torch-dreams

Updates:

• Added support for batch size>1 on a single image parameter + objective (AutoImageParam now has a new arg batch_size which defaults to 1). This would lead to higher quality visualizations with a smaller number of iterations.
• Fix unintended PytorchVersionError on torch 2.x. Thanks @matthiasgeihs 🙌

Interface change:
The input (layer_outputs) in custom objective functions is not anymore a list of tensors of shape (c, h, w). It's now a list of tensors of shape (n, c, h, w) where n = batch size. (The same applies for other shapes of intermediate layer outputs (...) -> (n, ...))

This is an example of an old v/s new objective function:

• Add support for batched image parameters + objective functions (speed++ 🚀)
• Adapt to latest torch version
• Fix casing + improved interface (breaking change)

• Add default support for any version of torch >= v1.8

Changes:

• (#48) Add support for torch 1.9.x. (Thanks to @seba-eng)

Updates:

• Added support for models with custom image normalizations with torch_dreams.dreamer.set_custom_normalization()

Bug fixes:

• Fixed corrupted image outputs on self.__array__() for torch_dreams.masked_image_param

changes:

• added support for caricatures, which was inspired from this issue

Changes:

• Added support for custom images with custom_image_param

Changes:

• Moved completely to torch v1.8.x
• FFT parameterization with shape [1, 3, height, width] from [1, 3, height, width//2, 2]
• Fixed failing tests due to incorrect device allocations

Changes:

• Fix random crashes on google colab caused due to torch.fft.irfftn on torch v1.8.x

Updates:

• auto_image_param instances can be saved as images with image_param.save('image.jpg')
• Fixed random crashes on google colab
• dreamer.get_snapshot() now in testing
• feedback loops now don't modify the original image parameter. Uses a deepcopy instead

• Optimization on multiple models simultaneously is now possible
• Ability to use custom torchvision.transforms
• Much better visualization quality thanks to torch.fft
• Supports both torch v1.7 and v1,8
• Faster optimizations