add kitchen related models

kitchen/dlp_kitchen_dataset_40kp_64kpp_4zdim.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:08b9eedcedce0f2f647178d41e61578afc391180d018dbbbfb665344f413351f
+size 772268935

latent_rep_chkpts/dlp_kitchen/hparams.json

@@ -0,0 +1,52 @@
+{
+  "ds": "franka_kitchen",
+  "batch_size": 64,
+  "lr": 0.0002,
+  "kp_activation": "tanh",
+  "pad_mode": "replicate",
+  "num_epochs": 250,
+  "n_kp": 1,
+  "recon_loss_type": "mse",
+  "sigma": 1.0,
+  "beta_kl": 0.1,
+  "beta_rec": 1.0,
+  "patch_size": 16,
+  "topk": 10,
+  "n_kp_enc": 40,
+  "eval_epoch_freq": 1,
+  "learned_feature_dim": 4,
+  "bg_learned_feature_dim": 1,
+  "n_kp_prior": 64,
+  "weight_decay": 0.0,
+  "kp_range": [
+    -1,
+    1
+  ],
+  "run_prefix": "_40kp_64kpp_4zdim",
+  "warmup_epoch": 1,
+  "iou_thresh": 0.15,
+  "anchor_s": 0.25,
+  "kl_balance": 0.001,
+  "milestones": [
+    20,
+    40,
+    80
+  ],
+  "image_size": 128,
+  "cdim": 3,
+  "enc_channels": [
+    32,
+    64,
+    128
+  ],
+  "prior_channels": [
+    16,
+    32,
+    64
+  ],
+  "scale_std": 0.3,
+  "offset_std": 0.2,
+  "obj_on_alpha": 0.1,
+  "obj_on_beta": 0.1,
+  "use_correlation_heatmaps": false
+}

latent_rep_chkpts/dlp_kitchen/saves/dlp_kitchen.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7aa98e16ee9794e948e8a373afad0e738b059e52057472e2bf32c7a31e4b308f
+size 55887527

latent_rep_chkpts/dlp_kitchen/saves/franka_kitchen_dlp_40kp_64kpp_4zdim.pth

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd8dfde0a1a0e8f12f5dc20dd7ef849260543503e2a858a0ee9ed8a297200739
+size 55886353

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/dataset_config.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04b75b06377d4239aa177e0f4ab0b906a43f526919bd699f1ea760ff778ac5e8
+size 498

