)
sgd = optimizers.SGD(lr=0.01, momentum=0.9, decay=0.0)
model.compile(loss=focal_loss(alpha=[1, 1, 1, 1], gamma=1),
optimizer="adam",
metrics=[myacc(threshold=0.5)])
# Save model callback
filepath = os.path.join(
args.out_model_dir,
"aed-batchsize_50-lr_0.01-{epoch:04d}-{val_Acc:.4f}.hdf5")
save_model = ModelCheckpoint(filepath=filepath,
monitor='val_Acc',
verbose=0,
save_best_only=False,
save_weights_only=False,
mode='auto',
period=1)
# Train
'''
history=model.fit( x=tr_x,
y=tr_y,
batch_size=50,
epochs=200,
verbose=1,
shuffle=True,
class_weight="auto",
callbacks=[save_model],
validation_data=(te_x,te_y)
)
'''
# Data generator
gen = Generator(batch_size=50, type='train')
history = model.fit_generator(
generator=gen.generate([tr_x], [tr_y]),
steps_per_epoch=300, # 100 iters is called an 'epoch'
epochs=100, # Maximum 'epoch' to train
verbose=1,
class_weight="auto",
callbacks=[save_model],
validation_data=(te_x, te_y))
with open('src/log.py', 'w') as f:
f.write("history=")
f.write(str(history.history))
def train(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
print(args)
workspace = args.workspace
tr_snr = args.tr_snr
te_snr = args.te_snr
lr = args.lr
# Load data.
t1 = time.time()
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "%ddb" % int(tr_snr), "data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "test", "%ddb" % int(te_snr), "data.h5")
(tr_x, tr_y) = pp_data.load_hdf5(tr_hdf5_path)
(te_x, te_y) = pp_data.load_hdf5(te_hdf5_path)
print(tr_x.shape, tr_y.shape)
print(te_x.shape, te_y.shape)
print("Load data time: %s s" % (time.time() - t1,))
batch_size = 128
print("%d iterations / epoch" % int(tr_x.shape[0] / batch_size))
# Build model
_, n_freq = tr_x.shape
# encode
T = 1
data = Input(shape=[n_freq])
x = Reshape([1, T, n_freq])(data)
x1 = Conv2D(10, (T, 11), strides=(10, 1), data_format='channels_first', padding='same')(x)
x1 = BatchNormalization(axis=-1)(x1)
x1 = Activation('relu')(x1)
x2 = Conv2D(12, (T, 7), strides=(10, 1), data_format='channels_first', padding='same')(x1)
x2 = BatchNormalization(axis=-1)(x2)
x2 = Activation('relu')(x2)
x3 = Conv2D(14, (T, 5), strides=(10, 1), data_format='channels_first', padding='same')(x2)
x3 = BatchNormalization(axis=-1)(x3)
x3 = Activation('relu')(x3)
x4 = Conv2D(15, (T, 5), strides=(10, 1), data_format='channels_first', padding='same')(x3)
x4 = BatchNormalization(axis=-1)(x4)
x4 = Activation('relu')(x4)
x5 = Conv2D(19, (1, 5), strides=(10, 1), data_format='channels_first', padding='same')(x4)
x5 = BatchNormalization(axis=-1)(x5)
x5 = Activation('relu')(x5)
x6 = Conv2D(21, (1, 5), strides=(10, 1), data_format='channels_first', padding='same')(x5)
x6 = BatchNormalization(axis=-1)(x6)
x6 = Activation('relu')(x6)
x7 = Conv2D(23, (1, 7), strides=(10, 1), data_format='channels_first', padding='same')(x6)
x7 = BatchNormalization(axis=-1)(x7)
x7 = Activation('relu')(x7)
x8 = Conv2D(25, (1, 11), strides=(10, 1), data_format='channels_first', padding='same')(x7)
x8 = BatchNormalization(axis=-1)(x8)
x8 = Activation('relu')(x8)
# decode
y1 = Conv2D(23, (1, 7), strides=(10, 1), data_format='channels_first', padding='same')(x8)
y1 = Add()([y1, x7])
y1 = BatchNormalization(axis=-1)(y1)
y1 = Activation('relu')(y1)
y2 = Conv2D(21, (1, 5), strides=(10, 1), data_format='channels_first', padding='same')(y1)
y2 = Add()([y2, x6])
y2 = BatchNormalization(axis=-1)(y2)
y2 = Activation('relu')(y2)
y3 = Conv2D(19, (1, 5), strides=(10, 1), data_format='channels_first', padding='same')(y2)
y3 = Add()([y3, x5])
y3 = BatchNormalization(axis=-1)(y3)
y3 = Activation('relu')(y3)
y4 = Conv2D(15, (1, 5), strides=(10, 1), data_format='channels_first', padding='same')(y3)
y4 = Add()([y4, x4])
y4 = BatchNormalization(axis=-1)(y4)
y4 = Activation('relu')(y4)
y5 = Conv2D(14, (1, 5), strides=(10, 1), data_format='channels_first', padding='same')(y4)
y5 = Add()([y5, x3])
y5 = BatchNormalization(axis=-1)(y5)
y5 = Activation('relu')(y5)
y6 = Conv2D(12, (1, 7), strides=(10, 1), data_format='channels_first', padding='same')(y5)
y6 = Add()([y6, x2])
y6 = BatchNormalization(axis=-1)(y6)
y6 = Activation('relu')(y6)
y7 = Conv2D(10, (1, 11), strides=(10, 1), data_format='channels_first', padding='same')(y6)
y7 = Add()([y7, x1])
y7 = BatchNormalization(axis=-1)(y7)
y7 = Activation('relu')(y7)
y8 = Conv2D(1, (1, n_freq), strides=(10, 1), data_format='channels_first', padding='same')(y7)
# y5 = BatchNormalization(axis=-1)(y5)
y8 = Activation('relu')(y8)
out = Reshape([n_freq])(y8)
model = Model(inputs=data, outputs=out)
adam = optimizers.Adam(lr=lr)
model.compile(loss='mean_absolute_error', optimizer=adam)
model.summary()
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
te_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=200)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models", "%ddb" % int(tr_snr))
pp_data.create_folder(model_dir)
# Train.
t1 = time.time()
model.fit_generator(tr_gen.generate(xs=[tr_x], ys=[tr_y]), validation_data=te_gen.generate(xs=[te_x], ys=[te_y]),
validation_steps=100, steps_per_epoch=200, epochs=200)
print("Training complete.")
