Skip to content

Data v2: GPU-CorruptNet exports with layer4.2.bn2

Latest

Choose a tag to compare

@tanaymihani tanaymihani released this 05 Nov 03:51
· 32 commits to main since this release

Exports from GPU-CorruptNet's trained ResNet-50 for the Tenstorrent runs.

  • corruptnet.npz: unchanged from data-v1 (head features, weights and logits; the input to layer4.2.bn3 for 32 frames).
  • corruptnet_bn2.npz: the input to layer4.2.bn2 for the same 32 frames, channels last (1,568 x 512), plus that BatchNorm folded into a scale and a shift. In ResNet-50, bn2 is followed by ReLU; bn3 is followed by the residual add. Made by model/corruptnet_bn2_export.py, which also checks that bn3's input matches data-v1.

Checksums are in data.sha256.