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/diff.txt

@@ -0,0 +1,261 @@
+diff --git a/analylize_trajectory.ipynb b/analylize_trajectory.ipynb
+index ee01840..6ac8c75 100644
+--- a/analylize_trajectory.ipynb
++++ b/analylize_trajectory.ipynb
+@@ -2,7 +2,7 @@
+  "cells": [
+   {
+    "cell_type": "code",
+-   "execution_count": 2,
++   "execution_count": 1,
+    "metadata": {},
+    "outputs": [],
+    "source": [
+@@ -180,19 +180,20 @@
+   },
+   {
+    "cell_type": "code",
+-   "execution_count": 5,
++   "execution_count": 7,
+    "metadata": {},
+    "outputs": [
+     {
+      "name": "stdout",
+      "output_type": "stream",
+      "text": [
+-      "t: 0 target_object: 51 max_obj_idxes: [20, 51, 100, 154, 202]\n",
+-      "t: 5 target_object: 7 max_obj_idxes: [7, 54, 106, 150, 212]\n",
+-      "t: 10 target_object: 1 max_obj_idxes: [1, 63, 111, 157, 215]\n",
+-      "t: 15 target_object: 1 max_obj_idxes: [1, 71, 111, 169, 203]\n",
+-      "t: 20 target_object: 10 max_obj_idxes: [10, 56, 122, 168, 212]\n",
+-      "t: 25 target_object: 4 max_obj_idxes: [4, 71, 115, 170, 220]\n"
++      "t: 0 target_object: 51 max_obj_idxes: [20, 13, 16, 10, 11, 51, 62, 70, 63, 57, 100, 109, 112, 119, 117, 154, 171, 161, 149, 151, 202, 212, 201, 200, 207] hit_percentage: 0.32\n",
++      "t: 5 target_object: 7 max_obj_idxes: [7, 17, 8, 4, 23, 54, 65, 63, 62, 69, 106, 110, 112, 99, 113, 150, 161, 158, 163, 166, 204, 207, 210, 201, 211] hit_percentage: 0.28\n",
++      "t: 10 target_object: 1 max_obj_idxes: [1, 23, 6, 4, 8, 63, 57, 54, 70, 66, 111, 101, 104, 120, 100, 157, 165, 148, 170, 166, 215, 212, 201, 207, 204] hit_percentage: 0.28\n",
++      "t: 15 target_object: 1 max_obj_idxes: [1, 8, 21, 3, 16, 71, 54, 51, 65, 69, 111, 107, 103, 106, 100, 169, 163, 151, 159, 162, 203, 218, 205, 215, 207] hit_percentage: 0.6\n",
++      "t: 20 target_object: 10 max_obj_idxes: [10, 16, 14, 23, 15, 56, 57, 50, 55, 64, 122, 106, 100, 103, 116, 168, 159, 148, 155, 153, 212, 202, 200, 198, 201] hit_percentage: 0.36\n",
++      "t: 25 target_object: 4 max_obj_idxes: [4, 5, 10, 6, 1, 71, 58, 65, 51, 73, 115, 111, 113, 107, 112, 170, 160, 148, 155, 167, 218, 220, 210, 211, 203] hit_percentage: 0.24\n",
++      "overall_hit_percentage: 0.3466666666666667\n"
+      ]
+     }
+    ],
+@@ -213,21 +214,29 @@
+     "    new_idx = i * 24 + j\n",
+     "    return new_idx\n",
+     "\n",
++    "overall_hit_percentage = []\n",
+     "for t in [0, 5, 10, 15, 20, 25]:\n",
+     "    with open(f'logs/panda_push/plans/3C_plan_3C_pintlarge_dlp_analysis_H5_T100/step_latest/particles_{t}.pkl', 'rb') as f:\n",
+     "        particles = pickle.load(f)\n",
+     "    particles = particles[0, :, :240]\n",
+     "    particles = particles.reshape(-1, 10)[:, 2:]\n",
+-    "    particle_similarity = cdist(particles, particles, metric='cosine')\n",
++    "    particle_similarity = cdist(particles, particles, metric='euclidean')\n",
+     "    particle_similarity = 1 - particle_similarity\n",
+     "\n",
+     "\n",
+     "    target_obj = objects_of_interest_dict[t][0]\n",
+     "    target_particle = convert_obj_idx_to_particle_idx(target_obj)\n",
+-    "    max_particle_idxes = np.argmax(particle_similarity[target_particle].reshape(5, -1), axis=1)\n",
+-    "    max_particle_idxes = [i*24 + j for i, j in enumerate(max_particle_idxes)]\n",
++    "    # max_particle_idxes = np.argmax(particle_similarity[target_particle].reshape(5, -1), axis=1)\n",
++    "    top_k_particle_indices = np.argsort(particle_similarity[target_particle].reshape(5, -1), axis=-1)[:, ::-1][:, :5]\n",