model_name = 'FullyCNN.h5'
model_path = os.path.join(model_dir, model_name)
model.save(model_path)
print("Training time: %s s" % (time.time() - t1,))
h5_train_list = [
f for f in sorted(os.listdir(conf1.data_train_dir)) if f.endswith('.h5')
]
h5_test_list = [
f for f in os.listdir(conf1.data_test_dir) if f.endswith('.h5')
]
tr_x = []
tr_y = []
te_x = []
te_y = []
for i in h5_test_list:
te_x_t, te_y_t = pp.load_hdf5(os.path.join(conf1.data_test_dir, i))
te_x.append(te_x_t)
te_y.append(te_y_t)
te_x = np.concatenate(te_x, axis=0)
te_y = np.concatenate(te_y, axis=0)
#scale test data
scaler = pickle.load(
open(os.path.join(conf1.packed_feature_dir, 'test', 'scaler.p'), 'rb'))
te_x = pp.scale_on_3d(te_x, scaler)
te_y = pp.scale_on_2d(te_y, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
print("Load data time: %s s" % (time.time() - t1, ))
def train(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
print(args)
workspace = args.workspace
tr_snr = args.tr_snr
te_snr = args.te_snr
lr = args.lr
iteration = args.iter
# Load data.
t1 = time.time()
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "%ddb" % int(tr_snr), "data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "test", "%ddb" % int(te_snr), "data.h5")
tr_adapt_utt_path = os.path.join(workspace, "adaptive_utterance", "train", "adaptive_utterance_spec.p")
te_adapt_utt_path = os.path.join(workspace, "adaptive_utterance", "test", "adaptive_utterance_spec.p")
tr_adapt_utt = cPickle.load(open(tr_adapt_utt_path, 'rb'))
te_adapt_utt = cPickle.load(open(te_adapt_utt_path, 'rb'))
tr_adapt_utt_len_path = os.path.join(workspace, "adaptive_utterance", "train", "adaptive_utterance_max_len.p")
te_adapt_utt_len_path = os.path.join(workspace, "adaptive_utterance", "test", "adaptive_utterance_max_len.p")
tr_adapt_utt_len = cPickle.load(open(tr_adapt_utt_len_path, 'rb'))
te_adapt_utt_len = cPickle.load(open(te_adapt_utt_len_path, 'rb'))
max_len = max(tr_adapt_utt_len, te_adapt_utt_len)
(tr_x1, tr_x2, tr_y1, tr_y2, tr_name) = pp_data.load_hdf5(tr_hdf5_path)
(te_x1, te_x2, te_y1, te_y2, te_name) = pp_data.load_hdf5(te_hdf5_path)
print(tr_x1.shape, tr_y1.shape, tr_x2.shape, tr_y2.shape)
print(te_x1.shape, te_y1.shape, te_x2.shape, te_y2.shape)
print("Load data time: %s s" % (time.time() - t1,))
batch_size = 500
print("%d iterations / epoch" % int(tr_x1.shape[0] / batch_size))
# Scale data.
if not True:
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "%ddb" % int(tr_snr),
"scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x1 = pp_data.scale_on_3d(tr_x1, scaler)
tr_y1 = pp_data.scale_on_2d(tr_y1, scaler)
te_x1 = pp_data.scale_on_3d(te_x1, scaler)
te_y1 = pp_data.scale_on_2d(te_y1, scaler)
tr_x2 = pp_data.scale_on_2d(tr_x2, scaler)
tr_y2 = pp_data.scale_on_2d(tr_y2, scaler)
te_x2 = pp_data.scale_on_2d(te_x2, scaler)
te_y2 = pp_data.scale_on_2d(te_y2, scaler)
print("Scale data time: %s s" % (time.time() - t1,))
# Debug plot.
if False:
plt.matshow(tr_x[0: 1000, 0, :].T, origin='lower', aspect='auto', cmap='jet')
plt.show()
pause
# Build model
(_, n_concat, n_freq) = tr_x1.shape
n_hid = 2048
input_dim1 = (257 + 40 + 30) * 2
input_dim2 = (257 + 40 + 30)
out_dim1 = (257 + 40 + 30) * 2
out_dim1_irm = 257 + 40 + 64
out_dim2 = (257 + 40 + 30)
out_dim2_irm = (257 + 40 + 64)
num_factorize = 30
def multiplication(pair_tensors):
'''
:param pair_tensors: x: (num_factorize,)
y: (num_factorize, n_hid)
:return: (n_hid,) sum(x[i]*y[i,:],axis=1)
'''
x, y = pair_tensors
return K.sum(tf.multiply(y, K.expand_dims(x, -1)), axis=1)
adapt_input = Input(shape=(None,), name='adapt_input')
layer = Reshape((-1, 257), name='reshape')(adapt_input)
layer = Dense(512, activation='relu', name='adapt_dense1')(layer)
layer = Dense(512, activation='relu', name='adapt_dense2')(layer)
layer = Dense(num_factorize, activation='softmax', name='adapt_out')(layer)
alpha = Lambda(lambda x: K.sum(x, axis=1), output_shape=(num_factorize,), name='sequence_sum')(layer)
input1 = Input(shape=(n_concat, input_dim1), name='input1')
layer = Flatten(name='flatten')(input1)
layer = Dense(n_hid * num_factorize, name='dense0')(layer)
layer = Reshape((num_factorize, n_hid), name='reshape2')(layer)
layer = Lambda(multiplication, name='multiply')([alpha, layer])
layer = Dense(n_hid, activation='relu', name='dense1')(layer)
layer = Dropout(0.2)(layer)
layer = Dense(n_hid, activation='relu', name='dense2')(layer)
layer = Dropout(0.2)(layer)
partial_out1 = Dense(out_dim1, name='1_out_linear')(layer)
partial_out1_irm = Dense(out_dim1_irm, name='1_out_irm', activation='sigmoid')(layer)
out1 = concatenate([partial_out1, partial_out1_irm], name='out1')
input2 = Input(shape=(input_dim2,), name='input2')
layer = concatenate([input2, out1], name='merge')
layer = Dense(n_hid, activation='relu', name='dense3')(layer)
layer = Dropout(0.2)(layer)
layer = Dense(n_hid, activation='relu', name='dense4')(layer)
layer = Dropout(0.2)(layer)
partial_out2 = Dense(out_dim2, name='2_out_linear')(layer)
partial_out2_irm = Dense(out_dim2_irm, name='2_out_irm', activation='sigmoid')(layer)
out2 = concatenate([partial_out2, partial_out2_irm], name='out2')
model = Model(inputs=[input1, input2, adapt_input], outputs=[out1, out2])
model.summary()
sys.stdout.flush()
model.compile(loss='mean_absolute_error',
optimizer=Adam(lr=lr, epsilon=1e-03))
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train', max_len=max_len)
eval_te_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100, max_len=max_len)
eval_tr_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100, max_len=max_len)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models", "%ddb" % int(tr_snr))
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats", "%ddb" % int(tr_snr))
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x1, tr_x2, tr_y1, tr_y2, tr_name, tr_adapt_utt)
te_loss = eval(model, eval_te_gen, te_x1, te_x2, te_y1, te_y2, te_name, te_adapt_utt)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate([tr_x1, tr_x2, tr_name], [tr_y1, tr_y2], tr_adapt_utt):
loss = model.train_on_batch(batch_x, batch_y)
iter += 1
# Validate and save training stats.
if iter % 100 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x1, tr_x2, tr_y1, tr_y2, tr_name, tr_adapt_utt)
te_loss = eval(model, eval_te_gen, te_x1, te_x2, te_y1, te_y2, te_name, te_adapt_utt)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
sys.stdout.flush()
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % (iteration / 20) == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == iteration + 1:
break
print("Training time: %s s" % (time.time() - t1,))
def train_tfrecords(args):
lr = args.lr
# Load data.
t1 = time.time()
tr_hdf5_dir = os.path.join("workspace", "tfrecords", "train", "crn_mixdb")
tr_hdf5_names = os.listdir(tr_hdf5_dir)
tr_path_list = [os.path.join(tr_hdf5_dir, i) for i in tr_hdf5_names]
te_hdf5_path = os.path.join("workspace", "packed_features", "spectrogram",
"test", "crn_mixdb", "data.h5")
(te_x, te_y) = pp_data.load_hdf5(te_hdf5_path)
print("test.h5 loaded ! ! !")
