def main():
start_time = time.time()
config = load_yaml(os.path.join(get_project_root(), 'experiments/drake_pusher_slider/env_config.yaml'))
config['dataset']['num_episodes'] = 1000 # half for train, half for valid
set_seed(500) # just randomly chosen
num_episodes = config['dataset']['num_episodes']
DATASET_NAME = "box_push_%d" %(num_episodes)
OUTPUT_DIR = os.path.join(get_data_ssd_root(), 'dataset', DATASET_NAME)
if not os.path.exists(OUTPUT_DIR):
os.makedirs(OUTPUT_DIR)
collect_episodes(
config,
output_dir=OUTPUT_DIR,
visualize=False,
debug=False)
elapsed = time.time() - start_time
print("Generating and saving dataset to disk took %d seconds" % (int(elapsed)))
def main():
start_time = time.time()
config = load_yaml(
os.path.join(get_project_root(),
'experiments/exp_20_mugs/config.yaml'))
config['dataset']['num_episodes'] = 10
set_seed(500) # just randomly chosen
DATASET_NAME = "mugs_%d" % (config['dataset']['num_episodes'])
OUTPUT_DIR = os.path.join(get_data_root(), 'sandbox', DATASET_NAME)
print("OUTPUT_DIR:", OUTPUT_DIR)
collect_episodes(config,
output_dir=OUTPUT_DIR,
visualize=False,
debug=False)
elapsed = time.time() - start_time
print("Generating and saving dataset to disk took %d seconds" %
(int(elapsed)))
def main():
start_time = time.time()
config = load_yaml(
os.path.join(get_project_root(),
'experiments/exp_18_box_on_side/config.yaml'))
# config['dataset']['num_episodes'] = 500 # half for train, half for valid
config['dataset']['num_episodes'] = 600 # half for train, half for valid
set_seed(500) # just randomly chosen
DATASET_NAME = "dps_box_on_side_%d" % (config['dataset']['num_episodes'])
OUTPUT_DIR = os.path.join(get_data_root(), "dev/experiments/18/data",
DATASET_NAME)
print("OUTPUT_DIR:", OUTPUT_DIR)
collect_episodes(config,
output_dir=OUTPUT_DIR,
visualize=False,
debug=False)
elapsed = time.time() - start_time
print("Generating and saving dataset to disk took %d seconds" %
(int(elapsed)))
def evaluate_mpc(
model_dy, # dynamics model
env, # the environment
episode, # OnlineEpisodeReader
mpc_input_builder, # DynamicsModelInputBuilder
planner, # RandomShooting planner
eval_indices=None,
goal_func=None, # function that gets goal from observation
config=None,
wait_for_user_input=False,
save_dir=None,
model_name="",
experiment_name="",
generate_initial_condition_func=None,
# (optional) function to generate initial condition, takes episode length N as parameter
):
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# must specify initial condition distribution
assert generate_initial_condition_func is not None
save_yaml(config, os.path.join(save_dir, 'config.yaml'))
writer = SummaryWriter(log_dir=save_dir)
pandas_data_list = []
for episode_length in config['eval']['episode_length']:
counter = 0
seed = 0
while counter < config['eval']['num_episodes']:
start_time = time.time()
seed += 1
set_seed(seed) # make it repeatable
# initial_cond = generate_initial_condition(config, N=episode_length)
initial_cond = generate_initial_condition_func(N=episode_length)
env.set_initial_condition_from_dict(initial_cond)
action_sequence_np = torch_utils.cast_to_numpy(
initial_cond['action_sequence'])
episode_data = mpc_single_episode(
model_dy=model_dy,
env=env,
action_sequence=action_sequence_np,
action_zero=np.zeros(2),
episode=episode,
mpc_input_builder=mpc_input_builder,
planner=planner,
eval_indices=eval_indices,
goal_func=goal_func,
config=config,
wait_for_user_input=wait_for_user_input,
)
# continue if invalid
if not episode_data['valid']:
print("invalid episode, skipping")
continue
pose_error = compute_pose_error(
obs=episode_data['obs_mpc_final'],
obs_goal=episode_data['obs_goal'],
)
object_delta = compute_pose_error(
obs=episode_data['obs_init'],
obs_goal=episode_data['obs_goal'])
print("object_delta\n", object_delta)
if wait_for_user_input:
print("pose_error\n", pose_error)
pandas_data = {
'episode_length': episode_length,
'seed': counter,
'model_name': model_name,
'experiment_name': experiment_name,
'object_pos_delta': object_delta['position_error'],
'object_angle_delta': object_delta['angle_error'],
'object_angle_delta_degrees':
object_delta['angle_error_degrees'],
}
pandas_data.update(pose_error)
pandas_data_list.append(pandas_data)
# log to tensorboard
for key, val in pose_error.items():
plot_name = "%s/episode_len_%d" % (key, episode_length)
writer.add_scalar(plot_name, val, counter)
writer.flush()
print("episode [%d/%d], episode_length %d, duration %.2f" %
(counter, config['eval']['num_episodes'], episode_length,
time.time() - start_time))
counter += 1
df_tmp = pd.DataFrame(pandas_data_list)
keys = ["angle_error_degrees", "position_error"]
for key in keys:
for i in range(10):
mean = df_tmp[key][df_tmp.episode_length ==
episode_length].mean()
median = df_tmp[key][df_tmp.episode_length ==
episode_length].median()
plot_name_mean = "mean/%s/episode_len_%d" % (key,
episode_length)
writer.add_scalar(plot_name_mean, mean, i)
plot_name_median = "median/%s/episode_len_%d" % (
key, episode_length)
writer.add_scalar(plot_name_median, median, i)
# save some data
df = pd.DataFrame(pandas_data_list)
df.to_csv(os.path.join(save_dir, "data.csv"))
def eval_dynamics(
config,
eval_dir, # str: directory to save output
multi_episode_dict=None,
n_rollout_list=None,
model_dy=None, # should already be in eval mode
phase_list=None, # typically it's
num_epochs=10,
):
assert n_rollout_list is not None
assert model_dy is not None
assert multi_episode_dict is not None
if phase_list is None:
phase_list = ["valid"]
# set random seed for reproduction
set_seed(config['train']['random_seed'])
tensorboard_dir = os.path.join(eval_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# save the config
save_yaml(config, os.path.join(eval_dir, "config.yaml"))
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
use_gpu = torch.cuda.is_available()
best_valid_loss = np.inf
global_iteration = 0
counters = {'train': 0, 'valid': 0}
epoch_counter_external = 0
stats = dict()
for n_rollout in n_rollout_list:
stats[n_rollout] = dict()
config_tmp = copy.copy(config)
config_tmp['train']['n_rollout'] = n_rollout
for phase in phase_list:
stats[n_rollout][phase] = dict()
print("Loading data for %s" % phase)
dataset = MultiEpisodeDataset(
config_tmp,
action_function=action_function,
observation_function=observation_function,
episodes=multi_episode_dict,
phase=phase)
dataloader = DataLoader(dataset,
batch_size=config['train']['batch_size'],
shuffle=True,
num_workers=config['train']['num_workers'],
drop_last=True)
loss_tensor_container = {"l2_avg": [], "l2_final_step": []}
step_duration_meter = AverageMeter()
global_iteration = 0
for epoch in range(num_epochs):
for i, data in enumerate(dataloader):
loss_container = dict() # store the losses for this step
# types of losses ["l2_avg", "l2_final_step"]
step_start_time = time.time()
global_iteration += 1
counters[phase] += 1
with torch.no_grad():
n_his = config['train']['n_history']
n_roll = n_rollout
n_samples = n_his + n_roll
if DEBUG:
print("global iteration: %d" % (global_iteration))
print("n_samples", n_samples)
# [B, n_samples, obs_dim]
observations = data['observations']
# [B, n_samples, action_dim]
actions = data['actions']
B = actions.shape[0]
if use_gpu:
observations = observations.cuda()
actions = actions.cuda()
# states, actions = data
assert actions.shape[1] == n_samples
loss_mse = 0.
