From: Reversible designs for extreme memory cost reduction of CNN training
Model | Accuracy | #Params | Channels | Pooling | \(M_{\theta }\) | \(M_{z}'+M_{g}'\) | \(\mathcal {M}\) |
---|---|---|---|---|---|---|---|
Resnet | \(94.7\%\) | 3.1M | \(32 - 64 - 128 - 256\) | Max Pooling | 12.5M | 1928 | 1.01G |
RevNet | \(94.5\%\) | 3.1M | \(40 - 80 - 256 - 320\) | Max Pooling | 12.7M | 640 | 348M |
i-RevNet | \(93.8\%\) | 42.8M | \(32 - 128 - 512 - 2048\) | \(\mathcal {P}_c - \mathcal {P}_c - \mathcal {P}_c\) | 171M | 640 | 500M |
RevNeXt (ours) | \(93.3\%\) | 3.7M | \(32 - 128 - 512 - 512\) | \([\mathcal {P}_c, \mathcal {P}_c, \mathcal {P}_b]\) | 14.8M | 352 | 200M |