++    "    max_particle_idxes = []\n",
++    "    for i in range(5):\n",
++    "        for idx in top_k_particle_indices[i]:\n",
++    "                max_particle_idxes.append(i*24 + idx)\n",
++    "    # max_particle_idxes = [i*24 + j for i, j in enumerate(max_particle_idxes)]\n",
+     "    max_obj_idxes = [convert_particle_idx_to_obj_idx(idx) for idx in max_particle_idxes]\n",
+-    "    print(\"t:\", t, \"target_object:\", target_obj, \"max_obj_idxes:\", max_obj_idxes)\n",
++    "    hit_percentage = np.mean([1 if obj_idx in objects_of_interest_dict[t] else 0 for obj_idx in max_obj_idxes])\n",
++    "    print(\"t:\", t, \"target_object:\", target_obj, \"max_obj_idxes:\", max_obj_idxes, \"hit_percentage:\", hit_percentage)\n",
++    "    overall_hit_percentage.append(hit_percentage)\n",
+     "    n_particles = 120\n",
+     "\n",
+     "    # fig, ax = plt.subplots()\n",
+@@ -238,33 +247,16 @@
+     "    # ax.imshow(particle_similarity)\n",
+     "    # ax.set_title(f\"Particle Similarity\", fontsize=12)\n",
+     "    # ax.set_xticks(range(n_particles), [f\"{i}\" for i in range(n_particles)])\n",
+-    "    # ax.set_yticks(range(n_particles), [f\"{i}\" for i in range(n_particles)])"
++    "    # ax.set_yticks(range(n_particles), [f\"{i}\" for i in range(n_particles)])\n",
++    "print(\"overall_hit_percentage:\", np.mean(overall_hit_percentage))"
+    ]
+   },
+   {
+    "cell_type": "code",
+-   "execution_count": 38,
++   "execution_count": null,
+    "metadata": {},
+-   "outputs": [
+-    {
+-     "data": {
+-      "text/plain": [
+-       "(5, 24)"
+-      ]
+-     },
+-     "execution_count": 38,
+-     "metadata": {},
+-     "output_type": "execute_result"
+-    }
+-   ],
+-   "source": [
+-    "t: 0 target_object: 51 max_obj_idxes: [20, 51, 100, 154, 202]\n",
+-    "t: 5 target_object: 7 max_obj_idxes: [7, 54, 106, 150, 204]\n",
+-    "t: 10 target_object: 1 max_obj_idxes: [1, 63, 111, 157, 215]\n",
+-    "t: 15 target_object: 1 max_obj_idxes: [1, 71, 111, 169, 203]\n",
+-    "t: 20 target_object: 10 max_obj_idxes: [10, 56, 122, 168, 212]\n",
+-    "t: 25 target_object: 4 max_obj_idxes: [4, 71, 115, 170, 218]"
+-   ]
++   "outputs": [],
++   "source": []
+   },
+   {
+    "cell_type": "code",
+diff --git a/config/TrainDLPConfig.yaml b/config/TrainDLPConfig.yaml
+index 6cfcd34..528c308 100644
+--- a/config/TrainDLPConfig.yaml
++++ b/config/TrainDLPConfig.yaml
+@@ -23,14 +23,14 @@ beta_kl: 0.1  # original
+ beta_rec: 1.0
+ scale_std: 0.3  # default
+ offset_std: 0.2  # default
+-n_kp_enc: 20  # total kp to output from the encoder / filter from prior
+-n_kp_prior: 32
+-patch_size: 16 # prior patch size need to be lower than posterior patch size: posterior is (image size * anchor_s)
++n_kp_enc: 50  # total kp to output from the encoder / filter from prior
++n_kp_prior: 128
++patch_size: 8 # prior patch size need to be lower than posterior patch size: posterior is (image size * anchor_s)
+ learned_feature_dim: 4  # latent visual features for each kp (excluding bg)
+ bg_learned_feature_dim: 1
+ topk: 10  # display top-10 kp with the smallest variance
+ recon_loss_type: "mse"
+-anchor_s: 0.25   # reduce this to 0.125 for small glimpses
++anchor_s: 0.125   # reduce this to 0.125 for small glimpses
+ kl_balance: 0.001
+   
+   
+diff --git a/diffuser/config/pandapush_pint.py b/diffuser/config/pandapush_pint.py
+index e18ff84..e96a7ee 100644
+--- a/diffuser/config/pandapush_pint.py
++++ b/diffuser/config/pandapush_pint.py
+@@ -32,8 +32,8 @@ mode_to_args = {
