visualtorch 1.0.0__tar.gz → 1.2.0__tar.gz

{visualtorch-1.0.0/visualtorch.egg-info → visualtorch-1.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: visualtorch
-Version: 1.0.0
+Version: 1.2.0
 Summary: Architecture visualization of Torch models
 Home-page: https://github.com/willyfh/visualtorch
 Author: Willy Fitra Hendria
@@ -30,6 +30,7 @@ Requires-Dist: sphinx-copybutton; extra == "dev"
 Requires-Dist: sphinx_design; extra == "dev"
 Requires-Dist: sphinx_gallery; extra == "dev"
 Requires-Dist: matplotlib; extra == "dev"
+Requires-Dist: torchvision; extra == "dev"
 Requires-Dist: pre-commit; extra == "dev"
 Requires-Dist: pytest; extra == "dev"
 Dynamic: author
@@ -53,19 +54,22 @@ Dynamic: summary
 
 </div>
 
-**VisualTorch** aims to help visualize Torch-based neural network architectures. It currently supports generating flow-style, graph-style, and LeNet-style architectures for PyTorch Sequential and Custom models. This tool is adapted from [visualkeras](https://github.com/paulgavrikov/visualkeras), [pytorchviz](https://github.com/szagoruyko/pytorchviz), and [pytorch-summary](https://github.com/sksq96/pytorch-summary).
+**VisualTorch** aims to help visualize Torch-based neural network architectures. It currently supports generating flow-style, graph-style, and LeNet-style architectures for PyTorch Sequential and Custom models. Its original visual styles were inspired by [visualkeras](https://github.com/paulgavrikov/visualkeras), [pytorchviz](https://github.com/szagoruyko/pytorchviz), [pytorch-summary](https://github.com/sksq96/pytorch-summary), and [torchview](https://github.com/mert-kurttutan/torchview); since then, it has grown its own unified tracing backend and architecture-handling logic well beyond its origins.
 
-**Note:** VisualTorch traces a real forward pass to build the diagram, which has two inherent
-limitations shared by any tracing-based approach (not bugs, and not fixable without full symbolic
-execution): (1) models with **data-dependent control flow** (e.g. a branch only taken if a tensor
-value crosses some threshold) only show whichever branch the traced dummy input happened to take;
-(2) a layer that returns **multiple meaningful output tensors** (e.g. a custom multi-task head)
-only has its first tensor's shape reflected in that node's size/label - its downstream connections
-are still correct either way. Contributions are welcome!
+**Note:** `1.0+` is a major release with breaking API changes, but with significantly better features and algorithms - upgrading is recommended. For the old API, use `0.2.5` or older.
+
+**Limitation:** VisualTorch traces a real forward pass to build the diagram, which has an inherent
+limitation shared by any tracing-based approach (not a bug, and not fixable without full symbolic
+execution): models with **data-dependent control flow** (e.g. a branch only taken if a tensor
+value crosses some threshold) only show whichever branch the traced dummy input happened to take.
+Separately, a layer that returns **multiple meaningful output tensors** (e.g. a custom multi-task
+head, or `nn.LSTM`'s `(output, (h_n, c_n))`) still has its node's size based on only its first
+tensor; with `show_dimension=True`, every output tensor's shape is shown in the label, not just
+the first. Downstream connections are correct either way. Contributions are welcome!
 
 <div align="center">
 
-![VisualTorch Examples](docs/source/_static/images/banners/readme-examples.png)
+![VisualTorch Examples](https://raw.githubusercontent.com/willyfh/visualtorch/e6ad79751e0f7412b1074beb45f9baeccd1419e4/docs/source/_static/images/banners/readme-examples.png)
 
 </div>
 
@@ -79,6 +83,12 @@ The docs include [usage examples](https://visualtorch.readthedocs.io/en/latest/u
 
 See the [Installation page](https://visualtorch.readthedocs.io/en/latest/markdown/get_started/installation.html).
 
+## Used in Research
+
+VisualTorch has been used in published research, including works published in Nature, IEEE, and MDPI.
+
+See the [Research Showcase page](https://visualtorch.readthedocs.io/en/latest/markdown/showcase/index.html) for the full list.
+
 ## Examples
 
 See the [Usage Examples page](https://visualtorch.readthedocs.io/en/latest/usage_examples/index.html).
@@ -91,16 +101,16 @@ Please feel free to send a pull request to contribute to this project by followi
 
 This poject is available as open source under the terms of the [MIT License](https://github.com/willyfh/visualtorch/blob/main/LICENSE).
 
