Use 1-dim feature vector extracted from pre-trained model to train a new CNN

[Kelvin-Zhong]

I have saw some other similar issues, but none of them can help me solve this problem.
I have used python to put all the feature extracted from the pre-train model in the leveldb, and use it as the source to train a new network.

I extracted the output vector of fc6 of imagenet network, which is a 1-dim 4096 vector.
Each value of my leveldb is a datum.SerializedtoString().
and datum.channel = 1, datum.height = 4096, datum.width = 1, and I have set the datum.label.

However, when I use this leveldb as input to train network, the loss function doesn't decrease, so as the Test score #1, and the Test #0 is stuck...

What is the issue? Can anyone provide a simple example on how to use this 1-dim vector to create a right leveldb to train a new network???

Many thanks~~

[Yangqing]

Just checking - are you using datum.data or datum.float_data?

Also, loss not decreasing is usually because you are not setting SGD right - a wrong network, or wrong learning rates, etc. Check those (maybe with a grid search on hyperparameters) to find out.

[Kelvin-Zhong]

Hi, Yangqing, I am using datum.float_data? Is it the reason that my training fails?
What's the difference between datum.data and datum.float_data?

[shelhamer]

There is now an example for HDF5 data modeling for vector features that should help you double-check the mechanics of what you're doing: http://nbviewer.ipython.org/github/BVLC/caffe/blob/dev/examples/hdf5_classification.ipynb

Further than that it is likely a matter of tuning the training as Yangqing advised.