Skip to content

RESNET18.md

Latest commit

 

History

History
91 lines (62 loc) · 3.19 KB

RESNET18.md

File metadata and controls

91 lines (62 loc) · 3.19 KB
Raw

ResNet-18 BasicBlock (resnet_basic_block)

Fused ResNet BasicBlock used in ResNet-18 and ResNet-34 (not the bottleneck block in ResNet-50+).

Layer kind 10. A full ResNet-18 backbone (stem + 8 blocks + classifier) can be expressed in one .nk file with kMaxLayers=100.

Block math (NHWC)

Matches the standard ResNet BasicBlock (inference BN folded into scale/bias):

  1. Conv 3×3 — in_channels → out_channels, stride s, pad 1
  2. BatchNorm + ReLU
  3. Conv 3×3 — out_channels → out_channels, stride 1, pad 1
  4. BatchNorm
  5. Residual — identity when stride == 1 and in_channels == out_channels; else 1×1 conv + BN projection with stride s
  6. ReLU

Runtime: conv uses Conv2D::forward → Kernels::; batch norm, ReLU, and residual merge use fused_kernel_ops.hpp → Kernels:: (CMSIS when enabled, reference fallback otherwise). Skip/residual is MatAddND — no separate skip kernel.

.nk layer descriptor (kind 10)

Field Type
in_channels uint32
out_channels uint32
stride uint8 (1 or 2)
reserved 3 bytes

Weights (tensor pairs: W then B)

Identity shortcut (stride=1, in == out):

  1. conv1 [out, 3, 3, in] + bias [out]; BN scale + beta [out] each
  2. conv2 [out, 3, 3, out] + bias; BN scale + beta

Projection shortcut (otherwise), append:

  1. shortcut [out, 1, 1, in] + bias; BN scale + beta

Fixture

models/resnet18_basic_block.nk — 4×4×4 input, identity BasicBlock, one TCAS case.

python tools/write_resnet18_basic_block_fixture.py

Full backbone

models/resnet18.nk — stem (7×7 conv, ReLU, max pool) + 8 BasicBlocks + avg pool + dense head. Input 56×56×3, 13 layers, ~11.2M weights. Regenerate:

python tools/write_resnet18_fixture.py

Builder: python/netkit/resnet18.py (build_resnet18_arch). Embedded TCAS tolerance 1e-4 (deep-network float accumulation).

Packager fusion: expand_resnet_basic_block_to_linear() + fuse_composite_blocks() in python/netkit/nk_fuse.py roundtrip the composite block from a linear conv/bn stack (same pattern as UIB). ONNX import with --no-fuse emits primitives; default convert optimize runs packager fuse — see models/import_resnet_basic_block.nk.

Pack PyTorch checkpoint

python -m netkit pack --arch resnet18 -o models/my_resnet18.nk --height 56 --width 56 --num-classes 10
# or
python tools/pack_resnet18_checkpoint.py -o models/my_resnet18.nk

Uses python/netkit/torch_backbone_pack.py to fold BatchNorm and map timm resnet18 weights into composite BasicBlock tensors.

Parity: python/tests/test_torch_backbone_pack.py (NumPy reference vs timm) and test_torch_backbone_runtime_parity.py (C++ tools/nk_infer vs timm) — see TESTING.md.

Python

arch = {
    "network": "cnn",
    "input": [H, W, C_in],
    "layers": [{
        "type": "resnet_basic_block",
        "in_channels": C_in,
        "out_channels": C_out,
        "stride": 1,
    }],
}

C++

  • Kernel: ResNetBasicBlock::forward() in src/resnet_basic_block.cpp
  • Loader: NkFormat::LayerKind::ResNetBasicBlock
  • Op registry: NkResNetBasicBlockOpDescriptor