# we don't have any visual observations, so states are observations
states = observations
# state_cur: B x n_his x state_dim
# state_cur = states[:, :n_his]
# [B, n_his, state_dim]
state_init = states[:, :n_his]
# We want to rollout n_roll steps
# actions = [B, n_his + n_roll, -1]
# so we want action_seq.shape = [B, n_roll, -1]
action_start_idx = 0
action_end_idx = n_his + n_roll - 1
action_seq = actions[:, action_start_idx:
action_end_idx, :]
if DEBUG:
print("states.shape", states.shape)
print("state_init.shape", state_init.shape)
print("actions.shape", actions.shape)
print("action_seq.shape", action_seq.shape)
# try using models_dy.rollout_model instead of doing this manually
rollout_data = rollout_model(state_init=state_init,
action_seq=action_seq,
dynamics_net=model_dy,
compute_debug_data=False)
# [B, n_roll, state_dim]
state_rollout_pred = rollout_data['state_pred']
# [B, n_roll, state_dim]
state_rollout_gt = states[:, n_his:]
if DEBUG:
print("state_rollout_gt.shape",
state_rollout_gt.shape)
print("state_rollout_pred.shape",
state_rollout_pred.shape)
# the loss function is between
# [B, n_roll, state_dim]
state_pred_err = state_rollout_pred - state_rollout_gt
# [B]
l2_avg_tensor = torch.mean(torch.norm(state_pred_err,
dim=-1),
dim=1).detach().cpu()
l2_avg = l2_avg_tensor.mean()
# [B]
l2_final_step_tensor = torch.norm(
state_pred_err[:, -1], dim=-1).detach().cpu()
l2_final_step = l2_final_step_tensor.mean()
loss_tensor_container["l2_avg"].append(l2_avg_tensor)
loss_container["l2_avg"] = l2_avg
loss_tensor_container["l2_final_step"].append(
l2_final_step_tensor)
loss_container["l2_final_step"] = l2_final_step
step_duration_meter.update(time.time() - step_start_time)
if (i % config['train']['log_per_iter']
== 0) or (global_iteration %
config['train']['log_per_iter'] == 0):
# print some logging information
log = ""
log += ', step time %.6f' % (step_duration_meter.avg)
# log data to tensorboard
for loss_type, loss_obj in loss_container.items():
plot_name = "%s/n_roll_%s/%s" % (loss_type, n_roll,
phase)
writer.add_scalar(plot_name, loss_obj.item(),
global_iteration)
log += " %s: %.6f," % (plot_name, loss_obj.item())
print(log)
writer.flush() # flush SummaryWriter events to disk
stats[n_rollout][phase] = dict()
for loss_type in loss_tensor_container:
t = torch.cat(loss_tensor_container[loss_type])
mean = t.mean()
median = t.median()
std = t.std()
stats[n_rollout][phase][loss_type] = {
'mean': mean,
'median': median,
'std': std
}
for stat_type, val in stats[n_rollout][phase][loss_type].items(
):
plot_name = "stats/%s/n_roll_%d/%s/%s" % (
loss_type, n_roll, phase, stat_type)
for idx_tmp in [0, 10, 100]:
writer.add_scalar(plot_name, val, idx_tmp)
def main():
d = load_model_and_data()
model_dy = d['model_dy']
dataset = d['dataset']
config = d['config']
multi_episode_dict = d['multi_episode_dict']
planner = d['planner']
planner_config = planner.config
idx_dict = get_object_and_robot_state_indices(config)
object_indices = idx_dict['object_indices']
robot_indices = idx_dict['robot_indices']
n_his = config['train']['n_history']
# save_dir = os.path.join(get_project_root(), 'sandbox/mpc/', get_current_YYYY_MM_DD_hh_mm_ss_ms())
save_dir = os.path.join(get_project_root(),
'sandbox/mpc/push_right_box_horizontal')
print("save_dir", save_dir)
if not os.path.exists(save_dir):
os.makedirs(save_dir)
# rotate
# episode_names = dataset.get_episode_names()
# print("len(episode_names)", len(episode_names))
# episode_name = episode_names[0]
# start_idx = 1
# n_roll = 15
# # straight + rotate
# episode_name = "2020-06-29-21-04-16"
# print('episode_name', episode_name)
# start_idx = 1
# n_roll = 15
# this is a nice straight push . . .
# push with box in horizontal position
episode_name = "2020-06-29-22-03-45"
start_idx = 2
n_roll = 10
# # validation set episodes
# episode_names = dataset.get_episode_names()
# print("len(episode_names)", len(episode_names))
# episode_name = episode_names[1]
# start_idx = 2
# n_roll = 15
camera_name = "d415_01"
episode = multi_episode_dict[episode_name]
print("episode_name", episode_name)
vis = meshcat_utils.make_default_visualizer_object()
vis.delete()
idx_list = list(range(start_idx, start_idx + n_roll + 1))
idx_list_GT = idx_list
goal_idx = idx_list[-1]
print("idx_list", idx_list)
# visualize ground truth rollout
if True:
for display_idx, episode_idx in enumerate(idx_list):
visualize_episode_data_single_timestep(
vis=vis,
dataset=dataset,
episode=episode,
camera_name=camera_name,
episode_idx=episode_idx,
display_idx=episode_idx,
)
data_goal = dataset._getitem(episode,
goal_idx,
rollout_length=1,
n_history=1)
states_goal = data_goal['observations_combined'][0]
z_states_goal = model_dy.compute_z_state(states_goal)['z']
print("states_goal.shape", states_goal.shape)
print("z_states_goal.shape", z_states_goal.shape)
##### VISUALIZE PREDICTED ROLLOUT ##########
data = dataset._getitem(episode, start_idx, rollout_length=n_roll)
states = data['observations_combined'].unsqueeze(0)
z = model_dy.compute_z_state(states)['z']
actions = data['actions'].unsqueeze(0)
idx_range_model_dy_input = data['idx_range']
print("data.keys()", data.keys())
print("data['idx_range']", data['idx_range'])
# z_init
z_init = z[:, :n_his]
# actions_init
action_start_idx = 0
action_end_idx = n_his + n_roll - 1
action_seq = actions[:, action_start_idx:action_end_idx]
print("action_seq GT\n", action_seq)
with torch.no_grad():
rollout_data = rollout_model(state_init=z_init.cuda(),
action_seq=action_seq.cuda(),
dynamics_net=model_dy,
compute_debug_data=False)
# [B, n_roll, state_dim]
# state_rollout_pred = rollout_data['state_pred']
z_rollout_pred = rollout_data['state_pred'].squeeze()
print("z_rollout_pred.shape", z_rollout_pred.shape)
if True:
for idx in range(len(z_rollout_pred)):
display_idx = data['idx_range'][idx + n_his]
visualize_model_prediction_single_timestep(
vis,
config,
z_pred=z_rollout_pred[idx],
display_idx=display_idx)
print("z_rollout_pred.shape", z_rollout_pred.shape)
# compute loss when rolled out using GT action sequence
eval_indices = object_indices
obs_goal = z_states_goal[object_indices].cuda()
reward_data = planner_utils.evaluate_model_rollout(
state_pred=rollout_data['state_pred'],
obs_goal=obs_goal,
eval_indices=eval_indices,
terminal_cost_only=planner_config['mpc']['mppi']['terminal_cost_only'],
p=planner_config['mpc']['mppi']['cost_norm'])
print("reward_data using action_seq_GT\n", reward_data['reward'])
##### MPC ##########
data = dataset._getitem(episode,
start_idx,
rollout_length=0,
n_history=config['train']['n_history'])
state_cur = data['observations_combined'].cuda()
z_state_cur = model_dy.compute_z_state(state_cur)['z']
action_his = data['actions'][:(n_his - 1)].cuda()
print("z_state_cur.shape", state_cur.shape)
print("action_his.shape", action_his.shape)
# don't seed with nominal actions just yet
action_seq_rollout_init = None
set_seed(SEED)
mpc_out = planner.trajectory_optimization(
state_cur=z_state_cur,
action_his=action_his,
obs_goal=obs_goal,
model_dy=model_dy,
action_seq_rollout_init=action_seq_rollout_init,
n_look_ahead=n_roll,
eval_indices=object_indices,
rollout_best_action_sequence=True,
verbose=True,
add_grid_action_samples=True,
)
print("\n\n------MPC output-------\n\n")
print("action_seq:\n", mpc_out['action_seq'])
mpc_state_pred = mpc_out['state_pred']
# current shape is [n_roll + 1, state_dim] but really should be
# [n_roll, state_dim] . . . something is up
print("mpc_state_pred.shape", mpc_state_pred.shape)
print("mpc_out['action_seq'].shape", mpc_out['action_seq'].shape)
print("n_roll", n_roll)
# visualize
for idx in range(n_roll):
episode_idx = start_idx + idx + 1
visualize_model_prediction_single_timestep(vis,
config,
z_pred=mpc_state_pred[idx],
display_idx=episode_idx,
name_prefix="mpc",
color=[255, 0, 0])
######## MPC w/ dynamics model input builder #############
print("\n\n-----DynamicsModelInputBuilder-----")
# dynamics model input builder
online_episode = OnlineEpisodeReader(no_copy=True)
ref_descriptors = d['spatial_descriptor_data']['spatial_descriptors']
ref_descriptors = torch_utils.cast_to_torch(ref_descriptors).cuda()
K_matrix = episode.image_episode.camera_K_matrix(camera_name)
T_world_camera = episode.image_episode.camera_pose(camera_name, 0)
visual_observation_function = \
VisualObservationFunctionFactory.descriptor_keypoints_3D(config=config,
camera_name=camera_name,
model_dd=d['model_dd'],
ref_descriptors=ref_descriptors,
K_matrix=K_matrix,
T_world_camera=T_world_camera,
)
input_builder = DynamicsModelInputBuilder(
observation_function=d['observation_function'],
visual_observation_function=visual_observation_function,
action_function=d['action_function'],
episode=online_episode)
compute_control_action_msg = dict()
compute_control_action_msg['type'] = "COMPUTE_CONTROL_ACTION"
for i in range(n_his):
episode_idx = idx_range_model_dy_input[i]
print("episode_idx", episode_idx)
# add image information to
data = add_images_to_episode_data(episode, episode_idx, camera_name)
online_episode.add_data(copy.deepcopy(data))
