[ChrisW09]

Fixes #537

Problem

ResNet.__init__ took block i's input width from layer_sizes[i] and always built the output layer as nn.Linear(layer_sizes[-1], num_classes). Only num_blocks equal to len(layer_sizes) - 1 or len(layer_sizes) worked:

• with more blocks, layer_sizes[i] raised IndexError, contrary to the model page's "later blocks stay at the final width";
• with fewer blocks (including num_blocks=0), the output layer didn't match the last block's width and the first forward pass failed.

Fix

deeptab/architectures/resnet.py tracks the running width:

• block i maps the current width to layer_sizes[min(i + 1, len(layer_sizes) - 1)];
• the output layer reads the width the last block produces, or layer_sizes[0] without blocks.

Configurations that worked before build the identical network.

Tests

tests/architectures/test_architecture_contracts.py::test_resnet_blocks_follow_layer_sizes_for_any_num_blocks[0-blocks|1-block|2-blocks|3-blocks|4-blocks|5-blocks] checks each block's widths, the output layer width, a forward pass and a backward pass with layer_sizes=[16, 8, 4]. The cases with 0, 1, 4 and 5 blocks fail on main; four blocks is the issue's IndexError case. The 2- and 3-block cases worked before and build the same state_dict as on main. All pass with this change.

This also makes the documented example in docs/api/configs/index.rst work. It uses ResNetConfig(layer_sizes=[256, 128, 32], num_blocks=4) and raises IndexError on main.

Docs

• ResNetConfig numpydoc: layer_sizes is described as the width schedule, and num_blocks explains blocks beyond it and the output width.
• docs/model_zoo/stable/resnet.md: the building-blocks table and the implementation notes describe the schedule and the output layer.

Checks

• just test: 3242 passed, 46 skipped, 28 xfailed; coverage 91.29%
• just docs: build succeeded (warnings treated as errors)
• just check: all hooks passed (ruff, prettier, pyright)