Skip to content

Fix #8462: embed patch sizes in einops pattern for einops >= 0.8 compatibility #8834

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
williams145 wants to merge 3 commits into
base: dev
Choose a base branch
from
Open

Fix #8462: embed patch sizes in einops pattern for einops >= 0.8 compatibility #8834

Changes from 2 commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
78756f8
Fix #8462: embed patch sizes in einops pattern for einops >= 0.8 comp…
williams145 Apr 18, 2026
093f0b0
Fix #8462: handle einops >= 0.8 Rearrange kwargs API change
williams145 Apr 18, 2026
f1cffe8
Fix ruff N806: rename uppercase vars B, C to batch, channels
williams145 Apr 18, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
36 changes: 32 additions & 4 deletions monai/networks/blocks/patchembedding.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,32 @@
SUPPORTED_POS_EMBEDDING_TYPES = {"none", "learnable", "sincos", "fourier"}


class _PatchRearrange(nn.Module):
"""Fallback patch rearrangement using pure PyTorch, for einops compatibility."""

def __init__(self, spatial_dims: int, patch_size: tuple) -> None:
super().__init__()
self.spatial_dims = spatial_dims
self.patch_size = patch_size

def forward(self, x: torch.Tensor) -> torch.Tensor:
B, C = x.shape[0], x.shape[1]
sp = x.shape[2:]
g = tuple(s // p for s, p in zip(sp, self.patch_size))
v: list[int] = [B, C]
for gi, pi in zip(g, self.patch_size):
v += [gi, pi]
x = x.view(*v)
n = self.spatial_dims
gdims = list(range(2, 2 + 2 * n, 2))
pdims = list(range(3, 3 + 2 * n, 2))
x = x.permute(0, *gdims, *pdims, 1).contiguous()
n_patches = 1
for gi in g:
n_patches *= gi
return x.reshape(B, n_patches, -1)
Copy link
Copy Markdown
Contributor

@coderabbitai coderabbitai Bot Apr 18, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

🛠️ Refactor suggestion | 🟠 Major

If the fallback is kept: tighten type hint, prefer reshape, add Google-style docstrings.

Three points on _PatchRearrange:

  1. x.view(*v) at line 47 will raise on non-contiguous inputs. reshape is safer and no slower here.
  2. patch_size: tuple is too loose — tuple[int, ...].
  3. Per coding guidelines, Google-style docstrings with Args:/Returns: are expected on __init__ and forward; the current one-line class docstring doesn't cover them.

As per coding guidelines: "Docstrings should be present for all definition which describe each variable, return value, and raised exception in the appropriate section of the Google-style of docstrings."

♻️ Proposed patch 
 class _PatchRearrange(nn.Module):
-    """Fallback patch rearrangement using pure PyTorch, for einops compatibility."""
-
-    def __init__(self, spatial_dims: int, patch_size: tuple) -> None:
+    """Fallback patch rearrangement in pure PyTorch, equivalent to the einops pattern
+    ``"b c (h p1) (w p2) (d p3) -> b (h w d) (p1 p2 p3 c)"`` (trimmed to ``spatial_dims``).
+    """
+
+    def __init__(self, spatial_dims: int, patch_size: tuple[int, ...]) -> None:
+        """
+        Args:
+            spatial_dims: number of spatial dimensions (2 or 3).
+            patch_size: per-axis patch size; length must equal ``spatial_dims``.
+        """
         super().__init__()
         self.spatial_dims = spatial_dims
         self.patch_size = patch_size
 
     def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x: input tensor of shape ``(B, C, *spatial)`` where each spatial dim is
+                divisible by the corresponding ``patch_size`` entry.
+
+        Returns:
+            Tensor of shape ``(B, n_patches, prod(patch_size) * C)``.
+        """
         B, C = x.shape[0], x.shape[1]
         sp = x.shape[2:]
         g = tuple(s // p for s, p in zip(sp, self.patch_size))
         v: list[int] = [B, C]
         for gi, pi in zip(g, self.patch_size):
             v += [gi, pi]
-        x = x.view(*v)
+        x = x.reshape(*v)
         n = self.spatial_dims
         gdims = list(range(2, 2 + 2 * n, 2))
         pdims = list(range(3, 3 + 2 * n, 2))
         x = x.permute(0, *gdims, *pdims, 1).contiguous()
         n_patches = 1
         for gi in g:
             n_patches *= gi
         return x.reshape(B, n_patches, -1)
🧰 Tools
🪛 Ruff (0.15.10)

[warning] 43-43: zip() without an explicit strict= parameter

Add explicit value for parameter strict=

(B905)

[warning] 45-45: zip() without an explicit strict= parameter

Add explicit value for parameter strict=

(B905)

🤖 Prompt for AI Agents 
Verify each finding against the current code and only fix it if needed.