train_path = os.path.join("workspace", "packed_features", "spectrogram",
"train", "crn_mixdb", "data.h5")
(tr_x, tr_y) = pp_data.load_hdf5(train_path)
print("train.h5 loaded ! ! !")
batch_size = 1024
# Scale data.
t1 = time.time()
input_x = Input(shape=(11, 161))
reshape_x = Reshape((1, 11, 161), input_shape=(11, 161))(input_x)
l1_input = ZeroPadding2D(padding=((1, 0), (0, 0)),
data_format="channels_first")(reshape_x)
l1 = Conv2D(filters=16,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l1_input)
l1 = BatchNormalization()(l1)
l1 = ELU()(l1)
l2_input = ZeroPadding2D(padding=((1, 0), (0, 0)),
data_format="channels_first")(l1)
l2 = Conv2D(filters=32,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l2_input)
l2 = BatchNormalization()(l2)
l2 = ELU()(l2)
l3_input = ZeroPadding2D(padding=((1, 0), (0, 0)),
data_format="channels_first")(l2)
l3 = Conv2D(filters=64,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l3_input)
l3 = BatchNormalization()(l3)
l3 = ELU()(l3)
l4_input = ZeroPadding2D(padding=((1, 0), (0, 0)),
data_format="channels_first")(l3)
l4 = Conv2D(filters=128,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l4_input)
l4 = BatchNormalization()(l4)
l4 = ELU()(l4)
l5_input = ZeroPadding2D(padding=((1, 0), (0, 0)),
data_format="channels_first")(l4)
l5 = Conv2D(filters=256,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l5_input)
l5 = BatchNormalization()(l5)
l5 = ELU()(l5)
reshape_x2 = Reshape((11, 4 * 256), input_shape=(11, 4, 256))(l5)
lstm1 = LSTM(units=4 * 256, activation='tanh',
return_sequences=True)(reshape_x2)
lstm2 = LSTM(units=4 * 256, activation='tanh',
return_sequences=True)(lstm1)
reshape_x3 = Reshape((256, 11, 4), input_shape=(11, 4 * 256))(lstm2)
l8_input = Concatenate(axis=1)([reshape_x3, l5])
l8 = Conv2DTranspose(filters=128,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l8_input)
l8 = Cropping2D(cropping=((1, 0), (0, 0)),
data_format="channels_first")(l8)
l8 = BatchNormalization()(l8)
l8 = ELU()(l8)
l9_input = Concatenate(axis=1)([l8, l4])
l9 = Conv2DTranspose(filters=64,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l9_input)
l9 = Cropping2D(cropping=((1, 0), (0, 0)),
data_format="channels_first")(l9)
l9 = BatchNormalization()(l9)
l9 = ELU()(l9)
l10_input = Concatenate(axis=1)([l9, l3])
l10 = Conv2DTranspose(filters=32,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l10_input)
l10 = Cropping2D(cropping=((1, 0), (0, 0)),
data_format="channels_first")(l10)
l10 = BatchNormalization()(l10)
l10 = ELU()(l10)
l11_input = Concatenate(axis=1)([l10, l2])
l11_input = ZeroPadding2D(padding=((0, 0), (1, 0)),
data_format="channels_first")(l11_input)
l11 = Conv2DTranspose(filters=16,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l11_input)
l11 = Cropping2D(cropping=((1, 0), (1, 0)),
data_format="channels_first")(l11)
l11 = BatchNormalization()(l11)
l11 = ELU()(l11)
l12_input = Concatenate(axis=1)([l11, l1])
l12 = Conv2DTranspose(filters=1,
kernel_size=(2, 3),
strides=(1, 2),
activation=None,
data_format="channels_first",
padding="valid")(l12_input)
l12 = Cropping2D(cropping=((1, 0), (0, 0)),
data_format="channels_first")(l12)
l12 = Reshape((11, 161), input_shape=(11, 161, 1))(l12)
l12 = Lambda(lambda x: keras.activations.softplus(x))(l12)
#l12 = keras.layers.Lambda(lambda x:keras.activations.softplus(x))(l12)
model = keras.models.Model(inputs=[input_x], outputs=l8)
model.summary()
#lr = 5e-5
#model_path = os.path.join(workspace, "models", "crn_mixdb", "md_%diters.h5" % 3935)
#model = load_model(model_path, custom_objects={'tf': tf})
#model = multi_gpu_model(model, 4)
model.compile(loss='mean_absolute_error',
optimizer=Adam(lr=lr, beta_1=0.9))
print("model is built ! ! !")
# Data generator.
eval_tr_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
eval_te_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
# Directories for saving models and training stats
model_dir = os.path.join("workspace", "models", "crn_mixdb")
pp_data.create_folder(model_dir)
stats_dir = os.path.join("workspace", "training_stats", "crn_mixdb")
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
print("start calculating initial loss.......")
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
sess = tf.Session()
x, y = load_tfrecord(batch=batch_size,
repeat=100000,
data_path=tr_path_list)
t1 = time.time()
for count in range(1000000000):
[tr_x, tr_y] = sess.run([x, y])
loss = model.train_on_batch(tr_x, tr_y)
iter += 1
# Validate and save training stats.
if iter % 1000 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
#te_loss = tr_loss
print("Iteration: %d, tr_loss: %f, te_loss: %f" %
(iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 5000 == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
print("Training time: %s s" % (time.time() - t1, ))
def train(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
print(args)
workspace = args.workspace
tr_snr = args.tr_snr
te_snr = args.te_snr
lr = args.lr
# Load data.
t1 = time.time()
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", "%ddb" % int(tr_snr), "data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram",
"test", "%ddb" % int(te_snr), "data.h5")
(tr_x, tr_y) = pp_data.load_hdf5(tr_hdf5_path)
(te_x, te_y) = pp_data.load_hdf5(te_hdf5_path)
print(tr_x.shape, tr_y.shape)
print(te_x.shape, te_y.shape)
print("Load data time: %s s" % (time.time() - t1, ))
batch_size = 500
print("%d iterations / epoch" % int(tr_x.shape[0] / batch_size))
# Scale data.
if True:
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", "%ddb" % int(tr_snr), "scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x = pp_data.scale_on_3d(tr_x, scaler)
tr_y = pp_data.scale_on_2d(tr_y, scaler)
te_x = pp_data.scale_on_3d(te_x, scaler)
te_y = pp_data.scale_on_2d(te_y, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
