Getting 0% accuracy while fine-tuning

[Vir Gandhi]

Hi, 

I'm new to caffe and trying different examples, one of which is:

I have to finetune ImageNet trained caffe model on a new dataset. So, I followed the notebook example given the caffe website: Fine-tune the ImageNet-trained CaffeNet on the "Flickr Style" dataset.

I modified my dataset accordingly and also changed the nodes in the final layer from 5 to 7 (as I have 7 categories and flickr had 5) by changing .prototxt files.

But after training I'm getting zero% accuracy, and I'm able to figure out where I'm making mistake.

I'm also attaching my prototxt files and the python code and last part of my output:

Running solvers for 200 iterations...

I0328 06:08:32.630990  5870 solver.cpp:228] Iteration 0, loss = 1.94591

I0328 06:08:32.631019  5870 solver.cpp:244]     Train net output #0: acc = 0

I0328 06:08:32.631037  5870 solver.cpp:244]     Train net output #1: loss = 1.94591 (* 1 = 1.94591 loss)

I0328 06:08:32.631059  5870 sgd_solver.cpp:106] Iteration 0, lr = 0.001

I0328 06:08:32.715519  5870 solver.cpp:228] Iteration 0, loss = 1.94591

I0328 06:08:32.715559  5870 solver.cpp:244]     Train net output #0: acc = 0

I0328 06:08:32.715579  5870 solver.cpp:244]     Train net output #1: loss = 1.94591 (* 1 = 1.94591 loss)

I0328 06:08:32.715596  5870 sgd_solver.cpp:106] Iteration 0, lr = 0.001

  0) pretrained: loss=1.946, acc= 0%; scratch: loss=1.946, acc= 0%

 10) pretrained: loss=0.000, acc=100%; scratch: loss=0.625, acc=100%

 20) pretrained: loss=0.000, acc=100%; scratch: loss=0.077, acc=100%

 30) pretrained: loss=0.000, acc=100%; scratch: loss=0.028, acc=100%

 40) pretrained: loss=0.000, acc=100%; scratch: loss=0.015, acc=100%

 50) pretrained: loss=0.000, acc=100%; scratch: loss=0.013, acc=100%

 60) pretrained: loss=0.000, acc=100%; scratch: loss=0.011, acc=100%

 70) pretrained: loss=0.000, acc=100%; scratch: loss=0.011, acc=100%

 80) pretrained: loss=0.000, acc=100%; scratch: loss=0.009, acc=100%

 90) pretrained: loss=0.000, acc=100%; scratch: loss=0.009, acc=100%

100) pretrained: loss=0.000, acc=100%; scratch: loss=0.009, acc=100%

110) pretrained: loss=0.000, acc=100%; scratch: loss=0.009, acc=100%

120) pretrained: loss=0.000, acc=100%; scratch: loss=0.007, acc=100%

130) pretrained: loss=0.000, acc=100%; scratch: loss=0.007, acc=100%

140) pretrained: loss=0.000, acc=100%; scratch: loss=0.007, acc=100%

150) pretrained: loss=0.000, acc=100%; scratch: loss=0.007, acc=100%

160) pretrained: loss=0.000, acc=100%; scratch: loss=0.006, acc=100%

170) pretrained: loss=0.000, acc=100%; scratch: loss=0.006, acc=100%

180) pretrained: loss=0.000, acc=100%; scratch: loss=0.006, acc=100%

190) pretrained: loss=0.000, acc=100%; scratch: loss=0.006, acc=100%

199) pretrained: loss=0.000, acc=100%; scratch: loss=0.006, acc=100%

Done.

[...]

Accuracy, trained from random initialization: 0.0%

[Jan]

To clarify: Your problem is not the training itself but that your code outputs a different accuracy than caffe itself during training, correct?

From a first look I don't see any problem with your code. What does the logging output look like,  are the weights actually loaded by caffe when you do the caffe.Net(..., weights, caffe.TEST) in eval_style_net ?

Jan