def train(img_file, mask_file, top_model_weights_path, epochs, batch_size):
train_gen, validation_gen, img_shape = load_data(img_file, mask_file)
img_height = img_shape[0]
img_width = img_shape[1]
lr_base = 0.01 * (float(batch_size) / 16)
model = MobileUNet(input_shape=(img_height, img_width, 3), alpha_up=0.25)
# model = MobileDeepLab(input_shape=(img_height, img_width, 3))
model.load_weights(os.path.expanduser(top_model_weights_path),
by_name=True)
# Freeze above conv_dw_12
for layer in model.layers[:70]:
layer.trainable = False
# Freeze above conv_dw_13
# for layer in model.layers[:76]:
# layer.trainable = False
model.summary()
model.compile(
optimizer=SGD(lr=0.0001, momentum=0.9),
# optimizer=Adam(lr=0.0001),
loss=dice_coef_loss,
metrics=[
dice_coef,
recall,
precision,
'binary_crossentropy',
],
)
# callbacks
scheduler = callbacks.LearningRateScheduler(
create_lr_schedule(epochs, lr_base=lr_base, mode='progressive_drops'))
tensorboard = callbacks.TensorBoard(log_dir='./logs')
checkpoint = callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
save_best_only=True)
model.fit_generator(
generator=train_gen(),
steps_per_epoch=nb_train_samples // batch_size,
epochs=epochs,
validation_data=validation_gen(),
validation_steps=nb_validation_samples // batch_size,
callbacks=[scheduler, tensorboard, checkpoint],
)
model.save(trained_model_path)
def train(img_file, mask_file, epochs, batch_size):
# train_gen, validation_gen, img_shape = load_data(img_file, mask_file)
img_height = 385
img_width = 385
lr_base = 0.01 * (float(batch_size) / 16)
model = MobileUNet(input_shape=(img_height, img_width, 3),
alpha=1,
alpha_up=0.25)
model.summary()
model.compile(
optimizer=optimizers.SGD(lr=0.0001, momentum=0.9),
# optimizer=Adam(lr=0.001),
# optimizer=optimizers.RMSprop(),
loss=dice_coef_loss,
metrics=[
dice_coef,
recall,
precision,
'binary_crossentropy',
],
)
#############
model.metrics_tensors('activation_1')
#########
# callbacks
scheduler = callbacks.LearningRateScheduler(
create_lr_schedule(epochs, lr_base=lr_base, mode='progressive_drops'))
tensorboard = callbacks.TensorBoard(log_dir='./logs')
csv_logger = callbacks.CSVLogger('logs/training.csv')
checkpoint = callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
save_best_only=True)
model.fit_generator(
generator=train_gen(),
steps_per_epoch=nb_train_samples // batch_size,
epochs=epochs,
validation_data=validation_gen(),
validation_steps=nb_validation_samples // batch_size,
# callbacks=[tensorboard, checkpoint, csv_logger],
callbacks=[scheduler, tensorboard, checkpoint, csv_logger],
)
model.save(trained_model_path)
def train(epochs, batch_size, img_size):
fresh_training = True
alpha_value = 0.9
img_file = 'data/id_pack/images-{}.npy'.format(img_size)
mask_file = 'data/id_pack/masks-{}.npy'.format(img_size)
trained_model_path = 'artifacts/model-{}.h5'.format(img_size)
print("training on image file:")
print(img_file)
print(mask_file)
# Load the data
train_gen, validation_gen, img_shape, train_len, val_len = load_data(
img_file, mask_file)
img_height = img_shape[0]
img_width = img_shape[1]
print(img_height, img_width)
lr_base = 0.01 * (float(batch_size) / 16)
if fresh_training:
model = MobileUNet(input_shape=(img_height, img_width, 3),
