Problem when I just run tests in mnist example

[bsar321]

Hello,
In the example "eth_mnist.py" when I run the tests after training network I get an accuracy of 0.75 (75%). So I don't always have to do the training, I save the weights, assignments and proportions variables at the end of the training. At a later moment, if I perform only the test (giving the previously mentioned variables to the network) I get an accuracy of only 0.62 (62%). What am I forgetting to save and what should I provide to the network to be able to replicate the results without having to run the training every time?
Thanks
Bruno

[Hananel-Hazan]

Its seems that you saving all that needed to recreate the experiment. Try to save the variable you mention to a file and load them again when needed. I suspect that the weight or the assignments and proportions variables change somehow. Also, make sure to use network.train(mode=False) in the end of the training.

[bsar321]

Hi, thank you for your reply.
What I do in the eth_mnist.py file is:

• After line 280 torch.save(network.connections[("X", "Ae")].w,'weights_INPUTtoEXCITATORY.pt') , and the same for assignments and proportions variables.
• Finishing the training, and after these saved files I do
  network.connections[("X","Ae")].w = torch.load(weights_INPUTtoEXCITATORY.pt') and similar for the other parameters.

What I see that might be missing is the layer bias, but how do I access the bias?
Thank you again

[bsar321]

Hi,
I solved question. The missing variable was "theta" (Adaptative thresholds).
Thank you

[Hananel-Hazan]

wonderful

[machuanfeng]

Comment options
Hello, may I ask how to save this parameter (theta) and where is it located