HR_inten_batch_input=tf.placeholder(tf.float32,HR_batch_dims)
LR_depth_batch_input=tf.placeholder(tf.float32,LR_batch_dims)
coar_inter_dep_batch=tf.image.resize_images(LR_depth_batch_input,tf.constant(HR_patch_size,dtype=tf.int32),tf.image.ResizeMethod.BICUBIC)
#gen_network construction
with tf.variable_scope("gen_inten"):
guided_ten=gslb.inten_feature_extraction_unit(HR_inten_batch_input,phase_train=phase_train)
with tf.variable_scope("gen_down1_inten"):
guided_4xto2x_gen=gslb.inten_downsample_unit(guided_ten, phase_train=phase_train)
with tf.variable_scope("gen_down2_inten"):
guided_8xto4x_gen=gslb.inten_downsample_unit(guided_4xto2x_gen, phase_train=phase_train)
with tf.variable_scope("gen_dep"):
dep_ten=gslb.LR_dep_feature_extraction_uint(LR_depth_batch_input,phase_train=phase_train)
dep_ten=gslb.LR_dep_upsampling_unit(dep_ten, phase_train=phase_train)
with tf.variable_scope("gen_up3_dep"):
dep_ten=gslb.LR_dep_fusion_unit(dep_ten, guided_8xto4x_gen, phase_train=phase_train)
dep_ten=gslb.LR_dep_upsampling_unit(dep_ten, phase_train=phase_train)
with tf.variable_scope("gen_up2_dep"):
dep_ten=gslb.LR_dep_fusion_unit(dep_ten, guided_4xto2x_gen, phase_train=phase_train)
dep_ten=gslb.LR_dep_upsampling_unit(dep_ten, phase_train=phase_train)
with tf.variable_scope("gen_8x_last4_layers"):
gen_ten=gslb.LR_dep_fusion_unit(dep_ten, guided_ten,phase_train=phase_train)
gen_ten=gslb.LR_recon_unit(gen_ten, coar_inter_dep_batch,phase_train=phase_train)
#define loss for gen
loss=tf.reduce_mean(tf.squared_difference(gen_ten,HR_depth_batch_input))
saver_full=tf.train.Saver()
#begin comp_gen testing
with tf.Session() as sess:
model_path="/media/kenny/Data/trained_models/residual_bn_csvt_models/noise_free/8x/full_model1/8x_nf_full_model.ckpt-98"
LR_depth_batch_input = tf.placeholder(tf.float32, LR_batch_dims)
coar_inter_dep_batch = tf.image.resize_images(
LR_depth_batch_input, tf.constant(HR_patch_size, dtype=tf.int32),
tf.image.ResizeMethod.BICUBIC)
#gen_network construction
with tf.variable_scope("gen_inten"):
guided_ten = gslb.inten_feature_extraction_unit(HR_inten_batch_input,
phase_train=phase_train)
with tf.variable_scope("gen_dep"):
dep_ten = gslb.LR_dep_feature_extraction_uint(LR_depth_batch_input,
phase_train=phase_train)
dep_ten = gslb.LR_dep_upsampling_unit(dep_ten, phase_train=phase_train)
with tf.variable_scope("gen_2x_last4_layers"):
gen_ten = gslb.LR_dep_fusion_unit(dep_ten,
guided_ten,
phase_train=phase_train)
gen_ten = gslb.LR_recon_unit(gen_ten,
coar_inter_dep_batch,
phase_train=phase_train)
#define loss for gen
loss = tf.reduce_mean(tf.squared_difference(gen_ten, HR_depth_batch_input))
train_op_small = tf.train.AdamOptimizer(1e-5).minimize(loss)
train_op_large = tf.train.AdamOptimizer(1e-4).minimize(loss)
#initial net and saver object
#var_list=tf.trainable_variables()
#gen_dep_vars = [v for v in var_list if v.name.startswith("gen_dep")]
#gen_inten_vars=[v for v in var_list if v.name.startswith("gen_inten")]
#saving_var_list=gen_dep_vars+gen_inten_vars