+             'device': 'cuda:1',
+             'droupout': 0.0,
+             'renderer': 'utils.ParticleRenderer',
+-            'eval_freq': 200,
+-            'n_train_steps': 1e6,
++            'eval_freq': 800,
++            'n_train_steps': 9e5,
+     },
+     '1C_dlp_pusht': {'env_config_dir': 'config/push_t_old',
+             'features_dim': 10,
+@@ -62,7 +62,7 @@ mode_to_args = {
+     '3C_dlp_pusht': {'env_config_dir': 'config/push_t',
+             'features_dim': 12,
+             'multiview': True,
+-            'n_diffusion_steps': 5,
++            'n_diffusion_steps': 50,
+             'model': 'models.AdaLNPINTDenoiser',
+             'particle_normalizer': 'ParticleLimitsNormalizer',
+             'horizon': 5,
+@@ -70,7 +70,12 @@ mode_to_args = {
+             'device': 'cuda:1',
+             'droupout': 0.0,
+             'renderer': 'utils.ParticleRenderer',
+-            'vis_freq': 999,
++            'vis_freq': 20,
++            'hidden_dim': 512,   # 512
++            'projection_dim': 512,  # 512
++            'n_heads': 8,  # 4,  8
++            'n_layers': 12, # 4,  6,  # 12
++            'n_saves': 5,
+     },
+     '1C_state': {'env_config_dir': 'config/n_cubes_state',
+             'features_dim': 4,
+diff --git a/diffuser/config/plan_config/plan_pandapush_pint.py b/diffuser/config/plan_config/plan_pandapush_pint.py
+index d742936..937c936 100644
+--- a/diffuser/config/plan_config/plan_pandapush_pint.py
++++ b/diffuser/config/plan_config/plan_pandapush_pint.py
+@@ -51,7 +51,8 @@ mode_to_args = {
+                   'n_diffusion_steps': 5,
+                   'horizon': 5,
+                   'device': 'cuda:1',
+-                  'diffusion_loadpath': 'diffusion/PushT_3C_dlp_adalnpint_new_H5_T5',
++                  # 'diffusion_loadpath': 'diffusion/PushT_3C_dlp_adalnpint_new_H5_T5',
++                  'diffusion_loadpath': 'diffusion/PushT_3C_dlp_pintlarge_H5_T5',
+                 #   'diffusion_loadpath': 'diffusion/PushT_1C_dlp_eit_H1_T5',
+                   # 'vis_freq': 999,
+                 #   'policy': 'sampling.GoalConditionedBCPolicy',
+diff --git a/diffuser/scripts/train_kitchen.py b/diffuser/scripts/train_kitchen.py
+index bb2f5de..3af843c 100644
+--- a/diffuser/scripts/train_kitchen.py
++++ b/diffuser/scripts/train_kitchen.py
+@@ -250,4 +250,4 @@ for i in range(n_epochs):
+ 
+     ## TODO: add evaluation code here:
+     if i % args.eval_freq == 0:
+-        evaluate_kitchen(policy, env, latent_rep_model, goal_fn, plan_args, video, i, plan_args.savepath, num_evals=40)
++        evaluate_kitchen(policy, env, latent_rep_model, goal_fn, plan_args, video, i, plan_args.savepath, num_evals=100)
+diff --git a/run_batch_plan.sh b/run_batch_plan.sh
+index 47315f9..ba59d13 100644
+--- a/run_batch_plan.sh
++++ b/run_batch_plan.sh
+@@ -1,10 +1,13 @@
+ #!/bin/bash
+ 
+-seeds=(42 188 288 388 488)
+-gpu_numbers=(1 2 3 1 2)
++# seeds=(42 188 288 388 488)
++# gpu_numbers=(3 2 3 1 2)
+ 
+-for i in "${!seeds[@]}"; do
+-    seed=${seeds[$i]}
+-    gpu_number=${gpu_numbers[$i]}
+-    CUDA_VISIBLE_DEVICES=0,$gpu_number python diffuser/scripts/plan_gc_pandapush.py --planning_only --config config.plan_config.plan_pandapush_pint --seed $seed --num_entity 1 --push_t --exp_note pint_action &
+-done
+
++# for i in "${!seeds[@]}"; do
++#     seed=${seeds[$i]}
++#     gpu_number=${gpu_numbers[$i]}
++#     CUDA_VISIBLE_DEVICES=0,$gpu_number python diffuser/scripts/plan_gc_pandapush.py --vis_traj_wandb --planning_only --config config.plan_config.plan_pandapush_pint --seed $seed --num_entity 6 --push_t --exp_note adalnpint &
