Add files using upload-large-folder tool

.gitattributes

@@ -199,3 +199,4 @@ wj867hap/eval_video/2_eval.mp4 filter=lfs diff=lfs merge=lfs -text
 wj867hap/eval_video/0_eval.mp4 filter=lfs diff=lfs merge=lfs -text
 3o6eq09y/episode_rosbags/episode_4_2024-12-31-21-17-59.bag filter=lfs diff=lfs merge=lfs -text
 nkdxms1x/episode_rosbags/episode_0_2024-12-31-21-03-48.bag filter=lfs diff=lfs merge=lfs -text
+3o6eq09y/episode_rosbags/episode_3_2024-12-31-21-17-17.bag filter=lfs diff=lfs merge=lfs -text

3o6eq09y/episode_rosbags/episode_3_2024-12-31-21-17-17.bag

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a57f5b875bacf8950e34cf0e2c271563d03d68bdb62322fcdba680f69ed3f65e
+size 1412768785

ksj6ie6d/.hydra/config.yaml

@@ -0,0 +1,350 @@
+root_dir: /home/${oc.env:USER}/fish_leon
+nstep: 3
+seed: 41
+dataset_shuffle_seed: ${seed}
+device: cuda
+save_video: true
+save_buffer: true
+use_tb: true
+baseline: false
+use_wandb: true
+eval: true
+process_contact_features: ${eval}
+obs_type: pixels
+use_color: true
+use_depth: true
+use_masks: false
+mask_list:
+- EE_obj_mask
+mask_representation: channels
+crop_hw:
+- 144
+- 144
+crop_down_offset: 48
+color_crop_type: null
+depth_crop_type: null
+segmask_crop_type: null
+add_crop_binary_mask: false
+add_coord_conv_map: false
+use_context_color: false
+use_context_depth: false
+use_context_segmask: false
+context_color_crop_type: null
+context_depth_crop_type: null
+context_segmask_crop_type: null
+context_add_crop_binary_mask: false
+context_add_coord_conv_map: false
+use_contact_map: false
+use_sdf_maps: false
+use_normals_maps: false
+which_objects: both
+max_contact_prob: 0.1
+max_depth: 2.0
+grasped_dtc_max_value: 0.105
+env_dtc_max_value: 0.425
+grasped_normals_mask_max_dtc_value: 0.105
+env_normals_mask_max_dtc_value: 0.425
+clamp_dtc: true
+dtc_adaptive_normalization: false
+mask_normals_within_sdf: true
+adaptive_normals_mask: true
+learnable_contact_preprocess_params: false
+contact_model_name: local_multitask_outhd64all_home_crop_h144w144d48_ctxt_seed_183386_epoch_9
+contact_estimation_model_ckpt_path: ~/fish_leon/contact_estimation/artifacts/175604_2/checkpoints/epoch=09-val_loss=0.00.ckpt
+num_eval: 5
+debug_timestamps: false
+open_loop: false
+action_trajectories: true
+stop_after_action: false
+interpolation_frequency: 25
+policy_frequency: 5
+wait_for_new_camera_frames: true
+random_start: false
+eval_starts: ${root_dir}/FISH/eval_starts/${suite.name}_${obs_type}/${task_name}
+train_demo_idxs_list_or_num: null
+num_valid_demos: null
+val_num_groups: 3
+name_of_expert_demo: 112_240x320_all_twodim_left_to_right_annotated_start_idx_5hz_zstd7_EE_pxl_coords_expert_demos_imp_act
+expert_dataset_dirpath: ${root_dir}/FISH/expert_demos/${suite.name}/${task_name}/${name_of_expert_demo}
+expert_dataset: ${expert_dataset_dirpath}/demos.zarr
+action_key: ${oc.if_else:${action_trajectories}, 'action_trajectory_${interpolation_frequency}hz',
+  'action'}
+semantic_demo_grouping_name: semantic_demo_grouping.yaml
+semantic_demo_grouping: ${expert_dataset_dirpath}/${semantic_demo_grouping_name}
+expert_dataset_config: ${expert_dataset_dirpath}/demo_config.yaml
+bc_regularize: false
+bc_weight_type: qfilter
+load_checkpoint: ${agent.load_checkpoint}
+wandb_run_id: '1007_0'
+true_action_history: false
+wandb_notes: null
+checkpoint_epoch: 12000
+load_residual_weight: false
+checkpoint_root_dir: /home/${oc.env:USER}/fish_leon/FISH
+checkpoint_weight_dir: ${checkpoint_root_dir}/exp_local/${suite.name}_${obs_type}/${task_name}/${wandb_run_id}
+residual_weight: ${root_dir}/FISH/weights/${suite.name}_${obs_type}/${task_name}/weight.pt
+experiment_dir: ./exp_local/${suite.name}_${obs_type}/${task_name}/${wandb_run_id}
+final_experiment_dir: ${experiment_dir}/${oc.generate_run_id:}
+agent:
+  _target_: agent.diffusion_policy.DiffusionPolicyAgent
+  name: diffusion_policy
+  load_checkpoint: ${eval}
+  device: ${device}
+  n_obs_steps: ${.config.policy_cfg.n_obs_steps}
+  suite_name: ${suite.name}
+  obs_type: ${obs_type}
+  enable_arm: ${eval}
+  enable_camera: ${eval}
+  use_tb: ${use_tb}
+  desired_image_shape:
+  - 13
+  - 180
+  - 240
+  orig_cam_shape:
+  - 3
+  - 240
+  - 320
+  config:
+    _target_: agent.diffusion_policy.DiffusionPolicyAgentConfig
+    compile: false
+    device: ${device}
+    cam_resize_shape: ${agent.desired_image_shape}
+    orig_cam_shape: ${agent.orig_cam_shape}
+    policy_frequency: ${policy_frequency}
+    interpolation_frequency: ${interpolation_frequency}
+    policy_cfg:
+      _target_: lerobot.common.policies.diffusion.configuration_diffusion.DiffusionConfig
+      n_obs_steps: 1
+      horizon: 36
+      n_action_steps: ${agent.config.policy_cfg.horizon}
+      input_shapes:
+        observation.image: ${agent.config.cam_resize_shape}
+        context_observation.image: ${agent.config.cam_resize_shape}
+        observation.state:
+        - 8
+        observation.action_history:
+        - 7
+      output_shapes:
+        action:
+        - 7
+      input_normalization_modes:
+        observation.image: mean_std
+        observation.state: min_max
+        observation.action_history: min_max
+      output_normalization_modes:
+        action: min_max
+      vision_backbone: resnet18
+      pretrained_backbone_weights: null
+      transforms:
+      - _target_: torchaug.transforms.RandomAffine
+        degrees:
+        - -5
+        - 5
+        translate:
+        - 0.05
+        - 0.05
+        batch_transform: true
+        num_chunks: -1
+        batch_inplace: true
+      - _target_: torchaug.transforms.RandomColorJitter
+        brightness: 0.3
+        contrast: 0.4
+        saturation: 0.5
+        hue: 0.08
+        batch_transform: true
+        num_chunks: -1
