kingfischer16 / srcnn-tf2 Goto Github PK

Add functions for preparing data for feeding into model. May be able to include augmentation in this. Should handle both training and test data.

Add functions for training and benchmarking. Should cover the basic training and benchmarking found in the paper, as well as make room for more advanced training (other loss functions, other models).

Add the file for the trained model for FSRCNN and FSRCNN-s, as well as the training meta info (data, preproc, metrics and loss, performance, timing, GPU). Compare scores to relevant articles.

Add the file for the trained SRCNN-Ex model, as well as meta info (data, preproc, training params, GPU, loss and metrics, performance). Compare performance scores to relevant benchmarks from articles.

Add the FSCRNN and FSCRNN-s model architectures to the library.

Add the file with the trained original SRCNN model, training meta info (data, epochs, preproc, GPU, timing, scores). Should be compared to the original paper by Dong et al.

Add the model builder function to the repo. Should be able to build the basic SRCNN model found in the paper as well as variants of layers (number and size).

Will likely need to add code to visualize the results and do side-by-side comparison, as well as matching zoom and pan operations for image inspection. Should also display the scores (PSNR, other metrics) of the real and generated images beside each other.

Add the improved SRCNN-Ex model variant. Find the reference paper for this.

Create a pretrained folder with a library of models. First should be the SRCNN model as found in the paper, with matching benchmark scores. Possibly add ensemble models, models trained with other loss functions, augmented data, or alternative datasets (with scores as well).

Build the processing code to use a pre-trained model in a basic Python-enabled and HTTP-enabled API using AWS Lambda.