torchsample.coord module#

torchsample.coord.feat_first(tensor)[source]#

Move the features (last dim) to dim1.

torchsample.coord.feat_last(tensor)[source]#

Move the features (dim1) to last dim.

torchsample.coord.full(batch, size, device=None, align_corners=False)[source]#

Generate 2D or 3D coordinates to fully n_samples an image.

Parameters

  • batch (int) – Batch size. If 0, don’t return a batch dimension.

  • size (tuple) – Size of field to generate pixel coordinates for. i.e. (x, y, ...).

  • device (torch.device) – The desired device of returned tensor

  • align_corners (bool) – if True, the corner pixels of the input and output tensors are aligned, and thus preserving the values at those pixels.

Returns

coords – (n, h, w, 2) Normalized coordinates for sampling functions.

Return type

torch.Tensor

torchsample.coord.full_like(tensor, *args, **kwargs)[source]#
torchsample.coord.normalize(coord, size, align_corners=False)[source]#

Normalize unnormalized coordinates.

unnormalize

Parameters

  • coord (torch.Tensor) – Values in range [0, size-1] to normalize.

  • size (int or tuple) – Unnormalized side length in pixels. If tuple, order is (x, y, ...) to match coord.

  • align_corners (bool) – if True, the corner pixels of the input and output tensors are aligned, and thus preserving the values at those pixels.

Returns

Normalized coordinates.

Return type

torch.Tensor

torchsample.coord.rand(batch, n_samples, dims=2, dtype=None, device=None)[source]#

Generate random coordinates in range [-1, 1].

Parameters

  • batch (int) – Batch size. If 0, don’t return a batch dimension.

  • n_samples (int) – Number of coordinates to generate per exemplar.

  • dims (int) – 2 for generating (x, y) coordinates. 3 for generating (x, y, z) coordinates. Defaults to 2.

  • dtype (torch.dtype) – The desired data type of returned tensor.

  • device (torch.device) – The desired device of returned tensor

Return type

(batch, n_samples, dims) random coordinates in range [-1, 1].

torchsample.coord.rand_biased(batch, n_samples, pred, k=3.0, beta=0.75, align_corners=False)[source]#

Uncertainty-based point sampling.

Described in training part of section 3.1 of:

PointRend: Image Segmentation as Rendering

Parameters

  • batch (int) – Batch size. If 0, don’t return a batch dimension.

  • n_samples (int) – Number of coordinates to generate per exemplar.

  • pred (torch.Tensor) – (b, c, h, w) or (b, c, d, h, w) prediction probabilities.

  • k (float) – Multiplier for oversampling for biased coordinates.

  • beta (float) – Portion of coordinates to be biased towards uncertain regions. Lower values of beta are more similar to a uniform random distribution, ignoring pred.

  • align_corners (bool) – if True, the corner pixels of the input and output tensors are aligned, and thus preserving the values at those pixels.

Returns

(b, n_samples, dim) Normalized coordinates.

Return type

torch.Tensor

torchsample.coord.randint(batch, n_samples, size, device=None, align_corners=False, replace=False)[source]#

Generate random pixel coordinates in range [-1, 1].

Unlike rand, the resulting coordinates will land exactly on pixels. Intended for sampling high resolution GT data during data loading.

Due to numerical precision, it is suggested to use mode="nearest" when sampling featuremaps of resolution size using these coordinates.

Parameters

  • batch (int) – Batch size. If 0, don’t return a batch dimension.

  • n_samples (int) –

  • size (tuple) – Size of field to generate pixel coordinates for. i.e. (x, y, ...).

  • device (torch.device) – The desired device of returned tensor

  • align_corners (bool) – if True, the corner pixels of the input and output tensors are aligned, and thus preserving the values at those pixels.

  • replace (bool) – Sample with or without replacement. Defaults to False.

Return type

(batch, n_samples, dims) random coordinates in range [-1, 1].

torchsample.coord.randint_like(n_samples, tensor, align_corners=False, replace=False)[source]#

Generate random pixel coordinates in range [-1, 1].

See randint.

Parameters

  • n_samples (int) –

  • tensor (tuple) – Tensor to generate random coordinates for. Coordinates will be placed on same device as tensor.

  • align_corners (bool) – if True, the corner pixels of the input and output tensors are aligned, and thus preserving the values at those pixels.

  • replace (bool) – Sample with or without replacement. Defaults to True.

Return type

(batch, n_samples, dims) random coordinates in range [-1, 1].

torchsample.coord.tensor_to_size(tensor)[source]#

Field size from tensor.

torchsample.coord.uncertain(pred, threshold=0.5, align_corners=False)[source]#

Report all coordinates with uncertain scores.

Used for inference with models trained with rand_biased.

features = feature_extractor(batch["image"])
ss_pred = decoder(batch["image"])
index_coords, norm_coords = ts.coord.uncertain(ss_pred)
features_sampled = ts.sample(norm_coords, features)
refined = mlp(features_sampled)
ss_pred_refined = ss_pred.clone()
ss_pred_refined[index_coords] = ts.feat_first(refined).flatten()
Parameters

  • pred (torch.Tensor) – (1, c, h, w) or (1, c, d, h, w) prediction probabilities. Can NOT handle >1 batch size; would result in ragged tensors.

  • threshold (float) – Coords with a max probability lower than this threshold will have coordinates generated.

  • align_corners (bool) – if True, the corner pixels of the input and output tensors are aligned, and thus preserving the values at those pixels.

Returns

  • index_coords (tuple) – Integer pixel coordinates that can be used to directly index into pred.

  • norm_coords – Normalized coordinates to be used with ts.sample. Does NOT have a batch dimension, as it would end up with a ragged tensor.

torchsample.coord.unnormalize(coord, size, align_corners=False, clip=False)[source]#

Unnormalize normalized coordinates.

Modified from PyTorch C source:

https://github.com/pytorch/pytorch/blob/c52290bad18d23d7decd37b581307f8241c0e8c5/aten/src/ATen/native/GridSampler.h#L26

Parameters

  • coord (torch.Tensor) – Values in range [-1, 1] to unnormalize.

  • size (int or tuple) – Unnormalized side length in pixels. If tuple, order is (x, y, ...) to match coord.

  • align_corners (bool) – if True, the corner pixels of the input and output tensors are aligned, and thus preserving the values at those pixels.

  • clip (bool) – If True, output will be clipped to range [0, size-1]

Returns

Unnormalized coordinates.

Return type

torch.Tensor