How many number filters for macro child to reproduce the results in the paper

[nott0]

You mention that the architecture in the paper has 12 layers with 2 fixed pooling layer (_factorized_reduction as written in the code), but didn't mention about how many out_filters is used to produce the result.

Indeed, macro_final.sh does give the result at around 96%, but it has 24 layers and 96 out_filters. How many out_filters for 12 layers child to reproduce the results in the paper?

[prabhant]

I had the same question, as the macro_search.sh produces only 12 layers of architecture and macro_final has a 24layer architecture.
But if you notice the parameters in scripts then you'd notice that in macro_search.sh they have only specified the --child_num_layers=12 and in macro_final its 24 same with the number of output filters so i assumed that there might be some inconsistency in the hyperparameters given in the script.