+        batch_inplace: true
+      use_group_norm: true
+      spatial_softmax_num_keypoints: 32
+      action_history_encoder_config:
+        _target_: lerobot.common.policies.diffusion.configuration_diffusion.Unet1dEncoderConfig
+        in_channels: 7
+        out_channels: 32
+        history_length: ${agent.config.policy_cfg.n_action_steps}
+        kernel_size: ${agent.config.policy_cfg.kernel_size}
+        downsample_kernel_size: 3
+        downsample_stride: 2
+        downsample_padding: 1
+      down_dims:
+      - 256
+      - 512
+      - 1024
+      kernel_size: 5
+      n_groups: 8
+      diffusion_step_embed_dim: 128
+      use_film_scale_modulation: true
+      noise_scheduler_type: DDIM
+      beta_schedule: squaredcos_cap_v2
+      beta_start: 0.0001
+      beta_end: 0.02
+      prediction_type: epsilon
+      clip_sample: true
+      clip_sample_range: 1.0
+      num_train_timesteps: 50
+      num_inference_steps: 10
+      do_mask_loss_for_padding: false
+    train_cfg:
+      _target_: utils.TrainConfig
+      lr: 0.0001
+      lr_scheduler: cosine
+      lr_warmup_steps: 500
+      adam_betas:
+      - 0.95
+      - 0.999
+      adam_eps: 1.0e-08
+      adam_weight_decay: 1.0e-06
+      grad_clip_norm: 10
+      offline_steps: ${num_train_frames_diffusion}
+      use_amp: true
+    observation_cfg:
+      _target_: agent.encoder.VisualFeatureSet
+      use_depth: ${use_depth}
+      use_color: ${use_color}
+      mask_input_dict:
+        _target_: agent.encoder.MaskInputDict
+        enable: ${use_masks}
+        representation: ${mask_representation}
+        mask_list: ${mask_list}
+      crop_input_config:
+        _target_: agent.encoder.CropInputConfig
+        color_crop_type: ${color_crop_type}
+        depth_crop_type: ${depth_crop_type}
+        segmask_crop_type: ${segmask_crop_type}
+        crop_hw: ${crop_hw}
+        crop_down_offset: ${crop_down_offset}
+        add_crop_binary_mask: ${add_crop_binary_mask}
+        add_coord_conv_map: ${add_coord_conv_map}
+      context_input_config:
+        _target_: agent.encoder.ContextInputConfig
+        use_color: ${use_context_color}
+        use_depth: ${use_context_depth}
+        mask_input_dict:
+          _target_: agent.encoder.MaskInputDict
+          enable: ${use_context_segmask}
+          representation: ${mask_representation}
+          mask_list: ${mask_list}
+        crop_input_config:
+          _target_: agent.encoder.CropInputConfig
+          color_crop_type: ${context_color_crop_type}
+          depth_crop_type: ${context_depth_crop_type}
+          segmask_crop_type: ${context_segmask_crop_type}
+          crop_hw: ${crop_hw}
+          crop_down_offset: ${crop_down_offset}
+          add_crop_binary_mask: ${context_add_crop_binary_mask}
+          add_coord_conv_map: ${context_add_coord_conv_map}
+      mask_soft_approx_scheduler_config:
+        _target_: agent.encoder.MaskSoftApproxSchedulerConfig
+        num_steps: 40000
+        initial_value: 10.0
+        final_value: 1000.0
+        interpolation_scheme: constant
+      use_contact_map: ${use_contact_map}
+      use_sdf_maps: ${use_sdf_maps}
+      use_normals_maps: ${use_normals_maps}
+      which_objects: ${which_objects}
+      grasped_dtc_max_value: ${grasped_dtc_max_value}
+      env_dtc_max_value: ${env_dtc_max_value}
+      grasped_normals_mask_max_dtc_value: ${grasped_normals_mask_max_dtc_value}
+      env_normals_mask_max_dtc_value: ${env_normals_mask_max_dtc_value}
+      clamp_dtc: ${clamp_dtc}
+      max_contact_prob: ${max_contact_prob}
+      mask_normals_within_sdf: ${mask_normals_within_sdf}
+      dtc_adaptive_normalization: ${dtc_adaptive_normalization}
+      adaptive_normals_mask: ${adaptive_normals_mask}
+      max_depth: ${max_depth}
+      image_shape: ${agent.desired_image_shape}
+      learnable_contact_preprocess_params: ${learnable_contact_preprocess_params}
+      learning_rate: ${agent.config.train_cfg.lr}
+      weight_decay: 0.0
+      contact_model_name: ${contact_model_name}
+      zero_centered: false
+suite:
+  suite: frankagym
+  name: frankagym
+  frame_stack: ${agent.n_obs_steps}
+  action_repeat: 1
+  discount: 0.99
+  hidden_dim: 1024
+  num_train_frames: 2010
+  num_seed_frames: 260
+  num_train_epochs: 5000
+  validate_every_epochs: 100
+  validate_diffusion_on_action_loss_every_epochs: 500
+  train_eval_diffusion_on_action_loss_every_epochs: 500
+  check_topk_every_epochs: 10
+  save_snapshot_every_epochs: 5000
+  eval_every_frames: 2000
+  num_eval_episodes: 5
+  save_snapshot: true
+  wait_for_user_to_start_episode: true
+  task_make_fn:
+    _target_: suite.frankagym.make
+    name: ${task_name}
+    height: 240
+    width: 320
+    frame_stack: ${suite.frame_stack}
+    action_repeat: ${suite.action_repeat}
+    seed: ${seed}
+    enable_arm: ${agent.enable_arm}
+    enable_gripper: ${enable_gripper}
+    start_with_gripper_open: ${start_with_gripper_open}
+    enable_camera: ${agent.enable_camera}
+    path_to_depth_extrinsics: ${path_to_depth_extrinsics}
+    contact_estimation_model_ckpt_path: ${contact_estimation_model_ckpt_path}
+    x_limit: ${x_limit}
+    y_limit: ${y_limit}
+    z_limit: ${z_limit}
+    device: ${device}
+    interpolation_frequency: ${interpolation_frequency}
+    policy_frequency: ${policy_frequency}
+    debug_timestamps: ${debug_timestamps}
+    stop_after_action: ${stop_after_action}
+    open_loop: ${open_loop}
+    wait_for_new_camera_frames: ${wait_for_new_camera_frames}
+    action_key: ${action_key}
+    action_trajectory_horizon: ${agent.config.policy_cfg.horizon}
+    action_trajectories: ${action_trajectories}
+    path_to_zarr_dataset: ${expert_dataset}
+    agent_policy_cfg: ???