compute_control_action_msg['data'] = data
# hack for seeing how much the history matters .. .
# online_episode.add_data(copy.deepcopy(data))
# save informatin for running zmq controller
save_pickle(compute_control_action_msg,
os.path.join(save_dir, 'compute_control_action_msg.p'))
goal_idx = idx_list_GT[-1]
goal_data = add_images_to_episode_data(episode, goal_idx, camera_name)
goal_data['observations']['timestamp_system'] = time.time()
plan_msg = {
'type': "PLAN",
'data': [goal_data],
'n_roll': n_roll,
'K_matrix': K_matrix,
'T_world_camera': T_world_camera,
}
save_pickle(plan_msg, os.path.join(save_dir, "plan_msg.p"))
print("len(online_episode)", len(online_episode))
# use this to construct input
# verify it's the same as what we got from using the dataset directly
idx = online_episode.get_latest_idx()
mpc_input_data = input_builder.get_dynamics_model_input(idx,
n_history=n_his)
# print("mpc_input_data\n", mpc_input_data)
state_cur_ib = mpc_input_data['states'].cuda()
action_his_ib = mpc_input_data['actions'].cuda()
z_state_cur_ib = model_dy.compute_z_state(state_cur_ib)['z']
set_seed(SEED)
mpc_out = planner.trajectory_optimization(
state_cur=z_state_cur_ib,
action_his=action_his_ib,
obs_goal=obs_goal,
model_dy=model_dy,
action_seq_rollout_init=None,
n_look_ahead=n_roll,
eval_indices=object_indices,
rollout_best_action_sequence=True,
verbose=True,
add_grid_action_samples=True,
)
# visualize
for idx in range(n_roll):
episode_idx = start_idx + idx + 1
visualize_model_prediction_single_timestep(
vis,
config,
z_pred=mpc_out['state_pred'][idx],
display_idx=episode_idx,
name_prefix="mpc_input_builder",
color=[255, 255, 0])
def train_explore_and_learn(
config,
train_dir, # str: directory to save output
data_dir,
visualize=False):
# set random seed for reproduction
set_seed(config['train_explore_and_learn']['random_seed'])
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# save the config
save_yaml(config, os.path.join(train_dir, "config.yaml"))
print(config)
num_exploration_rounds = config['train_explore_and_learn'][
'num_exploration_rounds']
num_episodes_per_exploration_round = config['train_explore_and_learn'][
'num_episodes_per_exploration_round']
num_timesteps = config['train_explore_and_learn']['num_timesteps']
# model_folder = os.path.join(train_dir, "../2020-04-05-23-00-30-887903")
# model_file = os.path.join(model_folder, "net_best_dy_model.pth")
# model_dy = torch.load(model_file)
model_dy = None
global_iteration = 0
##### setup to store the dataset
metadata = dict()
metadata['episodes'] = dict()
# data collector
data_collector = DrakePusherSliderEpisodeCollector(config)
##### explore and learn
for idx_exploration_round in range(num_exploration_rounds):
print("Exploration round %d / %d" %
(idx_exploration_round, num_exploration_rounds))
### exploration
if idx_exploration_round == 0:
# initial exploration
exploration_type = 'random'
else:
exploration_type = 'mppi'
collect_episodes(
config,
metadata,
data_collector,
num_episodes_per_exploration_round,
data_dir,
visualize,
exploration_type,
model_dy=None if exploration_type == 'random' else model_dy)
save_yaml(metadata, os.path.join(data_dir, 'metadata.yaml'))
### optimize the dynamics model
model_dy, global_iteration = train_dynamics(config, train_dir,
data_dir, model_dy,
global_iteration, writer)
def train_dynamics(
config,
train_dir, # str: directory to save output
):
# set random seed for reproduction
set_seed(config['train']['random_seed'])
st_epoch = config['train'][
'resume_epoch'] if config['train']['resume_epoch'] > 0 else 0
tee = Tee(os.path.join(train_dir, 'train_st_epoch_%d.log' % st_epoch), 'w')
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# save the config
save_yaml(config, os.path.join(train_dir, "config.yaml"))
print(config)
# load the data
episodes = load_episodes_from_config(config)
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(
config,
action_function=action_function,
observation_function=observation_function,
episodes=episodes,
phase=phase)
dataloaders[phase] = DataLoader(
datasets[phase],
batch_size=config['train']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train']['num_workers'])
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
# compute normalization parameters if not starting from pre-trained network . . .
'''
define model for dynamics prediction
'''
model_dy = None
if config['train']['resume_epoch'] >= 0:
# if resume from a pretrained checkpoint
state_dict_path = os.path.join(
train_dir, 'net_dy_epoch_%d_iter_%d_state_dict.pth' %
(config['train']['resume_epoch'], config['train']['resume_iter']))
print("Loading saved ckp from %s" % state_dict_path)
# why is this needed if we already do torch.load???
model_dy.load_state_dict(torch.load(state_dict_path))
# don't we also need to load optimizer state from pre-trained???
else:
# not starting from pre-trained create the network and compute the
# normalization parameters
model_dy = DynaNetMLP(config)
# compute normalization params
stats = datasets["train"].compute_dataset_statistics()
obs_mean = stats['observations']['mean']
obs_std = stats['observations']['std']
observations_normalizer = DataNormalizer(obs_mean, obs_std)
action_mean = stats['actions']['mean']
action_std = stats['actions']['std']
actions_normalizer = DataNormalizer(action_mean, action_std)
model_dy.action_normalizer = actions_normalizer
model_dy.state_normalizer = observations_normalizer
print("model_dy #params: %d" % count_trainable_parameters(model_dy))
# criterion
criterionMSE = nn.MSELoss()
# optimizer
params = model_dy.parameters()
optimizer = optim.Adam(params,
lr=config['train']['lr'],
betas=(config['train']['adam_beta1'], 0.999))
scheduler = ReduceLROnPlateau(optimizer,
'min',
factor=0.9,
patience=10,
verbose=True)
if use_gpu:
model_dy = model_dy.cuda()
best_valid_loss = np.inf
global_iteration = 0
epoch_counter_external = 0
try:
for epoch in range(st_epoch, config['train']['n_epoch']):
phases = ['train', 'valid']
epoch_counter_external = epoch
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
model_dy.train(phase == 'train')
meter_loss_rmse = AverageMeter()
# bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in enumerate(loader):
global_iteration += 1
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config['train']['n_history'], config[
'train']['n_rollout']
n_samples = n_his + n_roll
if config['env']['type'] in ['PusherSlider']:
states = data['observations']
actions = data['actions']
if use_gpu:
states = states.cuda()
actions = actions.cuda()
# states, actions = data
assert states.size(1) == n_samples
# normalize states and actions once for entire rollout
states = model_dy.state_normalizer.normalize(
states)
actions = model_dy.action_normalizer.normalize(
actions)
B = states.size(0)
loss_mse = 0.