-Originally, this project was based on the [visualkeras](https://github.com/paulgavrikov/visualkeras) (under the MIT license), with additional modifications inspired by [pytorchviz](https://github.com/szagoruyko/pytorchviz), and [pytorch-summary](https://github.com/sksq96/pytorch-summary), both of which are also licensed under the MIT license.
+Originally, this project was based on the [visualkeras](https://github.com/paulgavrikov/visualkeras) (under the MIT license), with additional modifications inspired by [pytorchviz](https://github.com/szagoruyko/pytorchviz), [pytorch-summary](https://github.com/sksq96/pytorch-summary), and [torchview](https://github.com/mert-kurttutan/torchview), all of which are also licensed under the MIT license.
 
 ## Citation
 
 Please cite this project in your publications if it helps your research.
 
-**Note:** the paper below describes the API as of its publication date (2024). VisualTorch has
-since had breaking API changes (see the [documentation](https://visualtorch.readthedocs.io/en/latest/)
-for the current API) - the DOI always resolves to what was actually reviewed and published, so
-it isn't updated to match.
+**Note:** the paper below describes VisualTorch as of its publication date (2024). The project has
+since been substantially refactored, including breaking API changes (see the
+[documentation](https://visualtorch.readthedocs.io/en/latest/) for the current API) - the DOI
+always resolves to what was actually reviewed and published.
 
 ```bibtex
 @article{Hendria2024,

{visualtorch-1.0.0 → visualtorch-1.2.0}/README.md

@@ -5,19 +5,22 @@
 
 </div>
 
-**VisualTorch** aims to help visualize Torch-based neural network architectures. It currently supports generating flow-style, graph-style, and LeNet-style architectures for PyTorch Sequential and Custom models. This tool is adapted from [visualkeras](https://github.com/paulgavrikov/visualkeras), [pytorchviz](https://github.com/szagoruyko/pytorchviz), and [pytorch-summary](https://github.com/sksq96/pytorch-summary).
+**VisualTorch** aims to help visualize Torch-based neural network architectures. It currently supports generating flow-style, graph-style, and LeNet-style architectures for PyTorch Sequential and Custom models. Its original visual styles were inspired by [visualkeras](https://github.com/paulgavrikov/visualkeras), [pytorchviz](https://github.com/szagoruyko/pytorchviz), [pytorch-summary](https://github.com/sksq96/pytorch-summary), and [torchview](https://github.com/mert-kurttutan/torchview); since then, it has grown its own unified tracing backend and architecture-handling logic well beyond its origins.
 
-**Note:** VisualTorch traces a real forward pass to build the diagram, which has two inherent
-limitations shared by any tracing-based approach (not bugs, and not fixable without full symbolic
-execution): (1) models with **data-dependent control flow** (e.g. a branch only taken if a tensor
-value crosses some threshold) only show whichever branch the traced dummy input happened to take;
-(2) a layer that returns **multiple meaningful output tensors** (e.g. a custom multi-task head)
-only has its first tensor's shape reflected in that node's size/label - its downstream connections
-are still correct either way. Contributions are welcome!
+**Note:** `1.0+` is a major release with breaking API changes, but with significantly better features and algorithms - upgrading is recommended. For the old API, use `0.2.5` or older.
