Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add logging for h5 #119

Merged
merged 14 commits into from
Dec 3, 2024
Merged

Add logging for h5 #119

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
14 commits
Select commit Hold shift + click to select a range
762b12f
add krec saving
alik-git Dec 2, 2024
19c1cf4
update h5 logging to use HDF5Logger, same as sim2sim.py
alik-git Dec 2, 2024
4752353
add parallel envs and handle saving for h5
alik-git Dec 2, 2024
9780d6e
add parallel env saving for krec as well
alik-git Dec 2, 2024
8bb1f9a
remove plotting and remove onnx export
alik-git Dec 2, 2024
f3b7b34
update logic
budzianowski Dec 3, 2024
2216adf
Merge branch 'master' of github.com:kscalelabs/sim into data_gen_and_…
budzianowski Dec 3, 2024
8aab407
add script to compare h5 files from mujoco vs isaac to ensure consist…
alik-git Dec 3, 2024
38a464b
update numpy requirement to <2.0.0
alik-git Dec 3, 2024
9137d88
change pillow req to > 6.2.0
alik-git Dec 3, 2024
76a9c65
scipy version > 1.10.0
alik-git Dec 3, 2024
f1cbc9e
lint
alik-git Dec 3, 2024
c78f3f7
organise imports
alik-git Dec 3, 2024
c6237e9
Merge branch 'master' into data_gen_and_krec
alik-git Dec 3, 2024
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
217 changes: 158 additions & 59 deletions sim/play.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -9,24 +9,28 @@
import copy
import logging
import os
import time
import uuid
from datetime import datetime
from typing import Any, Union

import cv2
import h5py
import numpy as np
from isaacgym import gymapi
import torch # isort: skip
from tqdm import tqdm

logger = logging.getLogger(__name__)

import krec
from sim.env import run_dir # noqa: E402
from sim.envs import task_registry # noqa: E402
from sim.model_export import ActorCfg, convert_model_to_onnx # noqa: E402
from sim.utils.helpers import get_args # noqa: E402
from sim.utils.logger import Logger # noqa: E402

import torch # isort: skip
from sim.h5_logger import HDF5Logger

logger = logging.getLogger(__name__)


def export_policy_as_jit(actor_critic: Any, path: Union[str, os.PathLike]) -> None:
Expand All @@ -41,9 +45,10 @@ def play(args: argparse.Namespace) -> None:
logger.info("Configuring environment and training settings...")
env_cfg, train_cfg = task_registry.get_cfgs(name=args.task)

# override some parameters for testing
env_cfg.env.num_envs = min(env_cfg.env.num_envs, 1)
env_cfg.sim.max_gpu_contact_pairs = 2**10
num_parallel_envs = 4
env_cfg.env.num_envs = num_parallel_envs
env_cfg.sim.max_gpu_contact_pairs = 2**10 * num_parallel_envs

if args.trimesh:
env_cfg.terrain.mesh_type = "trimesh"
else:
Expand Down Expand Up @@ -78,49 +83,94 @@ def play(args: argparse.Namespace) -> None:
export_policy_as_jit(ppo_runner.alg.actor_critic, path)
print("Exported policy as jit script to: ", path)

# export policy as a onnx module (used to run it on web)
if args.export_onnx:
path = ppo_runner.alg.actor_critic
convert_model_to_onnx(path, ActorCfg(), save_path="policy.onnx")
print("Exported policy as onnx to: ", path)
# # export policy as a onnx module (used to run it on web)
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why commented?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I think its outdated / something that was changed/handled in kinfer, when I run it I get this. Maybe wesley knows more, but low priority for now imo.

  loaded_dict = torch.load(path)
Exported policy as jit script to:  .
Traceback (most recent call last):
  File "/opt/anaconda3/envs/ksl/lib/python3.8/site-packages/torch/serialization.py", line 564, in _check_seekable
    f.seek(f.tell())
  File "/opt/anaconda3/envs/ksl/lib/python3.8/site-packages/torch/nn/modules/module.py", line 1729, in __getattr__
    raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
AttributeError: 'ActorCritic' object has no attribute 'seek'

During handling of the above exception, another exception occurred:

Traceback (most recent call last):
  File "sim/play.py", line 332, in <module>
    play(args)
  File "sim/play.py", line 94, in play
    convert_model_to_onnx(path, ActorCfg(), save_path="policy.onnx")
  File "/home/kasm-user/ali_repos/sim/sim/model_export.py", line 220, in convert_model_to_onnx
    all_weights = torch.load(model_path, map_location="cpu", weights_only=True)
  File "/opt/anaconda3/envs/ksl/lib/python3.8/site-packages/torch/serialization.py", line 1065, in load
    with _open_file_like(f, 'rb') as opened_file:
  File "/opt/anaconda3/envs/ksl/lib/python3.8/site-packages/torch/serialization.py", line 473, in _open_file_like
    return _open_buffer_reader(name_or_buffer)
  File "/opt/anaconda3/envs/ksl/lib/python3.8/site-packages/torch/serialization.py", line 458, in __init__
    _check_seekable(buffer)
  File "/opt/anaconda3/envs/ksl/lib/python3.8/site-packages/torch/serialization.py", line 567, in _check_seekable
    raise_err_msg(["seek", "tell"], e)
  File "/opt/anaconda3/envs/ksl/lib/python3.8/site-packages/torch/serialization.py", line 560, in raise_err_msg
    raise type(e)(msg)
AttributeError: 'ActorCritic' object has no attribute 'seek'. You can only torch.load from a file that is seekable. Please pre-load the data into a buffer like io.BytesIO and try to load from it instead.

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

lets remove it then

# if args.export_onnx:
# path = ppo_runner.alg.actor_critic
# convert_model_to_onnx(path, ActorCfg(), save_path="policy.onnx")
# print("Exported policy as onnx to: ", path)

# Prepare for logging
env_logger = Logger(env.dt)
robot_index = 0
joint_index = 1
stop_state_log = 1000
env_steps_to_run = 1000
now = datetime.now().strftime("%Y-%m-%d_%H-%M-%S")
if args.log_h5:
h5_file = h5py.File(f"data{now}.h5", "w")

# Create dataset for actions
max_timesteps = stop_state_log
num_dof = env.num_dof
dset_actions = h5_file.create_dataset("actions", (max_timesteps, num_dof), dtype=np.float32)

# Create dataset of observations
buf_len = len(env.obs_history) # length of observation buffer
dset_2D_command = h5_file.create_dataset(
"observations/2D_command", (max_timesteps, buf_len, 2), dtype=np.float32
) # sin and cos commands
dset_3D_command = h5_file.create_dataset(
"observations/3D_command", (max_timesteps, buf_len, 3), dtype=np.float32
) # x, y, yaw commands
dset_q = h5_file.create_dataset(
"observations/q", (max_timesteps, buf_len, num_dof), dtype=np.float32
) # joint positions
dset_dq = h5_file.create_dataset(
"observations/dq", (max_timesteps, buf_len, num_dof), dtype=np.float32
) # joint velocities
dset_obs_actions = h5_file.create_dataset(
"observations/actions", (max_timesteps, buf_len, num_dof), dtype=np.float32
) # actions
dset_ang_vel = h5_file.create_dataset(
"observations/ang_vel", (max_timesteps, buf_len, 3), dtype=np.float32
) # root angular velocity
dset_euler = h5_file.create_dataset(
"observations/euler", (max_timesteps, buf_len, 3), dtype=np.float32
) # root orientation
# Create directory for HDF5 files
h5_dir = run_dir() / "h5_out" / args.task / now
h5_dir.mkdir(parents=True, exist_ok=True)

# Get observation dimensions
num_joints = env.num_dof
obs_buffer = env.obs_history[0].tolist()[0]
obs_size = len(obs_buffer)

# Index mappings for observation buffer
# This is based on stompypro_env.py
# https://github.com/kscalelabs/sim/blob/54c40d55eab15a9e784e89fb47e64a668851a41b/sim/envs/humanoids/stompypro_env.py#L225
command_2d_start = 0
command_2d_end = 2
command_3d_start = 2
command_3d_end = 5
joint_pos_start = 5
joint_pos_end = joint_pos_start + num_joints
joint_vel_start = joint_pos_end
joint_vel_end = joint_vel_start + num_joints
prev_actions_start = joint_vel_end
prev_actions_end = prev_actions_start + num_joints
ang_vel_start = prev_actions_end
ang_vel_end = ang_vel_start + 3
euler_start = ang_vel_end
euler_end = euler_start + 3

