BC #02: MMOneSphere, ntrain100, PNet2_svd_pointwise_flow

3DoF demonstrators, testing pointwise loss using `flow` instead of `ee2flow`, which might be more realistic but may lose information related to collisions.
Daniel Seita
Created on April 30|Last edited on May 16
Comment
Contrast these results with using ee2flow instead of flow:
﻿https://wandb.ai/mooey5775/mixed_media/reports/BC-02-MMOneSphere-ntrain100-PNet2_svd_pointwise_ee2flow--VmlldzoxOTE0MDcx﻿
And the simple averaging layer (we want to do as good as this, I suppose, given that SVD should just give us the average of the flows for translation if the rotation is indeed the identity):
﻿https://wandb.ai/mooey5775/mixed_media/reports/BC-02-MMOneSphere-ntrain100-PNet2_avg_ee--VmlldzoxOTI1NjQx/edit﻿
Hmm ... this does seem pretty good in my view. The eval/info_done_final is showing at the end: (0.7639+0.7051+0.3441)/3 = 0.6043 and this is largley due to the weakness of 1 of the runs. 
Edit: ah that was with a different smoothing level, with the same smoothing of 0.99 to compare with others, I get:
(0.6118 + 0.5099 + 0.3641) / 3 = 0.4952 so it's not as good, all right that seems fair the learning curves for eval/info_done_final seem to show a slower initial learning for some reason.
﻿
Edit 05/14/2022: I don't think we can look at eval/MSE_loss as these were before the 05/14 fix where I put the entire evaluation under the model.eval() context.
Binary Success, Train/Eval MSEs﻿
Run set3
﻿
Example GIFs1st seed after 250 epochs (LGTM):
﻿
2nd seed after 250 epochs (LGTM):
﻿
3rd seed after 250 epochs (LGTM):
﻿
VariantUsing act_type: flow and action_type: svd as expected, with ladle_algorithmic_v02 (NOT v04) and translation action mode.
{
  "_hidden_keys": [],
  "act_type": "flow",
  "action_type": "svd",
  "actor_lr": 0.0001,
  "agent": "bc",
  "alg_policy": "ladle_algorithmic_v02",
  "algorithm": "BC",
  "batch_size": 24,
  "bc_data_dir": "/data/dseita/softgym_mm/data_demo/MMOneSphere_v01_BClone_filtered_ladle_algorithmic_v02_nVars_2000_obs_combo_act_translation",
  "bc_data_filtered": true,
  "data_buffer_capacity": 1000000,
  "encoder_type": "pointnet_svd_pointwise",
  "env_kwargs": {
    "action_mode": "translation",
    "action_repeat": 8,
    "camera_name": "top_down",
    "deterministic": false,
    "headless": true,
    "horizon": 100,
    "num_variations": 1000,
    "observation_mode": "cam_rgb",
    "render": true,
    "render_mode": "fluid"
  },
  "env_kwargs_action_mode": "translation",
  "env_kwargs_camera_height": 128,
  "env_kwargs_camera_width": 128,
  "env_kwargs_deterministic": false,
  "env_kwargs_num_variations": 2000,
  "env_kwargs_observation_mode": "point_cloud",
  "env_name": "MMOneSphere",
  "env_version": "v01",
  "exp_name": "MMOneSphere_v01_BC_ntrain_0100_PCL_PNet2_svd_pointwise_flow_2022_04_29_22_24_34_0001",
  "hidden_dim": 1024,
  "lambda_pos": 1.0,
  "lambda_rot": 100.0,
  "log_interval": 1,
  "n_epochs": 250,
  "n_eval_episodes": 10,
  "n_train_demos": 100,
  "num_filters": 32,
  "num_layers": 4,
  "project_axis_ang_y": false,
  "save_freq": 10,
  "save_model": true,
  "save_video": true,
  "seed": 100,
  "wandb_entity": "mooey5775",
  "wandb_project": "mixed_media",
  "weighted_MSE": false
}
﻿
Add a comment