def train():
evaluate_queue = queue.Queue()
evaluate_task = keras_util.EvaluateTask(evaluate_queue)
evaluate_task.setDaemon(True)
evaluate_task.start()
checkpoint = keras_util.EvaluateCallback(model_config, evaluate_queue)
start = time.time()
print("####### start train model")
for i in range(len(model_config.epoch)):
print("####### lr=%f, freeze layers=%2f epoch=%d" % (
model_config.lr[i], model_config.freeze_layers[i], model_config.epoch[i]))
clr = keras_util.CyclicLrCallback(base_lr=model_config.lr[i], max_lr=model_config.lr[i] * 5,
step_size=model_config.get_steps_per_epoch(i) / 2)
train_flow = data_loader.KerasGenerator(model_config=model_config,
featurewise_center=True,
featurewise_std_normalization=True,
width_shift_range=0.15,
height_shift_range=0.1,
horizontal_flip=True,
rotation_range=10,
rescale=1. / 256).flow_from_files(model_config.train_files, mode="fit",
target_size=model_config.image_size,
batch_size=
model_config.train_batch_size[i],
shuffle=True,
label_position=model_config.label_position)
if i == 0:
model = get_model(freeze_layers=model_config.freeze_layers[i], lr=model_config.lr[i],
output_dim=len(model_config.label_position))
model.fit_generator(generator=train_flow,
steps_per_epoch=model_config.get_steps_per_epoch(i),
epochs=model_config.epoch[i],
workers=16,
callbacks=[checkpoint])
else:
model = get_model(freeze_layers=model_config.freeze_layers[i], output_dim=len(model_config.label_position),
lr=model_config.lr[i], weights=None)
print("####### load weight file: %s" % model_config.get_weights_path(model_config.epoch[i - 1]))
model.load_weights(model_config.get_weights_path(model_config.epoch[i - 1]))
model.fit_generator(generator=train_flow,
steps_per_epoch=model_config.get_steps_per_epoch(i),
epochs=model_config.epoch[i],
initial_epoch=model_config.epoch[i - 1],
workers=16,
callbacks=[checkpoint])
print("####### train model spend %d seconds" % (time.time() - start))
print("####### train model spend %d seconds average" % ((time.time() - start) / model_config.epoch[-1]))
def train():
cb = []
evaluate_queue = queue.Queue()
evaluate_task = keras_util.EvaluateTask(evaluate_queue)
evaluate_task.setDaemon(True)
evaluate_task.start()
checkpoint = keras_util.EvaluateCallback(model_config, evaluate_queue)
cb.append(checkpoint)
start = time.time()
model_config.save_log("####### start train model")
init_stage = model_config.get_init_stage()
model_config.save_log("####### init stage is %d" % init_stage)
for i in range(init_stage, len(model_config.epoch)):
model_config.save_log(
"####### lr=%f, freeze layers=%2f epoch=%d" %
(model_config.lr[i], model_config.freeze_layers[i],
model_config.epoch[i]))
if model_config.clr:
clr = keras_util.CyclicLrCallback(
base_lr=model_config.lr[i],
max_lr=model_config.lr[i] * 5,
step_size=model_config.get_steps_per_epoch(i) / 2)
cb.append(clr)
train_flow = data_loader.KerasGenerator(model_config=model_config) \
.flow_from_files(model_config.train_files,
mode="fit",
target_size=model_config.image_size,
batch_size=
model_config.train_batch_size[i],
shuffle=True,
label_position=model_config.label_position)
if i == 0:
model_config.save_log(
"####### initial epoch is 0, end epoch is %d" %
model_config.epoch[i])
model = get_model(freeze_layers=model_config.freeze_layers[i],
lr=model_config.lr[i],
output_dim=len(model_config.label_position))
model.fit_generator(
generator=train_flow,
steps_per_epoch=model_config.get_steps_per_epoch(i),
epochs=model_config.epoch[i],
workers=16,
verbose=1,
callbacks=cb)
else:
model = get_model(freeze_layers=model_config.freeze_layers[i],
output_dim=len(model_config.label_position),
lr=model_config.lr[i],
weights=None)
if i == init_stage:
model_config.save_log(
"####### load weight file: %s" %
model_config.get_weights_path(model_config.initial_epoch))
model.load_weights(
model_config.get_weights_path(model_config.initial_epoch))
model_config.save_log(
"####### initial epoch is %d, end epoch is %d" %
(model_config.initial_epoch, model_config.epoch[i]))
model.fit_generator(
generator=train_flow,
steps_per_epoch=model_config.get_steps_per_epoch(i),
epochs=model_config.epoch[i],
initial_epoch=model_config.initial_epoch,
workers=16,
verbose=1,
callbacks=cb)
else:
model_config.save_log(
"####### load weight file: %s" %
model_config.get_weights_path(model_config.epoch[i - 1]))
model.load_weights(
model_config.get_weights_path(model_config.epoch[i - 1]))
model_config.save_log(
"####### initial epoch is %d, end epoch is %d" %
(model_config.epoch[i - 1], model_config.epoch[i]))
model.fit_generator(
generator=train_flow,
steps_per_epoch=model_config.get_steps_per_epoch(i),
epochs=model_config.epoch[i],
initial_epoch=model_config.epoch[i - 1],
workers=16,
verbose=1,
callbacks=cb)
model_config.save_log("####### train model spend %d seconds" %
(time.time() - start))
model_config.save_log("####### train model spend %d seconds average" %
((time.time() - start) / model_config.epoch[-1]))
# 等待最后一次预测结束
time.sleep(60)