In `@monai/networks/blocks/patchembedding.py` around lines 32 - 55, Update the
_PatchRearrange class: narrow the type of patch_size in __init__ to tuple[int,
...], replace the view call in forward (x.view(*v)) with reshape to handle
non-contiguous tensors safely, and add Google-style docstrings for the class
__init__ and forward methods describing Args and Returns (and any raised
exceptions) per project guidelines; reference the _PatchRearrange class, its
__init__ signature, and the forward method when making these changes.



class PatchEmbeddingBlock(nn.Module):
"""
A patch embedding block, based on: "Dosovitskiy et al.,
Expand Down Expand Up @@ -97,14 +123,16 @@ def __init__(
in_channels=in_channels, out_channels=hidden_size, kernel_size=patch_size, stride=patch_size
)
elif self.proj_type == "perceptron":
# for 3d: "b c (h p1) (w p2) (d p3)-> b (h w d) (p1 p2 p3 c)"
# for 3d: "b c (h p1) (w p2) (d p3) -> b (h w d) (p1 p2 p3 c)"
chars = (("h", "p1"), ("w", "p2"), ("d", "p3"))[:spatial_dims]
from_chars = "b c " + " ".join(f"({k} {v})" for k, v in chars)
to_chars = f"b ({' '.join([c[0] for c in chars])}) ({' '.join([c[1] for c in chars])} c)"
axes_len = {f"p{i + 1}": p for i, p in enumerate(patch_size)}
self.patch_embeddings = nn.Sequential(
Rearrange(f"{from_chars} -> {to_chars}", **axes_len), nn.Linear(self.patch_dim, hidden_size)
)
try:
rearrange_layer: nn.Module = Rearrange(f"{from_chars} -> {to_chars}", **axes_len)
except TypeError:
rearrange_layer = _PatchRearrange(spatial_dims, tuple(int(p) for p in patch_size))
self.patch_embeddings = nn.Sequential(rearrange_layer, nn.Linear(self.patch_dim, hidden_size))
Comment on lines +126 to +135
Copy link
Copy Markdown
Contributor

@coderabbitai coderabbitai Bot Apr 18, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

⚠️ Potential issue | 🟡 Minor

🧩 Analysis chain

🌐 Web query:

einops Rearrange pattern integer literals support 0.8

💡 Result:

No, einops Rearrange patterns do not support standalone integer literals (e.g., 'b 2 h w') as axis specifications. Patterns use symbolic axis names (letters like 'b', 'h'), anonymous axes (numbers in parentheses like '(b1 b2)'), ellipsis (...), or underscores (_ for skipping). Integer values are provided via the axes_lengths keyword argument (e.g., rearrange(x, 'b (h1 h) w c -> ...', h1=2)) or parsed from tensor shapes via parse_shape. This syntax has been consistent across versions, including 0.8.0+ (released 2024-04-28), with no changes introducing integer literal support in patterns per release notes, changelogs, and docs. Examples in official documentation and tutorials confirm named axes only.

Citations:

🏁 Script executed:

# Get the file and examine _PatchRearrange class and test coverage
head -150 monai/networks/blocks/patchembedding.py | tail -100

Repository: Project-MONAI/MONAI

Length of output: 4640

🏁 Script executed:

# Check for existing tests covering the fallback path
find . -name "*.py" -path "*/test*" -type f | xargs grep -l "patchembedding" | head -5

Repository: Project-MONAI/MONAI

Length of output: 130

🏁 Script executed:

# Get the _PatchRearrange class definition (around lines 32-55)
sed -n '32,56p' monai/networks/blocks/patchembedding.py

Repository: Project-MONAI/MONAI

Length of output: 958

🏁 Script executed:

# Check the test file for coverage of _PatchRearrange and the fallback path
cat monai/tests/networks/blocks/test_patchembedding.py

