Tips.txt

To Implement PyNET and extract each layer's parameters, do the followings:
	Go to Home/bashrc and add export PYTHONPATH=/localhome/mta188/local/lib/python3.10/dist-packages:$PYTHONPATH to the end of the file
	Open terminal
	git clone https://github.com/aiff22/PyNET-PyTorch.git
	Rename PyNET-PyTorch to PyNET
	cd PyNET
	pip install torch torchvision --target=/localhome/mta188/PyNET
	pip install scipy numpy imageio pillow tqdm --target=/localhome/mta188/PyNET
	Copy all the items under PyNET/PyNET into the main folder PyNET
	Copy a .py file and rename it to PyNET.py
	In PyNET.py write:
		import torch
		from model import PyNET  # model.py defines the architecture

		# Load model
		pynet = PyNET(level=5)  # You might need to specify the model level, e.g., PyNET(level=5)
		pynet.eval()

		# Load pre-trained weights
		checkpoint = torch.load('pynet_level5.pth', map_location='cpu')  # Replace with actual .pth file path
		pynet.load_state_dict(checkpoint)

		from PIL import Image
		from torchvision import transforms

		# Preprocess input
		transform = transforms.Compose([
		    transforms.Resize((1024, 1024)),  # Match training input size
		    transforms.ToTensor()
		])

		img = Image.open('sample.jpg').convert('RGB')
		input_tensor = transform(img).unsqueeze(0)  # Add batch dimension

		with torch.no_grad():
		    output = pynet(input_tensor)

		# Save or view output
		from torchvision.utils import save_image
		save_image(output, 'enhanced.jpg')

		# extract parameters (weights and biases) from each layer
		for name, module in pynet.named_modules():
		    if hasattr(module, 'weight'):
			print(f"\nLayer: {name}")
			print(f"  Weight shape: {module.weight.shape}")
			if hasattr(module, 'bias') and module.bias is not None:
			    print(f"  Bias shape: {module.bias.shape}")
        Then run "python3 PyNET.py" in terminal
	To train PyNET yourself, first download Zurich RAW to RGB dataset and put it in PyNET/raw_images, then run this commands in terminal:
		cd PyNET
		python3 train_model.py level=5 batch_size=50 num_train_epochs=8
	Training process is done by stochastic gradient decent algorithm (see "Art of Reinforcement Learning -> chapter 6 -> linear value approximation" to remind of the algorithm)
	To push to github, do the followings:
		First create a public repo under sfu-arch and name it PyNET
		Go to: https://github.com/settings/keys
		Click New SSH key
		Paste your public key (~/.ssh/id_ed25519.pub or ~/.ssh/id_rsa.pub)
		Click Add SSH key
		Open terminal
		cd PyNET
		git init
		git add .
		git commit -m "Adding PyNET Files"
		git remote set-url origin git@github.com:sfu-arch/PyNET.git
		git push origin main