# Debug plot.
if False:
plt.matshow(tr_x[0:1000, 0, :].T,
origin='lower',
aspect='auto',
cmap='jet')
plt.show()
pause
# Build model
(_, n_concat, n_freq) = tr_x.shape
n_hid = 2048
model = Sequential()
model.add(Flatten(input_shape=(n_concat, n_freq)))
model.add(Dense(n_hid, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(n_hid, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(n_hid, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(n_freq, activation='linear'))
model.summary()
model.compile(loss='mean_absolute_error', optimizer=Adam(lr=lr))
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
eval_te_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
eval_tr_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models", "%ddb" % int(tr_snr))
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats", "%ddb" % int(tr_snr))
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_y]):
loss = model.train_on_batch(batch_x, batch_y)
iter += 1
# Validate and save training stats.
if iter % 1000 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" %
(iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 5000 == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == 10001:
break
print("Training time: %s s" % (time.time() - t1, ))
def train(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
print(args)
workspace = args.workspace
tr_snr = args.tr_snr
te_snr = args.te_snr
lr = args.lr
snr_arr = [0, 5, 10, 15]
"""
workspace = "workspace"
tr_snr = 0
te_snr = 0
lr = 1e-4
"""
# Load data.
t1 = time.time()
for i in snr_arr:
tr_snr = i
te_snr = i
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "%ddb" % int(tr_snr), "data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "test", "%ddb" % int(te_snr), "data.h5")
(tr_x, tr_y, tr_n) = pp_data.load_hdf5(tr_hdf5_path) # zxy tr_n
(te_x, te_y, te_n) = pp_data.load_hdf5(te_hdf5_path) # zxy te_n
print(tr_x.shape, tr_y.shape)
# Scale data.
if True:
t2 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "%ddb" % int(tr_snr),
"scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x = pp_data.scale_on_3d(tr_x, scaler)
tr_y = pp_data.scale_on_2d(tr_y, scaler)
# tr_n = pp_data.scale_on_2d(tr_n, scaler)#zxy
te_x = pp_data.scale_on_3d(te_x, scaler)
te_y = pp_data.scale_on_2d(te_y, scaler)
# te_n = pp_data.scale_on_2d(te_n, scaler)#zxy
print("Scale data(%sdb) time: %s s" % (tr_snr, time.time() - t2,))
# append data
if i == 0:
tr_x_all = tr_x
tr_y_all = tr_y
te_x_all = te_x
te_y_all = te_y
else:
tr_x_all = np.concatenate((tr_x_all, tr_x), axis=0)
tr_y_all = np.concatenate((tr_y_all, tr_y), axis=0)
te_x_all = np.concatenate((te_x_all, te_x), axis=0)
te_y_all = np.concatenate((te_y_all, te_y), axis=0)
print(tr_x_all.shape, tr_y_all.shape)#zxy tr_n.shape
print(te_x_all.shape, te_y_all.shape)#zxy te_n.shape
print("Load data time: %s s" % (time.time() - t1,))
batch_size = 100
print("%d iterations / epoch" % int(tr_x.shape[0] / batch_size))
# Debug plot.
if False:
plt.matshow(tr_x[0 : 1000, 0, :].T, origin='lower', aspect='auto', cmap='jet')
plt.show()
pause
# Build model
(_, n_concat, n_freq) = tr_x.shape
# 1.Load Pre-model by Xu
model_path = os.path.join("premodel", "sednn_keras_logMag_Relu2048layer1_1outFr_7inFr_dp0.2_weights.75-0.00.hdf5")
pre_model = load_model(model_path)
#pre_model.summary()
# 2.Build train model
n_hid = 2048
#input:feature_x
main_input = Input(shape=(n_concat, n_freq), name='main_input')
x = Flatten(input_shape=(n_concat, n_freq))(main_input)
# 2.1Pre-train to get feature_x // should be called tranform learning 2018-7-8 experiment13
#x = pre_model(x)
#x = (pre_model.get_layer('input_1'))(x)
#x = (pre_model.get_layer('dense_1'))(x)
#x = (Dense(n_hid, activation='linear'))(x)
## model_mid = Model(inputs=pre_model.input, outputs=pre_model.get_layer('dense_1').output)
#model_mid.summary()
## x=model_mid(x)
x = (Dense(n_hid, activation='linear'))(x)
"""
x = (LSTM(n_hid,
activation='tanh',
recurrent_activation='hard_sigmoid',
use_bias=True,
kernel_initializer='glorot_uniform',
recurrent_initializer='orthogonal',
bias_initializer='zeros',
unit_forget_bias=True,
kernel_regularizer=None,
recurrent_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
recurrent_constraint=None,
bias_constraint=None,
dropout=0.0,
recurrent_dropout=0.3))(main_input)
x = (LSTM(n_hid,
activation='tanh',
recurrent_activation='hard_sigmoid',
use_bias=True,
kernel_initializer='glorot_uniform',
recurrent_initializer='orthogonal',
bias_initializer='zeros',
unit_forget_bias=True,
kernel_regularizer=None,
recurrent_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
recurrent_constraint=None,
bias_constraint=None,
dropout=0.0,
recurrent_dropout=0.3))(x)
x = (LSTM(n_hid,
activation='tanh',
recurrent_activation='hard_sigmoid',
use_bias=True,
kernel_initializer='glorot_uniform',
recurrent_initializer='orthogonal',
bias_initializer='zeros',
unit_forget_bias=True,
kernel_regularizer=None,
recurrent_regularizer=None,
bias_regularizer=None,
activity_regularizer=None,
kernel_constraint=None,
recurrent_constraint=None,
bias_constraint=None,
dropout=0.0,
recurrent_dropout=0.3))(x)
"""
#hidden1
x = (Dense(n_hid, name='hidden_1'))(x)
x = LeakyReLU(alpha=0.3)(x)
x = Dropout(0.3)(x)
x = (Dense(n_hid, activation='linear'))(x)
#hidden2
x = (Dense(n_hid, name='hidden_2'))(x)
x = LeakyReLU(alpha=0.3)(x)
x = Dropout(0.3)(x)
"""
x = (Dense(n_hid, activation='linear'))(x)
#hidden3
x = (Dense(n_hid, name='hidden_3'))(x)
x = LeakyReLU(alpha=0.3)(x)
x = Dropout(0.3)(x)
#x = (Dense(n_hid, activation='linear'))(x)
#hidden4
x = (Dense(n_hid, name='hidden_4'))(x)
x = LeakyReLU(alpha=0.3)(x)
x = Dropout(0.5)(x)
"""
#output1:^speech
output_y = Dense(n_freq, activation='linear', name='out_y')(x)
#define noisy_to_speech&noise model
model = Model(inputs=main_input, outputs=output_y)
#compile model with different loss and weights
model.compile(optimizer=Adam(lr=lr),
loss='mae',
metrics=['accuracy'])
#show model_summary
model.summary()
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
eval_te_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
eval_tr_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models") # , "%ddb" % int(tr_snr))
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats") # , "%ddb" % int(tr_snr))
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
#tr_n_loss = eval(model, eval_tr_gen, tr_x, tr_n)#zxy0523
#te_n_loss = eval(model, eval_te_gen, te_x, te_n)
#print("Iteration: %d, tr_n_loss: %f, te_n_loss: %f" % (iter, tr_n_loss, te_n_loss))
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_y]):
loss = model.train_on_batch(batch_x, batch_y)
