torchsr: Super Resolution Networks for PyTorch

Common errors

• RuntimeError: Expected all tensors to be on the same device, but found at least two devices, cpu and cuda:0!

  cause The input tensor you are passing to the `torchsr` model is on a different device (CPU or GPU) than the model itself.
  fix Move both the model and your input tensor to the same device using `.to(device)`. For example: `device = 'cuda' if torch.cuda.is_available() else 'cpu'; model = model.to(device); input_tensor = input_tensor.to(device)`.

• AttributeError: module 'torchsr.models' has no attribute 'my_non_existent_model'

  cause You are attempting to access a model that does not exist, or you have a typo in the model name within `torchsr.models`.
  fix Double-check the exact model name from the `torchsr` documentation or by inspecting the available models (e.g., `print(dir(torchsr.models))`). Ensure the name matches precisely.

• ValueError: Cannot find pretrained weights for model 'edsr_r16f64' at scale '4'

  cause The requested pretrained weights for the specified model and upscale factor (scale) are not available or could not be found/downloaded by the library.
  fix Verify that pretrained weights exist for your chosen model and scale combination by checking the `torchsr` documentation. If not available, you can instantiate the model without pretrained weights by setting `pretrained=False`, or choose a different model/scale combination.

Warnings

• gotcha PyTorch Version Compatibility: Pretrained models in `torchsr` were often trained with specific PyTorch versions. While the library generally aims for broad compatibility, significantly older or newer PyTorch versions might lead to issues when loading pretrained weights due to serialization changes. It's recommended to use a PyTorch version reasonably close to the latest stable release.
  Fix: If encountering errors with pretrained weights, try updating your PyTorch version or explicitly instantiating the model without `pretrained=True` to rule out compatibility issues with weights.
• gotcha Pretrained Model Availability: Not all models or all upscale factors (scales) within `torchsr` have pretrained weights available. Attempting to load a non-existent combination (e.g., `model(scale=5, pretrained=True)` where only `scale=2` is available) will raise a `ValueError`. Always verify the existence of pretrained weights for your desired configuration.
  Fix: Consult the `torchsr` documentation or repository for available models and their supported scales/pretrained weights. Alternatively, set `pretrained=False` to use an untrained model.
• gotcha Device Placement Mismatch: A common PyTorch issue where the input tensor and the model are on different devices (e.g., one on CPU, the other on GPU). This leads to `RuntimeError: Expected all tensors to be on the same device...`.
  Fix: Ensure both your model and your input tensors are explicitly moved to the same device using `.to(device)` (e.g., `model.to('cuda')`, `input_tensor.to('cuda')`).

Install

• pip install torchsr Install latest version

Imports

• models

  import torchsr.models

  Accesses a collection of Super Resolution models.
• datasets

  import torchsr.datasets

  Provides access to datasets relevant for Super Resolution tasks.

Quickstart

This quickstart demonstrates how to load a pretrained Super Resolution model (EDSR in this case) from `torchsr.models`, move it to the appropriate device, and perform inference on a dummy input tensor. The output shape confirms the upscale factor applied by the model.

import torch
import torchsr.models

# Instantiate an EDSR model with a 2x upscale factor, using pretrained weights
# Replace 'cpu' with 'cuda' if a GPU is available
device = 'cuda' if torch.cuda.is_available() else 'cpu'
model = torchsr.models.edsr_r16f64(scale=2, pretrained=True).to(device)
model.eval() # Set model to evaluation mode

# Create a dummy low-resolution input tensor (e.g., 1 channel, 64x64)
# Batch size is 1, so (1, C, H, W)
input_tensor = torch.randn(1, 3, 64, 64).to(device)

# Perform super-resolution inference
with torch.no_grad():
    output_tensor = model(input_tensor)

print(f"Input shape: {input_tensor.shape}")
print(f"Output shape: {output_tensor.shape}") # Should be 2x larger: (1, 3, 128, 128)