Merge pull request #4 from TerminalVelocityDPro/pad-fix-arjun

implement pad fix template

meddlr/data/transforms/transform.py

@@ -2,7 +2,7 @@
 """
 import math
 from functools import partial
-from typing import Optional, Tuple
+from typing import List, Optional, Sequence, Tuple
 
 import numpy as np
 import torch
@@ -11,7 +11,8 @@
 import meddlr.ops as F
 from meddlr.forward import SenseModel
 from meddlr.ops import complex as cplx
-from meddlr.transforms.gen.spatial import RandomAffine
+from meddlr.transforms.gen.spatial import RandomAffine, RandomTranslation
+from meddlr.transforms.transform_gen import TransformGen
 from meddlr.utils import transforms as T
 
 from .motion import MotionModel
@@ -26,7 +27,7 @@
 def affine_transform(
     image: torch.Tensor,
     nshots: int,
-    translation: Optional[RandomAffine] = None,
+    transforms: Sequence[TransformGen],
     trajectory: str = "blocked",
 ) -> torch.Tensor:
     """
@@ -46,9 +47,11 @@ def affine_transform(
         image: The complex-valued image. Shape [..., height, width].
         nshots: The number of shots in the image.
             This should be equivalent to ceil(phase_encode_dim / echo_train_length).
-        translation: This is the translation to augment images in the image
-            domain. This is either 'None' or 'RandomAffine' for now.
-        trajectory: One of 'interleaved' or 'consecutive'.
+        transforms: A sequence of random transform generators. These transforms
+            will be used to augment images in the image domain. We recommend using
+            [RandomTranslation, RandomAffine] in that order. This matches the MRAugment
+            augmentation strategy.
+        trajectory: One of 'interleaved' or 'blocked'.
 
     Returns:
         A motion corrupted image.
@@ -57,7 +60,11 @@ def affine_transform(
     offset = int(math.ceil(kspace.shape[-1] / nshots))
 
     for shot in range(nshots):
-        motion_image = translation.get_transform(image).apply_image(image)
+        # Apply sequence of random transforms to the image.
+        motion_image = image
+        for tfm in transforms:
+            motion_image = tfm.get_transform(motion_image).apply_image(motion_image)
+
         motion_kspace = F.fft2c(motion_image)
         if trajectory == "blocked":
             kspace[..., shot * offset : (shot + 1) * offset] = motion_kspace[
@@ -558,24 +565,33 @@ def __call__(self, kspace, maps, target, fname, slice_id, is_fixed, acceleration
 
         add_motion = self.add_motion and self._is_test
         if add_motion:
+            # Separate translation and affine transformations for proper padding.
+            translation = RandomTranslation(
+                p=1.0,
+                translate=self.translation,
+                pad_mode="reflect" if self.pad_like == "mraugment" else "constant",
+                pad_value=0.0,
+                ndim=2
+            )
+            affine = RandomAffine(
+                p=1.0, translate=None, angle=self.angle, pad_like=pad
+            )
+            transforms: List[TransformGen] = [translation, affine]
             # Motion seed should not be different for each slice for now.
             # TODO: Change this for 2D acquisitions.
+            for tfm_gen in transforms:
+                tfm_gen.seed(seed)
 
-            tfm_gen = RandomAffine(
-                p=1.0, translate=self.translation, angle=self.angle, pad_like=pad
-            )
-            tfm_gen.seed(seed)
-
-            image = image.permute(0, 3, 1, 2) # Shape: (B, 1, H, W)
+            image = image.permute(0, 3, 1, 2)  # Shape: (B, 1, H, W)
 
             motion_img = affine_transform(
                 image=image,
                 nshots=self.nshots,
-                translation=tfm_gen,
+                transforms=transforms,
                 trajectory=self.trajectory,
             )
 
-            motion_img = motion_img.permute(0, 2, 3, 1) # Shape: (B, H, W, 1)
+            motion_img = motion_img.permute(0, 2, 3, 1)  # Shape: (B, H, W, 1)
             sense = SenseModel(maps)
             kspace = sense(motion_img)