h5_loggers = []
for env_idx in range(env_cfg.env.num_envs):
h5_dir = run_dir() / "h5_out" / args.task / now / f"env_{env_idx}"
h5_dir.mkdir(parents=True, exist_ok=True)

h5_loggers.append(HDF5Logger(
data_name=f"{args.task}_env_{env_idx}",
num_actions=num_joints,
max_timesteps=env_steps_to_run,
num_observations=obs_size,
h5_out_dir=str(h5_dir)
))


if args.log_krec:
# Create directory for KRec files
krec_dir = run_dir() / "krec_out" / args.task / now
krec_dir.mkdir(parents=True, exist_ok=True)

start_time_ns = time.time_ns() # Current time in nanoseconds

krec_loggers = []
for env_idx in range(env_cfg.env.num_envs):

# Create KRec header for each environment
header = krec.KRecHeader(
uuid=str(uuid.uuid4()),
robot_platform=f"{args.task}-sim",
robot_serial="123",
task=args.task,
start_timestamp=start_time_ns,
end_timestamp=start_time_ns + int(env_steps_to_run * env_cfg.sim.dt * 1e9),
)

# Add actuator configs for each joint
for i in range(env.num_dof):
actuator_config = krec.ActuatorConfig(
i, # actuator id
kp=float(env.p_gains[env_idx, i].cpu()), # use env_idx instead of 0
ki=0.0,
kd=float(env.d_gains[env_idx, i].cpu()), # use env_idx instead of 0
max_torque=float(env.torque_limits[i].cpu()),
name=f"Joint{i}",
)
header.add_actuator_config(actuator_config)

# Create KRec object for this environment
krec_loggers.append(krec.KRec(header))

if args.render:
camera_properties = gymapi.CameraProperties()
Expand Down Expand Up @@ -151,17 +201,32 @@ def play(args: argparse.Namespace) -> None:
os.mkdir(experiment_dir)
video = cv2.VideoWriter(dir, fourcc, 50.0, (1920, 1080))

for t in tqdm(range(stop_state_log)):
for t in tqdm(range(env_steps_to_run)):
actions = policy(obs.detach())
if args.log_h5:
dset_actions[t] = actions.detach().numpy()
# Extract the current observation
for env_idx in range(env_cfg.env.num_envs):
cur_obs = env.obs_history[env_idx].tolist()[0]

h5_loggers[env_idx].log_data({
"t": np.array([t * env.dt], dtype=np.float32),
"2D_command": np.array(cur_obs[command_2d_start:command_2d_end], dtype=np.float32),
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you check if current and below are the same values?

                    "t": np.array([count_lowlevel * cfg.dt], dtype=np.float32),
                    "2D_command": np.array(
                        [
                            np.sin(2 * math.pi * count_lowlevel * cfg.dt / cfg.cycle_time),
                            np.cos(2 * math.pi * count_lowlevel * cfg.dt / cfg.cycle_time),
                        ],
                        dtype=np.float32,
                    ),
                    "3D_command": np.array([x_vel_cmd, y_vel_cmd, yaw_vel_cmd], dtype=np.float32),
                    ```

"3D_command": np.array(cur_obs[command_3d_start:command_3d_end], dtype=np.float32),
"joint_pos": np.array(cur_obs[joint_pos_start:joint_pos_end], dtype=np.float32),
"joint_vel": np.array(cur_obs[joint_vel_start:joint_vel_end], dtype=np.float32),
"prev_actions": np.array(cur_obs[prev_actions_start:prev_actions_end], dtype=np.float32),
"curr_actions": actions.detach().cpu().numpy()[0],
"ang_vel": np.array(cur_obs[ang_vel_start:ang_vel_end], dtype=np.float32),
"euler_rotation": np.array(cur_obs[euler_start:euler_end], dtype=np.float32),
"buffer": np.array(cur_obs, dtype=np.float32)
})

if args.fix_command:
env.commands[:, 0] = 0.5
env.commands[:, 1] = 0.0
env.commands[:, 2] = 0.0
env.commands[:, 3] = 0.0
obs, critic_obs, rews, dones, infos = env.step(actions.detach())
print(f"IMU: {obs[0, (3 * env.num_actions + 5) + 3 : (3 * env.num_actions + 5) + 2 * 3]}")