# state_cur: B x n_his x state_dim
state_cur = states[:, :n_his]
for j in range(n_roll):
state_des = states[:, n_his + j]
# action_cur: B x n_his x action_dim
action_cur = actions[:, j:j +
n_his] if actions is not None else None
# state_pred: B x state_dim
# state_cur: B x n_his x state_dim
# state_pred: B x state_dim
state_pred = model_dy(state_cur, action_cur)
loss_mse_cur = criterionMSE(
state_pred, state_des)
loss_mse += loss_mse_cur / n_roll
# update state_cur
# state_pred.unsqueeze(1): B x 1 x state_dim
state_cur = torch.cat([
state_cur[:, 1:],
state_pred.unsqueeze(1)
], 1)
meter_loss_rmse.update(np.sqrt(loss_mse.item()), B)
if phase == 'train':
optimizer.zero_grad()
loss_mse.backward()
optimizer.step()
if i % config['train']['log_per_iter'] == 0:
log = '%s [%d/%d][%d/%d] LR: %.6f' % (
phase, epoch, config['train']['n_epoch'], i,
data_n_batches[phase], get_lr(optimizer))
log += ', rmse: %.6f (%.6f)' % (np.sqrt(
loss_mse.item()), meter_loss_rmse.avg)
print(log)
# log data to tensorboard
# only do it once we have reached 500 iterations
if global_iteration > 500:
writer.add_scalar("Params/learning rate",
get_lr(optimizer),
global_iteration)
writer.add_scalar("Loss/train", loss_mse.item(),
global_iteration)
writer.add_scalar("RMSE average loss/train",
meter_loss_rmse.avg,
global_iteration)
if phase == 'train' and i % config['train'][
'ckp_per_iter'] == 0:
save_model(
model_dy, '%s/net_dy_epoch_%d_iter_%d' %
(train_dir, epoch, i))
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, config['train']['n_epoch'],
meter_loss_rmse.avg, best_valid_loss)
print(log)
if phase == 'valid':
scheduler.step(meter_loss_rmse.avg)
writer.add_scalar("RMSE average loss/valid",
meter_loss_rmse.avg, global_iteration)
if meter_loss_rmse.avg < best_valid_loss:
best_valid_loss = meter_loss_rmse.avg
save_model(model_dy, '%s/net_best_dy' % (train_dir))
writer.flush() # flush SummaryWriter events to disk
except KeyboardInterrupt:
# save network if we have a keyboard interrupt
save_model(
model_dy, '%s/net_dy_epoch_%d_keyboard_interrupt' %
(train_dir, epoch_counter_external))
writer.flush() # flush SummaryWriter events to disk
def main(dataset_name):
# sample from specific image
dataset_paths = None
episode_name = None
camera_name = None
episode_idx = None
if dataset_name == "dps_box_on_side_600":
camera_name = "camera_angled"
episode_name = "2020-05-13-21-55-01-487901_idx_33"
episode_idx = 22
dataset_paths = exp_18_utils.get_dataset_paths(dataset_name)
elif dataset_name == "correlle_mug-small_single_color_600":
camera_name = "camera_1_top_down"
episode_name = "2020-06-02-14-15-27-898104_idx_56"
episode_idx = 18
dataset_paths = exp_20_utils.get_dataset_paths(dataset_name)
elif dataset_name == "correlle_mug-small_many_colors_600":
camera_name = "camera_1_top_down"
episode_name = "2020-06-03-15-48-50-165064_idx_56"
episode_idx = 20
dataset_paths = exp_20_utils.get_dataset_paths(dataset_name)
else:
raise ValueError("unknown dataset")
dataset_root = dataset_paths['dataset_root']
dataset_name = dataset_paths['dataset_name']
## Load Model
model_name, model_file = get_DD_model_file(dataset_name)
model_train_dir = os.path.dirname(model_file)
print("model_train_dir", model_train_dir)
print("model_file", model_file)
model = torch.load(model_file)
model = model.cuda()
model = model.eval()
multi_episode_dict = DCDrakeSimEpisodeReader.load_dataset(
dataset_root, max_num_episodes=None)
output_dir = os.path.join(
model_train_dir,
'precomputed_vision_data/descriptor_keypoints/dataset_%s/' %
(dataset_name))
# compute descriptor confidence scores
if False:
print("\n\n---------Computing Descriptor Confidence Scores-----------")
metadata_file = os.path.join(output_dir, 'metadata.p')
if os.path.isfile(metadata_file):
answer = input(
"metadata.p file already exists, do you want to overwrite it? y/n\n"
)
if answer == "y":
shutil.rmtree(output_dir)
print("removing existing file and continuing")
else:
print("aborting")
quit()
set_seed(0)
compute_descriptor_confidences(
multi_episode_dict,
model,
output_dir,
batch_size=10,
num_workers=20,
model_file=model_file,
camera_name=camera_name,
num_ref_descriptors=50,
num_batches=10,
episode_name_arg=episode_name,
episode_idx=episode_idx,
)
if False:
confidence_score_data_file = os.path.join(output_dir, 'data.p')
confidence_score_data = load_pickle(confidence_score_data_file)
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
print(
"\n\n---------Selecting Spatially Separated Keypoints-----------")
score_and_select_spatially_separated_keypoints(
metadata,
confidence_score_data=confidence_score_data,
K=4,
position_diff_threshold=15,
output_dir=output_dir,
)
# visualize descriptors
if False:
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
episode_name = metadata['episode_name']
episode_idx = metadata['episode_idx']
camera_name = metadata['camera_name']
episode = multi_episode_dict[episode_name]
data = episode.get_image_data(camera_name, episode_idx)
rgb = data['rgb']
uv = metadata['indices']
print("uv.shape", uv.shape)
color = [0, 255, 0]
draw_reticles(rgb, uv[:, 0], uv[:, 1], label_color=color)
save_file = os.path.join(output_dir, 'sampled_descriptors.png')
plt.figure()
plt.imshow(rgb)
plt.savefig(save_file)
plt.show()
# visualize spatially separated descriptors
if False:
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
spatial_descriptor_file = os.path.join(output_dir,
'spatial_descriptors.p')
spatial_descriptors_data = load_pickle(spatial_descriptor_file)
des_idx = spatial_descriptors_data['spatial_descriptors_idx']
episode_name = metadata['episode_name']
episode_idx = metadata['episode_idx']
camera_name = metadata['camera_name']
episode = multi_episode_dict[episode_name]
data = episode.get_image_data(camera_name, episode_idx)
rgb = data['rgb']
uv = metadata['indices']
print("uv.shape", uv.shape)
color = [0, 255, 0]
draw_reticles(rgb, uv[des_idx, 0], uv[des_idx, 1], label_color=color)
save_file = os.path.join(output_dir,
'spatially_separated_descriptors.png')
plt.figure()
plt.imshow(rgb)
plt.savefig(save_file)
plt.show()
if True:
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
# metadata_file = "/media/hdd/data/key_dynam/dev/experiments/09/precomputed_vision_data/dataset_2020-03-25-19-57-26-556093_constant_velocity_500/model_name_2020-04-07-14-31-35-804270_T_aug_dataset/2020-04-09-20-51-50-624799/metadata.p"
# metadata = load_pickle(metadata_file)
print("\n\n---------Precomputing Descriptor Keypoints-----------")
descriptor_keypoints_output_dir = os.path.join(output_dir,
"descriptor_keypoints")
precompute_descriptor_keypoints(multi_episode_dict,
model,
descriptor_keypoints_output_dir,
ref_descriptors_metadata=metadata,
batch_size=8,
num_workers=20,
camera_names=[camera_name])
print("Data saved at: ", output_dir)
print("Finished Normally")
def train_dynamics(
config,
train_dir, # str: directory to save output
multi_episode_dict=None,
visual_observation_function=None,
metadata=None,
spatial_descriptors_data=None,
):
assert multi_episode_dict is not None
# assert spatial_descriptors_idx is not None
# set random seed for reproduction
set_seed(config['train']['random_seed'])
st_epoch = config['train'][
'resume_epoch'] if config['train']['resume_epoch'] > 0 else 0
tee = Tee(os.path.join(train_dir, 'train_st_epoch_%d.log' % st_epoch), 'w')
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# save the config
save_yaml(config, os.path.join(train_dir, "config.yaml"))
if metadata is not None:
save_pickle(metadata, os.path.join(train_dir, 'metadata.p'))
if spatial_descriptors_data is not None:
save_pickle(spatial_descriptors_data,
os.path.join(train_dir, 'spatial_descriptors.p'))
training_stats = dict()
training_stats_file = os.path.join(train_dir, 'training_stats.yaml')
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(
config,
action_function=action_function,
observation_function=observation_function,
episodes=multi_episode_dict,
phase=phase,
visual_observation_function=visual_observation_function)
print("len(datasets[phase])", len(datasets[phase]))
dataloaders[phase] = DataLoader(
datasets[phase],
batch_size=config['train']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train']['num_workers'],
drop_last=True)
