search

MSRDL / Deep4Cast

Probabilistic Multivariate Time Series Forecast using Deep Learning
BSD 3-Clause "New" or "Revised" License
94 stars 22 forks source link

Integrated hyperopt #36

Closed bischtob closed 6 years ago

bischtob commented 6 years ago

I integrated hyperopt and wrote a simple tutorial

austingross commented 6 years ago

Found a bug in the notebook: best_values, trials = optimizer.fit() produces error: <keras.optimizers.RMSprop object at 0x000002AC271EB TypeError: 'generator' object is not subscriptable

austingross commented 6 years ago

Checked out the most recent commit but am getting new error in hyperoptimization.ipynb:

best_values, trials = optimizer.fit() returns TypeError: 'generator' object is not subscriptable

The traceback points here: c:\users\augross\desktop\deep4cast\deep4cast\optimizers.py in fit(self) 123 algo=tpe.suggest, 124 trials=trials, --> 125 max_evals=self.n_iter 126 ) 127

and finally here

~\Envs\deep4cast\lib\site-packages\hyperopt\pyll\base.py in toposort(expr) 713 G.add_edges_from([(n_in, node) for n_in in node.inputs()]) 714 order = nx.topological_sort(G) --> 715 assert order[-1] == expr 716 return order 717

austingross commented 6 years ago

@bischtob long-story-short, reinstalling requirements.txt introduced a new error, this time in cntk. Changing the keras backend to tensorflow works...but we need to diagnose the cntk bug. I'll approve the PR so we can continue working on this, but I'll file an issue for the future.

bischtob commented 6 years ago

Got it. I will test on CNTK backend.