Repository: Project-MONAI/MONAI

Length of output: 145

🏁 Script executed:

# Find the test file with correct path
find . -name "test_patchembedding.py" -type f 2>/dev/null

Repository: Project-MONAI/MONAI

Length of output: 109

🏁 Script executed:

# Also check for any tests that import or test _PatchRearrange
rg "_PatchRearrange" --type py

Repository: Project-MONAI/MONAI

Length of output: 275

🏁 Script executed:

# Read the test file
cat ./tests/networks/blocks/test_patchembedding.py

Repository: Project-MONAI/MONAI

Length of output: 7510

🏁 Script executed:

# Check einops version constraints and any comments about the fallback logic
head -30 monai/networks/blocks/patchembedding.py

Repository: Project-MONAI/MONAI

Length of output: 1308

🏁 Script executed:

# Verify if x.view() in _PatchRearrange could fail on non-contiguous tensors
rg "x\.view\(" monai/networks/blocks/patchembedding.py -A 2 -B 2

Repository: Project-MONAI/MONAI

Length of output: 233

Retract the simplification suggestion; the current try/except approach is necessary.

einops Rearrange does not support integer literals in patterns—numeric axis values must be passed via axes_lengths (the current axes_len dict). The suggested approach of embedding integers as literals (e.g., "b c (h 16) (w 16) (d 16)") is not feasible. The fallback with _PatchRearrange is the correct strategy for version compatibility.

However, address these remaining issues in _PatchRearrange:

  1. Missing Google-style docstrings: Add docstrings to __init__ and forward methods describing arguments, return values, and behavior per coding guidelines.
  2. Type hint specificity: Change patch_size: tuple to patch_size: tuple[int, ...].
  3. Use reshape() instead of view(): Line 47 uses x.view(*v), which fails on non-contiguous tensors; reshape() is safer.
  4. Incomplete test coverage: The test suite only exercises the Rearrange path (since einops is installed). The fallback is never deterministically validated. Add a test that directly instantiates and tests _PatchRearrange independently.
🤖 Prompt for AI Agents 
Verify each finding against the current code and only fix it if needed.

In `@monai/networks/blocks/patchembedding.py` around lines 126 - 135, The
_PatchRearrange fallback needs fixes: add Google-style docstrings to the class
methods __init__ and forward describing arguments, return values, and behavior;
change the type hint patch_size: tuple to patch_size: tuple[int, ...]; replace
any use of x.view(*v) with x.reshape(*v) to avoid errors on non-contiguous
tensors; and add a deterministic unit test that directly instantiates and
exercises _PatchRearrange (independent of einops/Rearrange) to validate its
behavior for representative spatial_dims/patch_size combinations.

Comment on lines +131 to +135
Copy link
Copy Markdown
Contributor

@coderabbitai coderabbitai Bot Apr 18, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

⚠️ Potential issue | 🟡 Minor

Fallback path isn't deterministically covered by tests.

_PatchRearrange only executes when Rearrange(..., **axes_len) raises TypeError, i.e. only on einops ≥ 0.8. Existing test_shape cases (tests/networks/blocks/test_patchembedding.py) exercise whichever branch the installed einops version selects, so CI never compares both paths in the same run. Suggest a targeted test that forces the fallback — e.g. monkey-patch Rearrange to raise TypeError, or instantiate _PatchRearrange directly and compare its output against the Rearrange path for a known input.

As per coding guidelines: "Ensure new or modified definitions will be covered by existing or new unit tests."

🤖 Prompt for AI Agents 
Verify each finding against the current code and only fix it if needed.

In `@monai/networks/blocks/patchembedding.py` around lines 131 - 135, Tests don't
deterministically exercise the fallback _PatchRearrange because current tests
run whichever einops version is installed; add a unit test that forces the
fallback by monkey-patching the Rearrange symbol to raise TypeError (or by
directly instantiating _PatchRearrange) and assert that outputs of the fallback
match the normal Rearrange path for a representative input shape; target names
to modify/assert are _PatchRearrange, Rearrange and the patch embedding behavior
(e.g., the patch_embeddings sequence) in the existing
tests/networks/blocks/test_patchembedding.py so both branches are covered and
compared.

self.position_embeddings = nn.Parameter(torch.zeros(1, self.n_patches, hidden_size))
self.dropout = nn.Dropout(dropout_rate)

Expand Down
Loading