Very high (number of directories / number of code files) ratio
Organization is inconsistent: top-level directories are variously named for type of task (classify), type of architecture (rnn), or model itself (gpt-2).
Within classify directory, inconsistency in directory structure (img/ has an imagenet subdirectory, but everything else (mnist, cifar) is at top level, including another imagenet model.
A number of utilities (I think) are not documented with their purpose.
classify also has a semi_supervised subdirectory. Unless all semi-supervised examples are going to be classification examples, it should probably not be here.
There are a lot of directories with only one subdirectory.
New Guidelines
Fewer dirs with only one subdir.
Separate folders for common tutorial tasks (images, text)
Separate folders for specific research work (one folder per paper); open to future consolidation if there are several similar papers
Current Issues:
New Guidelines