if args.render:
env.gym.fetch_results(env.sim, True)
Expand All @@ -187,17 +252,6 @@ def play(args: argparse.Namespace) -> None:
base_vel_yaw = env.base_ang_vel[robot_index, 2].item()
contact_forces_z = env.contact_forces[robot_index, env.feet_indices, 2].cpu().numpy()

if args.log_h5:
for i in range(buf_len):
cur_obs = env.obs_history[i].tolist()[0]
dset_2D_command[t, i] = cur_obs[0:2] # sin and cos commands
dset_3D_command[t, i] = cur_obs[2:5] # x, y, yaw commands
dset_q[t, i] = cur_obs[5 : 5 + num_dof] # joint positions
dset_dq[t, i] = cur_obs[5 + num_dof : 5 + 2 * num_dof] # joint velocities
dset_obs_actions[t, i] = cur_obs[5 + 2 * num_dof : 5 + 3 * num_dof] # actions
dset_ang_vel[t, i] = cur_obs[5 + 3 * num_dof : 8 + 3 * num_dof] # root angular velocity
dset_euler[t, i] = cur_obs[8 + 3 * num_dof : 11 + 3 * num_dof] # root orientation

env_logger.log_states(
{
"dof_pos_target": dof_pos_target,
Expand All @@ -219,15 +273,60 @@ def play(args: argparse.Namespace) -> None:
if num_episodes > 0:
env_logger.log_rewards(infos["episode"], num_episodes)

if args.log_krec:
# Log data for each environment
for env_idx in range(env_cfg.env.num_envs):
frame = krec.KRecFrame(
video_timestamp=start_time_ns + int(t * env.dt * 1e9),
frame_number=t,
inference_step=t // env_cfg.control.decimation,
)

# Add actuator states and commands for each joint
for i in range(env.num_dof):
state = krec.ActuatorState(
actuator_id=i,
online=True,
position=env.dof_pos[env_idx, i].item(),
velocity=env.dof_vel[env_idx, i].item(),
torque=env.torques[env_idx, i].item(),
)
command = krec.ActuatorCommand(
i, # actuator id
position=actions[env_idx, i].item(),
velocity=0.0,
torque=actions[env_idx, i].item(),
)
frame.add_actuator_state(state)
frame.add_actuator_command(command)

# Add IMU data
imu_values = krec.IMUValues(
gyro=krec.Vec3(x=obs[env_idx, 0], y=obs[env_idx, 1], z=obs[env_idx, 2]),
quaternion=krec.IMUQuaternion(x=obs[env_idx, 3], y=obs[env_idx, 4], z=obs[env_idx, 5], w=obs[env_idx, 6]),
)
frame.set_imu_values(imu_values)
krec_loggers[env_idx].add_frame(frame)

env_logger.print_rewards()
env_logger.plot_states()
# env_logger.plot_states()
budzianowski marked this conversation as resolved.
Show resolved Hide resolved

if args.render:
video.release()

if args.log_h5:
print("Saving data to " + os.path.abspath(f"data{now}.h5"))
h5_file.close()
# print(f"Saving HDF5 file to {h5_logger.h5_file_path}") # TODO use code from kdatagen
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

proper todo - remove print ?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yes proper TODO, will remove as soon as its done.

for h5_logger in h5_loggers:
h5_logger.close()
print(f"HDF5 file(s) saved!")

if args.log_krec:
for env_idx, krec_logger in enumerate(krec_loggers):
krec_file_path = krec_dir / f"env_{env_idx}" / f"walking_{str(uuid.uuid4())[:8]}.krec"
krec_file_path.parent.mkdir(parents=True, exist_ok=True)
print(f"Saving KRec file to {krec_file_path}")
krec_logger.save(str(krec_file_path))
print("KRec files saved!")


if __name__ == "__main__":
Expand Down
6 changes: 6 additions & 0 deletions sim/utils/helpers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -232,6 +232,12 @@ def get_args() -> argparse.Namespace:
"default": False,
"help": "Enable HDF5 logging",
},
{
"name": "--log_krec",
"action": "store_true",
"default": False,
"help": "Enable KRec logging",
},
{
"name": "--trimesh",
"action": "store_true",
Expand Down
Loading