+    true_action_history: ${true_action_history}
+num_train_frames_bc: 50000
+num_train_frames_drq: 1100000
+stddev_schedule_drq: linear(1.0,0.1,100000)
+task_name: FrankaInsertion-v1
+num_train_frames_vinn: 25000
+num_train_frames_diffusion: 1000000
+num_train_epochs_bc: 5000
+num_train_epochs_diffusion: 5000
+validate_every_epochs_bc: 5
+validate_every_epochs_diffusion: 25
+validate_diffusion_on_action_loss_every_epochs: 50
+train_eval_diffusion_on_action_loss_every_epochs: 500
+check_topk_every_epochs: 5
+check_topk_every_epochs_diffusion: ${validate_diffusion_on_action_loss_every_epochs}
+save_snapshot_every_epochs_diffusion: 5000
+x_limit:
+- 0.2
+- 0.7
+y_limit:
+- -0.4
+- 0.4
+z_limit:
+- -0.05
+- 0.55
+home_displacement:
+- 0.55
+- 0.0
+- 0.55
+- 180.0
+- 0.0
+- 0.0
+enable_gripper: true
+start_with_gripper_open: true
+offset_mask:
+- 1
+- 1
+- 1
+- 1
+- 1
+- 1
+path_to_depth_extrinsics: ~/fish_leon/FISH/cfgs/camera_poses/camera_poses_L515/20240904-122305/color_tf_world.npy

ksj6ie6d/.hydra/hydra.yaml

@@ -0,0 +1,169 @@
+hydra:
+  run:
+    dir: ${final_experiment_dir}
+  sweep:
+    dir: ${final_experiment_dir}
+    subdir: ${hydra.job.num}
+  launcher:
+    submitit_folder: ${final_experiment_dir}/.slurm
+    timeout_min: 60
+    cpus_per_task: null
+    gpus_per_node: null
+    tasks_per_node: 1
+    mem_gb: null
+    nodes: 1
+    name: ${hydra.job.name}
+    stderr_to_stdout: false
+    _target_: hydra_plugins.hydra_submitit_launcher.submitit_launcher.LocalLauncher
+  sweeper:
+    _target_: hydra._internal.core_plugins.basic_sweeper.BasicSweeper
+    max_batch_size: null
+    params: null
+  help:
+    app_name: ${hydra.job.name}
+    header: '${hydra.help.app_name} is powered by Hydra.
+
+      '
+    footer: 'Powered by Hydra (https://hydra.cc)
+
+      Use --hydra-help to view Hydra specific help
+
+      '
+    template: '${hydra.help.header}
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (group=option)
+
+
+      $APP_CONFIG_GROUPS
+
+
+      == Config ==
+
+      Override anything in the config (foo.bar=value)
+
+
+      $CONFIG
+
+
+      ${hydra.help.footer}
+
+      '
+  hydra_help:
+    template: 'Hydra (${hydra.runtime.version})
+
+      See https://hydra.cc for more info.
+
+
+      == Flags ==
+
+      $FLAGS_HELP
+
+
+      == Configuration groups ==
+
+      Compose your configuration from those groups (For example, append hydra/job_logging=disabled
+      to command line)
+
+
+      $HYDRA_CONFIG_GROUPS
+
+
+      Use ''--cfg hydra'' to Show the Hydra config.
+
+      '
+    hydra_help: ???
+  hydra_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][HYDRA] %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+    root:
+      level: INFO
+      handlers:
+      - console
+    loggers:
+      logging_example:
+        level: DEBUG
+    disable_existing_loggers: false
+  job_logging:
+    version: 1
+    formatters:
+      simple:
+        format: '[%(asctime)s][%(name)s][%(levelname)s] - %(message)s'
+    handlers:
+      console:
+        class: logging.StreamHandler
+        formatter: simple
+        stream: ext://sys.stdout
+      file:
+        class: logging.FileHandler
+        formatter: simple
+        filename: ${hydra.runtime.output_dir}/${hydra.job.name}.log
+    root:
+      level: INFO
+      handlers:
+      - console
+      - file
+    disable_existing_loggers: false
+  env: {}
+  mode: RUN
+  searchpath: []
+  callbacks: {}
+  output_subdir: .hydra
+  overrides:
+    hydra:
+    - hydra.mode=RUN
+    task:
+    - agent=diffusion
+    - suite=frankagym
+    - suite/frankagym_task@_global_=insertion
+  job:
+    name: eval_robot
+    chdir: true
+    override_dirname: agent=diffusion,suite/frankagym_task@_global_=insertion,suite=frankagym
+    id: ???
+    num: ???