alpha=alpha_value,
alpha_up=0.25)
else:
with CustomObjectScope(custom_objects()):
model = load_model(SAVED_MODEL)
model.summary()
model.compile(
# optimizer=optimizers.SGD(lr=0.0001, momentum=0.9),
optimizer=optimizers.Adam(lr=0.0001),
# optimizer=Adam(lr=0.001),
# optimizer=optimizers.RMSprop(),
#loss=dice_coef_loss,
loss='mean_absolute_error',
#loss = loss_gu,
metrics=[
dice_coef, recall, precision, dice_coef_loss, 'mean_absolute_error'
],
)
# callbacks
scheduler = callbacks.LearningRateScheduler(
create_lr_schedule(epochs, lr_base=lr_base, mode='progressive_drops'))
tensorboard = callbacks.TensorBoard(log_dir='./logs')
csv_logger = callbacks.CSVLogger('logs/training.csv')
checkpoint = callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
save_best_only=True)
'''
looks like we do have some legacy support issue
the steps_per_epoch and validation_steps is actually the number of sample
legacy_generator_methods_support = generate_legacy_method_interface(
allowed_positional_args=['generator', 'steps_per_epoch', 'epochs'],
conversions=[('samples_per_epoch', 'steps_per_epoch'),
('val_samples', 'steps'),
('nb_epoch', 'epochs'),
('nb_val_samples', 'validation_steps'),
'''
nb_train_samples = train_len
nb_validation_samples = val_len
print("training sample is ", nb_train_samples)
if fresh_training:
cb_list = [scheduler, tensorboard, checkpoint, csv_logger]
else:
cb_list = [tensorboard, checkpoint, csv_logger]
model.fit_generator(
generator=train_gen(),
steps_per_epoch=nb_train_samples // batch_size,
epochs=epochs,
validation_data=validation_gen(),
validation_steps=nb_validation_samples // batch_size,
callbacks=cb_list,
)
model.save(trained_model_path)
def train(img_file, mask_file, epochs, batch_size):
train_gen, validation_gen, img_shape, train_len, val_len = load_data(img_file, mask_file)
img_height = img_shape[0]
img_width = img_shape[1]
lr_base = 0.01 * (float(batch_size) / 16)
print(img_height, img_width)
model = MobileUNet(input_shape=(img_height, img_width, 3),
alpha=1,
alpha_up=0.25)
model.summary()
model.compile(
optimizer=optimizers.SGD(lr=0.0001, momentum=0.9),
# optimizer=Adam(lr=0.001),
# optimizer=optimizers.RMSprop(),
#loss=dice_coef_loss,
loss='mean_squared_error',
metrics=[
dice_coef,
recall,
precision,
'binary_crossentropy',
],
)
# callbacks
scheduler = callbacks.LearningRateScheduler(
create_lr_schedule(epochs, lr_base=lr_base, mode='progressive_drops'))
tensorboard = callbacks.TensorBoard(log_dir='./logs')
csv_logger = callbacks.CSVLogger('logs/training.csv')
checkpoint = callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
save_best_only=True)
'''
looks like we do have some legacy support issue
the steps_per_epoch and validation_steps is actually the number of sample
legacy_generator_methods_support = generate_legacy_method_interface(
allowed_positional_args=['generator', 'steps_per_epoch', 'epochs'],
conversions=[('samples_per_epoch', 'steps_per_epoch'),
('val_samples', 'steps'),
('nb_epoch', 'epochs'),
('nb_val_samples', 'validation_steps'),
'''
nb_train_samples = train_len
nb_validation_samples = val_len
model.fit_generator(
generator=train_gen(),
steps_per_epoch=nb_train_samples // batch_size,
epochs=epochs,
validation_data=validation_gen(),
