search

doomie / HessianFree

An implementation of the Hessian-free optimization algorithm in Theano
61 stars 23 forks source link

broken on MNIST? #1

Open lzamparo opened 9 years ago

lzamparo commented 9 years ago

Hi,

I wanted to try out your HF implementation, but cannot get it to run in your example as given in the README.md.

Even after fixing a few small errors in deep_autoencoder (272: rbm.get_cost_updates() no longer takes weight_decay param, 401:402 valid_batch_size/test_batch_size params should be swapped), I still get nans in the CG code. There were also a bunch of warnings in the compilation of theano.function for unused inputs. This gist contains a subset of the output, basically the ridge just keeps growing until it reaches inf.

appleyuchi commented 5 years ago

I'll check it in the following days@lzamparo

appleyuchi commented 5 years ago

@lzamparo Hi,same as yours,I cannot run it either,for no util.py file.