class_weight=None,
max_queue_size=10,
initial_epoch=iter00,
verbose=1,
workers=workers_num)
iter00 = iter00 + 1
''' Test after N*(display_status_ratio) iteration.'''
weight_tmp1 = model.get_weights()
model_512.set_weights(weight_tmp1)
train_output = model_512.predict(
[traindata_90d, traindata_0d, traindata_45d, traindata_m45d],
batch_size=1)
''' Save prediction image(disparity map) in 'current_output/' folder '''
train_error, train_bp = display_current_output(train_output,
traindata_label, iter00,
directory_t)
training_mean_squared_error_x100 = 100 * np.average(
np.square(train_error))
training_bad_pixel_ratio = 100 * np.average(train_bp)
save_path_file_new = (directory_ckp +
'/iter%04d_trainmse%.3f_bp%.2f.hdf5' %
(iter00, training_mean_squared_error_x100,
training_bad_pixel_ratio))
"""
Save bad pixel & mean squared error
"""
print(save_path_file_new)
f1 = open(txt_name, 'a')
model.fit_generator(my_generator, steps_per_epoch = steps_per_epoch,
epochs = iter00+1, class_weight=None, max_queue_size=10,
initial_epoch=iter00, verbose=1,workers=workers_num,callbacks=[reduce_lr])
iter00=iter00+1
''' Test after N*(display_status_ratio) iteration.'''
weight_tmp1=model.get_weights()
model_512.set_weights(weight_tmp1)
train_output=model_512.predict([train512_data["90d"],train512_data["0d"],train512_data["45d"],train512_data["m45d"]],batch_size=1)
val_output=model_512.predict([val512_data["90d"],val512_data["0d"],val512_data["45d"],val512_data["m45d"]],batch_size=1)
''' Save prediction image(disparity map) in 'current_output/' folder '''
train_error, train_bp = display_current_output(train_output, train512_data["label"], iter00, output_dir, split="train")
val_error, val_bp = display_current_output(val_output, val512_data["label"], iter00, output_dir, split='val')
# training_mean_squared_error_x100=100*np.average(np.square(train_error))
# training_bad_pixel_ratio=100*np.average(train_bp)
train_mse = np.average(np.square(train_error))
val_mse = np.average(np.square(val_error))
val_mean_squared_error_x100 = 100*val_mse
val_bad_pixel_ratio=100*np.average(val_bp)
val_loss.append(val_mse)
train_loss.append(train_mse)
if plot_losses:
plt.plot(train_loss)
plt.plot(val_loss)
k=tcind2k[test_corner_ind],
axes=(1, 2))
backrot_val_label = np.rot90(val512_data[test_corner]["label"],
k=tcind2k[test_corner_ind],
axes=(1, 2))
backrot_train_output = np.rot90(train_output,
k=tcind2k[test_corner_ind],
axes=(1, 2))
backrot_train_label = np.rot90(train512_data[test_corner]["label"],
k=tcind2k[test_corner_ind],
axes=(1, 2))
train_error, train_bp = display_current_output(backrot_train_output,
backrot_train_label,
iter00,
output_dir,
split='train',
corner=test_corner)
val_error, val_bp = display_current_output(backrot_val_output,
backrot_val_label,
iter00,
output_dir,
split='val',
corner=test_corner)
# training_mean_squared_error_x100=100*np.average(np.square(train_error))
# training_bad_pixel_ratio=100*np.average(train_bp)
train_mse = np.average(np.square(train_error))
val_mse = np.average(np.square(val_error))
val_mean_squared_error_x100 = 100 * val_mse
val_bad_pixel_ratio = 100 * np.average(val_bp)
def main():
parser = argparse.ArgumentParser()
parser.add_argument('--dir',
type=str,
default='epinet_output/mytrain/',
help='save directory')
args = parser.parse_args()
'''
settings
'''
directory_t = args.dir
if not os.path.exists(directory_t):
os.makedirs(directory_t)
Setting02_AngualrViews = np.array([0, 1, 2, 3, 4, 5, 6, 7, 8])
'''
Define Model parameters
first layer: 3 convolutional blocks,
second layer: 7 convolutional blocks,
last layer: 1 convolutional block
'''
model_conv_depth = 7 # 7 convolutional blocks for second layer
model_filt_num = 70
model_learning_rate = 0.1**7
'''
Load Train data from LF .png files
'''
print('Load training data...')
dir_LFimages = [
'additional/antinous', 'additional/boardgames', 'additional/dishes',
'additional/greek', 'additional/kitchen', 'additional/medieval2',
'additional/museum', 'additional/pens', 'additional/pillows',
'additional/platonic', 'additional/rosemary', 'additional/table',
'additional/tomb', 'additional/tower', 'additional/town',
'additional/vinyl'
]
traindata_all, traindata_label = load_LFdata(dir_LFimages)
traindata_90d, traindata_0d, traindata_45d, traindata_m45d, _ = generate_traindata512(
traindata_all, traindata_label, Setting02_AngualrViews)
'''
Load Test data from LF .png files
'''
print('Load test data...')
dir_LFimages = [
'stratified/backgammon', 'stratified/dots', 'stratified/pyramids',
'stratified/stripes', 'training/boxes', 'training/cotton',
'training/dino', 'training/sideboard'
]
valdata_all, valdata_label = load_LFdata(dir_LFimages)
valdata_90d, valdata_0d, valdata_45d, valdata_m45d, valdata_label = generate_traindata512(
valdata_all, valdata_label, Setting02_AngualrViews)
# (valdata_90d, 0d, 45d, m45d) to validation or test
print('Load test data... Complete')
'''
Model for predicting full-size LF images
'''
image_w = 512
image_h = 512
model_512 = define_epinet(image_w, image_h, Setting02_AngualrViews,
model_conv_depth, model_filt_num,
model_learning_rate)
'''
load weights
'''
# model_512.load_weights('epinet_checkpoints/pretrained_9x9.hdf5')
model_512.load_weights(
'epinet_checkpoints/EPINET_train_ckp/iter0198_trainmse0.827_bp2.56.hdf5'
)
'''
show train results
'''
train_output = model_512.predict(
[traindata_90d, traindata_0d, traindata_45d, traindata_m45d],
batch_size=1)
train_error, train_bp = display_current_output(train_output,
traindata_label, 0,
directory_t)
training_mean_squared_error_x100 = 100 * np.average(np.square(train_error))
training_bad_pixel_ratio = 100 * np.average(train_bp)
print('train mse%.3f,bp%.2f' %
(training_mean_squared_error_x100, training_bad_pixel_ratio))
val_output = model_512.predict(
[valdata_90d, valdata_0d, valdata_45d, valdata_m45d], batch_size=1)
val_error, val_bp = display_current_output(val_output, valdata_label, 1,
directory_t)
val_mean_squared_error_x100 = 100 * np.average(np.square(val_error))
val_bad_pixel_ratio = 100 * np.average(val_bp)
print('val mse%.3f,bp%.2f' %
(val_mean_squared_error_x100, val_bad_pixel_ratio))