cjerry1243 / mobilenetv2_tf_to_caffe Goto Github PK

When I add the code : compare_layer_output(net, 'conv1', checkpoint, 'MobilenetV2/Conv/Conv2D:0', image_file) in converter_v2.py
error occurs:
('Save converted caffemodel to', 'caffemodel_fromckpt/mobilenet_v2_1.0_224.caffemodel')
Traceback (most recent call last):
File "converter_v2.py", line 297, in
compare_layer_output(net, 'conv1', checkpoint, 'MobilenetV2/Conv/Conv2D:0', image_file)
File "converter_v2.py", line 278, in compare_layer_output
tf_output = sess.run(tensor, feed_dict={images: image})
File "/home/lyq/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 900, in run
run_metadata_ptr)
File "/home/lyq/.local/lib/python2.7/site-packages/tensorflow/python/client/session.py", line 1111, in _run
str(subfeed_t.get_shape())))
ValueError: Cannot feed value of shape (3, 224, 224) for Tensor u'Placeholder:0', which has shape '(?, 224, 224, 3)'

I solved ther problem by change the channl, but there is alos problems,,
Traceback (most recent call last):
File "run_ckpt.py", line 126, in
compare_layer_output(net, layer_name, checkpoint_path, tensor_name, image_file)
File "run_ckpt.py", line 107, in compare_layer_output
err = square_error(tf_output[0, :, :, i], caffe_output[0, :, :, i])
File "run_ckpt.py", line 75, in square_error
return np.sum(np.square(x-x_))
ValueError: operands could not be broadcast together with shapes (112,112) (113,113)

why the channel output of tf and caffe are not equal, that is vey strange

Thanks for your job. I have test your tool, it's great.
And when I check the prototxt which you provided, I found there is not exist "conv2_1/expand" block? But when I check other's implement, there is a "conv2_1/expand" block used 1*1 conv operate.