+    config_name: config_eval
+    env_set: {}
+    env_copy: []
+    config:
+      override_dirname:
+        kv_sep: '='
+        item_sep: ','
+        exclude_keys: []
+  runtime:
+    version: 1.3.2
+    version_base: '1.1'
+    cwd: /home/leonmkim/fish_leon/FISH
+    config_sources:
+    - path: hydra.conf
+      schema: pkg
+      provider: hydra
+    - path: /home/leonmkim/fish_leon/FISH/cfgs
+      schema: file
+      provider: main
+    - path: ''
+      schema: structured
+      provider: schema
+    output_dir: /home/leonmkim/fish_leon/FISH/exp_local/frankagym_pixels/FrankaInsertion-v1/1007_0/ksj6ie6d
+    choices:
+      suite: frankagym
+      suite/frankagym_task@_global_: insertion
+      agent: diffusion
+      hydra/env: default
+      hydra/callbacks: null
+      hydra/job_logging: default
+      hydra/hydra_logging: default
+      hydra/hydra_help: default
+      hydra/help: default
+      hydra/sweeper: basic
+      hydra/launcher: submitit_local
+      hydra/output: default
+  verbose: false

ksj6ie6d/.hydra/overrides.yaml

@@ -0,0 +1,3 @@
+- agent=diffusion
+- suite=frankagym
+- suite/frankagym_task@_global_=insertion

ksj6ie6d/tb/events.out.tfevents.1735697318.leonmkim-ROG-Strix-G15CS-G15CS.1741133.0

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95740a1c217bacdbf8973af251ea8461fbed9e4a24fd87b124da865d5b9dbc17
+size 88

ksj6ie6d/wandb/run-20241231_210837-ksj6ie6d/files/code/FISH/eval_robot.py

@@ -0,0 +1,599 @@
+#%%
+import warnings
+import os
+
+os.environ['MKL_SERVICE_FORCE_INTEL'] = '1'
+os.environ['MUJOCO_GL'] = 'egl'
+from pathlib import Path
+#%%
+import hydra
+import numpy as np
+import torch
+
+import utils
+from utils import get_feature_dirname_from_configs
+
+from video import VideoRecorder
+import pickle
+import time
+import threading
+import shutil
+from logger import Logger
+
+import wandb 
+from omegaconf import OmegaConf, open_dict
+
+from replay_buffer_robot import RosbagEvalReplayBufferStorage
+from lerobot.common.utils.utils import _relative_path_between
+
+torch.backends.cudnn.benchmark = True
+warnings.filterwarnings('ignore', category=DeprecationWarning)
+
+# import specs for replay buffer
+from dm_env import specs
+
+import sys, signal
+import yaml
+
+import binomial_cis as bc
+
+# get path of current file
+current_path = os.path.dirname(os.path.realpath(__file__))
+sys.path.append(os.path.join(current_path, os.pardir))
+# from contact_estimation.src.utils.viz_utils import normalized_surface_normal_to_rgb, depth_map_to_im, grasped_env_dtc_map_to_im, contact_prob_map_to_im, desaturate_color_image, masked_overlay_im_list
+
+def make_agent(obs_spec, action_spec, cfg):
+	cfg.obs_shape = obs_spec['pixels'].shape
+	dataset_statistics = None # this will be loaded from the checkpoint
+	try:
+		cfg.action_shape = action_spec.shape
+	except:
+		pass
+	return hydra.utils.instantiate(cfg, dataset_statistics)
+
+class Workspace:
+	def __init__(self, cfg):
+		self.work_dir = Path.cwd()
+		print(f'workspace: {self.work_dir}')
+		
+		signal.signal(signal.SIGINT, self.signal_handler)
+		
+		self.cfg = cfg
+		self.loading_uncompiled_checkpoint_with_compile = False
+		self.loading_compiled_checkpoint_with_no_compile = False
+
+		snapshot_path = Path(self.cfg.checkpoint_weight_dir) / f'snapshot_{self.cfg.checkpoint_epoch}.pt' 
+		self.load_checkpoint_conf(snapshot_path=snapshot_path)
+
+		# load config for action trajectories
+		utils.set_seed_everywhere(self.cfg.seed)
+		self.device = torch.device(self.cfg.device)
+		self.setup()
+
+		# self.agent = make_agent(self.eval_env.observation_spec(),
+		# 						self.eval_env.action_spec(), self.cfg.agent)
+		self.timer = utils.Timer()
+		# self._global_step = 0
+		self._global_episode = 0
+		self._global_epoch = 0
+		self.num_episode_successes = 0
+
+		self.alpha_range = [.01, .025, .05, .1]
+
+		# Need to convert hydra config to primitive container for wandb https://docs.wandb.ai/guides/integrations/hydra
+		with open_dict(self.cfg):
+			self.cfg.feature_type = get_feature_dirname_from_configs(
+				hydra.utils.instantiate(self.cfg.agent.config.observation_cfg), 
+				self.cfg.agent.config.policy_cfg.input_shapes,
+				hydra.utils.instantiate(self.cfg.agent.config.policy_cfg.action_history_encoder_config) if 'observation.action_history' in self.cfg.agent.config.policy_cfg.input_shapes else None,
+			)
+
+		wandb_config = OmegaConf.to_container(
+			self.cfg, resolve=True, throw_on_missing=True
+		)
+		# must be called before any tf summary writer is created
+		if self.cfg.use_wandb:
+			# get the run id from the final_experiment_dir directory
+			run_id = os.path.basename(os.path.normpath(self.cfg.final_experiment_dir))
+			wandb.init(project='extrinsic_contact_downstream', entity='serialexperimentsleon', job_type='eval', sync_tensorboard=self.cfg.use_tb, config=wandb_config, id=run_id)
+
+		self.logger = Logger(self.work_dir, use_tb=self.cfg.use_tb, use_wandb=self.cfg.use_wandb)
+
+		# if not self.loading_uncompiled_checkpoint_with_compile and self.cfg.agent.config.compile:
+		# 	self.agent.compile_modules()
+			
+		# self.load_checkpoint(snapshot_path=snapshot_path)
+
+		# if self.loading_uncompiled_checkpoint_with_compile: # need to call compile after loading the checkpoint
+		# 	self.agent.compile_modules()
+
+		print(f"loaded agent with feature_type: {self.cfg.feature_type}")
+
+	def check_for_key_press(self):
+		while self.continue_keypress_thread:
+			inp = input("Press 'r' to restart current episode, 'n' to stop current episode and skip to next, 'q' to break entire eval\n")
+			if inp == 'n':
+				self.preempt_episode = True
+				print("preempting episode")
+			elif inp in ['', '0', '1']: # enter key
+				if inp in ['0', '1']:
+					self.num_episode_successes += int(inp)
+				self.proceed_after_env_reset_event.set()
+				print("proceeding to start episode!")
