A SuperResolution training program for creating/training upscaling machine learning model, developed for educational purposes only. The result may vary between training data and hyperparameter, the example are from own trained model.
- EDSR - Enhanced Deep Residual Networks for Single Image Super-Resolution
- VDR - Very Deep Convolutional Network
- AE - AutoEncoder Super Resolution
- DCNN - Deep Convolutional Super Resolution Neural Network
- Resnet - Residual Network Deep Convolutional Super Resolution Neural Network
- SRGAN - GAN (Generative Adversarial Network) based Super Resolution Network
- SSIM - Structural similarity index measure
- MSA - Mean Square Absolute
- MSE - Mean Square Error
- VGG16 - Perceptible Loss Error Function
- VGG19 - Perceptible Loss Error Function
The most basic command line. Using the default option.
python superresolution/SuperResolution.py --data-set-directory /path_to_training_data/python superresolution/SuperResolution.py --batch-size 16 --epochs 10 --image-size 128 128 --model edsr --learning-rate 0.0003 --decay-rate 0.9 --decay-step 10000 --color-space rgb --loss-fn msa --shuffle-data-set-size 1024 --show-psnr --data-set-directory /path_to_training_data/ --output-dir image-super-resolution-result/
python superresolution/SuperResolution.py --batch-size 16 --epochs 10 --image-size 128 128 --model vdr --learning-rate 0.0003 --color-space rgb --loss-fn msa --shuffle-data-set-size 512 --show-psnr --data-set-directory /path_to_training_data/ --output-dir image-super-resolution-result/
python superresolution/SuperResolution.py --batch-size 16 --epochs 10 --image-size 128 128 --model dcsr-ae --learning-rate 0.0003 --color-space rgb --loss-fn msa --shuffle-data-set-size 512 --show-psnr --data-set-directory /path_to_training_data/ --output-dir image-super-resolution-result/
python superresolution/SuperResolution.py --batch-size 16 --epochs 10 --image-size 128 128 --model cnnsr --learning-rate 0.002 --color-space rgb --loss-fn msa --shuffle-data-set-size 512 --show-psnr --data-set-directory /path_to_training_data/ --output-dir image-super-resolution-result/
python superresolution/SuperResolution.py --batch-size 16 --epochs 10 --image-size 128 128 --model dcsr-resnet --learning-rate 0.0003 --color-space rgb --loss-fn msa --shuffle-data-set-size 512 --show-psnr --data-set-directory /path_to_training_data/ --output-dir image-super-resolution-result/ usage: SuperResolution [-h] [--epochs EPOCHS] [--batch-size BATCH_SIZE] [--use-checkpoint] [--checkpoint-filepath CHECKPOINT_DIR] [--checkpoint-every-epoch CHECKPOINT_EVERY_NTH_EPOCH] [--learning-rate LEARNING_RATE] [--device DEVICES] [--cpu] [--gpu]
[--distribute-strategy {mirror}] [--verbosity VERBOSITY] [--use-float16] [--cache-ram] [--cache-file CACHE_PATH] [--shuffle-data-set-size DATASET_SHUFFLE_SIZE] [--data-set-directory TRAIN_DIRECTORY_PATHS]
[--validation-data-directory VALIDATION_DIRECTORY_PATHS] [--test-data-directory TEST_DIRECTORY_PATHS] [--image-size INPUT_IMAGE_SIZE INPUT_IMAGE_SIZE] [--output-image-size OUTPUT_IMAGE_SIZE OUTPUT_IMAGE_SIZE] [--seed SEED]
[--color-space {rgb,lab}] [--color-channels {1,3,4}] [--optimizer {adam,rmsprop,sgd,adadelta}] [--disable-validation] [--config CONFIG] [--model-filename MODEL_FILEPATH] [--output-dir OUTPUT_DIR] [--example-batch EXAMPLE_NTH_BATCH]
[--example-batch-grid-size EXAMPLE_NTH_BATCH_GRID_SIZE] [--show-psnr] [--metrics {psnr,ssim}] [--decay-rate LEARNING_RATE_DECAY] [--decay-step LEARNING_RATE_DECAY_STEP] [--model {dcsr,dscr-post,dscr-pre,edsr,dcsr-ae,dcsr-resnet,vdsr,srgan}]
[--loss-fn {mse,ssim,msa,psnr,vgg16,vgg19,none}]
Super Resolution Training Model Program
optional arguments:
-h, --help show this help message and exit
--epochs EPOCHS Set the number of passes that the training set will be trained against.
--batch-size BATCH_SIZE
number of training element per each batch, during training.