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
# compute normalization parameters if not starting from pre-trained network . . .
if False:
dataset = datasets["train"]
data = dataset[0]
print("data['observations_combined'].shape",
data['observations_combined'].shape)
print("data.keys()", data.keys())
print("data['observations_combined']",
data['observations_combined'][0])
print("data['observations_combined'].shape",
data['observations_combined'].shape)
print("data['actions'].shape", data['actions'].shape)
print("data['actions']\n", data['actions'])
quit()
'''
Build model for dynamics prediction
'''
model_dy = build_dynamics_model(config)
if config['dynamics_net'] == "mlp_weight_matrix":
raise ValueError("can't use weight matrix with standard setup")
# criterion
criterionMSE = nn.MSELoss()
l1Loss = nn.L1Loss()
smoothL1 = nn.SmoothL1Loss()
# optimizer
params = model_dy.parameters()
lr = float(config['train']['lr'])
optimizer = optim.Adam(params,
lr=lr,
betas=(config['train']['adam_beta1'], 0.999))
# setup scheduler
sc = config['train']['lr_scheduler']
scheduler = None
if config['train']['lr_scheduler']['enabled']:
if config['train']['lr_scheduler']['type'] == "ReduceLROnPlateau":
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=sc['factor'],
patience=sc['patience'],
threshold_mode=sc['threshold_mode'],
cooldown=sc['cooldown'],
verbose=True)
elif config['train']['lr_scheduler']['type'] == "StepLR":
step_size = config['train']['lr_scheduler']['step_size']
gamma = config['train']['lr_scheduler']['gamma']
scheduler = StepLR(optimizer, step_size=step_size, gamma=gamma)
else:
raise ValueError("unknown scheduler type: %s" %
(config['train']['lr_scheduler']['type']))
if use_gpu:
print("using gpu")
model_dy = model_dy.cuda()
# print("model_dy.vision_net._ref_descriptors.device", model_dy.vision_net._ref_descriptors.device)
# print("model_dy.vision_net #params: %d" %(count_trainable_parameters(model_dy.vision_net)))
best_valid_loss = np.inf
valid_loss_type = config['train']['valid_loss_type']
global_iteration = 0
counters = {'train': 0, 'valid': 0}
epoch_counter_external = 0
loss = 0
try:
for epoch in range(st_epoch, config['train']['n_epoch']):
phases = ['train', 'valid']
epoch_counter_external = epoch
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
# only validate at a certain frequency
if (phase == "valid") and (
(epoch % config['train']['valid_frequency']) != 0):
continue
model_dy.train(phase == 'train')
average_meter_container = dict()
step_duration_meter = AverageMeter()
# bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in enumerate(loader):
loss_container = dict() # store the losses for this step
step_start_time = time.time()
global_iteration += 1
counters[phase] += 1
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config['train']['n_history'], config[
'train']['n_rollout']
n_samples = n_his + n_roll
if DEBUG:
print("global iteration: %d" % (global_iteration))
print("n_samples", n_samples)
# [B, n_samples, obs_dim]
states = data['observations_combined']
# [B, n_samples, action_dim]
actions = data['actions']
B = actions.shape[0]
if use_gpu:
states = states.cuda()
actions = actions.cuda()
# state_cur: B x n_his x state_dim
# state_cur = states[:, :n_his]
# [B, n_his, state_dim]
state_init = states[:, :n_his]
# We want to rollout n_roll steps
# actions = [B, n_his + n_roll, -1]
# so we want action_seq.shape = [B, n_roll, -1]
action_start_idx = 0
action_end_idx = n_his + n_roll - 1
action_seq = actions[:, action_start_idx:
action_end_idx, :]
if DEBUG:
print("states.shape", states.shape)
print("state_init.shape", state_init.shape)
print("actions.shape", actions.shape)
print("action_seq.shape", action_seq.shape)
# try using models_dy.rollout_model instead of doing this manually
rollout_data = rollout_model(state_init=state_init,
action_seq=action_seq,
dynamics_net=model_dy,
compute_debug_data=False)
# [B, n_roll, state_dim]
state_rollout_pred = rollout_data['state_pred']
# [B, n_roll, state_dim]
state_rollout_gt = states[:, n_his:]
if DEBUG:
print("state_rollout_gt.shape",
state_rollout_gt.shape)
print("state_rollout_pred.shape",
state_rollout_pred.shape)
# the loss function is between
# [B, n_roll, state_dim]
state_pred_err = state_rollout_pred - state_rollout_gt
# everything is in 3D space now so no need to do any scaling
# all the losses would be in meters . . . .
loss_mse = criterionMSE(state_rollout_pred,
state_rollout_gt)
loss_l1 = l1Loss(state_rollout_pred, state_rollout_gt)
loss_l2 = torch.norm(state_pred_err, dim=-1).mean()
loss_smoothl1 = smoothL1(state_rollout_pred,
state_rollout_gt)
loss_smoothl1_final_step = smoothL1(
state_rollout_pred[:, -1], state_rollout_gt[:, -1])
# compute losses at final step of the rollout
mse_final_step = criterionMSE(
state_rollout_pred[:, -1], state_rollout_gt[:, -1])
l2_final_step = torch.norm(state_pred_err[:, -1],
dim=-1).mean()
l1_final_step = l1Loss(state_rollout_pred[:, -1],
state_rollout_gt[:, -1])
loss_container['mse'] = loss_mse
loss_container['l1'] = loss_l1
loss_container['mse_final_step'] = mse_final_step
loss_container['l1_final_step'] = l1_final_step
loss_container['l2_final_step'] = l2_final_step
loss_container['l2'] = loss_l2
loss_container['smooth_l1'] = loss_smoothl1
loss_container[
'smooth_l1_final_step'] = loss_smoothl1_final_step
# compute the loss
loss = 0
for key, val in config['loss_function'].items():
if val['enabled']:
loss += loss_container[key] * val['weight']
loss_container['loss'] = loss
for key, val in loss_container.items():
if not key in average_meter_container:
average_meter_container[key] = AverageMeter()
average_meter_container[key].update(val.item(), B)
step_duration_meter.update(time.time() - step_start_time)
if phase == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i % config['train']['log_per_iter']
== 0) or (global_iteration %
config['train']['log_per_iter'] == 0):
log = '%s [%d/%d][%d/%d] LR: %.6f' % (
phase, epoch, config['train']['n_epoch'], i,
data_n_batches[phase], get_lr(optimizer))
log += ', l2: %.6f' % (loss_container['l2'].item())
log += ', l2_final_step: %.6f' % (
loss_container['l2_final_step'].item())
log += ', step time %.6f' % (step_duration_meter.avg)
step_duration_meter.reset()
print(log)
# log data to tensorboard
# only do it once we have reached 100 iterations
if global_iteration > 100:
writer.add_scalar("Params/learning rate",
get_lr(optimizer),
global_iteration)
writer.add_scalar("Loss_train/%s" % (phase),
loss.item(), global_iteration)
for loss_type, loss_obj in loss_container.items():
plot_name = "Loss/%s/%s" % (loss_type, phase)
writer.add_scalar(plot_name, loss_obj.item(),
counters[phase])
if phase == 'train' and global_iteration % config['train'][
'ckp_per_iter'] == 0:
save_model(
model_dy, '%s/net_dy_epoch_%d_iter_%d' %
(train_dir, epoch, i))
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, config['train']['n_epoch'],
average_meter_container[valid_loss_type].avg,
best_valid_loss)
print(log)
# record all average_meter losses
for key, meter in average_meter_container.items():
writer.add_scalar("AvgMeter/%s/%s" % (key, phase),
meter.avg, epoch)
if phase == "train":
if (scheduler is not None) and (
config['train']['lr_scheduler']['type']
== "StepLR"):
scheduler.step()
if phase == 'valid':
if (scheduler is not None) and (
config['train']['lr_scheduler']['type']
== "ReduceLROnPlateau"):
scheduler.step(
average_meter_container[valid_loss_type].avg)
if average_meter_container[
valid_loss_type].avg < best_valid_loss:
best_valid_loss = average_meter_container[
valid_loss_type].avg
training_stats['epoch'] = epoch
training_stats['global_iteration'] = counters['valid']
save_yaml(training_stats, training_stats_file)
save_model(model_dy, '%s/net_best_dy' % (train_dir))
writer.flush() # flush SummaryWriter events to disk
except KeyboardInterrupt:
# save network if we have a keyboard interrupt
save_model(
model_dy, '%s/net_dy_epoch_%d_keyboard_interrupt' %
(train_dir, epoch_counter_external))
writer.flush() # flush SummaryWriter events to disk
def train_dynamics(config,
train_dir, # str: directory to save output
multi_episode_dict, # multi_episode_dict
):
use_precomputed_keypoints = config['dataset']['visual_observation']['enabled'] and config['dataset']['visual_observation']['descriptor_keypoints']
# set random seed for reproduction
set_seed(config['train']['random_seed'])
st_epoch = config['train']['resume_epoch'] if config['train']['resume_epoch'] > 0 else 0
tee = Tee(os.path.join(train_dir, 'train_st_epoch_%d.log' % st_epoch), 'w')
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# save the config
save_yaml(config, os.path.join(train_dir, "config.yaml"))
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(config,
action_function=action_function,
observation_function=observation_function,
episodes=multi_episode_dict,
phase=phase)
dataloaders[phase] = DataLoader(
datasets[phase], batch_size=config['train']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train']['num_workers'], drop_last=True)