+			elif inp == 'q':
+				self.proceed_after_env_reset_event.set()
+				self.preempt_episode = True
+				self.exit_eval = True
+				self.continue_keypress_thread = False # will stop the keypress thread
+				print("quitting eval")
+				break
+			elif inp == 'r':
+				print('restarting episode')
+				self.preempt_episode = True
+				self.restart_episode = True
+			else:
+				print("Invalid key press, try again")
+		
+		# self.keypress_input_thread.join() # wait for the keypress thread to finish
+
+	def signal_handler(self, signal, frame):
+		print("\nprogram exiting gracefully")
+		self.proceed_after_env_reset_event.set()
+		self.preempt_episode = True
+		self.exit_eval = True
+		self.continue_keypress_thread = False # will stop the keypress thread
+		self.keypress_input_thread.join() # wait for the keypress thread to finish
+		video_filepath = self.video_recorder.save()
+		# get the video file and convert to video tensor to log 
+		self.logger.log_video('eval/video', video_filepath, self.global_step)
+		wandb.finish()
+		sys.exit(0)
+		
+	def setup(self):
+		# create envs
+		self.eval_env = hydra.utils.call(self.cfg.suite.task_make_fn)
+		# expert_demo_config_path = os.path.join(os.path.dirname(self.cfg.expert_dataset), 'demo_config.yaml')
+		# self.expert_demo_config = yaml.load(open(expert_demo_config_path, 'r'), Loader=yaml.FullLoader)
+		# self.eval_env._env.action_trans_norm = expert_demo_config['max_translation_action_norm']
+		# self.eval_env._env.action_rot_norm = expert_demo_config['max_rotation_action_norm']
+		# self.eval_env._env.action_period = expert_demo_config['sample_period']
+		# print(f"setting max_translation_action_norm to {expert_demo_config['max_translation_action_norm']} and sample_period to {expert_demo_config['sample_period']}")
+		# print(f"setting max_rotation_action_norm to {expert_demo_config['max_rotation_action_norm']}")
+
+		# self.eval_env.set_demo_params(self.cfg.expert_dataset)
+
+		# Turn off random start
+		self.eval_env.random_start = False
+
+		# create replay buffer
+		# data_specs = [
+		# 	{
+		# 		'observation': self.eval_env.observation_spec(),
+		# 	},
+		# 	# self.eval_env.observation_spec()['features'],
+		# 	self.eval_env.action_spec(),
+		# 	specs.Array(self.eval_env.action_spec().shape, self.eval_env.action_spec().dtype, 'vinn_action'), 
+		# 	specs.Array((1, ), np.float32, 'reward'),
+		# 	specs.Array((1, ), np.float32, 'discount'),
+		# ]
+
+		# self.eval_replay_storage = ZarrEvalReplayBufferStorage(data_specs, self.work_dir / 'eval_buffer', debug_timestamps=self.cfg.debug_timestamps, save_buffer=self.cfg.save_buffer, debug_info_data_specs=self.eval_env.debug_info_data_specs, camera_info_dict=self.eval_env.get_camera_info_dict())
+		self.eval_replay_storage = RosbagEvalReplayBufferStorage(self.work_dir)
+
+		self.video_recorder = VideoRecorder(
+			self.work_dir if self.cfg.save_video else None,
+			ros_enabled=True,
+			fps=self.cfg.agent.config.policy_frequency,
+			)
+		
+		print('workspace setup complete')
+		
+	@property
+	def global_step(self):
+		# return self._global_step
+		return self.eval_env.get_global_step()
+
+	@property
+	def global_episode(self):
+		return self._global_episode
+
+	@property
+	def global_frame(self):
+		return self.global_step * self.cfg.action_repeat
+
+	@property
+	def global_epoch(self):
+		return self._global_epoch
+	
+	def reset(self, eval_idx):
+		if not self.eval_env.enable_arm:
+			return np.array([0,0,0], dtype=np.float32)
+		self.eval_env.arm_refresh(reset=False)
+		# Set start position
+		try:
+			self.eval_env.set_position(self.start_pos[eval_idx])
+		except:
+			self.eval_env.arm.set_position(self.start_pos[eval_idx])
+			if self.eval_env.arm.keep_gripper_closed:
+				self.eval_env.arm.close_gripper_fully()
+			else:
+				self.eval_env.arm.open_gripper_fully()
+		time.sleep(0.1)
+		time_step = self.eval_env.step(np.zeros(self.eval_env.action_spec().shape[0], dtype=np.float32),
+									   np.zeros(self.eval_env.action_spec().shape[0], dtype=np.float32))	
+		return time_step
+
+	def eval(self):
+		# before evals start, prompt user for name of grasped object and the left book of the slot location
+		grasped_obj_name = input("Enter the name of the grasped object: ")
+		left_book_slot = input("Enter the left book slot location: ")
+		# update wandb config
+		if self.cfg.use_wandb:
+			wandb.config.update({'grasped_obj_name': grasped_obj_name, 'left_book_slot': left_book_slot})
+
+		self.preempt_episode = False
+		self.exit_eval = False
+		self.restart_episode = False
+
+		self.continue_keypress_thread = True
+		self.proceed_after_env_reset_event = threading.Event()
+		self.keypress_input_thread = threading.Thread(target=self.check_for_key_press)
+		self.keypress_input_thread.start()
+
+		# # Set model to eval mode
+		# self.agent.train(False)
+		
+		eval_until_episode = utils.Until(self.cfg.num_eval)
+
+		self.use_action_history = False
+		# if "dp" in repr(self.agent) and "observation.action_history" in self.cfg.agent.config.policy_cfg.input_shapes:
+		if "observation.action_history" in self.cfg.agent.config.policy_cfg.input_shapes:
+			self.use_action_history = True
+
+		# self.eval_replay_storage._new_eval_step(0)
+
+		# if 'vinn' in repr(self.agent) or 'openloop' in repr(self.agent):
+		# 	with open(self.cfg.expert_dataset, 'rb') as f:
+		# 		if self.cfg.obs_type == 'pixels':
+		# 			self.expert_demo, _, self.expert_action, self.expert_reward = pickle.load(f)
+		# 		elif self.cfg.obs_type == 'features':
+		# 			_, self.expert_demo, self.expert_action, self.expert_reward = pickle.load(f)
+				
+		# 		if self.cfg.action_trajectories:
+		# 			with open(self.cfg.expert_action_trajectories, 'rb') as f:
+		# 				self.expert_action = pickle.load(f)
+
+		# 		if isinstance(self.cfg.train_demo_idxs_list_or_num, int):
+		# 			if self.cfg.train_demo_idxs_list_or_num == -1:
+		# 				self.cfg.train_demo_idxs_list_or_num = len(self.expert_demo)
+		# 			train_demo_idxs_list_or_num = list(range(self.cfg.train_demo_idxs_list_or_num))
+				