++# done
++
++CUDA_VISIBLE_DEVICES=0,1 python diffuser/scripts/plan_gc_pandapush.py --vis_traj_wandb --planning_only --config config.plan_config.plan_pandapush_pint --seed 188 --num_entity 4 --push_t --exp_note pintlarge_3color &
++# CUDA_VISIBLE_DEVICES=0,3 python diffuser/scripts/plan_gc_pandapush.py --vis_traj_wandb --planning_only --config config.plan_config.plan_pandapush_pint --seed 188 --num_entity 4 --push_t --exp_note pintlarge_2color
+
+diff --git a/run_batch_train.sh b/run_batch_train.sh
+index 4af8c5a..f51b78b 100644
+--- a/run_batch_train.sh
++++ b/run_batch_train.sh
+@@ -1,10 +1,10 @@
+ #!/bin/bash
+ 
+-seeds=(42 188 288 388 488)
+-gpu_numbers=(1 2 3 1 2)
++seeds=(188 288 388 488)
++gpu_numbers=(1 2 3 3)
+ 
+ for i in "${!seeds[@]}"; do
+     seed=${seeds[$i]}
+     gpu_number=${gpu_numbers[$i]}
+-    CUDA_VISIBLE_DEVICES=0,$gpu_number python diffuser/scripts/train.py --seed $seed --num_entity 3 --input_type vqvae --exp_note adalnpint &
++    CUDA_VISIBLE_DEVICES=0,$gpu_number python diffuser/scripts/train_kitchen.py --seed $seed --num_entity 1 --kitchen --exp_note 40kp_0.25anchor_s &
+ done
+
+diff --git a/vq_bet_official/examples/configs/train_pandapush.yaml b/vq_bet_official/examples/configs/train_pandapush.yaml
+index 45b6976..ac8559a 100644
+--- a/vq_bet_official/examples/configs/train_pandapush.yaml
++++ b/vq_bet_official/examples/configs/train_pandapush.yaml
+@@ -16,8 +16,9 @@ sequentially_select: false
+ # vqvae_load_dir: "vq_bet_official/checkpoints/pandapush/2024-09-25/21-18-19/silvery-wind-20/trained_vqvae.pt" # 2C
+ # vqvae_load_dir: "vq_bet_official/checkpoints/pandapush/2024-09-25/21-35-15/apricot-haze-25/trained_vqvae.pt" # 3C
+ # vqvae_load_dir: "vq_bet_official/checkpoints/pandapush/2024-09-26/00-37-23/floral-frost-33/trained_vqvae.pt" # 1T
+-vqvae_load_dir: "vq_bet_official/checkpoints/pandapush/2024-09-26/15-23-53/classic-capybara-39/trained_vqvae.pt"
+-num_entity: 2
++# vqvae_load_dir: "vq_bet_official/checkpoints/pandapush/2024-09-26/15-23-53/classic-capybara-39/trained_vqvae.pt"
++vqvae_load_dir: "vq_bet_official/checkpoints/pandapush/2024-09-26/22-36-48/woven-music-45/trained_vqvae.pt"
++num_entity: 3
+ env_config_dir: config/vqvae_push_t
+ # env_config_dir: config/n_cubes_raw
+ push_t: true

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/diffusion_config.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a093c84c6e03ffb58464bffdaf3c2a41da282703b7c8fd20bb16d36599decbf
+size 341

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/eval_800_0.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:30a460bc9515a5911ab5b577ae473e6343843af02757e9ab638207c3643b6fda
+size 2938530

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/model_config.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1485a1fc1ce71da742f9629ae7e82142e1eeac50568d7787622b31ca7c3a7f2f
+size 329

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/render_config.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:534ea1b7376c27b5bc94d8185cafddfb86361d6151d6a266ae9aaac98a5ab7f3
+size 157

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/state_810000.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1989c9f09d77db7583ad8dc61dc1ed45388fb99a5153ce52a30ee7aa344d19c0
+size 64892725

pretrained_models/kitchen/diffusion/kitchen_1C_dlp_40kp_0.25anchor_s_H5_T5/trainer_config.pkl

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2497149f77e0498c9f1b096ddc5ceebb4b0cf7619a05e89358d0d21463767aba
+size 401