--use-checkpoint Set the path the checkpoint will be saved/loaded.
--checkpoint-filepath CHECKPOINT_DIR
Set the path the checkpoint will be saved/loaded.
--checkpoint-every-epoch CHECKPOINT_EVERY_NTH_EPOCH
Set how often the checkpoint will be update, per epoch.
--learning-rate LEARNING_RATE
Set the initial Learning Rate
--device DEVICES Select the device explicitly that will be used.
--cpu Explicit use the CPU as the compute device.
--gpu Explicit use of GPU
--distribute-strategy {mirror}
Select Distribute Strategy.
--verbosity VERBOSITY
Set the verbosity level of the program
--use-float16 Hint the usage of Float 16 (FP16) in the model.
--cache-ram Use System Memory (RAM) as Cache storage.
--cache-file CACHE_PATH
Set the cache file path that will be used to store dataset cached data.
--shuffle-data-set-size DATASET_SHUFFLE_SIZE
Set the size of the shuffle buffer size, zero disables shuffling.
--data-set-directory TRAIN_DIRECTORY_PATHS
Directory path where the images are located dataset images
--validation-data-directory VALIDATION_DIRECTORY_PATHS
Directory path where the images are located dataset images
--test-data-directory TEST_DIRECTORY_PATHS
Directory path where the images are located dataset images
--image-size INPUT_IMAGE_SIZE INPUT_IMAGE_SIZE
Set the size of the images in width and height for the model.
--output-image-size OUTPUT_IMAGE_SIZE OUTPUT_IMAGE_SIZE
Set the size of the images in width and height for the model.
--seed SEED Set the random seed
--color-space {rgb,lab}
Select Color Space used in the model.
--color-channels {1,3,4}
Select Number of channels in the color space. GrayScale, RGB and RGBA.
--optimizer {adam,rmsprop,sgd,adadelta}
Select optimizer to be used
--disable-validation Select if use data validation step.
--config CONFIG Config File - Json.
--model-filename MODEL_FILEPATH
Define file path that the generator model will be saved at.
--output-dir OUTPUT_DIR
Set the output directory that all the models and results will be stored at
--example-batch EXAMPLE_NTH_BATCH
Set the number of train batches between saving work in progress result.
--example-batch-grid-size EXAMPLE_NTH_BATCH_GRID_SIZE
Set the grid size of number of example images.
--show-psnr Set the grid size of number of example images.
--metrics {psnr,ssim}
Set what metric to capture.
--decay-rate LEARNING_RATE_DECAY
Set Learning rate Decay.
--decay-step LEARNING_RATE_DECAY_STEP
Set Learning rate Decay Step.
--model {dcsr,dscr-post,dscr-pre,edsr,dcsr-ae,dcsr-resnet,vdsr,srgan}
Set which model type to use.
--loss-fn {mse,ssim,msa,psnr,vgg16,vgg19,none}
Set Loss FunctionUpscaling images using pre-trained upscale model.
The following allows to upscale a single image.
python3 superresolution/UpScaleUtil.py --model super-resolution-model-2113109.h5 --input-file low_res.png --save-output high_res.png --batch 32 --color-space rgbThe following allows to upscale a whole directory.
python3 superresolution/UpScaleUtil.py --model super-resolution-model-2113109.h5 --save-output high_output_dir/ --input-file low_input_dir/ --batch 32 --color-space rgbThe Evolution Program allow sto try multiple variable combination in order to find a good set of variable configuration. Simliar to hyperparameter testing.
python3 superresolution/super-resolution-evolution-test.py --epochs 8 --batch 32 rgb --image-size 128 128 --data-set-directory /path_to_training_data/ --validation-data-directory /path_to_validation_data/ --output-dir evolution_test/Argument options
usage: SuperResolution Model Evolution [-h] [--epochs EPOCHS] [--batch-size BATCH_SIZE] [--use-checkpoint] [--checkpoint-filepath CHECKPOINT_DIR] [--checkpoint-every-epoch CHECKPOINT_EVERY_NTH_EPOCH] [--learning-rate LEARNING_RATE]
[--device DEVICES] [--cpu] [--gpu] [--distribute-strategy {mirror}] [--verbosity VERBOSITY] [--use-float16] [--cache-ram] [--cache-file CACHE_PATH] [--shuffle-data-set-size DATASET_SHUFFLE_SIZE]
[--data-set-directory TRAIN_DIRECTORY_PATHS] [--validation-data-directory VALIDATION_DIRECTORY_PATHS] [--test-data-directory TEST_DIRECTORY_PATHS] [--image-size INPUT_IMAGE_SIZE INPUT_IMAGE_SIZE]
[--output-image-size OUTPUT_IMAGE_SIZE OUTPUT_IMAGE_SIZE] [--seed SEED] [--color-space {rgb,lab}] [--color-channels {1,3,4}] [--optimizer {adam,rmsprop,sgd,adadelta}] [--disable-validation] [--config CONFIG]
[--output-dir OUTPUT_DIR] [--models [{cnnsr,dcsr,edsr,dcsr-ae,dcsr-resnet,vdsr,srgan,esrgan} ...]] [--loss-functions [{mse,ssim,msa,vgg16,vgg19} ...]] [--optimizer-evolution [{adam,rmsprop,sgd,adadelta} ...]]