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
# compute normalization parameters if not starting from pre-trained network . . .
'''
define model for dynamics prediction
'''
model_dy = build_visual_dynamics_model(config)
K = config['vision_net']['num_ref_descriptors']
print("model_dy.vision_net._reference_descriptors.shape", model_dy.vision_net._ref_descriptors.shape)
print("model_dy.vision_net.descriptor_dim", model_dy.vision_net.descriptor_dim)
print("model_dy #params: %d" % count_trainable_parameters(model_dy))
camera_name = config['vision_net']['camera_name']
W = config['env']['rgbd_sensors']['sensor_list'][camera_name]['width']
H = config['env']['rgbd_sensors']['sensor_list'][camera_name]['height']
diag = np.sqrt(W**2 + H**2) # use this to scale the loss
# sample reference descriptors unless using precomputed keypoints
if not use_precomputed_keypoints:
# sample reference descriptors
episode_names = list(datasets["train"].episode_dict.keys())
episode_names.sort()
episode_name = episode_names[0]
episode = datasets["train"].episode_dict[episode_name]
episode_idx = 0
camera_name = config["vision_net"]["camera_name"]
image_data = episode.get_image_data(camera_name, episode_idx)
des_img = torch.Tensor(image_data['descriptor'])
mask_img = torch.Tensor(image_data['mask'])
ref_descriptor_dict = sample_descriptors(des_img,
mask_img,
config['vision_net']['num_ref_descriptors'])
model_dy.vision_net._ref_descriptors.data = ref_descriptor_dict['descriptors']
model_dy.vision_net.reference_image = image_data['rgb']
model_dy.vision_net.reference_indices = ref_descriptor_dict['indices']
else:
metadata_file = os.path.join(get_data_root(), config['dataset']['descriptor_keypoints_dir'], 'metadata.p')
descriptor_metadata = load_pickle(metadata_file)
# [32, 2]
ref_descriptors = torch.Tensor(descriptor_metadata['ref_descriptors'])
# [K, 2]
ref_descriptors = ref_descriptors[:K]
model_dy.vision_net._ref_descriptors.data = ref_descriptors
model_dy.vision_net._ref_descriptors_metadata = descriptor_metadata
# this is just a sanity check
assert model_dy.vision_net.num_ref_descriptors == K
print("reference_descriptors", model_dy.vision_net._ref_descriptors)
# criterion
criterionMSE = nn.MSELoss()
l1Loss = nn.L1Loss()
# optimizer
params = model_dy.parameters()
lr = float(config['train']['lr'])
optimizer = optim.Adam(params, lr=lr, betas=(config['train']['adam_beta1'], 0.999))
# setup scheduler
sc = config['train']['lr_scheduler']
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=sc['factor'],
patience=sc['patience'],
threshold_mode=sc['threshold_mode'],
cooldown= sc['cooldown'],
verbose=True)
if use_gpu:
print("using gpu")
model_dy = model_dy.cuda()
print("model_dy.vision_net._ref_descriptors.device", model_dy.vision_net._ref_descriptors.device)
print("model_dy.vision_net #params: %d" %(count_trainable_parameters(model_dy.vision_net)))
best_valid_loss = np.inf
global_iteration = 0
epoch_counter_external = 0
try:
for epoch in range(st_epoch, config['train']['n_epoch']):
phases = ['train', 'valid']
epoch_counter_external = epoch
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
model_dy.train(phase == 'train')
meter_loss_rmse = AverageMeter()
step_duration_meter = AverageMeter()
# bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in enumerate(loader):
step_start_time = time.time()
global_iteration += 1
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config['train']['n_history'], config['train']['n_rollout']
n_samples = n_his + n_roll
if DEBUG:
print("global iteration: %d" %(global_iteration))
# visual_observations = data['visual_observations']
visual_observations_list = data['visual_observations_list']
observations = data['observations']
actions = data['actions']
if use_gpu:
observations = observations.cuda()
actions = actions.cuda()
# states, actions = data
assert actions.size(1) == n_samples
B = actions.size(0)
loss_mse = 0.
# compute the output of the visual model for all timesteps
visual_model_output_list = []
for visual_obs in visual_observations_list:
# visual_obs is a dict containing observation for a single
# time step (of course across a batch however)
# visual_obs[<camera_name>]['rgb_tensor'] has shape [B, 3, H, W]
# probably need to cast input to cuda
dynamics_net_input = None
if use_precomputed_keypoints:
# note precomputed descriptors stored on disk are of size
# K = 32. We need to trim it down to the appropriate size
# [B, K_disk, 2] where K_disk is num keypoints on disk
keypoints = visual_obs[camera_name]['descriptor_keypoints']
# [B, 32, 2] where K is num keypoints
keypoints = keypoints[:,:K]
if DEBUG:
print("keypoints.shape", keypoints.shape)
dynamics_net_input = keypoints.flatten(start_dim=1)
else:
out_dict = model_dy.vision_net.forward(visual_obs)
# [B, vision_model_out_dim]
dynamics_net_input = out_dict['dynamics_net_input']
visual_model_output_list.append(dynamics_net_input)
# concatenate this into a tensor
# [B, n_samples, vision_model_out_dim]
visual_model_output = torch.stack(visual_model_output_list, dim=1)
# cast this to float so it can be concatenated below
visual_model_output = visual_model_output.type_as(observations)
if DEBUG:
print('visual_model_output.shape', visual_model_output.shape)
print("observations.shape", observations.shape)
print("actions.shape", actions.shape)
# states is gotten by concatenating visual_observations and observations
# [B, n_samples, vision_model_out_dim + obs_dim]
states = torch.cat((visual_model_output, observations), dim=-1)
# state_cur: B x n_his x state_dim
state_cur = states[:, :n_his]
if DEBUG:
print("states.shape", states.shape)
for j in range(n_roll):
if DEBUG:
print("n_roll j: %d" %(j))
state_des = states[:, n_his + j]
# action_cur: B x n_his x action_dim
action_cur = actions[:, j : j + n_his] if actions is not None else None
# state_pred: B x state_dim
# state_pred: B x state_dim
input = {'observation': state_cur,
'action': action_cur,
}
if DEBUG:
print("state_cur.shape", state_cur.shape)
print("action_cur.shape", action_cur.shape)
state_pred = model_dy.dynamics_net(input)
# normalize by diag to ensure the loss is in [0,1] range
loss_mse_cur = criterionMSE(state_pred/diag, state_des/diag)
loss_mse += loss_mse_cur / n_roll
# l1Loss
loss_l1 = l1Loss(state_pred, state_des)
# update state_cur
# state_pred.unsqueeze(1): B x 1 x state_dim
# state_cur: B x n_his x state_dim
state_cur = torch.cat([state_cur[:, 1:], state_pred.unsqueeze(1)], 1)
meter_loss_rmse.update(np.sqrt(loss_mse.item()), B)
step_duration_meter.update(time.time() - step_start_time)
if phase == 'train':
optimizer.zero_grad()
loss_mse.backward()
optimizer.step()
if (i % config['train']['log_per_iter'] == 0) or (global_iteration % config['train']['log_per_iter'] == 0):
log = '%s [%d/%d][%d/%d] LR: %.6f' % (
phase, epoch, config['train']['n_epoch'], i, data_n_batches[phase],
get_lr(optimizer))
log += ', rmse: %.6f (%.6f)' % (
np.sqrt(loss_mse.item()), meter_loss_rmse.avg)
log += ', step time %.6f' %(step_duration_meter.avg)
step_duration_meter.reset()
print(log)
# log data to tensorboard
# only do it once we have reached 100 iterations
if global_iteration > 100:
writer.add_scalar("Params/learning rate", get_lr(optimizer), global_iteration)
writer.add_scalar("Loss_MSE/%s" %(phase), loss_mse.item(), global_iteration)
writer.add_scalar("L1/%s" %(phase), loss_l1.item(), global_iteration)
writer.add_scalar("L1_fraction/%s" %(phase), loss_l1.item()/diag, global_iteration)
writer.add_scalar("RMSE average loss/%s" %(phase), meter_loss_rmse.avg, global_iteration)
if phase == 'train' and i % config['train']['ckp_per_iter'] == 0:
save_model(model_dy, '%s/net_dy_epoch_%d_iter_%d' % (train_dir, epoch, i))
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, config['train']['n_epoch'], meter_loss_rmse.avg, best_valid_loss)
print(log)
if phase == 'valid':
if config['train']['lr_scheduler']['enabled']:
scheduler.step(meter_loss_rmse.avg)
# print("\nPhase == valid")
# print("meter_loss_rmse.avg", meter_loss_rmse.avg)
# print("best_valid_loss", best_valid_loss)
if meter_loss_rmse.avg < best_valid_loss:
best_valid_loss = meter_loss_rmse.avg
save_model(model_dy, '%s/net_best_dy' % (train_dir))
writer.flush() # flush SummaryWriter events to disk
except KeyboardInterrupt:
# save network if we have a keyboard interrupt
save_model(model_dy, '%s/net_dy_epoch_%d_keyboard_interrupt' % (train_dir, epoch_counter_external))
writer.flush() # flush SummaryWriter events to disk
def collect_episodes(config,
output_dir=None,
visualize=True,
debug=False,
run_from_thread=False,
seed=None):
# gets a random seed for each thread/process independently
if seed is None:
seed = np.random.RandomState().randint(0, 10000)
set_seed(seed)
if output_dir is None:
output_dir = os.path.join(os.getcwd(), 'data')
if not os.path.exists(output_dir):
os.makedirs(output_dir)
# save the config
config_save_file = os.path.join(output_dir, 'config.yaml')
save_yaml(config, config_save_file)
# initialize config for DataCollector
num_episodes = config['dataset']['num_episodes']
# record some metadata
metadata = dict()
metadata['episodes'] = dict()
while (len(metadata['episodes']) < num_episodes):
i = len(metadata['episodes'])
if debug:
input("Press Enter to continue...")
