ImageSR Machine Learning Troubleshooting Expert

I've been implementing a paper on Image Super-Resolution using Pytorch, but my model seems to be failing to learn. I'm looking for an experienced consultant in Machine Learning, particularly in the field of Image Super-Resolution, to help me understand what could be going wrong with my current implementation.

ImageSR paper: https://ieeexplore.ieee.org/document/8489155
Author's MATLAB code: https://github.com/junxuan-li/A-frequency-domain-neural-network-for-fast-image-super-resolution

Note that, since this is an image super-resolution problem that improves the input image, without changing the image dimensions, I have computed the PSNR just providing the input as model output to have a baseline, which yields 29.5, this is a baseline I first adopted.

I am having some difficulties getting the model to learn beyond this value. I am wondering if you would have any insights on the reason for that.

I am using the VOC2012 to train the data and the SET14 to test it.
I am training up to 50 epochs. But previously I had long runs, up to 200 epochs without any improvement.
I implemented your model described in the `cnn_init.m` in your MATLAB code, that is:
--> 6 layers, each be defined as
====> `conv(w(x+b))`.
====> the convolution uses a kernel of size 5
====> 1st part: the w(x+b) part takes an input of 1 single channel and expands it to 9 channels
====> 2nd part: the convolution takes 9 channels as input, and outputs a single channel.
====> the backpropagation of the 1st part applies a function to make the derivatives w.r.t. w and b are symmetric.
--> Each layer above produces a result that feeds to the next.
--> All the outputs from the layers above are summed, in a final layer.
--> Gradients are clipped to be in the range [-1e3, 1e3].

I believe I covered all the details presented in that work (paper + Matlab code), but still, I am unable to have the model learning.

Key Responsibilities:
- Meet with me in a video call online for me to show you my code and have you to help me getting it to learn.
- Identify potential reasons for the failure of my implementation of the paper.
- Suggest and potentially implement modifications to the code, if necessary.

Additional Details:
- So far, I haven't encountered any errors or warnings during the training process.
- I evaluate the performance of my model by comparing it to the paper's reported results, using a validation dataset, and monitoring loss and accuracy metrics during training.

Ideal Skills and Experience:
- Strong background in Machine Learning, specifically in the field of Image Super-Resolution.
- Experience in debugging and optimizing machine learning models.
- Proficiency in Pytorch is a must.
- Excellent problem-solving skills.