iter += 1
# Validate and save training stats.
if iter % 50 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 3000 == 0:
model_path = os.path.join(model_dir, "md_dnn2_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == 3001:
break
#zxy
resultz = model.evaluate(tr_x, tr_y)
print ("/nTrain Acc:" )
print(resultz)
resultz = model.evaluate(te_x, te_y)
print ("/nTest Acc:" )
print(resultz)
print(model.metrics_names) #zxy
print("Training time: %s s" % (time.time() - t1,))
def train_noise(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
print(args)
workspace = args.workspace
tr_snr = args.tr_snr
te_snr = args.te_snr
lr = args.lr
"""
workspace = "workspace"
tr_snr = 0
te_snr = 0
lr = 1e-4
"""
# Load data.
t1 = time.time()
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "%ddb" % int(tr_snr), "data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "test", "%ddb" % int(te_snr), "data.h5")
(tr_x, tr_y, tr_n) = pp_data.load_hdf5(tr_hdf5_path)#zxy tr_n
(te_x, te_y, te_n) = pp_data.load_hdf5(te_hdf5_path)#zxy te_n
print(tr_x.shape, tr_y.shape, tr_n.shape)#zxy tr_n.shape
print(te_x.shape, te_y.shape, te_n.shape)#zxy te_n.shape
print("Load data time: %s s" % (time.time() - t1,))
batch_size = 500
print("%d iterations / epoch" % int(tr_x.shape[0] / batch_size))
# Scale data.
if True:
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "%ddb" % int(tr_snr), "scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x = pp_data.scale_on_3d(tr_x, scaler)
#tr_y = pp_data.scale_on_2d(tr_y, scaler)
tr_n = pp_data.scale_on_2d(tr_n, scaler)#zxy
te_x = pp_data.scale_on_3d(te_x, scaler)
#te_y = pp_data.scale_on_2d(te_y, scaler)
te_n = pp_data.scale_on_2d(te_n, scaler)#zxy
print("Scale data time: %s s" % (time.time() - t1,))
# Debug plot.
if False:
plt.matshow(tr_x[0 : 1000, 0, :].T, origin='lower', aspect='auto', cmap='jet')
plt.show()
pause
# Build model
(_, n_concat, n_freq) = tr_x.shape
# 1.Load Pre-model by Xu
model_path = os.path.join("premodel", "sednn_keras_logMag_Relu2048layer1_1outFr_7inFr_dp0.2_weights.75-0.00.hdf5")
pre_model = load_model(model_path)
# 2.Build train model
n_hid = 2048
#input:feature_x
main_input = Input(shape=(n_concat, n_freq), name='main_input')
x = Flatten(input_shape=(n_concat, n_freq))(main_input)
# 2.1Pre-train to get feature_x
x = pre_model(x)
#hidden1
x = (Dense(n_hid))(x)
x = LeakyReLU(alpha=0.3)(x)
x = Dropout(0.3)(x)
#hidden2
x = (Dense(n_hid))(x)
x = LeakyReLU(alpha=0.3)(x)
x = Dropout(0.3)(x)
#hidden3
x = (Dense(n_hid))(x)
x = LeakyReLU(alpha=0.3)(x)
x = Dropout(0.3)(x)
#output1:^speech
output_y = Dense(n_freq, activation='linear', name='out_y')(x)
#define noisy_to_speech&noise model
model = Model(inputs=main_input, outputs=output_y)
#compile model with different loss and weights
model.compile(optimizer=Adam(lr=lr),
loss='mae',
metrics=['accuracy'])
#show model_summary
model.summary()
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
eval_te_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
eval_tr_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models", "%ddb_n" % int(tr_snr))
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats", "%ddb_n" % int(tr_snr))
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x, tr_n)
te_loss = eval(model, eval_te_gen, te_x, te_n)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
#tr_n_loss = eval(model, eval_tr_gen, tr_x, tr_n)#zxy0523
#te_n_loss = eval(model, eval_te_gen, te_x, te_n)
#print("Iteration: %d, tr_n_loss: %f, te_n_loss: %f" % (iter, tr_n_loss, te_n_loss))
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_n) in tr_gen.generate(xs=[tr_x], ys=[tr_n]):
loss = model.train_on_batch(batch_x, batch_n)
iter += 1
# Validate and save training stats.
if iter % 100 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x, tr_n)
te_loss = eval(model, eval_te_gen, te_x, te_n)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 1000 == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == 3001:
break
#zxy
resultz = model.evaluate(tr_x, tr_n)
print ("/nTrain Acc:" )
print(resultz)
resultz = model.evaluate(te_x, te_n)
print ("/nTest Acc:" )
print(resultz)
print(model.metrics_names) #zxy
print("Training time: %s s" % (time.time() - t1,))
def train(workspace,
tr_snr,
te_snr,
lr,
model_name=None,
force=False,
iters=100000):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
# Directories for saving models and training stats
if model_name is None:
model_name = '_'.join([str(snr) for snr in tr_snr]) + 'ddbs'
model_dir = os.path.join(workspace, "models", model_name)
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats", model_name)
pp_data.create_folder(stats_dir)
model_path = os.path.join(model_dir, f"md_{iters}iters.h5")
if os.path.isfile(model_path) and not force:
print(f'Model already trained ({model_path})')
return
# Load data.
t1 = time.time()
tr_x = None
tr_y = None
for snr in tr_snr:
tr_hdf5_path = os.path.join(workspace, "packed_features",
"spectrogram", "train", "%ddb" % int(snr),
"data.h5")
(X, y) = pp_data.load_hdf5(tr_hdf5_path)
if tr_x is None:
tr_x = X
tr_y = y
else:
tr_x = np.concatenate((tr_x, X))
tr_y = np.concatenate((tr_y, y))
te_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram",
"test", "%ddb" % int(te_snr), "data.h5")
(te_x, te_y) = pp_data.load_hdf5(te_hdf5_path)
print(tr_x.shape, tr_y.shape)
print(te_x.shape, te_y.shape)
print("Load data time: %s s" % (time.time() - t1, ))
batch_size = 500
print("%d iterations / epoch" % int(tr_x.shape[0] / batch_size))
# Scale data.
if True:
t1 = time.time()
scaler = read_combined_scaler(workspace, tr_snr)
tr_x = pp_data.scale_on_3d(tr_x, scaler)
#tr_y = pp_data.scale_on_2d(tr_y, scaler)
te_x = pp_data.scale_on_3d(te_x, scaler)
#te_y = pp_data.scale_on_2d(te_y, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
# Debug plot.
if False:
plt.matshow(tr_x[0:1000, 0, :].T,
origin='lower',
aspect='auto',
cmap='jet')
plt.show()
pause
print(tf.test.is_gpu_available())
# Build model
(_, n_concat, n_freq) = tr_x.shape
n_hid = 2048
model = Sequential()
model.add(Flatten(input_shape=(n_concat, n_freq)))
model.add(Dense(n_hid, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(n_hid, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(n_hid, activation='relu'))
model.add(Dropout(0.2))
model.add(Dense(n_freq, activation='linear'))
model.summary()
model.compile(loss='mean_absolute_error', optimizer=Adam(lr=lr))
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
eval_te_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=50)
eval_tr_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=50)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
pickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=pickle.HIGHEST_PROTOCOL)
# Train.h
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_y]):
loss = model.train_on_batch(batch_x, batch_y)