Super Resolution Training Model Evolution Program
optional arguments:
-h, --help show this help message and exit
--epochs EPOCHS Set the number of passes that the training set will be trained against.
--batch-size BATCH_SIZE
number of training element per each batch, during training.
--use-checkpoint Set the path the checkpoint will be saved/loaded.
--checkpoint-filepath CHECKPOINT_DIR
Set the path the checkpoint will be saved/loaded.
--checkpoint-every-epoch CHECKPOINT_EVERY_NTH_EPOCH
Set how often the checkpoint will be update, per epoch.
--learning-rate LEARNING_RATE
Set the initial Learning Rate
--device DEVICES Select the device explicitly that will be used.
--cpu Explicit use the CPU as the compute device.
--gpu Explicit use of GPU
--distribute-strategy {mirror}
Select Distribute Strategy.
--verbosity VERBOSITY
Set the verbosity level of the program
--use-float16 Hint the usage of Float 16 (FP16) in the model.
--cache-ram Use System Memory (RAM) as Cache storage.
--cache-file CACHE_PATH
Set the cache file path that will be used to store dataset cached data.
--shuffle-data-set-size DATASET_SHUFFLE_SIZE
Set the size of the shuffle buffer size, zero disables shuffling.
--data-set-directory TRAIN_DIRECTORY_PATHS
Directory path where the images are located dataset images
--validation-data-directory VALIDATION_DIRECTORY_PATHS
Directory path where the images are located dataset images
--test-data-directory TEST_DIRECTORY_PATHS
Directory path where the images are located dataset images
--image-size INPUT_IMAGE_SIZE INPUT_IMAGE_SIZE
Set the input training images size. Low Resolution (LR).
--output-image-size OUTPUT_IMAGE_SIZE OUTPUT_IMAGE_SIZE
Set the size of the images in width and height for the model (HR).
--seed SEED Set the random seed
--color-space {rgb,lab}
Select Color Space used in the model.
--color-channels {1,3,4}
Select Number of channels in the color space. GrayScale, RGB and RGBA.
--optimizer {adam,rmsprop,sgd,adadelta}
Select optimizer to be used
--disable-validation Disable validation if validation data is present.
--config CONFIG Config File - Json.
--output-dir OUTPUT_DIR
Set the output directory that all the models and results will be stored at
--models [{cnnsr,dcsr,edsr,dcsr-ae,dcsr-resnet,vdsr,srgan,esrgan} ...]
Overide what Model to include in training evolution.
--loss-functions [{mse,ssim,msa,vgg16,vgg19} ...]
Overide what Loss functions to include in training evolution.
--optimizer-evolution [{adam,rmsprop,sgd,adadelta} ...]
Select optimizer to be usedpython3.9 or higher
python3 -m venv venv
source venv/bin/activatepip install -r requirements.txt requirements_cuda.txt
pip install tensorflow[and-cuda]==2.14.1pip install -r requirements.txt requirements_rocm.txtdocker build -t super-resolution-rocm -f Dockerfile.rocm .
docker run --network=host --device=/dev/kfd --device=/dev/dri --ipc=host --shm-size 16G --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --name sr-rocm super-resolution-rocm sudo apt-get install -y nvidia-container-toolkitdocker build -t super-resolution-cuda -f Dockerfile.cuda .
docker run --network=host --gpus all --name sr-cuda super-resolution-cuda python3 superresolution/generate_tflite.py --model super-resolution-model.keras --output model-lite.tflitepython3 -m tf2onnx.convert --saved-model tensorflow-model-path --output model.onnxWhen running python script as background process, it will still be terminated if closing the terminal window. However, with the nohup it can be run in the background as well close the terminal window.
nohup python3 superresolution/SuperResolution.py ...your arguments... &This project is licensed under the GPL+3 License - see the LICENSE file for details.