It would be useful to have some functions on the dataset and the model to help name the logs. Something like
"runs/" + dataset.oneline_description() + "/" + model.oneline_description() + other_info
These need to incorporate all that is needed to have uniqueness and it would also be helpful if they helped organize the comparisons.
By using /, we separate into folders, so the rightmost folders are easier to be compared together (although tensorboard is helpful enough for this).
So I think it is better to start with everything related to the dataset.
It might get crowded, so maybe some very short description tag is better? Like hash-ME650-MN50-MP500-...?
For the model, describing which layers (including sizes) would be necessary. It would be better to have it automatically, but it doesn't seem easy. Also, the model is small enough to be easy to change the string generator. That being said, at least the layer size needs to be given dynamically. Something like
def str_description(self):
return f"in-GC{len(self.conv1)}-RELU-GCout-mean"
It would be useful to have some functions on the dataset and the model to help name the logs. Something like
These need to incorporate all that is needed to have uniqueness and it would also be helpful if they helped organize the comparisons.
By using
/, we separate into folders, so the rightmost folders are easier to be compared together (although tensorboard is helpful enough for this).So I think it is better to start with everything related to the dataset.
It might get crowded, so maybe some very short description tag is better? Like
hash-ME650-MN50-MP500-...?For the model, describing which layers (including sizes) would be necessary. It would be better to have it automatically, but it doesn't seem easy. Also, the model is small enough to be easy to change the string generator. That being said, at least the layer size needs to be given dynamically. Something like