print("\n")
start_time = time.time()
print("collecting episode %d of %d" % (i + 1, num_episodes))
name = "%s_idx_%d" % (get_current_YYYY_MM_DD_hh_mm_ss_ms(), i)
n_his = config['train_dynamics']['n_history']
ic = generate_initial_condition(
config=config,
T_aug_enabled=True,
n_his=n_his,
randomize_velocity=True,
randomize_sdf=True,
randomize_color=True,
)
env = DrakeMugsEnv(ic['config'], visualize=visualize)
if debug:
print("initial condition\n", ic)
# set initial condition on environment
if visualize:
print("setting target realtime rate 1.0")
env.simulator.set_target_realtime_rate(1.0)
env.reset()
context = env.get_mutable_context()
env.set_object_position(context, ic['q_slider'])
env.set_pusher_position(context, ic['q_pusher'])
print("ic['action_sequence'].shape", ic['action_sequence'].shape)
# simulate for 10 seconds to let the mug stabilize
action_zero = env.get_zero_action()
env.step(action_zero, dt=10.0)
episode = collect_single_episode(
env, action_seq=ic['action_sequence'])['episode_container']
# potentially discard it if the object didn't move during the data collection
if len(episode._data['trajectory']) < 10:
print("trajectory was too short, skipping")
continue
obs_start = episode._data['trajectory'][0]['observation']
obs_end = episode._data['trajectory'][-1]['observation']
q_slider_start = obs_start['slider']['position']['translation']
q_slider_end = obs_end['slider']['position']['translation']
dq_slider = obs_start['slider']['position']['translation'] - obs_end[
'slider']['position']['translation']
if debug:
print("len(episode._data['trajectory'])",
len(episode._data['trajectory']))
print("q_slider_start", q_slider_start)
print("q_slider_end", q_slider_end)
print("dq_slider", dq_slider)
print("np.linalg.norm(dq_slider)", np.linalg.norm(dq_slider))
pose_error = compute_pose_error(obs_start, obs_end)
# if slider didn't move by at least 1 mm then discard this episode
if (pose_error['position_error'] <
0.01) and (pose_error['angle_error_degrees'] < 10):
print(
"discarding episode since slider didn't move sufficiently far")
continue
print("saving to disk")
metadata['episodes'][name] = dict()
image_data_file = episode.save_images_to_hdf5(output_dir)
non_image_data_file = episode.save_non_image_data_to_pickle(output_dir)
print("output_dir:", output_dir)
print("non_image_data.keys()", episode.non_image_data.keys())
metadata['episodes'][name]['non_image_data_file'] = non_image_data_file
metadata['episodes'][name]['image_data_file'] = image_data_file
print("done saving to disk")
elapsed = time.time() - start_time
print("single episode took: %.2f seconds" % (elapsed))
if not run_from_thread:
save_yaml(metadata, os.path.join(output_dir, 'metadata.yaml'))
print("Finished collecting episodes")
return {'metadata': metadata}
def multiprocess_main(num_episodes=1000, num_threads=4):
set_seed(500) # just randomly chosen
start_time = time.time()
config = load_yaml(
os.path.join(get_project_root(),
'experiments/exp_20_mugs/config.yaml'))
num_episodes_per_thread = math.ceil(num_episodes / num_threads)
num_episodes = num_threads * num_episodes_per_thread
# DATASET_NAME = "mugs_random_colors_%d" % (num_episodes)
# DATASET_NAME = "single_mug_%d"
# DATASET_NAME = "correlle_mug-small_single_color_%d" %(num_episodes)
# DATASET_NAME = "single_corelle_mug_%d" %(num_episodes)
# DATASET_NAME = "correlle_mug-small_many_colors_%d" %(num_episodes)
DATASET_NAME = "correlle_mug-small_many_colors_random_%d" % (num_episodes)
# OUTPUT_DIR = os.path.join(get_data_root(), 'sandbox', DATASET_NAME)
OUTPUT_DIR = os.path.join(get_data_ssd_root(), 'dataset', DATASET_NAME)
print("OUTPUT_DIR:", OUTPUT_DIR)
output_dir = OUTPUT_DIR
if not os.path.exists(output_dir):
os.makedirs(output_dir)
def f(q_tmp):
config = load_yaml(
os.path.join(get_project_root(),
'experiments/exp_20_mugs/config.yaml'))
config['dataset']['num_episodes'] = num_episodes_per_thread
out = collect_episodes(config,
output_dir=OUTPUT_DIR,
visualize=False,
debug=False,
run_from_thread=True)
q_tmp.put(out)
q = Queue()
process_list = []
for i in range(num_threads):
p = Process(target=f, args=(q, ))
p.start()
process_list.append(p)
metadata = {'episodes': {}}
for p in process_list:
while p.is_alive():
p.join(timeout=1)
# empty out the queue
while not q.empty():
out = q.get()
metadata['episodes'].update(out['metadata']['episodes'])
# double check
for p in process_list:
p.join()
time.sleep(1.0)
print("All threads joined")
elapsed = time.time() - start_time
# collect the metadata.yaml files
while not q.empty():
out = q.get()
metadata['episodes'].update(out['metadata']['episodes'])
save_yaml(metadata, os.path.join(OUTPUT_DIR, 'metadata.yaml'))
print("Generating and saving dataset to disk took %d seconds" %
(int(elapsed)))
def mpc_w_learned_dynamics(config,
train_dir,
mpc_dir,
state_dict_path=None,
keypoint_observation=False):
# set random seed for reproduction
set_seed(config['train']['random_seed'])
tee = Tee(os.path.join(mpc_dir, 'mpc.log'), 'w')
print(config)
use_gpu = torch.cuda.is_available()
'''
model
'''
if config['dynamics']['model_type'] == 'mlp':
model_dy = DynaNetMLP(config)
else:
raise AssertionError("Unknown model type %s" %
config['dynamics']['model_type'])
# print model #params
print("model #params: %d" % count_trainable_parameters(model_dy))
if state_dict_path is None:
if config['mpc']['mpc_dy_epoch'] == -1:
state_dict_path = os.path.join(train_dir, 'net_best_dy.pth')
else:
state_dict_path = os.path.join(
train_dir, 'net_dy_epoch_%d_iter_%d.pth' % \
(config['mpc']['mpc_dy_epoch'], config['mpc']['mpc_dy_iter']))
print("Loading saved ckp from %s" % state_dict_path)
model_dy.load_state_dict(torch.load(state_dict_path))
model_dy.eval()
if use_gpu:
model_dy.cuda()
criterionMSE = nn.MSELoss()
# generate action/observation functions
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
# planner
planner = planner_from_config(config)
'''
env
'''
# set up goal
obs_goals = np.array([[
262.9843, 267.3102, 318.9369, 351.1229, 360.2048, 323.5128, 305.6385,
240.4460, 515.4230, 347.8708
],
[
381.8694, 273.6327, 299.6685, 331.0925, 328.7724,
372.0096, 411.0972, 314.7053, 517.7299, 268.4953
],
[
284.8728, 275.7985, 374.0677, 320.4990, 395.4019,
275.4633, 306.2896, 231.4310, 507.0849, 312.4057
],
[
313.1638, 271.4258, 405.0255, 312.2325, 424.7874,
266.3525, 333.6973, 225.7708, 510.1232, 305.3802
],
[
308.6859, 270.9629, 394.2789, 323.2781, 419.7905,
280.1602, 333.8901, 228.1624, 519.1964, 321.5318
],
[
386.8067, 284.8947, 294.2467, 323.2223, 313.3221,
368.9970, 405.9415, 330.9298, 495.9970, 268.9920
],
