torch_enhance.models

class torch_enhance.models.BaseModel

Base Super-Resolution module

denormalize01(x: torch.Tensor) → torch.Tensor

Normalize from [0, 1] -> [0, 255]

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Normalized image tensor Return type:torch.Tensor

denormalize11(x: torch.Tensor) → torch.Tensor

Normalize from [-1, 1] -> [0, 255]

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Normalized image tensor Return type:torch.Tensor

static download(url: str, weights_path: str) → None

Download pretrained weights

Parameters:weights_path (str) – Path to save pretrained weights. Returns: Return type:None

enhance(x: torch.Tensor) → torch.Tensor

Super-resolve x and cast as image

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Super-Resolved image as tensor Return type:torch.Tensor

load_pretrained(weights_url: str, weights_path: str) → None

Download pretrained weights and load as state dict

Parameters:

• weights_url (str) – Base URL to pretrained weights.
• weights_path (str) – Path to save pretrained weights.

Returns:

Return type:

None

loss = MSELoss()

normalize01(x: torch.Tensor) → torch.Tensor

Normalize from [0, 255] -> [0, 1]

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Normalized image tensor Return type:torch.Tensor

normalize11(x: torch.Tensor) → torch.Tensor

Normalize from [0, 255] -> [-1, 1]

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Normalized image tensor Return type:torch.Tensor

class torch_enhance.models.Bicubic(scale_factor: int)

Bicubic Interpolation Upsampling module

Parameters:scale_factor (int) – Super-Resolution scale factor. Determines Low-Resolution downsampling.

forward(x: torch.Tensor) → torch.Tensor

Super-resolve Low-Resolution input tensor

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Super-Resolved image as tensor Return type:torch.Tensor

class torch_enhance.models.SRCNN(scale_factor: int)

Super-Resolution Convolutional Neural Network https://arxiv.org/pdf/1501.00092v3.pdf

Parameters:scale_factor (int) – Super-Resolution scale factor. Determines Low-Resolution downsampling.

forward(x: torch.Tensor) → torch.Tensor

Super-resolve Low-Resolution input tensor

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Super-Resolved image as tensor Return type:torch.Tensor

class torch_enhance.models.VDSR(scale_factor)

Very Deep Super Resolution https://arxiv.org/pdf/1511.04587.pdf

Parameters:scale_factor (int) – Super-Resolution scale factor. Determines Low-Resolution downsampling.

forward(x)

Super-resolve Low-Resolution input tensor

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Super-Resolved image as tensor Return type:torch.Tensor

class torch_enhance.models.EDSR(scale_factor: int)

Enhanced Deep Residual Networks for Single Image Super-Resolution https://arxiv.org/pdf/1707.02921v1.pdf

Parameters:scale_factor (int) – Super-Resolution scale factor. Determines Low-Resolution downsampling.

forward(x: torch.Tensor) → torch.Tensor

Super-resolve Low-Resolution input tensor

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Super-Resolved image as tensor Return type:torch.Tensor

class torch_enhance.models.ESPCN(scale_factor: int)

Efficient Sub-Pixel Convolutional Neural Network https://arxiv.org/pdf/1609.05158v2.pdf

Parameters:

• scale_factor (int) – Super-Resolution scale factor. Determines Low-Resolution downsampling.
• pretrained (bool) – If True download and load pretrained weights

forward(x: torch.Tensor) → torch.Tensor

Super-resolve Low-Resolution input tensor

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Super-Resolved image as tensor Return type:torch.Tensor

class torch_enhance.models.SRResNet(scale_factor: int)

Super-Resolution Residual Neural Network https://arxiv.org/pdf/1609.04802v5.pdf

Parameters:scale_factor (int) – Super-Resolution scale factor. Determines Low-Resolution downsampling.

forward(x: torch.Tensor) → torch.Tensor

Super-resolve Low-Resolution input tensor

Parameters:x (torch.Tensor) – Input Low-Resolution image as tensor Returns:Super-Resolved image as tensor Return type:torch.Tensor