validation_steps=nb_validation_samples // batch_size,
# callbacks=[tensorboard, checkpoint, csv_logger],
callbacks=[scheduler, tensorboard, checkpoint, csv_logger],
)
model.save(trained_model_path)
def train(coco_path, checkpoint_path, log_path, epochs=100, batch_size=50):
cat_nms = ['book', 'apple', 'keyboard']
BATCH_SIZE = batch_size
NUM_EPOCH = epochs
IMAGE_W = 224
IMAGE_H = 224
model = MobileUNet(input_shape=(IMAGE_H, IMAGE_W, 3),
alpha_up=0.25,
num_classes=(len(cat_nms) + 1))
# model.load_weights(os.path.expanduser(mobilenet_weights_path.format(img_height)),
# by_name=True)
# # Freeze mobilenet original weights
# for layer in model.layers[:82]:
# layer.trainable = False
seed = 1
np.random.seed(seed)
training_generator = coco_generator(cat_nms,
coco_path,
batch_size=BATCH_SIZE)
validation_generator = coco_generator(cat_nms,
coco_path,
subset='val',
batch_size=BATCH_SIZE)
model.summary()
if os.path.exists(checkpoint_path):
model.load_weights(checkpoint_path, by_name=True)
model.compile(
optimizer=optimizers.SGD(lr=0.0001, momentum=0.9),
# optimizer=Adam(lr=0.001),
# optimizer=optimizers.RMSprop(),
loss=dice_coef_loss,
metrics=[
dice_coef,
recall,
precision,
'binary_crossentropy',
],
)
lr_base = 0.01 * (float(BATCH_SIZE) / 16)
# callbacks
scheduler = callbacks.LearningRateScheduler(
create_lr_schedule(NUM_EPOCH,
lr_base=lr_base,
mode='progressive_drops'))
tensorboard = callbacks.TensorBoard(log_dir=log_path)
csv_logger = callbacks.TensorBoard(log_path)
checkpoint = callbacks.ModelCheckpoint(filepath=checkpoint_path,
save_weights_only=True,
save_best_only=True)
# Train model on dataset
model.fit_generator(
generator=training_generator,
validation_data=validation_generator,
epochs=NUM_EPOCH,
callbacks=[scheduler, tensorboard, checkpoint, csv_logger],
)
def main(name, dataset, epochs, batch_size, learning_rate, attention, n_iter,
enc_dim, dec_dim, z_dim, oldmodel, live_plotting):
if name is None:
name = dataset
channels, height, width = 1, 28, 28
x_dim = channels * height * width
# Configure attention mechanism
if attention != "":
read_N, write_N = attention.split(",")
read_N = int(read_N)
write_N = int(write_N)
reader = AttentionReader(channels, height, width, read_N)
writer = AttentionWriter(channels, height, width, write_N)
attention_tag = "r%d-w%d" % (read_N, write_N)
else:
reader = Reader(channels, height, width)
writer = Writer(x_dim, channels, height, width)
attention_tag = "full"
lr_str = "%1.0e" % learning_rate
result_dir = os.path.join("./", "result")
if not os.path.exists(result_dir):
os.makedirs(result_dir)
longname = "%s-%s-t%d-enc%d-dec%d-z%d-lr%s" % \
(dataset, attention_tag, n_iter, enc_dim, dec_dim, z_dim, lr_str)
print("\nRunning experiment %s" % longname)
print(" dataset: %s" % dataset)
print(" subdirectory: %s" % result_dir)
print(" learning rate: %g" % learning_rate)
print(" attention: %s" % attention)
print(" n_iterations: %d" % n_iter)
print(" encoder dimension: %d" % enc_dim)
print(" z dimension: %d" % z_dim)
print(" decoder dimension: %d" % dec_dim)
print(" batch size: %d" % batch_size)
print(" epochs: %d" % epochs)
print()
# ----------------------------------------------------------------------
encoder_rnn = tf.nn.rnn_cell.LSTMCell(enc_dim, state_is_tuple=True)
