Use Caffe-based mobilenet to classify pit memos

First help Caffe shake a pot: Caffe image processing carried out a very clever package, in order to protobuffer form of organized building block-style network construction is also very flexible and convenient, where the pits are their own walking instability twisted into.

1. A ITER in Caffe is a batch, not an epoch.

2. When using an existing model archive to fine_tune the network, the last layer must be renamed to no longer use the last-level archive parameter because it changes the number of the output. And TensorFlow is not the same time, if the last layer of the parameter number is inconsistent, training will not throw an exception, actually can run normally, and then ... Convergence is just a dream.

3. For the Caffe classification label file, you need to strictly follow the rules starting from 0, if like I shoot the head from 1, you can converge, I live eat mobile phone. (According to colleague experience: categories need to be continuous, 0,1,2,3,4 .... is OK; 0,1,2,4,6 ...). It will be KO's.

4. The choice of learning rate requires the patience of alchemy, it is recommended to start from 0.001 trial and error, each descending an order of magnitude. From the large and small learning rate to the initial training, will greatly improve the time cost (and your sense of resentment).

5. Because I do the classification of the label file in the earlier to do a lot of numerical processing, type conversion and so on, each process/link negligence will cause the final label file problems, and these problems because the data volume is too large, difficult to detect by the naked eye, very error-prone, for the creation of a label file , Be careful!

(The estimate is not complete, can only be continued)

Use Caffe-based mobilenet to classify pit memos