Sketchpad(ImageEditor) passes {"background","layers","composite"}
dict to predict(); handle composite with white-background alpha
compositing. Also quiet launch + UTF-8 stdout guards for the
space container's cp949 locale.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Add MNIST drawing demo: Flask web app and HF Space source
web/: local Flask app serving a canvas UI backed by the
depth5 model. space/: Gradio app (Sketchpad input) deployed as
https://huggingface.co/spaces/shyo2/a0-digit-demo. Shared
MNIST-style centering preprocessing lives in src/preprocess.py.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
YAML frontmatter (license, tags, pipeline_tag) plus model
description, results matrix, usage snippet, and repro commands
so the repo renders as a proper model card on the Hub.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
HF rejects non-LFS binaries; route results/*.png through LFS
alongside the model weights.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
RESULTS.md documents the full experiment matrix and findings:
plain CNNs saturate ~99.5% with BN but destabilize at depth40,
and collapse entirely without BN (11.35%), while same-depth
residual nets hold ~99.4%. DATA.md records MNIST source and the
torchvision auto-download path. results/*.png are the generated
comparison plots.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Add trained weights and history for 13 experiment runs
Grid over depth {5,10,15,20,40} x {plain,residual} x {BN,no-BN}.
Each run dir holds model.pt (state_dict, ~0.4-3M params) and
history.json with per-epoch loss/accuracy. Weights tracked via
git-lfs.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Reads runs/*/history.json and renders val-acc / train-loss curves
(with and without BN) plus a best-accuracy summary bar chart into
results/*.png.
Generated with [Devin](https://devin.ai)
Co-Authored-By: Devin <158243242+devin-ai-integration[bot]@users.noreply.github.com>
Add CNN/ResNet model builder and MNIST training pipeline