我正在用喀拉拉邦编写一个自定义的神经网络，必须将输入传递给一个函数，该函数计算帧列表中各帧之间的光流。但是，当我将代码包装在tensorflow tf.py_func（）中时，该函数不会求值并返回None。似乎无法找出一个简单的解决方法，将不胜感激。以下是我正在使用的代码，由于我是Tensorflow的新手，所以希望获得帮助，并想知道如何直接在keras中或在tensorflow中执行上述程序，然后使其与keras兼容：
def optical_flow(frame_list):
mid = len(frame_list)//2
base_frame = frame_list[mid]
new_list = frame_list.copy()
new_list = np.delete(new_list, mid, axis=0)
def calculate_optical_flow(one, two):
flow = cv2.calcOpticalFlowFarneback(one,
two,
flow=None,
pyr_scale=0.5, levels=1, winsize=15,
iterations=2,
poly_n=5, poly_sigma=1.1, flags=0)
return flow
def move_image_based_on_flow(prev_img, flow):
# Generate a cartesian grid
height, width = flow.shape[0], flow.shape[1]
R2 = np.dstack(np.meshgrid(np.arange(width), np.arange(height)))
# desired mapping is simply the addition of this grid with the flow
pixel_map = R2 + flow
pixel_map = pixel_map.astype(np.float32)
#perform the remapping of each pixel in the original image
warped_frame = cv2.remap(prev_img, pixel_map[:, :, 0], pixel_map[:, :, 1], cv2.INTER_LINEAR)
return warped_frame
warped_frames = []
for frame in new_list:
warped_bps_in_frame = []
for bp_ind, bp_heatmap in enumerate(frame):
flow = calculate_optical_flow(base_frame[:][bp_ind][:, :, np.newaxis], bp_heatmap[:, :, np.newaxis])
new_frame = move_image_based_on_flow(bp_heatmap[:, :, np.newaxis], flow)
warped_bps_in_frame.append(new_frame)
warped_frames.append(warped_bps_in_frame)
warped_frames = np.asanyarray(warped_frames, dtype=np.uint8)
base_frame = base_frame[np.newaxis, :, :, :]
warped_frames = np.insert(warped_frames, mid, [base_frame], axis=0)
return warped_frames
def optical_flow_n(x):
return tf.py_func(optical_flow, [x], tf.float32)
def optical_flow_net(x):
x = layers.Lambda(optical_flow_n, output_shape=(total_frames, num_bodyparts, frame_height, frame_width))(x)
x = layers.Conv3D(1, kernel_size=(1,1,total_frames), padding='same')(x)
return x
optical_flow_input = layers.Input(shape=(total_frames, num_bodyparts, frame_height, frame_width))
optical_flow_output = optical_flow_net(optical_flow_input)
model_ofn = models.Model(inputs=[optical_flow_input], outputs=[optical_flow_output])
运行上面的代码时出现以下错误：
---------------------------------------------------------------------------
ValueError Traceback (most recent call last)
<ipython-input-10-1a6519e3ddb3> in <module>()
1 optical_flow_input = layers.Input(shape=(total_frames, num_bodyparts, frame_height, frame_width))
----> 2 optical_flow_output = optical_flow_net(optical_flow_input)
3 model_ofn = models.Model(inputs=[optical_flow_input], outputs=[optical_flow_output])
<ipython-input-9-d6cf34c238c9> in optical_flow_net(x)
46 def optical_flow_net(x):
47 x = layers.Lambda(optical_flow_n, output_shape=(total_frames, num_bodyparts, frame_height, frame_width))(x)
---> 48 x = layers.Conv3D(1, kernel_size=(1,1,total_frames), padding='same')(x)
49 return x
~\AppData\Local\conda\conda\envs\DLCdependencies\lib\site-packages\keras\engine\base_layer.py in __call__(self, inputs, **kwargs)
412 # Raise exceptions in case the input is not compatible
413 # with the input_spec specified in the layer constructor.
--> 414 self.assert_input_compatibility(inputs)
415
416 # Collect input shapes to build layer.
~\AppData\Local\conda\conda\envs\DLCdependencies\lib\site-packages\keras\engine\base_layer.py in assert_input_compatibility(self, inputs)
309 self.name + ': expected ndim=' +
310 str(spec.ndim) + ', found ndim=' +
--> 311 str(K.ndim(x)))
312 if spec.max_ndim is not None:
313 ndim = K.ndim(x)
ValueError: Input 0 is incompatible with layer conv3d_1: expected ndim=5, found ndim=None