Misc.md

This is quick evaluation of everything, which not fit in any other category on ImageNet-2012.

The architecture is similar to CaffeNet, but has differences:

1. Images are resized to small side = 128 for speed reasons.
2. fc6 and fc7 layers have 2048 neurons instead of 4096.
3. Networks are initialized with LSUV-init

Other

ReLU non-linearity, fc6 and fc7 layer only

Name	Accuracy	LogLoss	Comments
Default	0.471	2.36	bias lr_rate = 2x weights lr_rate
1x	0.470	2.37	bias lr_rate = 1x weights lr_rate
5x	0.472	2.35	bias lr_rate = 5x weights lr_rate
NoBias	0.445	2.50	Biases initialized with zeros, lr_rate = 0

Prototxt, logs

P.S. Logs are merged from lots of "save-resume", because were trained at nights, so plot "Accuracy vs. seconds" will give weird results.