+
+**Limitation:** VisualTorch traces a real forward pass to build the diagram, which has an inherent
+limitation shared by any tracing-based approach (not a bug, and not fixable without full symbolic
+execution): models with **data-dependent control flow** (e.g. a branch only taken if a tensor
+value crosses some threshold) only show whichever branch the traced dummy input happened to take.
+Separately, a layer that returns **multiple meaningful output tensors** (e.g. a custom multi-task
+head, or `nn.LSTM`'s `(output, (h_n, c_n))`) still has its node's size based on only its first
+tensor; with `show_dimension=True`, every output tensor's shape is shown in the label, not just
+the first. Downstream connections are correct either way. Contributions are welcome!
 
 <div align="center">
 
-![VisualTorch Examples](docs/source/_static/images/banners/readme-examples.png)
+![VisualTorch Examples](https://raw.githubusercontent.com/willyfh/visualtorch/e6ad79751e0f7412b1074beb45f9baeccd1419e4/docs/source/_static/images/banners/readme-examples.png)
 
 </div>
 
@@ -31,6 +34,12 @@ The docs include [usage examples](https://visualtorch.readthedocs.io/en/latest/u
 
 See the [Installation page](https://visualtorch.readthedocs.io/en/latest/markdown/get_started/installation.html).
 
+## Used in Research
+
+VisualTorch has been used in published research, including works published in Nature, IEEE, and MDPI.
+
+See the [Research Showcase page](https://visualtorch.readthedocs.io/en/latest/markdown/showcase/index.html) for the full list.
+
 ## Examples
 
 See the [Usage Examples page](https://visualtorch.readthedocs.io/en/latest/usage_examples/index.html).
@@ -43,16 +52,16 @@ Please feel free to send a pull request to contribute to this project by followi
 
 This poject is available as open source under the terms of the [MIT License](https://github.com/willyfh/visualtorch/blob/main/LICENSE).
 
-Originally, this project was based on the [visualkeras](https://github.com/paulgavrikov/visualkeras) (under the MIT license), with additional modifications inspired by [pytorchviz](https://github.com/szagoruyko/pytorchviz), and [pytorch-summary](https://github.com/sksq96/pytorch-summary), both of which are also licensed under the MIT license.
+Originally, this project was based on the [visualkeras](https://github.com/paulgavrikov/visualkeras) (under the MIT license), with additional modifications inspired by [pytorchviz](https://github.com/szagoruyko/pytorchviz), [pytorch-summary](https://github.com/sksq96/pytorch-summary), and [torchview](https://github.com/mert-kurttutan/torchview), all of which are also licensed under the MIT license.
 
 ## Citation
 
 Please cite this project in your publications if it helps your research.
 
-**Note:** the paper below describes the API as of its publication date (2024). VisualTorch has
-since had breaking API changes (see the [documentation](https://visualtorch.readthedocs.io/en/latest/)
-for the current API) - the DOI always resolves to what was actually reviewed and published, so
-it isn't updated to match.
+**Note:** the paper below describes VisualTorch as of its publication date (2024). The project has
+since been substantially refactored, including breaking API changes (see the
+[documentation](https://visualtorch.readthedocs.io/en/latest/) for the current API) - the DOI
+always resolves to what was actually reviewed and published.
 
