Fast Super Resolution Convolutional Nets

PS. Its my first attempt to implement a paper, so please bear with me in case i do/say something really stupid
Can someone help me with this one?
Hi all,
I was trying to implement the model as defined here in this FSRCNN(scroll a bit down until the headline says FSRCNN), but i am pretty sure i messed up with the structure in the end,
Can someone invest their spare time to help me as to where i am wrong?


The problem is that the loss is not decreasing at all…(due to structural defects at the end…)

Also I am getting list index out of range error while trying to apply Deconv op

The dataset

here's the traceback
IndexError                                Traceback (most recent call last)
<ipython-input-10-5c97c0961611> in <module>()
----> 1 train_cnn(X)

<ipython-input-9-da512eedc9e1> in train_cnn(X)
      1 def train_cnn(X):
----> 2     prediction = CNN(X)
      3     #mse
      4     loss = tf.reduce_mean(tf.pow(tf.subtract(prediction, Y), 2.0))
      5     #optimizer

<ipython-input-8-34dbf6cd2f6a> in CNN(X)
     74     have a look at the link below...
     75     '''
---> 76     op_image = tf.nn.conv2d_transpose(c0, filter = [9, 9, 56, 12], output_shape = [1,320,480,1], strides=[1,2,2,1],padding='same',name='deconv')
     78     return op_image

C:\ProgramData\Anaconda3\lib\site-packages\tensorflow\python\ops\ in conv2d_transpose(value, filter, output_shape, strides, padding, data_format, name)
   1017     filter = ops.convert_to_tensor(filter, name="filter")
   1018     axis = 3 if data_format == "NHWC" else 1
-> 1019     if not value.get_shape()[axis].is_compatible_with(filter.get_shape()[3]):
   1020       raise ValueError("input channels does not match filter's input channels, "
   1021                        "{} != {}".format(value.get_shape()[3], filter.get_shape(

C:\ProgramData\Anaconda3\lib\site-packages\tensorflow\python\framework\ in __getitem__(self, key)
    498         return TensorShape(self._dims[key])
    499       else:
--> 500         return self._dims[key]
    501     else:
    502       if isinstance(key, slice):

IndexError: list index out of range

I tried to implement it as given in the stack’s link…

Can you show your code?

there’s a link to my code

Why is the dataset small? Superresolution is one of the few image problems where data is abundant. All you need are the larger sized images.

Actually I was just trying it out on few images

Actually the problem lies with the few lines when I am writing

c7= TF.add(c0,c6)
And what follows thereafter..

Should I comment my code and re-upload?

Yes that would be good. You’ll get more attention / help if you also upload to Github.

  1. Dataset Link
  2. The research paper link
  3. Implementation of re-search paper (its working)
  4. Implementation of the Intel Link
  5. Previous folder link