+		# 		self.expert_demo = self.expert_demo[train_demo_idxs_list_or_num]
+		# 		self.expert_action = self.expert_action[train_demo_idxs_list_or_num]
+		# 		self.expert_reward = self.expert_reward[train_demo_idxs_list_or_num]
+		# 		# if self.cfg.action_plans:
+		# 		# 	self.expert_action_plans = self.expert_action_plans[self.cfg.train_demo_idxs_list_or_num]
+		# 		# self.expert_demo = self.expert_demo[:self.cfg.num_demos]
+		# 		# self.expert_action = self.expert_action[:self.cfg.num_demos]
+		# 		# self.expert_reward = self.expert_reward[:self.cfg.num_demos]
+
+		# 		self.expert_demo = np.concatenate(self.expert_demo, axis=0)
+		# 		self.expert_rgb_obs = np.ascontiguousarray(np.transpose(self.expert_demo, (0,2,3,1))[:, :,:,:3].astype(np.uint8))
+		# 		self.expert_action = np.concatenate(self.expert_action, axis=0)
+				
+		# 		self.agent.save_representations(self.expert_demo, self.expert_action, 128, config=self.expert_demo_config)
+		
+		# Get start points
+		if self.cfg.random_start:
+			eval_starts = Path(self.cfg.eval_starts) / 'starts.pkl'
+			if eval_starts.exists():
+				with eval_starts.open('rb') as f:
+					self.start_pos = pickle.load(f)
+			else:
+				eval_starts = Path(self.cfg.eval_starts)
+				eval_starts.mkdir(parents=True, exist_ok=True)
+				
+				# Generate start points
+				self.start_pos = []
+				try:
+					for _ in range(self.cfg.num_eval):
+						self.start_pos.append(self.eval_env.get_random_pos())
+				except:
+					for _ in range(self.cfg.num_eval):
+						self.start_pos.append(self.eval_env.arm.get_random_pos())
+
+				# Save start points for the task
+				eval_starts = eval_starts / 'starts.pkl'
+				with eval_starts.open('wb') as f:
+					pickle.dump(self.start_pos, f)
+
+		time_step = self.eval_env.reset()
+		# replay_thread = None
+		while eval_until_episode(self.global_episode) and not self.exit_eval:
+			# self.video_recorder.init(self.eval_env, video_filename=f'{self.global_episode}_eval.mp4')
+			print(f"Starting episode {self.global_episode}")
+			time_step = self.eval_env.reset() #Leon: need to call reset twice in case objects are trapped
+			self.video_recorder.init(self.eval_env, video_filename=f'{self.global_episode}_eval.mp4')
+			# x = input("Press Enter to continue... after reseting env")
+			print("Press Enter to continue... after reseting env. To rate prev episode, press 0 for failure and 1 for success")
+			self.proceed_after_env_reset_event.clear() # clear the event flag
+			self.proceed_after_env_reset_event.wait() # blocking wait for the event flag to be set
+			if self.global_episode > 0:
+				self.logger.log_metrics({'num_success': self.num_episode_successes}, self.global_step, 'eval', episode=self.global_episode)
+				self.logger.log_metrics({'success_rate': self.num_episode_successes/self.global_episode}, self.global_step, 'eval', episode=self.global_episode)
+
+				# log confidence intervals for success rate
+				k = self.num_episode_successes # number of successes
+				n = self.global_episode # number of trials
+
+				table_columns = []
+				table_data = []
+				for alpha in self.alpha_range:
+					lb = bc.binom_ci(k, n, alpha, 'lb')
+					ub = bc.binom_ci(k, n, alpha, 'ub')
+
+					self.logger.log_metrics({f'success_rate_lb_{alpha}': lb}, self.global_step, 'eval', episode=self.global_episode)
+					self.logger.log_metrics({f'success_rate_ub_{alpha}': ub}, self.global_step, 'eval', episode=self.global_episode)
+
+			time_step = self.eval_env.reset()
+			# debug_info_dict = self.eval_env.debug_info_dict
+			# if replay_thread is not None:
+			# 	# wait for the last replay thread to finish
+			# 	replay_thread.join()
+
+			# self.eval_replay_storage.add(time_step._replace(observation=time_step.observation[self.cfg.obs_type]), debug_info_dict)
+			# replay_thread = threading.Thread(target=self.eval_replay_storage.add, args=(time_step._replace(observation=time_step.observation[self.cfg.obs_type]), debug_info_dict))
+			# replay_thread = threading.Thread(target=self.eval_replay_storage.add, args=(time_step, debug_info_dict))
+
+			# replay_thread.start()
+			if self.cfg.random_start:
+				time_step = self.reset(self.global_episode)
+				time.sleep(2) #5)
+			# if 'vinn' in repr(self.agent):
+			# 	self.agent.reset()
+			# 	# self.agent.buffer.reset()
+			# 	# if self.cfg.open_loop:
+			# 	# 	self.agent.current_step = 0
+			# if 'openloop' in repr(self.agent):
+			# 	self.agent.curr_step = 0
+			# at start of each episode, provide zero action for policies that use action history
+			# shape should be (T_o, T_a, action_dim)
+
+			# while not time_step.last() and not self.preempt_episode:
+			self.video_recorder.ros_start_recording()
+			self.eval_replay_storage.start_episode()
+			self.eval_env.start_policy_timer()
+			while not self.eval_env.episode_done() and not self.preempt_episode:
+				# with torch.no_grad(), utils.eval_mode(self.agent):
+				# 	# if self.cfg.agent.provide_topk:
+				# 	# 	action, vinn_action, topk = self.agent.act(
+				# 	# 		time_step.observation['pixels'],
+				# 	# 		self.global_step,
+				# 	# 		eval_mode=True)
+				# 	# elif self.cfg.agent.provide_obs:
+				# 	# 	action, vinn_action, obs = self.agent.act(
+				# 	# 		time_step.observation['pixels'],
+				# 	# 		self.global_step,
+				# 	# 		eval_mode=True)
+				# 	# else:
+				# 	action, vinn_action = self.agent.act(
+				# 		time_step.observation,
+				# 		self.global_step,
+				# 		eval_mode=True,
+				# 		obs_timestamp=time_step.observation['timestamp'],
+				# 		obs_seq=time_step.observation['seq'],
+				# 		action_history=action_history,
+				# 		action_history_start_timestamp=action_history_start_timestamp,
+				# 		)
+					# DONT WAIT FOR POLICY TO GET AN ACTION 
+					# we dont want to slow down grabbing obs and passing to sam/contact features
+
+				self.eval_env.run_policy_threads() # this just does a rospy sleep
+					
+				# if self.use_action_history:
+				# 	action_history_start_timestamp = time_step.observation['timestamp']
+				# 	# action_history = action[:self.cfg.agent.config.policy_cfg.action_history_encoder_config.history_length, ...]		