decoder_rnn = tf.nn.rnn_cell.LSTMCell(dec_dim, state_is_tuple=True)
sampler = Qsampler(z_dim)
draw = DrawModel(batch_size,
x_dim,
z_dim,
dec_dim,
n_iter=n_iter,
reader=reader,
encoder_rnn=encoder_rnn,
sampler=sampler,
decoder_rnn=decoder_rnn,
writer=writer)
# -------------------------------------------------------------------------
x = tf.placeholder(tf.float32, shape=(batch_size, x_dim))
lr = tf.placeholder(tf.float32, shape=[])
x_recons, kl_term, xs, c_ws = draw.reconstruct(x)
def binary_crossentropy(t, o):
eps = 1E-8
return -(t * tf.log(o + eps) + (1.0 - t) * tf.log(1.0 - o + eps))
Lx = tf.reduce_mean(
tf.reduce_sum(binary_crossentropy(x, tf.nn.sigmoid(x_recons)), 1))
# Lx = tf.losses.sigmoid_cross_entropy(x, logits=x_recons)
Lz = tf.reduce_mean(kl_term)
cost = Lx + Lz
# -----------------------------------------------------------------------
optim = tf.train.AdamOptimizer(lr, beta1=0.5)
gradients_and_vars = optim.compute_gradients(cost)
gradients_and_vars = [(grad_var[0] if grad_var[0] is None else
tf.clip_by_norm(grad_var[0], 5), grad_var[1])
for grad_var in gradients_and_vars]
train_opt = optim.apply_gradients(gradients_and_vars)
sample_x, sample_rg = draw.sample()
# --------------------------------------------------------------
data_directory = os.path.join("./", "mnist")
if not os.path.exists(data_directory):
os.makedirs(data_directory)
train_data = mnist.input_data.read_data_sets(
data_directory, one_hot=True).train # (samples, height, width)
# ----------------------------------------------------
sess_config = tf.ConfigProto(allow_soft_placement=True,
log_device_placement=True)
sess_config.gpu_options.allow_growth = True
sess_config.gpu_options.per_process_gpu_memory_fraction = 0.9
saver = tf.train.Saver(max_to_keep=2)
process = Process()
with tf.Session(config=sess_config) as sess:
'''
Load from checkpoint or start a new session
'''
sess.run(tf.global_variables_initializer())
losses = []
lr_schedule = create_lr_schedule(lr_base=learning_rate,
decay_rate=0.1,
decay_epochs=5000,
truncated_epoch=15000,
mode="exp")
for epoch in range(epochs):
process.start_epoch()
lr_epoch = lr_schedule(epoch)
# Given data batch to Graph
x_batch, _ = train_data.next_batch(batch_size)
x_batch = (x_batch > 0.5).astype(np.float32)
lx, lz, tl, _ = sess.run([Lx, Lz, cost, train_opt],
feed_dict={
x: x_batch,
lr: lr_epoch
})
losses.append([lx, lz, tl])
if epoch % 10 == 0 or epoch == epochs - 1:
process.format_meter(epoch, epochs, {
"Lx": lx,
"Lz": lz,
"Cost": tl
})
np.save(result_dir + "/draw_losses.npy", np.array(losses))
saver.save(sess,
os.path.join(result_dir, "model_ckpt"),
global_step=epochs)
print("Results have been saved")
# -----------------------------------------------------------
# canvases: shape(n_iter, batch_size, x_dim)
# rg: shape(n_iter, batch_size, 4)
x_batch, _ = train_data.next_batch(batch_size)
x_batch = (x_batch > 0.5).astype(np.float32)
canvases_gen, rg_gen = sess.run([sample_x, sample_rg], feed_dict={})
canvases_trn, rg_trn = sess.run([xs, c_ws], feed_dict={x: x_batch})
canvases_gen, rg_gen = np.array(canvases_gen), np.array(rg_gen)
canvases_trn, rg_trn = np.array(canvases_trn), np.array(rg_trn)
pickle_save([canvases_gen, rg_gen, canvases_trn, rg_trn],
["Generate", "Train"], "./result/draw.pkl")
