Testing monochromatic images on your code
iucurmak opened this issue · 6 comments
How can i apply to testing your code for monochromatic images. Is it possible? Or what i need to change from your code? Could you please let me know?
The network is not trained for monochromatic images. In theory, you can feed it a monochromatic image by repeating it across the channel dimension. Meaning if you have a monochromatic image x with shape (H, W, 1), you can do tf.repeat it into (H, W, 3). The resulting bitrate will however be suboptimal, since the network is not trained to leverage the redundancies now.
In theory you could train a new network however, where the input and output is (H,W,1) dimensional instead.
Thank you for informing me. Which files i need to rearrange if i train the code for monochromatic images? And also which files need rearrangement for testing code?
Probably requires quite some changes -- essentially everywhere where the code assumes a channel dimension of 3. Some locations:
- In
inputpipeline.pytheimages_decodedand_preprocessfunctions - In
autoencoder.pytheget_mean_varfunction (must calculate new mean/variance or just use average of the means/variances there)
Starting from the changes in inputpipeline.py, you can probably run it and follow the errors.
Appreciate for your help. Have great day.
Let me know if you manage to train a model for monochromatic, would be interesting!