[deeplab] Requesting Documentation for Auto-deeplab / hnasnet

[lm-b]

System information

• What is the top-level directory of the model you are using: models-master\research\deeplab
• Have I written custom code (as opposed to using a stock example script provided in TensorFlow): no
• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Windows 7
• TensorFlow installed from (source or binary): binary
• TensorFlow version (use command below): 1.9
• Bazel version (if compiling from source): N/A
• CUDA/cuDNN version: 7.6
• GPU model and memory: Nvidia GP100 16 GB
• Exact command to reproduce:
  To run the NAS: python train.py --logtostderr --training_number_of_steps=1000 --train_split="train" --model_variant="nas_hnasnet" --output_stride=8 --decoder_output_stride=4 --train_crop_size="385,385" train_batch_size=1 --dataset="SDML" --train_logdir="testing\outputlogs" --dataset_dir="testing\SDML" --fine_tune_batch_norm=False --atrous_rates=12 --atrous_rates=24 --atrous_rates=36 --initialize_last_layer=False
  To extract the architecture: Unknown

Describe the problem

I'd like to request documentation on how to deploy the auto-deeplab or hnasnet. I'm able to perform the NAS training using the above code, but after that point there appears to be no indication on how to decode the ideal network from the graph or if /where that decoder script is included in the deeplab files. The paper says that the NAS uses the DARTs decoder to ID the best cells and a Viterbi algorithm to ID the best macroscale architecture, but I'm finding the implementation of those steps to be non-obvious from the available files, and it doesn't appear to be triggered automatically by the train.py script. I'm admittedly new to tensorflow, but based on other issues (e.g. #7222 , #7201 , #7453 ) there seems to be significant general confusion on which commands and/or .py files to use to deploy the NAS and how to extract the best found network. Any info in the right direction would be great, and adding it to the deeplab README in some form might alleviate opening new issues in this vein in the future.

Source code / logs

N/A