iter += 1
# Validate and save training stats.
if iter % 1000 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" %
(iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
pickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=pickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 10000 == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == iters + 1:
break
print("Training time: %s s" % (time.time() - t1, ))
def continue_train_tfrecord():
workspace = "workspace"
lr = 1e-5
iter = 220000
data_type = "IRM"
# Load model.
if data_type == "DM":
model_path = os.path.join(workspace, "models", "elu_mixdb",
"md_%diters.h5" % iter)
else:
model_path = os.path.join(workspace, "models", "mask_mixdb",
"md_%diters.h5" % iter)
model = load_model(model_path)
#model = multi_gpu_model(model, 4)
model.compile(loss='mean_absolute_error',
optimizer=Adam(lr=lr, beta_1=0.2))
# Load data.
if data_type == "DM":
tr_hdf5_dir = os.path.join(workspace, "tfrecords", "train", "mixdb")
tr_hdf5_names = os.listdir(tr_hdf5_dir)
tr_path_list = [os.path.join(tr_hdf5_dir, i) for i in tr_hdf5_names]
te_hdf5_path = os.path.join(workspace, "packed_features",
"spectrogram", "test", "mixdb", "data.h5")
else:
tr_hdf5_dir = os.path.join(workspace, "tfrecords", "train",
"mask_mixdb")
tr_hdf5_names = os.listdir(tr_hdf5_dir)
tr_path_list = [os.path.join(tr_hdf5_dir, i) for i in tr_hdf5_names]
te_hdf5_path = os.path.join(workspace, "packed_features",
"spectrogram", "test", "mask_mixdb",
"data.h5")
#(tr_x1, tr_y1) = pp_data.load_hdf5("workspace/packed_features/spectrogram/train/mixdb/data100000.h5")
(te_x, te_y) = pp_data.load_hdf5(te_hdf5_path)
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", "mixdb", "scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
te_x = pp_data.scale_on_3d(te_x, scaler)
#tr_x1 = pp_data.scale_on_3d(tr_x1, scaler)
if data_type == "DM":
te_y = pp_data.scale_on_2d(te_y, scaler)
tr_y1 = pp_data.scale_on_2d(tr_y1, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
# Directories for saving models and training stats
if data_type == "DM":
model_dir = os.path.join(workspace, "models", "elu_mixdb", "continue")
stats_dir = os.path.join(workspace, "training_stats", "elu_mixdb",
"continue")
else:
model_dir = os.path.join(workspace, "models", "mask_mixdb", "continue")
stats_dir = os.path.join(workspace, "training_stats", "mask_mixdb",
"continue")
pp_data.create_folder(model_dir)
pp_data.create_folder(stats_dir)
# Print loss before training.
batch_size = 1024 * 4
#eval_tr_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
eval_te_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
#tr_loss = eval(model, eval_tr_gen, tr_x1, tr_y1)
tr_loss = 0
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" %
(iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
sess = tf.Session()
x, y = load_tfrecord(batch=batch_size,
repeat=100000,
data_path=tr_path_list)
t1 = time.time()
for count in range(1000000000):
[tr_x, tr_y] = sess.run([x, y])
loss = model.train_on_batch(tr_x, tr_y)
iter += 1
# Validate and save training stats.
if iter % 1000 == 0:
#tr_loss = eval(model, eval_tr_gen, tr_x1, tr_y1)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" %
(iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 5000 == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == 100001:
break
print("Training time: %s s" % (time.time() - t1, ))
def continue_train(args):
workspace = args.workspace
lr = args.lr
iter = args.iteration
data_type = "IRM"
# Load model.
if data_type == "DM":
model_path = os.path.join(workspace, "models", "mixdb",
"md_%diters.h5" % iter)
else:
model_path = os.path.join(workspace, "models", "mask_mixdb",
"md_%diters.h5" % iter)
model = load_model(model_path)
#model = multi_gpu_model(model, 4)
model.compile(loss='mean_absolute_error',
optimizer=Adam(lr=lr, beta_1=0.2))
# Load data.
t1 = time.time()
if data_type == "DM":
tr_hdf5_path = os.path.join(workspace, "packed_features",
"spectrogram", "train", "mixdb", "data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features",
"spectrogram", "test", "mixdb", "data.h5")
else:
tr_hdf5_path = os.path.join(workspace, "packed_features",
"spectrogram", "train", "mask_mixdb",
"data.h5")
te_hdf5_path = os.path.join(workspace, "packed_features",
"spectrogram", "test", "mask_mixdb",
"data.h5")
tr_hdf5_dir = os.path.join(workspace, "packed_features", "spectrogram",
"train", "mask_mixdb")
tr_hdf5_names = os.listdir(tr_hdf5_dir)
tr_hdf5_names = [i for i in tr_hdf5_names if i.endswith(".h5")]
tr_path_list = [os.path.join(tr_hdf5_dir, i) for i in tr_hdf5_names]
(tr_x, tr_y) = pp_data.load_hdf5(tr_hdf5_path)
(te_x, te_y) = pp_data.load_hdf5(te_hdf5_path)
print(tr_x.shape, tr_y.shape)
print(te_x.shape, te_y.shape)
print("Load data time: %s s" % (time.time() - t1, ))
batch_size = 2048
print("%d iterations / epoch" % int(tr_x.shape[0] / batch_size))
# Scale data.
if True:
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", "mixdb", "scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x = pp_data.scale_on_3d(tr_x, scaler)
te_x = pp_data.scale_on_3d(te_x, scaler)
if data_type == "DM":
tr_y = pp_data.scale_on_2d(tr_y, scaler)
te_y = pp_data.scale_on_2d(te_y, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
#scaler_path = os.path.join(workspace, "packed_features", "spectrogram", "train", "mixdb", "scaler.p")
#scaler = pickle.load(open(scaler_path, 'rb'))
tr_gen = DataGenerator(batch_size=batch_size, type='train')
eval_te_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
eval_tr_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
#tr_gen = DataGenerator_h5py(batch_size=batch_size, type='train', scaler = scaler)
#eval_te_gen = DataGenerator_h5py(batch_size=batch_size, type='test', te_max_iter=100, scaler =scaler)
#eval_tr_gen = DataGenerator_h5py(batch_size=batch_size, type='test', te_max_iter=100, scaler =scaler)
# Directories for saving models and training stats
if data_type == "DM":
model_dir = os.path.join(workspace, "models", "chinese_mixdb",
"continue")
stats_dir = os.path.join(workspace, "training_stats", "chinese_mixdb",
"continue")
else:
model_dir = os.path.join(workspace, "models", "mask_mixdb", "continue")
stats_dir = os.path.join(workspace, "training_stats", "mask_mixdb",
"continue")
pp_data.create_folder(model_dir)
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
#tr_loss = eval_h5py(model, eval_tr_gen, tr_path_list)
#te_loss = eval_h5py(model, eval_te_gen, [te_hdf5_path])
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_y]):
#for (batch_x, batch_y) in tr_gen.generate(tr_path_list):
loss = model.train_on_batch(batch_x, batch_y)
