model_learning_rate = 0.1**5 model_512 = define_epinet(round(img_scale * image_h), round(img_scale * image_w), Setting02_AngualrViews, model_conv_depth, model_filt_num, model_learning_rate) ''' Model Initialization ''' model_512.load_weights(path_weight) dum_sz = model_512.input_shape[0] dum = np.zeros((1, dum_sz[1], dum_sz[2], dum_sz[3]), dtype=np.float32) dummy = model_512.predict([dum, dum, dum, dum], batch_size=1) """ Depth Estimation """ for image_path in dir_LFimages: (val_90d, val_0d, val_45d, val_M45d) = make_multiinput(image_path, image_h, image_w, Setting02_AngualrViews) # print(val_90d.shape) # exit() start = time.time() # predict val_output_tmp = model_512.predict([ val_90d[:, ::img_scale_inv, ::img_scale_inv], val_0d[:, ::img_scale_inv, ::img_scale_inv], val_45d[:, ::img_scale_inv, ::img_scale_inv], val_M45d[:, ::img_scale_inv, ::img_scale_inv] ], batch_size=1) runtime = time.time() - start
center_offset = (0, 0) # center_os_x = center_x - (an - 1)//2 # center_os_y = center_y - (an - 1)//2 # center_offset = (center_os_x, center_os_y) #if subsample_type == "center": print("center offset: ") print(center_offset) for image_path in dir_LFimages: (val_90d, val_0d, val_45d, val_M45d, impaths) = make_multiinput(image_path, image_h, image_w, Setting02_AngualrViews, ds=ds, center_offsets=center_offset, net_type="original", corner_code=None) start = time.clock() all_outputs = model.predict([ val_90d[:, ::img_scale_inv, ::img_scale_inv], val_0d[:, ::img_scale_inv, ::img_scale_inv], val_45d[:, ::img_scale_inv, ::img_scale_inv], val_M45d[:, ::img_scale_inv, ::img_scale_inv] ], batch_size=1) # predict val_output_tmp = all_outputs[-1]
# write_graph=True, write_images=True) # #tbCallback.set_model(model_512) """ Depth Estimation """ for center_coord in center_coords: center_offset = infer_cos_from_cc(center_coord,ds) #for center_offset in center_offsets: subsample = ("center",center_offset) #"center" #"upperleft" print ( "center offset: " ) print ( center_offset ) # for image_path in dir_LFimages: for sample in sample_names: (val_90d , val_0d, val_45d, val_M45d,impaths)=make_multiinput(sample, Setting02_AngualrViews, ds=ds, center_offsets = center_offset, net_type='original', crop_size=crop_size, crop_or_resize=crop_or_resize) start=time.clock() # predict val_output_tmp=model_512.predict([ val_90d[:,::img_scale_inv,::img_scale_inv], val_0d[:,::img_scale_inv,::img_scale_inv], val_45d[:,::img_scale_inv,::img_scale_inv], val_M45d[:,::img_scale_inv,::img_scale_inv]], batch_size=1); runtime=time.clock() - start plt.imshow(val_output_tmp[0,:,:,0]) iar_center, iac_center = center_coord
center_offset = infer_cos_from_corner(corner_coords, corner_code, an=ds.iar_max, model_cw=model_cw) print("center offset: ") print(center_offset) print("corner code:") print(corner_code) for sample in samples: (val_90d, val_0d, val_M45d, impaths) = make_multiinput(sample, Setting02_AngualrViews, ds=ds, center_offsets=center_offset, net_type="corner", corner_code=corner_code) start = time.clock() # predict val_output_tmp = model_512.predict([ val_90d[:, ::img_scale_inv, ::img_scale_inv], val_0d[:, ::img_scale_inv, ::img_scale_inv], val_M45d[:, ::img_scale_inv, ::img_scale_inv] ], batch_size=1) if corner_code == "NE": val_output_tmp = np.rot90(val_output_tmp, axes=(2, 1))
model_learning_rate = 0.1**5 model_512 = define_epinet(int(img_scale_inv * image_h), int(img_scale_inv * image_w), angular_views, model_conv_depth, model_filt_num, model_learning_rate) ''' Model Initalization ''' model_512.load_weights(path_weight) dum_sz = model_512.input_shape[0] dum = np.zeros((1, dum_sz[1], dum_sz[2], dum_sz[3]), dtype=np.float32) dummy = model_512.predict([dum, dum, dum, dum], batch_size=1) """ Depth Estimation """ for image_path in dir_LFimages: (val_90d, val_0d, val_45d, val_M45d) = make_multiinput(image_path, image_h, image_w, angular_views) start = time.clock() # predict val_output_tmp = model_512.predict([ val_90d[:, ::img_scale_inv, ::img_scale_inv], val_0d[:, ::img_scale_inv, ::img_scale_inv], val_45d[:, ::img_scale_inv, ::img_scale_inv], val_M45d[:, ::img_scale_inv, ::img_scale_inv] ], batch_size=1) runtime = time.clock() - start plt.imshow(val_output_tmp[0, :, :, 0]) print("%.5f(s)" % runtime)