+				# 	# add n_obs_steps dimension to action_history, for now we assume n_obs_steps = 1
+				# 	# TODO: handle n_obs_steps > 1
+				# 	action_history = action[np.newaxis, ...]
+
+				# time_step = self.eval_env.step(action, vinn_action) # obs, reward after action has been taken
+				# debug_info_dict = self.eval_env.debug_info_dict
+				
+				# time_step = self.eval_env.ros_step()
+				
+				# replay_thread.join()
+
+				# time how long it takes to execute the step
+				# time_before_add = time.perf_counter()
+				# self.eval_replay_storage.add(time_step._replace(observation=time_step.observation[self.cfg.obs_type]), debug_info_dict)
+				# use thread to call the add function in a separate thread
+				# replay_thread = threading.Thread(target=self.eval_replay_storage.add, args=(time_step._replace(observation=time_step.observation[self.cfg.obs_type]), debug_info_dict))
+
+				# replay_thread = threading.Thread(target=self.eval_replay_storage.add, args=(time_step, debug_info_dict))
+				# replay_thread.start()
+				
+				# print(f"Time to add to replay buffer: {time.perf_counter() - time_before_add}")
+
+				# self.video_recorder.record(self.eval_env)
+				# self._global_step += 1
+			
+			self.eval_env.stop_policy_timer()
+			
+			if self.restart_episode:
+				# means we should delete the current episode and start again
+				self.restart_episode = False
+				self.eval_replay_storage.reset_current_episode()
+				self.video_recorder.reset_current_episode()
+
+			else:
+				self.eval_replay_storage.store_current_episode()
+				video_filepath = self.video_recorder.save()
+				self.logger.log_video(f"eval/{video_filepath.name.rstrip('.mp4')}", video_filepath, self.global_step)
+				self._global_episode += 1
+			
+			self.preempt_episode = False # reset preempt_episode flag
+		
+		# self.video_recorder.save(f'{episode}_eval.mp4')
+		# get the video file and convert to video tensor to log 
+
+		self.eval_env.reset()
+		
+		print("Evaluation finished. To wrap up, rate prev episode, press 0 for failure and 1 for success")
+		self.proceed_after_env_reset_event.clear() # clear the event flag
+		self.proceed_after_env_reset_event.wait() # blocking wait for the event flag to be set
+		if self.global_episode > 0:
+			# self.logger.log_metrics({'num_success': self.num_episode_successes}, self.global_step, 'eval', episode=self.global_episode)
+			self.logger.log_metrics({'num_success': self.num_episode_successes}, self.global_step, 'eval', episode=self.global_episode)
+			self.logger.log_metrics({'success_rate': self.num_episode_successes/self.global_episode}, self.global_step, 'eval', episode=self.global_episode)
+			
+			# log confidence intervals for success rate
+			k = self.num_episode_successes # number of successes
+			n = self.global_episode # number of trials
+
+			table_columns = ['success_rate']
+			table_data = [self.num_episode_successes/self.global_episode]
+			for alpha in self.alpha_range:
+				lb = bc.binom_ci(k, n, alpha, 'lb')
+				ub = bc.binom_ci(k, n, alpha, 'ub')
+
+				self.logger.log_metrics({f'success_rate_lb_{alpha}': lb}, self.global_step, 'eval', episode=self.global_episode)
+				self.logger.log_metrics({f'success_rate_ub_{alpha}': ub}, self.global_step, 'eval', episode=self.global_episode)
+
+				table_columns.extend([f'success_rate_lb_{alpha}', f'success_rate_ub_{alpha}'])
+				table_data.extend([lb, ub])
+
+			table_data = [table_data]
+
+			# seperately log as a table
+			wandb.log({
+				"eval/success_rate_ci": wandb.Table(data=table_data, columns=table_columns)
+			})
+
+			# also accumulate eval metrics across previous eval runs 
+			# TODO: change wandb init to resume from an existing run!!!