iter += 1
# Validate and save training stats.
if iter % 500 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
#tr_loss = eval_h5py(model, eval_tr_gen, tr_path_list)
#te_loss = eval_h5py(model, eval_te_gen, [te_hdf5_path])
print("Iteration: %d, tr_loss: %f, te_loss: %f" %
(iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss,
}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 5000 == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == 100001:
break
print("Training time: %s s" % (time.time() - t1, ))
def train(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
print(args)
workspace = args.workspace
model_name = args.model_name
lr = args.lr
tr_dir_name = args.tr_dir_name
va_dir_name = args.va_dir_name
iter_training = args.iteration
dropout = args.dropout
# Load data.
t1 = time.time()
tr_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", tr_dir_name, "data.h5")
# va_hdf5_path = os.path.join(workspace, "packed_features", "spectrogram", "validation", va_dir_name, "data.h5")
(tr_x, tr_y) = pp_data.load_hdf5(tr_hdf5_path)
# (va_x, va_y) = pp_data.load_hdf5(va_hdf5_path)
print(tr_x.shape, tr_y.shape)
# print(va_x.shape, va_y.shape)
print("Load data time: %s s" % (time.time() - t1, ))
batch_size = 500
print("%d iterations / epoch" % int(tr_x.shape[0] / batch_size))
# Scale data.
if True:
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "spectrogram",
"train", tr_dir_name, "scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x = pp_data.scale_on_3d(tr_x, scaler)
tr_y = pp_data.scale_on_2d(tr_y, scaler)
# va_x = pp_data.scale_on_3d(va_x, scaler)
# va_y = pp_data.scale_on_2d(va_y, scaler)
print("Scale data time: %s s" % (time.time() - t1, ))
# Debug plot.
if False:
plt.matshow(tr_x[0:1000, 0, :].T,
origin='lower',
aspect='auto',
cmap='jet')
plt.show()
pause
# Build model
(_, n_concat, n_freq) = tr_x.shape
n_hid = 2048
with tf.Session() as sess:
model = DNN(sess,
lr,
batch_size, (n_concat, n_freq),
n_freq,
dropouts=dropout,
training=True)
model.build()
sess.run(tf.global_variables_initializer())
merge_op = tf.summary.merge_all()
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
# eval_te_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
eval_tr_gen = DataGenerator(batch_size=batch_size,
type='test',
te_max_iter=100)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models", model_name)
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats", model_name)
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(sess, model, eval_tr_gen, tr_x, tr_y)
# te_loss = eval(model, eval_te_gen, te_x, te_y)
# print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
print("Iteration: %d, tr_loss: %f" % (iter, tr_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
}
# 'te_loss': te_loss,}
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
pickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=pickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_y]):
feed_dict = {model.x_noisy: batch_x, model.y_clean: batch_y}
_, loss, summary_str = sess.run(
[model.optimizer, model.loss, merge_op], feed_dict=feed_dict)
iter += 1
# Validate and save training stats.
if iter % 1000 == 0:
tr_loss = eval(sess, model, eval_tr_gen, tr_x, tr_y)
# te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f" % (iter, tr_loss))
# print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {
'iter': iter,
'tr_loss': tr_loss,
}
# 'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
pickle.dump(stat_dict,
open(stat_path, 'wb'),
protocol=pickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 5000 == 0:
ckpt_file_path = os.path.join(model_dir, model_name)
# if os.path.isdir(model_dir) is False:
# os.makedirs(model_dir)
tf.train.Saver().save(sess,
ckpt_file_path,
write_meta_graph=True)
print("Saved model to %s" % ckpt_file_path)
if iter == iter_training + 1:
break
print("Training time: %s s" % (time.time() - t1, ))
def train(args):
"""Train the neural network. Write out model every several iterations.
Args:
workspace: str, path of workspace.
tr_snr: float, training SNR.
te_snr: float, testing SNR.
lr: float, learning rate.
"""
class MetricsHistory(Callback):
def on_epoch_end(self, epoch, logs={}):
file_logger.write([str(epoch),
str(logs['loss']),
str(logs['val_loss'])
])
print(args)
workspace = args.workspace
#tr_snr = args.tr_snr
#te_snr = args.te_snr
lr = args.lr
#TF = args.TF
model_name = args.model_name
#model_save_dir = os.path.join(args.workspace, 'saved_models')
# Load data
t1 = time.time()
print("Loading the train and vallidation dataset")
tr_hdf5_path = os.path.join(workspace, "packed_features", "train", "mag.h5")
te_hdf5_path = os.path.join(workspace, "packed_features", "val", "mag.h5")
(tr_x, tr_y) = pp_data.load_hdf5(tr_hdf5_path)
(te_x, te_y) = pp_data.load_hdf5(te_hdf5_path)
print('train_x shape:')
print(tr_x.shape, tr_y.shape)
print('test_x shape:')
print(te_x.shape, te_y.shape)
print("Load data time: %f s" % (time.time() - t1))
print('\n')
# Scale data
if True:
print("Scaling train and test dataset. This will take some time, please wait patiently...")
t1 = time.time()
scaler_path = os.path.join(workspace, "packed_features", "train", "mag_scaler.p")
scaler = pickle.load(open(scaler_path, 'rb'))
tr_x = pp_data.scale_on_3d(tr_x, scaler)
tr_y = pp_data.scale_on_2d(tr_y, scaler)
te_x = pp_data.scale_on_3d(te_x, scaler)
te_y = pp_data.scale_on_2d(te_y, scaler)
print("Scale data time: %f s" % (time.time() - t1))
