v0.7 (#26)

* update from scratch configs

* update gym pretraining configs - use fewer epochs

* update robomimic pretraining configs - use fewer epochs

* allow trajectory plotting in eval agent

* add simple vit unet

* update avoid pretraining configs - use fewer epochs

* update furniture pretraining configs - use same amount of epochs as before

* add robomimic diffusion unet pretraining configs

* update robomimic finetuning configs - higher lr

* add vit unet checkpoint urls

* update pretraining and finetuning instructions as configs are updated

cfg/d3il/eval/avoid_m1/eval_diffusion_mlp.yaml

@@ -0,0 +1,68 @@
+defaults:
+  - _self_
+hydra:
+  run:  
+    dir: ${logdir}
+_target_: agent.eval.eval_diffusion_agent.EvalDiffusionAgent
+
+name: ${env_name}_eval_diffusion_mlp_ta${horizon_steps}_td${denoising_steps}
+logdir: ${oc.env:DPPO_LOG_DIR}/d3il-eval/${name}/${now:%Y-%m-%d}_${now:%H-%M-%S}_${seed}
+base_policy_path:
+normalization_path: ${oc.env:DPPO_DATA_DIR}/d3il/avoid_m1/normalization.npz
+
+seed: 42
+device: cuda:0
+env_name: avoiding-m5
+obs_dim: 4
+action_dim: 2
+denoising_steps: 20
+cond_steps: 1
+horizon_steps: 4
+act_steps: 4
+
+n_steps: 25
+render_num: 40
+
+plotter:
+  _target_: env.plot_traj.TrajPlotter
+  env_type: avoid
+  normalization_path: ${normalization_path}
+
+env:
+  n_envs: 40
+  name: ${env_name}
+  max_episode_steps: 100
+  reset_at_iteration: True
+  save_video: False
+  best_reward_threshold_for_success: 2
+  save_full_observations: True
+  wrappers:
+    d3il_lowdim:
+      normalization_path: ${normalization_path}
+    multi_step:
+      n_obs_steps: ${cond_steps}
+      n_action_steps: ${act_steps}
+      max_episode_steps: ${env.max_episode_steps}
+      pass_full_observations: ${env.save_full_observations}
+      reset_within_step: False
+
+model:
+  _target_: model.diffusion.diffusion.DiffusionModel
+  predict_epsilon: True
+  denoised_clip_value: 1.0
+  #
+  network_path: ${base_policy_path}
+  network:
+    _target_: model.diffusion.mlp_diffusion.DiffusionMLP
+    time_dim: 16
+    mlp_dims: [512, 512, 512]
+    activation_type: ReLU
+    residual_style: True
+    cond_dim: ${eval:'${obs_dim} * ${cond_steps}'}
+    horizon_steps: ${horizon_steps}
+    action_dim: ${action_dim}
+  horizon_steps: ${horizon_steps}
+  obs_dim: ${obs_dim}
+  action_dim: ${action_dim}
+  denoising_steps: ${denoising_steps}
+  device: ${device}

cfg/d3il/pretrain/avoid_m1/pre_diffusion_mlp.yaml

@@ -25,12 +25,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 15000
+  n_epochs: 5000
   batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 15000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m1/pre_gaussian_mlp.yaml

@@ -24,12 +24,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 10000
+  n_epochs: 5000
   batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 10000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m1/pre_gmm_mlp.yaml

@@ -25,12 +25,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 10000
-  batch_size: 32
+  n_epochs: 5000
+  batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 10000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m2/pre_diffusion_mlp.yaml

@@ -25,12 +25,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 15000
+  n_epochs: 5000
   batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 15000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m2/pre_gaussian_mlp.yaml

@@ -24,12 +24,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 10000
+  n_epochs: 5000
   batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 10000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m2/pre_gmm_mlp.yaml

@@ -25,12 +25,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 10000
-  batch_size: 32
+  n_epochs: 5000
+  batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 10000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m3/pre_diffusion_mlp.yaml

@@ -25,12 +25,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 15000
+  n_epochs: 5000
   batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 15000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m3/pre_gaussian_mlp.yaml

@@ -24,12 +24,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 10000
+  n_epochs: 5000
   batch_size: 16
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 10000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500

cfg/d3il/pretrain/avoid_m3/pre_gmm_mlp.yaml

@@ -25,12 +25,12 @@ wandb:
   run: ${now:%H-%M-%S}_${name}
 
 train:
-  n_epochs: 10000
+  n_epochs: 5000
   batch_size: 32
   learning_rate: 1e-4
   weight_decay: 1e-6
   lr_scheduler:
-    first_cycle_steps: 10000
+    first_cycle_steps: 5000
     warmup_steps: 100
     min_lr: 1e-5
   save_model_freq: 500