Deep Image Prior

usage: python deepimprior.py ./data/im03.png --outpath ./data/outimages/ --outprefix out_denoised --verbose 1 --epochs 10000 --learning_rate 0.1 --save_model_path ./data/outmodels/model --period 20Denoise a given image using deep image prior algorithm Beware of the followingissues before proceeding with the usage of this script: - Convolution basedalgorithms are sensible to image size. Please use a square image. Ex: 800x800,or 600x600. - Image size significantly effects the training. Either make sureyou have enough computation power, or adjust the image size appropriately. -Lastly as with all the gradient based methods we are using a stable learningrate. Feel free to adjust it before the training phase. I am thinking ofadding decay learning rate optionin the future..positional arguments:  imagepath             path to the imageoptional arguments:  -h, --help            show thishelp message andexit  --outpath OUTPATH     pathfor saving outputs  --outprefix OUTPREFIX                        prefix that will e prepended to intermediate files  --epochs EPOCHS       number of training epochs  --verbose {0,1}       verbose output during training  --learning_rate LEARNING_RATE                        learning ratefor the optimizer  --save_model_path SAVE_MODEL_PATH                        Save model to path at each period of epochs  --load_model_path LOAD_MODEL_PATH                        Load model from path to resume training  --period PERIOD       Periodic activity number, saving images, models etc at the end of each                        period/epoch number

fromdeepimpriorimportDeepImPriorManagerfromPILimportImageverbose=Trueimage_path="./data/my_noisy_image.png"period=200# interval of epochs, used for scheduling callbackslearning_rate=0.01# the value is taken from the paperepochs=2400# the value is taken from the paperout_folder="./data/outimages"out_prefix="my_denoised_"save_model_path="./data/outmodels/model"# save model here at the end of a period# if you want to resum trainingload_model_path:Optional[str]="./data/outmodels/model_1000"# in verbose output you can save the model plot and its summary to a fileplot_path="plot_model.png",summary_path="model_summary.txt",deep_prior=DeepImPriorManager(noisy_image=Image.open(image_path),verbose=verbose,period=period,learning_rate=learning_rate,epochs=epochs,out_folder=out_folder,out_prefix=out_prefix,save_model_path=save_model_path,load_model_path=load_model_path,plot_path=plot_path,summary_path=summary_path)# fits the model and saves it to "out_folder/outp_prefix-result.png"deep_prior.run_save()# or you can just fit the model and do something else with it like# evaluation etc.deep_prior.run()model=deep_prior.model# do other stuff with the model

defapply_layers(inlayer,lst:List[tf.keras.layers.Layer]):"""!\brief apply layers consecutively    \param inlayer input either a result of a previous evaluation or    tf.keras.layers.Input. Notice that it is not tf.keras.layers.InputLayer    """x=inlayerforlayerinlst:x=layer(x)returnx