# Debug plot.
if False:
plt.matshow(tr_x[0 : 1000, 0, :].T, origin='lower', aspect='auto', cmap='jet')
plt.show()
#time.sleep(secs)
os.system("pause")
# Build model
batch_size = 150
epoch = 100
print("The neural networks you have chosed is %s" % model_name)
print("The training batch is set to %d and the %s will be training for at most %d epoches" % (batch_size, model_name.upper(), epoch))
print("======iteration of one epoch======" )
iter_each_epoch = int(tr_x.shape[0] / batch_size)
#val_each_epoch = int(te_x.shape[0] / batch_size)
#print("There are %d iterations / epoch" % int(tr_x.shape[0] / batch_size))
print("There are %d iterations / epoch" % iter_each_epoch)
log_save_dir = os.path.join(workspace, 'log')
if not os.path.isdir(log_save_dir):
os.makedirs(log_save_dir)
log_path = os.path.join(log_save_dir, 'out_{}.csv'.format(model_name))
#log_path = os.path.join(log_save_dir, 'out_%ddb_%s.csv' %(int(snr[0]), model_name))
file_logger = FileLogger(log_path, ['epoch', 'train_loss', 'val_loss'])
(_, n_concat, n_freq) = tr_x.shape
#temp_tr_x = tr_x[:, 3, :][:, np.newaxis, :]
#print(temp_tr_x.shape)
#np.axis
n_hid = 2048
#data_gen = DataGenerator(batch_size=batch_size, type='train')
#tr_gen = data_gen.generate(xs=[tr_x], ys=[tr_y])
#te_gen = data_gen.generate(xs=[te_x], ys=[te_y])
#temp_tr_x = tr_gen[:, 3, :][:, np.newaxis, :]
'''
model = Sequential()
model.add(Flatten(input_shape=(n_concat, n_freq)))
model.add(BatchNormalization())
model.add(Dense(n_hid, activation='relu', kernel_regularizer=regularizers.l2(l=0.0001)))
model.add(Dropout(0.2))
model.add(BatchNormalization())
model.add(Dense(n_hid, activation='relu', kernel_regularizer=regularizers.l2(l=0.0001)))
model.add(Dropout(0.2))
model.add(BatchNormalization())
model.add(Dense(n_hid, activation='relu', kernel_regularizer=regularizers.l2(l=0.0001)))
model.add(Dropout(0.2))
model.add(Dense(n_freq, activation='linear'))
#model.summary()
'''
print('Model selected:', model_name.lower())
if model_name == 'dnn':
model = dnn(n_hid, n_concat, n_freq)
elif model_name == 'sdnn1':
model = sdnn1(n_hid, n_concat, n_freq)
elif model_name == 'sdnn2':
model = sdnn2(n_hid, n_concat, n_freq)
elif model_name == 'sdnn3':
model = sdnn3(n_hid, n_concat, n_freq)
elif model_name == 'fcn':
model = fcn(n_concat, n_freq)
elif model_name == 'fcn1':
model = fcn1(n_concat, n_freq)
elif model_name == 'fcn1':
model = fcn1_re(n_concat, n_freq)
elif model_name == 'fcn2':
model = fcn2(n_concat, n_freq)
elif model_name == 'fcn3':
model = fcn3(n_concat, n_freq)
elif model_name == 'fcn4':
model = fcn4(n_concat, n_freq)
elif model_name == 'm_vgg':
model = m_vgg(n_concat, n_freq)
elif model_name == 'm_vgg1':
model = m_vgg1(n_concat, n_freq)
elif model_name == 'm_vgg2':
model = m_vgg2(n_concat, n_freq)
elif model_name == 'm_vgg3':
model = m_vgg3(n_concat, n_freq)
elif model_name == 'm_vgg4':
model = m_vgg3(n_concat, n_freq)
elif model_name == 'CapsNet':
model = CapsNet(n_concat, n_freq, 3)
elif model_name == 'brnn' :
recur_layers = 7
unit = 256
output_dim = n_freq
model = brnn(n_concat, n_freq, unit, recur_layers, output_dim)
elif model_name == 'rnn' :
output_dim = n_freq
model = rnn(n_concat, n_freq, output_dim)
elif model_name == 'tcn' :
input_dim = n_freq
model = tcn(n_concat, input_dim)
if model is None:
exit('Please choose a valid model: [dnn, sdnn, sdnn1, cnn, scnn1]')
#mean_squared_error
model.compile(loss = 'mean_squared_error',
optimizer=Adam(lr=lr))
print(model.summary())
#plot model
#plot_model(model, to_file=args.save_dir+'/model.png', show_shapes=True)
#plot_model(model, to_file='%s/%s_model.png' % (log_save_dir, model_name), show_shapes=True)
# Save model and weights
model_save_dir = os.path.join(workspace, 'saved_models', "%s" % model_name)
model_save_name = "weights-checkpoint-{epoch:02d}-{val_loss:.2f}.h5"
if not os.path.isdir(model_save_dir):
os.makedirs(model_save_dir)
model_path = os.path.join(model_save_dir, model_save_name)
checkpoint = ModelCheckpoint(model_path, monitor='val_loss', verbose=1, save_best_only=True, mode='min')
print('Saved trained model at %s' % model_save_dir)
#reduce_lr = ReduceLROnPlateau(monitor='val_loss', factor=0.5, patience=4, min_lr=0.00001, verbose=1)
lr_decay = LearningRateScheduler(schedule=lambda epoch: lr * (0.9 ** epoch))
metrics_history = MetricsHistory()
hist = model.fit(x=tr_x,
y=tr_y,
batch_size=batch_size,
epochs=epoch,
verbose=1,
shuffle=True,
validation_data=(te_x, te_y),
#validation_split=0.1,
callbacks=[metrics_history, checkpoint, lr_decay])
'''
hist = model.fit_generator(tr_gen,
steps_per_epoch=iter_each_epoch,
epochs=epoch,
verbose=1,
validation_data=te_gen,
validation_steps=val_each_epoch,
callbacks=[metrics_history, checkpoint, reduce_lr])
'''
print(hist.history.keys())
# list all data in history
#print(hist.history.keys())
'''
# summarize history for accuracy
plt.plot(hist.history['acc'])
plt.plot(hist.history['val_acc'])
plt.title('model accuracy')
plt.ylabel('accuracy')
plt.xlabel('epoch')
plt.legend(['train', 'test'], loc='upper left')
plt.show()
'''
# summarize history for loss
model_png = "train_test_loss"
loss_fig_dir = os.path.join(log_save_dir, '%s_%s.png' % (model_name, model_png))
plt.plot(hist.history['loss'])
plt.plot(hist.history['val_loss'])
plt.title('model loss')
plt.ylabel('loss')
plt.xlabel('epoch')
plt.legend(['train', 'val'], loc='upper right')
plt.savefig(loss_fig_dir)
#plt.show()
'''
fig = plt.gcf()
plt.show()
fig.savefig('tessstttyyy.png', dpi=100)
'''
file_logger.close()
'''
# Data generator.
tr_gen = DataGenerator(batch_size=batch_size, type='train')
eval_te_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
eval_tr_gen = DataGenerator(batch_size=batch_size, type='test', te_max_iter=100)
# Directories for saving models and training stats
model_dir = os.path.join(workspace, "models", "%ddb" % int(tr_snr))
pp_data.create_folder(model_dir)
stats_dir = os.path.join(workspace, "training_stats", "%ddb" % int(tr_snr))
pp_data.create_folder(stats_dir)
# Print loss before training.
iter = 0
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Train.
t1 = time.time()
for (batch_x, batch_y) in tr_gen.generate(xs=[tr_x], ys=[tr_y]):
#loss = model.train_on_batch(batch_x, batch_y)
if iter % 2000 == 0:
lr *= 0.1
model.train_on_batch(batch_x, batch_y)
iter += 1
# Validate and save training stats.
if iter % 1000 == 0:
tr_loss = eval(model, eval_tr_gen, tr_x, tr_y)
te_loss = eval(model, eval_te_gen, te_x, te_y)
print("Iteration: %d, tr_loss: %f, te_loss: %f" % (iter, tr_loss, te_loss))
# Save out training stats.
stat_dict = {'iter': iter,
'tr_loss': tr_loss,
'te_loss': te_loss, }
stat_path = os.path.join(stats_dir, "%diters.p" % iter)
cPickle.dump(stat_dict, open(stat_path, 'wb'), protocol=cPickle.HIGHEST_PROTOCOL)
# Save model.
if iter % 5000 == 0:
model_path = os.path.join(model_dir, "md_%diters.h5" % iter)
model.save(model_path)
print("Saved model to %s" % model_path)
if iter == 10001:
break
'''
print("Training time: %s s" % (time.time() - t1,))