 ```bibtex
 @article{Hendria2024,

{visualtorch-1.0.0 → visualtorch-1.2.0}/setup.py

@@ -21,7 +21,7 @@ def _read_requirements(file: str) -> list:
 
 setuptools.setup(
     name="visualtorch",
-    version="1.0.0",
+    version="1.2.0",
     author="Willy Fitra Hendria",
     author_email="willyfitrahendria@gmail.com",
     description="Architecture visualization of Torch models",

visualtorch-1.2.0/tests/test_connectors.py

@@ -0,0 +1,187 @@
+"""Tests for the frontend-agnostic connector-routing helpers."""
+
+# Copyright (C) 2024 Willy Fitra Hendria
+# SPDX-License-Identifier: MIT
+
+from visualtorch.connectors import _segment_intersects_rect, compute_skip_levels
+
+
+def _no_bbox(_node_id: str) -> tuple[float, float, float, float] | None:
+    return None
+
+
+# ---- compute_skip_levels: span/content gating ----
+
+
+def test_compute_skip_levels_ignores_span_le_1_edges() -> None:
+    """An edge with column span <= 1 should never be assigned a detour level."""
+    id_to_column = {"a": 0, "b": 1}
+
+    edge_to_level, num_levels = compute_skip_levels([("a", "b")], id_to_column, lambda *_: True, _no_bbox)
+
+    assert edge_to_level == {}
+    assert num_levels == 0
+
+
+def test_compute_skip_levels_assigns_distinct_levels_to_overlapping_skips() -> None:
+    """Two skip edges whose column spans genuinely overlap must get different levels."""
+    id_to_column = {"a": 0, "b": 3, "c": 2, "d": 5}
+    edges = [("a", "b"), ("c", "d")]
+
+    edge_to_level, num_levels = compute_skip_levels(edges, id_to_column, lambda *_: True, _no_bbox)
+
+    assert num_levels == 2
+    assert edge_to_level[("a", "b")] != edge_to_level[("c", "d")]
+
+
+def test_compute_skip_levels_allows_touching_intervals_to_share_a_level() -> None:
+    """Two skip edges that only touch at a shared column boundary can share a level."""
+    id_to_column = {"a": 0, "b": 3, "c": 3, "d": 5}
+    edges = [("a", "b"), ("c", "d")]
+
+    edge_to_level, num_levels = compute_skip_levels(edges, id_to_column, lambda *_: True, _no_bbox)
+
+    assert num_levels == 1
+    assert edge_to_level[("a", "b")] == edge_to_level[("c", "d")] == 0
+
+
+def test_compute_skip_levels_ignores_edges_with_no_content() -> None:
+    """A skip edge that `edge_has_content` reports as empty shouldn't consume a level."""
+    id_to_column = {"a": 0, "b": 3}
+
+    edge_to_level, num_levels = compute_skip_levels([("a", "b")], id_to_column, lambda *_: False, _no_bbox)
+
+    assert edge_to_level == {}
+    assert num_levels == 0
+
+
+# ---- compute_skip_levels: collision-awareness ----
+
+
+def test_compute_skip_levels_keeps_level_when_intervening_box_collides() -> None:
+    """A span>1 edge whose straight line genuinely crosses an intervening same-row box."""
+    id_to_column = {"a": 0, "mid": 1, "b": 2}
+    bboxes = {
+        "a": (0.0, 0.0, 10.0, 10.0),
+        "mid": (20.0, 0.0, 30.0, 10.0),  # same row (y=0..10) as a and b - directly in the path
+        "b": (40.0, 0.0, 50.0, 10.0),
+    }
+
+    edge_to_level, num_levels = compute_skip_levels(
+        [("a", "b")],
+        id_to_column,
+        lambda *_: True,
+        lambda node_id: bboxes[node_id],
+    )
+
+    assert num_levels == 1
+    assert edge_to_level[("a", "b")] == 0
+
+
+def test_compute_skip_levels_drops_level_when_intervening_box_is_a_different_row() -> None:
+    """A span>1 edge whose straight line passes clear of an intervening box in a different row."""
+    id_to_column = {"a": 0, "mid": 1, "b": 2}
+    bboxes = {
+        "a": (0.0, 0.0, 10.0, 10.0),
+        "mid": (20.0, 100.0, 30.0, 110.0),  # a different row entirely - well below the line
+        "b": (40.0, 0.0, 50.0, 10.0),
+    }
+
+    edge_to_level, num_levels = compute_skip_levels(
+        [("a", "b")],
+        id_to_column,
