What it does This repo implements channel pruning: instead of zeroing out individual weights (which hardware mostly ignores), it removes entire channels from convolution layers, then finetunes to recover accuracy. It targets old-school Caffe models—VGG-16, ResNet-50, even Faster R-CNN—with published speed-ups of 2×–5× and pre-pruned model weights ready to download.

The interesting bit The “3C” method fuses two older techniques—spatial low-rank decomposition plus channel decomposition—into a single pipeline. The author also maintains a patched Caffe fork with bicubic interpolation and a specific 256-then-crop preprocessing, so you are locked into their toolchain end-to-end.

Key highlights

• Pre-built pruned models available for 4× VGG-16, 2× ResNet-50, and 2×/4× Faster R-CNN
• Top-1 accuracy drops modestly: VGG-16 goes from ~71.5% to 70.6% at 4× speed, or 67.8% at 5×
• Requires 4 GPUs (~11 GB each) for finetuning at batch size 128
• Python 3 scaffolding around a Caffe C++ core; dependencies include scipy, sklearn, easydict
• Paper authors later evolved this into AMC (ECCV 2018), which swaps manual pruning for reinforcement learning

Caveats

• Hard dependency on the author’s Caffe fork; not PyTorch or TensorFlow
• Setup is involved: clone recursive, build Caffe from source, download ImageNet, hand-edit prototxt paths
• README contains a broken “next section” link and some mangled step numbering

Verdict Worth a look if you are maintaining legacy Caffe pipelines or studying structured pruning fundamentals. Everyone else should probably jump straight to the authors’ later AMC work or modern PyTorch pruning toolkits.