A major (performance) update on the submodule: srl-zoo

[ncble]

Image rotated by 90 degrees

• see issue

Image preprocessing

Instead of using specific image preprocessing (mode='imagenet') as

    x[..., 0] -= 0.485
    x[..., 1] -= 0.456
    x[..., 2] -= 0.406
    # Scaling
    x[..., 0] /= 0.229
    x[..., 1] /= 0.224
    x[..., 2] /= 0.225

the mode='tf' that rescale pixel value to [-1,1] is preferable and could avoid unwanted effect when applying training with the other datasets/networks.

[ncble]

Since the srl_zoo is less popular than robotic-rl-srl, I decided to post the changelog here:

Highlights

1. 2~5 times speed-up (overall) (srl_zoo) compared to the current version of origin/master
2. Better SRL training mechanism (more intuitive, better modularity) that support sophisticated update (e.g. GAN)
3. Add GAN to srl_zoo
4. Scalable SRL models (support any image shape, see the detail below)
5. ~9 times faster DataLoader which is also simpler, since it's natively supported by pytorch.
6. Remove several redundant codes to speed-up training.
7. Fix several issues of (robotics-rl_srl): #41, #42, #43, #46, #47, #48, #49...

Note: Before, we need to modify 8 scripts in order to add one new model, now only two scripts (at most three): models/modules.py and models/my_custom_model.py (see the template models/new_model_template.py)

Changelog

SRL part

• Support any image resolution for the entire toolbox (--img-shape="(3,128,128)"), including the DataLoader, Environment, SRL models. Before, the SRL models are not scalable with respect to image shape, and it's not sufficient to modify only the input shape (e.g. need to manually calculate each layers' shape, size, etc). Now, all models function more like keras.

• Support adversarial state representation learning. (e.g. GAN)

• [New scripts] models/base_trainer.py, models/new_model_template.py, models/gan.py

  • base_trainer.py: new trainig pipeline for better modularity.
  • new_model_template.py: a simple example of "how to add new model to srl_zoo".
  • gan.py: adversarial state representation learning.

• Better (simpler, ~10 times faster) plots

• Support new monitor mode (--monitor) "loss" (before, there is only "pbar" progressbar): monitor losses during training, calculate GTC per epoch.

• Support control of number of CPU for dataloader (--num-worker).

• Support "anytime training": load the previous trained SRL model weights to continue the training --srl-pre-weights (weights path)

• Change validation mechanism to the classic one (i.e. within one epoch: train then valid). Before, we alternate between training and validation mode at batch level.

• Support specific GPU number. (by --gpu-num=0, --gpu_num1, etc)

• [Remove]: preprocessing/preprocess.py (it's useless), models/custom_layers.py

• [Rename] the models/models.py is renamed to models/base_models.py, since it's more intuitive for the outsider. Currently, there are several confusing names "custom_layers.py", "modules.py", "learner.py" "models.py

  RL part

• support any image shape.

• support specific GPU number. (by --gpu-num=0, --gpu_num1, etc)

• support --srl-model-path indicate the SRL model weights path. Before, we can only load either the latest (by calling --latest) or manually change the config/srl_model.yaml model weights path.

• register new srl models

• fix issues:

  • image rotated by 90 degrees
  • --log-folder folder doesn't exist.

• [New] replay/plot_pipeline.py: aggregate all losses and plot on one figure. (draft code to be refined, merged with replay/aggregate_plots.py, compare_plots.py, gather_results.py)

[araffin]

Btw, why did you use a branch on your repo where you could create a branch on the original repo? (you are part of the collaborators), same for srl-robotics repo

[ncble]

The original version (master) of srl_zoo has 5719 lines of code. My pull-request has already added and removed +4145/-1453 lines of code, thus 5598 lines of code have been modified (~97% of code). There is no need to be merged to the master branch.