network model depends on the input size

[guoyejun]

Hi,

i'm trying to use this method in a c program which invokes tensorflow C API, so my way is to first save the network model in python program:
graph = tf.graph_util.convert_variables_to_constants(sess, sess.graph_def, ["y"])
tf.train.write_graph(graph, '.', 'graph.pb', as_text=False)

But, the saved network model depends on the size of the input image, due to the following code which requires output_shape. It means we have to prepare a model file (130M+) for each possible input size. IMHO, both conv and deconv are not depend on the size of the input image.

network = tl.layers.DeConv2dLayer(input_layer,
                            shape = [filter_size, filter_size, num_out_channels, num_in_channels],
                            output_shape = [sz[0], sz[1]*scale, sz[2]*scale, num_out_channels],
                            strides=[1, scale, scale, 1],
                            W_init=init_matrix,
                            padding='SAME',
                            act=tf.identity,
                            name=str)

Is there another API to be choosen? or we need to ask tensorflow to introduce a new API without output_shape but (sz[0], scale, num_out_channles)? thanks

[gabrieleilertsen]

Hi! And sorry for the late reply. I'm afraid that I'm unable to help you. I have no experience with tensorflow in C, but it indeed seems limiting that an exported model is tied to a specific resolution. Since the model describes a fully convolutional network, it would should be able to take different resolutions.

I'll leave this issue open, in case someone else can help with the problem.