sess = tf.Session() K.set_session(sess) model = RCNet_THdim_model() # model = RCNet_THdim_dropout_model() # model = RCNet_shortdense_THdim_model() model.summary() from yolo.training_v1 import darkeras_loss, _TRAINER from yolo.datacenter.data import shuffle, test_shuffle inp_x = model.input net_out = model.output say("Building {} loss function".format(cfg.model_name), verbalise=verbalise) loss_ph, loss_op = darkeras_loss(net_out) say("Building {} train optimizer".format(cfg.model_name), verbalise=verbalise) optimizer = _TRAINER[cfg.trainer](cfg.lr) gradients = optimizer.compute_gradients(loss_op, var_list=model.trainable_weights) train_op = optimizer.apply_gradients(gradients) sess.run(tf.global_variables_initializer()) model.load_weights(pretrain_weight_path, by_name=True) # End User Custom Training Function collect_enduser_trainset() batches = shuffle()
show_trainable_state = False # 여기를 True 로 바꾸면, conv layer 와 dense layer 의 학습별 weigths 가 변하는지 안변하는지를 확인할 수 있다. trained_save_weights_prefix = 'models/train/{}-'.format(cfg.model_name) sess = tf.Session() K.set_session(sess) model = yolo_vgg16_TFdim_model(is_top=True, is_new_training=True) model.summary() from yolo.training_v1 import darkeras_loss, _TRAINER from yolo.datacenter.data import shuffle inp_x = model.input net_out = model.output say("Building {} loss function".format(cfg.model_name), verbalise=verbalise) loss_ph, loss_op = darkeras_loss(net_out) say("Building {} train optimizer".format(cfg.model_name), verbalise=verbalise) optimizer = _TRAINER[cfg.trainer](cfg.lr) gradients = optimizer.compute_gradients(loss_op, var_list=model.trainable_weights) train_op = optimizer.apply_gradients(gradients) sess.run(tf.global_variables_initializer()) model.load_weights(pretrain_weight_path, by_name=True) batches = shuffle() for i, (x_batch, datum) in enumerate(batches): train_feed_dict = { loss_ph[key]:datum[key] for key in loss_ph }
model.add(LeakyReLU(alpha=0.1)) model.add(Dense(1470)) return model model = make_yolotiny_network(is_freeze) model.summary() from yolo.training_v1 import darkeras_loss, _TRAINER from yolo.datacenter.data import shuffle inp_x = model.input net_out = model.output sess = K.get_session() say("Building {} loss function".format(cfg.model_name), verbalise=verbalise) loss_ph, loss_op = darkeras_loss(net_out) say("Building {} train optimizer".format(cfg.model_name), verbalise=verbalise) optimizer = _TRAINER[cfg.trainer](cfg.lr) gradients = optimizer.compute_gradients(loss_op) train_op = optimizer.apply_gradients(gradients) sess.run(tf.global_variables_initializer()) model.load_weights(weights_path) say("Setting weigths : {}", format(weights_path), verbalise=verbalise) model.summary() print(model.output_shape) batches = shuffle() for i, (x_batch, datum) in enumerate(batches):
# In[6]: from yolo.training_v1 import darkeras_loss, _TRAINER from yolo.datacenter.data import shuffle # In[7]: inp_x = model.input net_out = model.output sess = K.get_session() say("Building {} loss function".format(cfg.model_name), verbalise=verbalise) loss_ph, loss_op = darkeras_loss(net_out) say("Building {} train optimizer".format(cfg.model_name), verbalise=verbalise) optimizer = _TRAINER[cfg.trainer](cfg.lr) gradients = optimizer.compute_gradients(loss_op) train_op = optimizer.apply_gradients(gradients) sess.run(tf.global_variables_initializer()) model.load_weights(weights_path) say("Setting weigths : {}",format(weights_path), verbalise=verbalise) pop_layer = model.pop() # dense_25 say("{} layer poped".format(pop_layer), verbalise=verbalise) pop_layer = model.pop() # leakyrelu_34 say("{} layer poped".format(pop_layer), verbalise=verbalise) pop_layer = model.pop() # dense_24
sess = tf.Session() K.set_session(sess) model = RCNet_THdim_model() # model = RCNet_THdim_dropout_model() # model = RCNet_shortdense_THdim_model() model.summary() from yolo.training_v1 import darkeras_loss, _TRAINER from yolo.datacenter.data import shuffle, test_shuffle inp_x = model.input net_out = model.output say("Building {} loss function".format(cfg.model_name), verbalise=verbalise) loss_ph, loss_op = darkeras_loss(net_out) say("Building {} train optimizer".format(cfg.model_name), verbalise=verbalise) optimizer = _TRAINER[cfg.trainer](cfg.lr) gradients = optimizer.compute_gradients(loss_op, var_list=model.trainable_weights) train_op = optimizer.apply_gradients(gradients) sess.run(tf.global_variables_initializer()) model.load_weights(pretrain_weight_path, by_name=True) batches = shuffle() train_histories = {} train_histories['train_loss'] = []