+        lambda *_: True,
+        lambda node_id: bboxes[node_id],
+    )
+
+    assert edge_to_level == {}
+    assert num_levels == 0
+
+
+def test_compute_skip_levels_missing_bbox_conservatively_assumes_collision() -> None:
+    """If an intervening id's bbox can't be resolved, assume a collision rather than guessing clear."""
+    id_to_column = {"a": 0, "mid": 1, "b": 2}
+    bboxes = {"a": (0.0, 0.0, 10.0, 10.0), "b": (40.0, 0.0, 50.0, 10.0)}
+
+    edge_to_level, num_levels = compute_skip_levels(
+        [("a", "b")],
+        id_to_column,
+        lambda *_: True,
+        lambda node_id: bboxes.get(node_id),
+    )
+
+    assert num_levels == 1
+    assert edge_to_level[("a", "b")] == 0
+
+
+def test_compute_skip_levels_missing_endpoint_bbox_conservatively_assumes_collision() -> None:
+    """If the edge's own start/end bbox can't be resolved, also assume a collision."""
+    id_to_column = {"a": 0, "b": 2}
+
+    edge_to_level, num_levels = compute_skip_levels([("a", "b")], id_to_column, lambda *_: True, _no_bbox)
+
+    assert num_levels == 1
+    assert edge_to_level[("a", "b")] == 0
+
+
+# ---- _segment_intersects_rect ----
+
+
+def test_segment_intersects_rect_segment_fully_inside() -> None:
+    """A segment entirely within the rect counts as intersecting."""
+    assert _segment_intersects_rect(2, 2, 8, 8, 0, 0, 10, 10) is True
+
+
+def test_segment_intersects_rect_segment_crosses_through_middle() -> None:
+    """A segment passing straight through the rect's interior intersects."""
+    assert _segment_intersects_rect(-5, 5, 15, 5, 0, 0, 10, 10) is True
+
+
+def test_segment_intersects_rect_segment_passes_above() -> None:
+    """A segment entirely above the rect doesn't intersect."""
+    assert _segment_intersects_rect(-5, -20, 15, -20, 0, 0, 10, 10) is False
+
+
+def test_segment_intersects_rect_segment_passes_below() -> None:
+    """A segment entirely below the rect doesn't intersect."""
+    assert _segment_intersects_rect(-5, 50, 15, 50, 0, 0, 10, 10) is False
+
+
+def test_segment_intersects_rect_touches_one_corner() -> None:
+    """Touching exactly one corner counts as intersecting (conservative)."""
+    assert _segment_intersects_rect(-5, -5, 0, 0, 0, 0, 10, 10) is True
+
+
+def test_segment_intersects_rect_touches_one_edge() -> None:
+    """Touching exactly one edge counts as intersecting (conservative)."""
+    assert _segment_intersects_rect(-5, 5, 0, 5, 0, 0, 10, 10) is True
+
+
+def test_segment_intersects_rect_horizontal_segment_clear() -> None:
+    """A horizontal segment clear of the rect doesn't intersect."""
+    assert _segment_intersects_rect(-5, -5, 15, -5, 0, 0, 10, 10) is False
+
+
+def test_segment_intersects_rect_horizontal_segment_through_rect() -> None:
+    """A horizontal segment passing through the rect intersects."""
+    assert _segment_intersects_rect(-5, 5, 15, 5, 0, 0, 10, 10) is True
+
+
+def test_segment_intersects_rect_vertical_segment_through_rect() -> None:
+    """A vertical segment passing through the rect intersects."""
+    assert _segment_intersects_rect(5, -5, 5, 15, 0, 0, 10, 10) is True
+
+
+def test_segment_intersects_rect_vertical_segment_clear() -> None:
+    """A vertical segment clear of the rect doesn't intersect."""
+    assert _segment_intersects_rect(50, -5, 50, 15, 0, 0, 10, 10) is False
+
+
+def test_segment_intersects_rect_degenerate_point_inside() -> None:
+    """A zero-length segment (a point) inside the rect counts as intersecting."""
+    assert _segment_intersects_rect(5, 5, 5, 5, 0, 0, 10, 10) is True
+
+
+def test_segment_intersects_rect_degenerate_point_outside() -> None:
+    """A zero-length segment (a point) outside the rect doesn't intersect."""
+    assert _segment_intersects_rect(50, 50, 50, 50, 0, 0, 10, 10) is False

