CoSTAR Dataset Joints vector embeddings

[priyankahubli]

No description provided.

[ahundt]

Wow this code looks pretty clean overall, thanks!

• Could you merge the changes from the branch called sharper, and then add instructions to run training and embedding visualization to the README.md?

  • I'll need the command line-by-line instructions I can copy paste in myself to create the embeddings myself and visualize them.

• Were you able to train any new embeddings with these? Did it work ok? Could you create a screenshot and add it to the README.md to show what is expected?

• Is there also a specific version of the costar_dataset repository I'll need?

[ahundt]

Am I correct that this code goes with ahundt/costar_dataset#9?

[priyankahubli]

Yes, now you would need the latest version of the costar dataset from the feature_embeddings branch - 4cd22cb.
I have added and made the changes you requested. Please let me know if this is okay.
Thanks!

[priyankahubli]

I also merged it with the sharper branch.

[ahundt]

Edit: Sorry the comments here were for the wrong pull request. I'm testing this PR now.

[ahundt]

I tried running the commands as specified in your README.md changes and it crashed right away:

ahundt@femur|~/src/darts/cnn on priyanka_hard_youtube?
± export CUDA_VISIBLE_DEVICES="1" && python3 train_costar.py --data ~/.keras/datasets/costar_block_stacking_dataset_v0.4 --arch TDC --epochs 200 --batch_size 128 --feature_mode time_difference_images --num_images_per_example 200
Warning:  apex was installed without --cuda_ext. Fused syncbn kernels will be unavailable.  Python fallbacks will be used instead.
Warning:  apex was installed without --cuda_ext.  FusedAdam will be unavailable.
Warning:  apex was installed without --cuda_ext.  FusedLayerNorm will be unavailable.
/usr/local/lib/python3.5/dist-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.
  from ._conv import register_converters as _register_converters
Tensorflow is not installed. Skipping tf related imports
Experiment dir : eval-20190712-164157-EXP-stacking-TDC-0
2019_07_12_16_41_57 loading op dict: operations.OPS
2019_07_12_16_41_57 loading primitives:genotypes.PRIMITIVES
2019_07_12_16_41_57 primitives: ['none', 'max_pool_3x3', 'skip_connect', 'sep_conv_3x3', 'dil_conv_3x3', 'flood_conv_3x3', 'dil_flood_conv_3x3', 'choke_conv_3x3', 'dil_choke_conv_3x3']
Getting stacking data
Using CoSTAR Dataset
2019_07_12_16_41_59 Initial Learning Rate: 0.004690909090909092                                                                                                                                                                                                                            
Traceback (most recent call last):                                                                                                                                                                                                                                                         
  File "train_costar.py", line 796, in <module>
    main()
  File "train_costar.py", line 366, in main
    train_stats = train(train_loader, model, criterion, optimizer, int(epoch), args)
  File "train_costar.py", line 500, in train
    prefetcher = data_prefetcher(train_loader, cutout=args.cutout, cutout_length=args.cutout_length)
  File "train_costar.py", line 459, in __init__
    self.preload()
  File "train_costar.py", line 463, in preload
    self.next_input_img, self.next_input_vec, self.next_target = next(self.loader)
ValueError: not enough values to unpack (expected 3, got 2)
-> [1]

I'm guessing that's in the following lines of train_costar.py?

    prefetcher = data_prefetcher(train_loader, cutout=args.cutout, cutout_length=args.cutout_length)

    cart_error, angle_error = [], []
    input_img, input_vec, target = prefetcher.next()

[ahundt]

This currently seems to crash right away when I follow the instructions, see notes above and below.

[ahundt]

1. All of your additions look like they are under the original DARTS readme section, not in a separate section or below sharpDARTS. We need to create a separate section before the DARTS section for your new changes. The DARTS section begins at "The sharpDARTS Repository is Based on Differentiable Architecture Search (DARTS)"
2. Notes here must link to the original paper and to the costar dataset, explain what this part of the code is/does so people can figure out what a "Temporal Distance Classifier" etc is.