if platform.system() == 'Darwin':
os.environ['KMP_DUPLICATE_LIB_OK'] = 'True'
if model_path:
model = models.load_model(
model_path,
custom_objects={'DivByRedSum': DivByRedSum,
'mean_pred' : mean_pred,
'max_pred': max_pred,
'min_pred' : min_pred,
'mean_true': mean_true,
'max_true': max_true,
'min_true':min_true})
else:
print('Using normal model.')
model = vgg16_enc_dec_model((img_height, img_width, img_channels),
freeze_vgg = False,
output_choice = 'sigmoid',
**{'reg':None,'dropout':0.1})
model.load_weights(weights_path)
if tr_or_val == 'tr':
image_paths = get_img_paths(path_train+"/images")
mask_paths = get_img_paths(path_train+"/masks")
else:
image_paths = get_img_paths(path_test+"/images")
mask_paths = get_img_paths(path_test+"/masks")
#TODO: If there is a problem with NaN values, one can try this:
#for n,layer in enumerate(model.layers):
# weights = layer.get_weights()
# new_weights = [np.nan_to_num(w) for w in weights]
# layer.set_weights(new_weights)
custom_objects={
'DivByRedSum': DivByRedSum,
'mean_pred': mean_pred,
'max_pred': max_pred,
'min_pred': min_pred,
'mean_true': mean_true,
'max_true': max_true,
'min_true': min_true
},
compile=True)
print('Loaded model from {}'.format(args.load_model))
else:
print('Using new model with {}'.format(args.model_output_choice))
template_model = vgg16_enc_dec_model(
(img_height, img_width, img_channels),
freeze_vgg=False,
output_choice=args.model_output_choice,
**counter_overfit[args.counter_overfit])
n_gpu = len(get_available_gpus())
print('Number of gpu to use: {}'.format(n_gpu))
if n_gpu > 1 and not do_not_use_mult_gpu:
model = multi_gpu_model(template_model, gpus=n_gpu)
print("Training using multiple GPUs..")
else:
model = template_model
print("Training using single GPU or CPU..")
print('Using {} as loss function'.format(loss_dict[args.loss]))
model.compile(optimizer=keras.optimizers.Adam(clipnorm=1.0),