{visualtorch-1.0.0 → visualtorch-1.2.0}/tests/test_flow.py

@@ -174,16 +174,16 @@ def test_rnn_model_flow_view_runs(rnn_model: nn.Module) -> None:
 
 
 @pytest.mark.parametrize("orientation", ["x", "y", "z"])
-def test_flow_view_one_dim_orientation(classifier_model: nn.Module, orientation: str) -> None:
+def test_flow_view_low_dim_orientation(classifier_model: nn.Module, orientation: str) -> None:
     """Test flow view on a model with a 1D output, for every supported orientation."""
-    img = flow_view(classifier_model, input_shape=(1, 3, 16, 16), one_dim_orientation=orientation)
+    img = flow_view(classifier_model, input_shape=(1, 3, 16, 16), low_dim_orientation=orientation)
     assert img is not None
 
 
-def test_flow_view_invalid_one_dim_orientation_raises(classifier_model: nn.Module) -> None:
-    """An unsupported one_dim_orientation should raise a clear ValueError."""
+def test_flow_view_invalid_low_dim_orientation_raises(classifier_model: nn.Module) -> None:
+    """An unsupported low_dim_orientation should raise a clear ValueError."""
     with pytest.raises(ValueError, match="unsupported orientation"):
-        flow_view(classifier_model, input_shape=(1, 3, 16, 16), one_dim_orientation="bad")
+        flow_view(classifier_model, input_shape=(1, 3, 16, 16), low_dim_orientation="bad")
 
 
 def test_flow_view_with_type_ignore(sequential_model: nn.Sequential) -> None:
@@ -232,6 +232,13 @@ def test_flow_view_output_size_matches_pre_refactor_baseline(sequential_model: n
     `extract_architecture`/`layout_columns` backend - confirmed via a `git worktree` comparison
     at the time of the rewrite (see the graph_view equivalent test for why size, not an exact
     pixel hash: aggdraw's anti-aliasing isn't portable across platforms, but layout math is).
+
+    Updated after `show_input` defaulted to True (an intentional, deliberate visual change - a
+    single-consumer input box is now shown by default, unlike flow_view's original look), so
+    these sizes are no longer literally pre-refactor but reflect the new intended default.
+
+    "legend" was updated again after the synthetic input class was renamed from `InputDummyLayer`
+    to `Input`, changing that legend patch's text width by 1px.
     """
     cases = {
         "default": flow_view(sequential_model, input_shape=(1, 3, 32, 32)),
@@ -242,12 +249,12 @@ def test_flow_view_output_size_matches_pre_refactor_baseline(sequential_model: n
         "no_funnel": flow_view(sequential_model, input_shape=(1, 3, 32, 32), draw_funnel=False),
     }
     expected_sizes = {
-        "default": (124, 42),
-        "no_volume": (116, 32),
-        "show_dimension": (303, 59),
-        "legend": (124, 136),
-        "type_ignore": (82, 42),
-        "no_funnel": (124, 42),
+        "default": (144, 42),
+        "no_volume": (136, 32),
+        "show_dimension": (353, 59),
+        "legend": (148, 136),
+        "type_ignore": (102, 42),
+        "no_funnel": (144, 42),
     }
 
     for name, img in cases.items():
@@ -440,10 +447,13 @@ def test_flow_view_funnels_survive_large_de_differences_between_layers() -> None
     # pre-rewrite main (1ee630e) via a git-worktree comparison for this exact model. The buggy
     # intermediate version rendered thousands fewer non-background pixels here (missing funnel
     # segments), so a wide but real tolerance still catches a regression of that class.
-    assert img.size == (153, 336)
+    #
+    # Updated after `show_input` defaulted to True - the input box is now shown, adding a
+    # consistent amount of extra canvas/content on top of the original baseline above.
+    assert img.size == (171, 364)
     non_bg = _non_background_pixel_count(img)
     error_msg = f"non-background pixel count {non_bg} outside expected range - funnel likely broken"
-    assert 21000 <= non_bg <= 24000, error_msg
+    assert 27000 <= non_bg <= 31000, error_msg
 
 
 def test_flow_view_shows_all_input_boxes_for_multi_input_model() -> None:
@@ -451,7 +461,7 @@ def test_flow_view_shows_all_input_boxes_for_multi_input_model() -> None:
     hiding one would make it ambiguous which arrow originates from which named input.
     """  # noqa: D205
     from visualtorch.backend import extract_architecture
-    from visualtorch.utils.layer_utils import InputDummyLayer
+    from visualtorch.utils.layer_utils import Input
 
     class TwoInputNet(nn.Module):
         def __init__(self) -> None:
@@ -464,10 +474,115 @@ def test_flow_view_shows_all_input_boxes_for_multi_input_model() -> None:
             return self.head(torch.cat([self.a(x), self.b(y)], dim=1))
 
     architecture = extract_architecture(TwoInputNet(), ((1, 4), (1, 4)))
-    input_labels = {
-        layer.module.name() for layer in architecture.columns[0] if isinstance(layer.module, InputDummyLayer)
-    }
+    input_labels = {layer.module.name() for layer in architecture.columns[0] if isinstance(layer.module, Input)}
     assert input_labels == {"input_0", "input_1"}
 