+			run_filter={
+				"jobType": "eval",
+				"config.wandb_run_id": self.cfg.wandb_run_id, 
+				"summary_metrics.episode": {"$gte": 5},
+				"config.checkpoint_epoch": self.cfg.checkpoint_epoch,
+				"state": "finished",
+				"config.grasped_obj_name": grasped_obj_name,
+				"config.left_book_slot": left_book_slot,
+			}
+
+			api = wandb.Api()
+			filtered_runs = api.runs("serialexperimentsleon/extrinsic_contact_downstream", filters=run_filter)
+			total_num_successes = self.num_episode_successes
+			total_num_episodes = self.global_episode
+			list_of_historical_run_ids = []
+			if len(filtered_runs) > 0:
+				for filtered_run in filtered_runs:
+					total_num_successes += filtered_run.summary_metrics['eval/num_success']
+					total_num_episodes += filtered_run.summary_metrics['episode']
+					list_of_historical_run_ids.append(filtered_run.id)
+
+			wandb.summary['total_num_successes'] = total_num_successes
+			wandb.summary['total_num_episodes'] = total_num_episodes
+			wandb.summary['total_success_rate'] = total_num_successes/total_num_episodes
+				
+			# log the accumulated metrics as a table
+			total_table_columns = ['total_num_successes', 'total_num_episodes', 'total_success_rate']
+			total_table_data = [total_num_successes, total_num_episodes, total_num_successes/total_num_episodes]
+			for alpha in self.alpha_range:
+				lb = bc.binom_ci(total_num_successes, total_num_episodes, alpha, 'lb')
+				ub = bc.binom_ci(total_num_successes, total_num_episodes, alpha, 'ub')
+				total_table_columns.extend([f'total_success_rate_lb_{alpha}', f'total_success_rate_ub_{alpha}'])
+				total_table_data.extend([lb, ub])
+				wandb.summary[f'total_success_rate_lb_{alpha}'] = lb
+				wandb.summary[f'total_success_rate_ub_{alpha}'] = ub
+				
+			total_table_data = [total_table_data]
+			wandb.log({
+				'eval/total_success_rate_ci': wandb.Table(data=total_table_data, columns=total_table_columns)
+			})
+
+		self.continue_keypress_thread = False # will stop the keypress thread
+		self.keypress_input_thread.join() # wait for the keypress thread to finish
+
+	def load_checkpoint_conf(self, snapshot_path):
+		config_path = snapshot_path.parent / 'config.yaml'
+		if not config_path.exists():
+			raise FileNotFoundError(f'No snapshot conf found at {config_path}')
+		else:
+			# load the omegaconf config
+			hydra.core.global_hydra.GlobalHydra.instance().clear()
+			hydra.initialize(
+				str(_relative_path_between(Path(config_path).absolute().parent, Path(__file__).absolute().parent)),
+			)
+			cfg = hydra.compose(Path(config_path).stem)
+			from deepdiff import DeepDiff
+			from omegaconf import open_dict
+			diff = DeepDiff(OmegaConf.to_container(cfg), OmegaConf.to_container(self.cfg)) # old, new
+			# import re
+			overwriteable_keys = [f"root{overwritable_key}" for overwritable_key in ["['use_wandb']", "['path_to_depth_extrinsics']", "['eval']", "['root_dir']", "['wandb_notes']", "['agent']['config']['train_cfg']['use_amp']", "['agent']['config']['compile']", "['agent']['config']['policy_cfg']['num_inference_steps']"]]
+			if "values_changed" in diff:
+				# top_k_checkpoints, wandb_notes, agent.config.train_cfg.use_amp, save_snapshot_every_epochs_diffusion, check_topk_every_epochs_diffusion, validate_diffusion_on_action_loss_every_epochs, train_eval_diffusion_on_action_loss_every_epochs, validate_every_epochs_diffusion
+				# for keys above, overwrite the old config with the new config
+				for k, v in diff['values_changed'].items():
+					# replace any keys that are under "root['suite']"
+					if k in overwriteable_keys or k.startswith("root['suite']"):
+						print(f"Found changed key {k} with value {v}. Overwriting old checkpoint config")
+						if k == "root['agent']['config']['compile']":
+							if diff['values_changed'][k]['new_value']:
+								self.loading_uncompiled_checkpoint_with_compile = True
+							elif not diff['values_changed'][k]['new_value']:
+								# raise ValueError("Cannot load a compiled checkpoint without compile")
+								self.loading_compiled_checkpoint_with_no_compile = True
+						exec(f"{k.replace('root[', 'cfg[')} = {k.replace('root[', 'self.cfg[')}")
+			# for any new values, update the old checkpoint config
+			if "dictionary_item_added" in diff:
+				for new_key in diff['dictionary_item_added']: # this is a list
+					# if new_key == "root['suite']['task_make_fn']['observation_cfg']":
+					if new_key == "root['suite']['task_make_fn']['agent_policy_cfg']":
+						# pass the agents observation_cfg to the suite task_make_fn
+						with open_dict(cfg): # to allow addition of non-existing keys
+							# cfg.suite.task_make_fn.observation_cfg = cfg.agent.config.observation_cfg
+							cfg.suite.task_make_fn.agent_policy_cfg = cfg.agent.config
+						continue
+					elif "['agent']['config']['policy_cfg']['input_shapes']" in new_key:
+						# skip adding the new key if it is the input_shapes of the policy_cfg
+						continue
+					else:
+						print(f"Found new key {new_key} with value {eval(new_key.replace('root[', 'self.cfg['))}. Adding to checkpoint config")
+					# eval(new_key.replace('root', 'cfg')) = eval(new_key.replace('root', 'self.cfg'))
+					if new_key == "root['agent']['config']['compile']":
+						if self.cfg.agent.config.compile:
+							self.loading_uncompiled_checkpoint_with_compile = True
+					
+					with open_dict(cfg):
+						exec(f"{new_key.replace('root[', 'cfg[')}={new_key.replace('root[', 'self.cfg[')}")
+			self.cfg = cfg
+
+	def load_checkpoint(self, snapshot_path, bc=False):
+		print(f'resuming {repr(self.agent)}: {snapshot_path}')
+		with snapshot_path.open('rb') as f:
+			payload = torch.load(f)
+		agent_payload = {}
+		for k, v in payload.items():
+			if k not in self.__dict__:
+				agent_payload[k] = v
+			elif k == '_global_epoch':
+				self._global_epoch = v
+				print(f'loaded epoch: {v}')
+				if self.cfg.use_wandb:
+					# add to config of wandb
+					wandb.config.update({'epoch': v})
+
+		# self.agent.load_snapshot_eval(agent_payload, bc)
+
+@hydra.main(config_path='cfgs', config_name='config_eval')
+def main(cfg):
+	from eval_robot import Workspace as W
+	root_dir = Path.cwd()
+	workspace = W(cfg)
+
+	workspace.eval()
+
+if __name__ == '__main__':
+	main()