Looks good -- I wonder if we could have a basic test though?

Exposing TestAll in a different PR sounds good to me.

Sorry for delay on this. I added a test that runs solver.snapshot() and checks if the expected files were generated.

Thanks for the pycaffe extension @gustavla!

Is there also a way to load a snapshot into the solver with pycaffe?

Hi @gustavla , could you please provide an example on how to use this feature?
in #3077 you mentioned that you wrote a SIGTERM interrupt handler?
Thanks!

@mpkuse Just use the restore method that accompanies the solver. The input is the path to your .solverstate file that gets saved during snapshots --> solver.restore('your_solverstate_file.solverstate'). It's also mentioned in the Python section under Interfaces in the Caffe documentation: http://caffe.berkeleyvision.org/tutorial/interfaces.html

Is there a way to set snapshot file name from inside a Python program, or is it necessary to have the name inside the solver prototxt from the beginning? Sorry if I ask something obvious, but there is no documentation, apart from this page and http://caffe.berkeleyvision.org/tutorial/interfaces.html.

@VasLem
This example demonstrates the Python interface to it's full extent. However do note some options may only be available on the opencl branch and not (yet) mainline. But it should give you a good idea of it:
https://github.com/naibaf7/opencl_caffe_examples/blob/master/mnist_lenet/mnist_lenet.ipynb

@naibaf7 :The current snapshot seems to have a problem, or at least I cant find a way around it.
When someone tries to train using Pycaffe and save the best model, one way to do so is to have a main loop like this:

for it in range (max_iter):
    solver.step(1)

    #check if its the best so far, if so , save it  
    solver.snapshot()

However, at this point ,solver.iter has increased by one already and when a solver.snapshot() is issued, it saves the "Next iterations model".
Is there a way around this? or is it me doing something wrong here?
Trying to get the accuracy according to the examples in IPython notebooks also doesn't work. Simply doing :

def run_test(solver, test_iter):
    acc = 0
    #batch_size_test =  solver.test_nets[0].blobs['data'].data.shape[0]
    
    for i in range(test_iter):
        #testing the network on all test set and calculate the test accuracy
        solver.test_nets[0].forward()
        acc += solver.test_nets[0].blobs['accuracy'].data
        # or 
        # corrects += sum(solver.test_nets[0].blobs['ip1'].data.argmax(1) == 
        # solver.test_nets[0].blobs['label'].data)
    
    final_result= '{0:.4f}'.format( acc / test_iter )
    # or 
    # final_result= '{0:.4f}'.format( corrects / test_iter * batch_size_test )
  return float(final_result)

Won't capture the best model, it captures the next immediate one instead.
So Simply put, it if you have set your test_interval to 500, and you get a model with x accuracy.
you wont be able to save that specific model, in that specific iteration.
All you can do is to save the next immediate one.
Any idea how to fix this?