     img = flow_view(TwoInputNet(), input_shape=((1, 4), (1, 4)))
     assert img is not None
+
+
+def test_flow_view_mismatched_depth_siamese_branches_needs_no_detour() -> None:
+    """Sibling branches of different depths merging shouldn't trigger a routed detour."""
+
+    class SiameseNet(nn.Module):
+        def __init__(self) -> None:
+            super().__init__()
+            self.image_branch = nn.Sequential(
+                nn.Conv2d(3, 8, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(),
+                nn.AdaptiveAvgPool2d(1),
+                nn.Flatten(),
+            )
+            self.vector_branch = nn.Sequential(
+                nn.Linear(10, 8),
+                nn.ReLU(),
+            )
+            self.head = nn.Linear(16, 4)
+
+        def forward(self, image: torch.Tensor, vector: torch.Tensor) -> torch.Tensor:
+            """Run each branch on its own input tensor, then concatenate and project."""
+            image_features = self.image_branch(image)
+            vector_features = self.vector_branch(vector)
+            merged = torch.cat([image_features, vector_features], dim=1)
+            return self.head(merged)
+
+    class SiameseNetDepthMatched(nn.Module):
+        def __init__(self) -> None:
+            super().__init__()
+            self.image_branch = nn.Sequential(
+                nn.Conv2d(3, 8, kernel_size=3, stride=1, padding=1),
+                nn.ReLU(),
+                nn.AdaptiveAvgPool2d(1),
+                nn.Flatten(),
+            )
+            self.vector_branch = nn.Sequential(
+                nn.Linear(10, 8),
+                nn.ReLU(),
+                nn.Linear(8, 8),
+                nn.ReLU(),
+            )
+            self.head = nn.Linear(16, 4)
+
+        def forward(self, image: torch.Tensor, vector: torch.Tensor) -> torch.Tensor:
+            """Run each branch on its own input tensor, then concatenate and project."""
+            image_features = self.image_branch(image)
+            vector_features = self.vector_branch(vector)
+            merged = torch.cat([image_features, vector_features], dim=1)
+            return self.head(merged)
+
+    input_shape = ((1, 3, 16, 16), (1, 10))
+    img_mismatched = flow_view(SiameseNet(), input_shape=input_shape)
+    img_matched = flow_view(SiameseNetDepthMatched(), input_shape=input_shape)
+
+    assert img_mismatched.size[1] == img_matched.size[1]
+
+
+def test_flow_view_low_dim_orientation_affects_2d_shapes() -> None:
+    """A 2D shape (e.g. an RNN's (seq_len, hidden_size)) should now respond to
+    low_dim_orientation too, not just genuine 1D shapes.
+    """  # noqa: D205
+
+    class SequenceClassifier(nn.Module):
+        def __init__(self, hidden_size: int) -> None:
+            super().__init__()
+            self.lstm = nn.LSTM(input_size=8, hidden_size=hidden_size, batch_first=True)
+
+        def forward(self, x: torch.Tensor) -> torch.Tensor:
+            out, _ = self.lstm(x)
+            return out
+
+    model = SequenceClassifier(hidden_size=64)
+    input_shape = (1, 5, 8)
+
+    sizes = {
+        orientation: flow_view(model, input_shape=input_shape, low_dim_orientation=orientation).size
+        for orientation in ("x", "y", "z")
+    }
+
+    assert len(set(sizes.values())) == 3, f"expected all 3 orientations to differ, got {sizes}"
+
+
+def test_flow_view_2d_shape_seq_len_is_discarded() -> None:
+    """The positional-like dim (e.g. seq_len) of a 2D shape shouldn't affect box size -
+    only the feature-like dim (e.g. hidden_size) should.
+    """  # noqa: D205
+
+    class SequenceClassifier(nn.Module):
+        def __init__(self, hidden_size: int) -> None:
+            super().__init__()
+            self.lstm = nn.LSTM(input_size=8, hidden_size=hidden_size, batch_first=True)
+
+        def forward(self, x: torch.Tensor) -> torch.Tensor:
+            out, _ = self.lstm(x)
+            return out
+
+    model = SequenceClassifier(hidden_size=64)
+    img_short_seq = flow_view(model, input_shape=(1, 5, 8))
+    img_long_seq = flow_view(model, input_shape=(1, 50, 8))
+
+    assert img_short_seq.tobytes() == img_long_seq.tobytes()
+
+    model_bigger_hidden = SequenceClassifier(hidden_size=256)
+    img_bigger_hidden = flow_view(model_bigger_hidden, input_shape=(1, 5, 8))
+
+    assert img_short_seq.tobytes() != img_bigger_hidden.tobytes()