A pytorch implementation of a ACSC model based on Lerned Convolutional Sparse Coding model proposed here and or here.

For a linux with a cuda simply run

pip install pipenv && pipenv install --dev

This will install pipenv and pipenv will install the rest of the rest of dependencies specified in the Pipfile. If Pipfile does not work then use Python version >= 3.6 and install using pip all dependencies (mostly just pytorch)

Their are 3 main entry points:

1. train_denoise.py
2. test_denoise.py
3. analyze_model.py

all entry points are run in the same fashion: if PipFile is used

pipenv run python xx.py --args_file params.json

or if pip is used

python3 xx.py --args_file params.json

Example for how to set params.json file can be found in ./saved_models/acsc[1-4].

train will build model and run train session using parameters that are givin in params.json. params.json is updated and coped to log-path specified in params.json. train will automatically run test and analyze.

pipenv run python train_denoise.py --args_file params.json

Download npz PASCAL VOC images provided google drive make sure to specify dataset_path in params.json.

test will build the model specified in params.json. Two types of tests are run:

1. PASCAL VOC2012 testset (must be downloaded)
2. famous images

all result are saved in log_dir provided in params.json.

pipenv run python test_denoise.py --args_file params.json

Evaluate model beyond test accuracy. Sparsity factors etc.

pipenv run python analyze_denoise.py --args_file params.json

Param file is first set with model + train + test parameter. During train and test the Params file is constantly updated. When A model is done training it saved its weights and the final state of the params.json. Thus running train on the saved params.json will load the last state of the saved model proceed to continue training session. [Exmaple of initial params file]

{
    "train_args":
    {
        "noise": 25,
        "epoch": 30,
        "batch_size": 15,
        "learning_rate": 1e-4,
        "dataset_path": "./pascal_120.npz",
        "log_dir": "saved_models/",
        "name": "acsc"
    },
    "model_args": {
        "num_input_channels": 1,
        "num_output_channels": 1,
        "kc": 64,
        "ks": 7,
        "ista_iters": 20,
        "iter_weight_share": true,
        "share_decoder": false
    },
    "test_args": {
        "noise": 25,
        "load_path": "",
        "name": "acsc",
        "testset_famous_path": "test_images/",
        "testset_pascal_path": "VOCdevkit/VOC2010/JPEGImages/"
    }
}