mridc.collections.common.parts package

Submodules

mridc.collections.common.parts.fft module

mridc.collections.common.parts.fft.fft2(data: Tensor, centered: bool = True, normalization: str = 'ortho', spatial_dims: Optional[Sequence[int]] = None) → Tensor[source]

Apply 2 dimensional Fast Fourier Transform.

Parameters

  • data (Complex valued input data containing at least 3 dimensions: dimensions -2 & -1 are spatial dimensions. All) –

  • dimensions. (other dimensions are assumed to be batch) –

  • centered (Whether to center the fft.) –

  • normalization ("ortho" is the default normalization used by PyTorch. Can be changed to "ortho" or None.) –

  • spatial_dims (dimensions to apply the FFT) –

Return type

The FFT of the input.

mridc.collections.common.parts.fft.ifft2(data: Tensor, centered: bool = True, normalization: str = 'ortho', spatial_dims: Optional[Sequence[int]] = None) → Tensor[source]

Apply 2 dimensional Inverse Fast Fourier Transform.

Parameters

  • data (Complex valued input data containing at least 3 dimensions: dimensions -2 & -1 are spatial dimensions. All) –

  • dimensions. (other dimensions are assumed to be batch) –

  • centered (Whether to center the fft.) –

  • normalization ("ortho" is the default normalization used by PyTorch. Can be changed to "ortho" or None.) –

  • spatial_dims (dimensions to apply the FFT) –

Return type

The FFT of the input.

mridc.collections.common.parts.patch_utils module

mridc.collections.common.parts.ptl_overrides module

class mridc.collections.common.parts.ptl_overrides.MRIDCNativeMixedPrecisionPlugin(init_scale: float = 4294967296, growth_interval: int = 1000)[source]

Bases: NativeMixedPrecisionPlugin

Native Mixed Precision Plugin for MRIDC.

mridc.collections.common.parts.rnn_utils module

mridc.collections.common.parts.rnn_utils.rnn_weights_init(module, std_init_range=0.02, xavier=True)[source]

# TODO: check if this is the correct way to initialize RNN weights Initialize different weights in Transformer model.

Parameters

  • module (torch.nn.Module to be initialized) –

  • std_init_range (standard deviation of normal initializer) –

  • xavier (if True, xavier initializer will be used in Linear layers as was proposed in AIAYN paper, otherwise normal) –

  • paper) (initializer will be used (like in BERT) –

mridc.collections.common.parts.training_utils module

mridc.collections.common.parts.training_utils.avoid_bfloat16_autocast_context()[source]

If the current autocast context is bfloat16, cast it to float32.

mridc.collections.common.parts.training_utils.avoid_float16_autocast_context()[source]

If the current autocast context is float16, cast it to bfloat16 if available or float32.

mridc.collections.common.parts.utils module

mridc.collections.common.parts.utils.check_stacked_complex(data: Tensor) → Tensor[source]

Check if tensor is stacked complex (real & imag parts stacked along last dim) and convert it to a combined complex tensor.

Parameters

data (A complex valued tensor, where the size of the final dimension might be 2.) –

Return type

A complex valued tensor.

mridc.collections.common.parts.utils.coil_combination(data: Tensor, sensitivity_maps: Tensor, method: str = 'SENSE', dim: int = 0) → Tensor[source]

Coil combination.

Parameters

  • data (The input tensor.) –

  • sensitivity_maps (The sensitivity maps.) –

  • method (The coil combination method.) –

  • dim (The dimensions along which to apply the coil combination transform.) –

Return type

Coil combined data.

mridc.collections.common.parts.utils.complex_abs(data: Tensor) → Tensor[source]

Compute the absolute value of a complex valued input tensor.

Parameters

data (A complex valued tensor, where the size of the final dimension should be 2.) –

Return type

Absolute value of data.

mridc.collections.common.parts.utils.complex_abs_sq(data: Tensor) → Tensor[source]

Compute the squared absolute value of a complex tensor.

Parameters

data (A complex valued tensor, where the size of the final dimension should be 2.) –

Return type

Squared absolute value of data.

mridc.collections.common.parts.utils.complex_conj(x: Tensor) → Tensor[source]

Complex conjugate.

This applies the complex conjugate assuming that the input array has the last dimension as the complex dimension.

Parameters

x (A PyTorch tensor with the last dimension of size 2.) –

Return type

A PyTorch tensor with the last dimension of size 2.

mridc.collections.common.parts.utils.complex_mul(x: Tensor, y: Tensor) → Tensor[source]

Complex multiplication.

This multiplies two complex tensors assuming that they are both stored as real arrays with the last dimension being the complex dimension.

Parameters

  • x (A PyTorch tensor with the last dimension of size 2.) –

  • y (A PyTorch tensor with the last dimension of size 2.) –

Return type

A PyTorch tensor with the last dimension of size 2.

mridc.collections.common.parts.utils.is_none(x: Optional[Any]) → bool[source]

Check if a string is None.

Parameters

x (The string to check.) –

Return type

True if x is None, False otherwise.

mridc.collections.common.parts.utils.rss(data: Tensor, dim: int = 0) → Tensor[source]

Compute the Root Sum of Squares (RSS).

RSS is computed assuming that dim is the coil dimension.

Parameters

  • data (The input tensor) –

  • dim (The dimensions along which to apply the RSS transform) –

Return type

The RSS value.

mridc.collections.common.parts.utils.rss_complex(data: Tensor, dim: int = 0) → Tensor[source]

Compute the Root Sum of Squares (RSS) for complex inputs.

RSS is computed assuming that dim is the coil dimension.

Parameters

  • data (The input tensor) –

  • dim (The dimensions along which to apply the RSS transform) –

Return type

The RSS value.

mridc.collections.common.parts.utils.save_reconstructions(reconstructions: Dict[str, ndarray], out_dir: Path)[source]

Save reconstruction images.

This function writes to h5 files that are appropriate for submission to the leaderboard.

Parameters

  • reconstructions (A dictionary mapping input filenames to corresponding reconstructions.) –

  • out_dir (Path to the output directory where the reconstructions should be saved.) –

mridc.collections.common.parts.utils.sense(data: Tensor, sensitivity_maps: Tensor, dim: int = 0) → Tensor[source]

The SENSitivity Encoding (SENSE) transform 1.

References

1

Pruessmann KP, Weiger M, Scheidegger MB, Boesiger P. SENSE: Sensitivity encoding for fast MRI. Magn Reson Med 1999; 42:952-962.

Parameters

  • data (The input tensor) –

  • sensitivity_maps (The sensitivity maps) –

  • dim (The coil dimension) –

Return type

A coil-combined image.

mridc.collections.common.parts.utils.tensor_to_complex_np(data: Tensor) → ndarray[source]

Converts a torch tensor to a numpy array.

Parameters

data (Input torch tensor to be converted to numpy.) –

Return type

Complex Numpy array version of data.

mridc.collections.common.parts.utils.to_tensor(data: ndarray) → Tensor[source]

Converts a numpy array to a torch tensor.

For complex arrays, the real and imaginary parts are stacked along the last dimension.

Parameters

data (Input numpy array to be converted to torch.) –

Return type

Torch tensor version of data.

Module contents