Closed tsrobinson closed 2 years ago
An untested version of this function was added in f8e7551f1bf7dfd96f3b720fc315b072bff85c10. @ayn2 to test.
Re. #11, we could create a default option for evaluating model performance using critic score when model mode="wgan" or "cgan"
We could also add functionality to evaluate performance mid-training, and adjust the output dataframe to include metrics per reporting interval, rather than per tested model.
I.e.
Trial | Epoch | Dropout | ... | Layer_struc | Test output |
---|---|---|---|---|---|
1 | 10 | 0.1 | ... | [256,256] | 35 |
... | |||||
1 | 100 | 0.1 | ... | [256,256] | 32 |
2 | 10 | 0.3 | ... | [512,256] | 47 |
Progress:
critic_loss()
function for assessing W-loss as hyperparameter metrictune()
pipelineStill to do:
critic_loss
a default version, perhaps by having test_func = "w"
as an aliastune()
With respect to checkpointing @ayn2, I thought the basic pipeline would be something like:
# User specifies:
epochs = 100
checkpoints = 4
# Calculate epochs per cycle
cycle_epochs = epochs // checkpoints
# Then we loop through the cycles:
for cycle in range(run_cycles):
...
model.fit(..., epochs = cycle_epochs)
...
We need to make sure this works with the k-fold validation -- you don't want these models to "contaminate" each other. I presume we can do this by having the cycle
loop inside the k
loop.
Create function that performs hyperparameter tuning for a GAN given input data.
Should have the following arguments:
hyp_list
-- list of hyperparameters to tunehyp_vals
-- dictionary of options for each hyperparameter (maybe also supply a default dictionary)fixed_opts
-- option to specify fixed aspects of modelrestrict_epochs
-- limit epochstune_mode
-- option to use different search algorithm (random sampling, exhaustive, maybe even latin hypercube etc.)Should also come with a warning about time constraints -- could takes hours/days!