[
432.0219, 299.6021, 340.8581, 339.4676, 360.2354,
384.5515, 451.4394, 345.2190, 514.6357, 291.2043
],
[
351.3389, 264.5325, 267.5279, 318.2321, 293.7460,
360.0423, 378.4428, 306.9586, 516.4390, 259.7810
],
[
521.1902, 254.0693, 492.7884, 349.7861, 539.6320,
364.5190, 569.2258, 268.8824, 506.9431, 286.9752
],
[
264.8554, 275.9547, 338.1317, 345.3435, 372.7012,
308.4648, 299.3454, 239.9245, 506.2117, 373.8413
]])
for mpc_idx in range(config['mpc']['num_episodes']):
if keypoint_observation:
mpc_episode_keypoint_observation(config,
mpc_idx,
model_dy,
mpc_dir,
planner,
obs_goals[mpc_idx],
action_function,
observation_function,
use_gpu=use_gpu)
else:
# not supported for now
raise AssertionError("currently only support keypoint observation")
def main():
# load dynamics model
model_dict = load_model_state_dict()
model = model_dict['model_dy']
model_dd = model_dict['model_dd']
config = model.config
env_config = load_yaml(
os.path.join(get_project_root(),
'experiments/exp_18_box_on_side/config.yaml'))
env_config['env']['observation']['depth_int16'] = True
n_history = config['train']['n_history']
# enable the right observations
camera_name = model_dict['metadata']['camera_name']
spatial_descriptor_data = model_dict['spatial_descriptor_data']
ref_descriptors = spatial_descriptor_data['spatial_descriptors']
K = ref_descriptors.shape[0]
ref_descriptors = torch.Tensor(
ref_descriptors).cuda() # put them on the GPU
print("ref_descriptors\n", ref_descriptors)
print("ref_descriptors.shape", ref_descriptors.shape)
# create the environment
# create the environment
env = DrakePusherSliderEnv(env_config)
env.reset()
T_world_camera = env.camera_pose(camera_name)
camera_K_matrix = env.camera_K_matrix(camera_name)
# create another environment for doing rollouts
env2 = DrakePusherSliderEnv(env_config, visualize=False)
env2.reset()
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.drake_pusher_position_3D(
config)
visual_observation_function = \
VisualObservationFunctionFactory.descriptor_keypoints_3D(config=config,
camera_name=camera_name,
model_dd=model_dd,
ref_descriptors=ref_descriptors,
K_matrix=camera_K_matrix,
T_world_camera=T_world_camera,
)
episode = OnlineEpisodeReader()
mpc_input_builder = DynamicsModelInputBuilder(
observation_function=observation_function,
visual_observation_function=visual_observation_function,
action_function=action_function,
episode=episode)
vis = meshcat_utils.make_default_visualizer_object()
vis.delete()
initial_cond = get_initial_state()
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
obs_init = env.get_observation()
#### ROLLOUT USING LEARNED MODEL + GROUND TRUTH ACTIONS ############
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
# add just some large number of these
episode.clear()
for i in range(n_history):
action_zero = np.zeros(2)
obs_tmp = env.get_observation()
episode.add_observation_action(obs_tmp, action_zero)
def goal_func(obs_tmp):
state_tmp = mpc_input_builder.get_state_input_single_timestep(
{'observation': obs_tmp})['state']
return model.compute_z_state(
state_tmp.unsqueeze(0))['z_object'].flatten()
#
idx = episode.get_latest_idx()
obs_raw = episode.get_observation(idx)
z_object_goal = goal_func(obs_raw)
z_keypoints_init_W = keypoints_3D_from_dynamics_model_output(
z_object_goal, K)
z_keypoints_init_W = torch_utils.cast_to_numpy(z_keypoints_init_W)
z_keypoints_obj = keypoints_world_frame_to_object_frame(
z_keypoints_init_W, T_W_obj=slider_pose_from_observation(obs_init))
# keypoints_W
# color = [1, 0, 0]
# meshcat_utils.visualize_points(vis=vis,
# name="keypoints_W",
# pts=z_keypoints_init_W,
# color=color,
# size=0.02,
# )
# rollout single action sequence using the simulator
gt_rollout_data = env_utils.rollout_action_sequence(
env, initial_cond['action_sequence'].cpu().numpy())
env_obs_rollout_gt = gt_rollout_data['observations']
gt_rollout_episode = gt_rollout_data['episode_reader']
action_state_gt = mpc_input_builder.get_action_state_tensors(
start_idx=0, num_timesteps=N, episode=gt_rollout_episode)
state_rollout_gt = action_state_gt['states']
action_rollout_gt = action_state_gt['actions']
z_object_rollout_gt = model.compute_z_state(
state_rollout_gt)['z_object_flat']
print('state_rollout_gt.shape', state_rollout_gt.shape)
print("z_object_rollout_gt.shape", z_object_rollout_gt.shape)
# using the vision model to get "goal" keypoints
z_object_goal = goal_func(env_obs_rollout_gt[-1])
z_object_goal_np = torch_utils.cast_to_numpy(z_object_goal)
z_keypoints_goal = keypoints_3D_from_dynamics_model_output(
z_object_goal, K)
z_keypoints_goal = torch_utils.cast_to_numpy(z_keypoints_goal)
# visualize goal keypoints
# color = [0, 1, 0]
# meshcat_utils.visualize_points(vis=vis,
# name="goal_keypoints",
# pts=z_keypoints_goal,
# color=color,
# size=0.02,
# )
# input("press Enter to continue")
#### ROLLOUT USING LEARNED MODEL + GROUND TRUTH ACTIONS ############
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
# add just some large number of these
episode.clear()
for i in range(n_history):
action_zero = np.zeros(2)
obs_tmp = env.get_observation()
episode.add_observation_action(obs_tmp, action_zero)
# [n_history, state_dim]
idx = episode.get_latest_idx()
dyna_net_input = mpc_input_builder.get_dynamics_model_input(
idx, n_history=n_history)
state_init = dyna_net_input['states'].cuda() # [n_history, state_dim]
action_init = dyna_net_input['actions'] # [n_history, action_dim]
print("state_init.shape", state_init.shape)
print("action_init.shape", action_init.shape)
action_seq_gt_torch = initial_cond['action_sequence']
action_input = torch.cat(
(action_init[:(n_history - 1)], action_seq_gt_torch), dim=0).cuda()
print("action_input.shape", action_input.shape)
# rollout using the ground truth actions and learned model
# need to add the batch dim to do that
z_init = model.compute_z_state(state_init)['z']
rollout_pred = rollout_model(state_init=z_init.unsqueeze(0),
action_seq=action_input.unsqueeze(0),
dynamics_net=model,
compute_debug_data=True)
state_pred_rollout = rollout_pred['state_pred']
print("state_pred_rollout.shape", state_pred_rollout.shape)
for i in range(N):
# vis GT for now
name = "GT_3D/%d" % (i)
T_W_obj = slider_pose_from_observation(env_obs_rollout_gt[i])
# print("T_W_obj", T_W_obj)
# green
color = np.array([0, 1, 0]) * get_color_intensity(i, N)
meshcat_utils.visualize_points(vis=vis,
name=name,
pts=z_keypoints_obj,
color=color,
size=0.01,
T=T_W_obj)
# red
color = np.array([0, 0, 1]) * get_color_intensity(i, N)
state_pred = state_pred_rollout[:, i, :]
pts_pred = keypoints_3D_from_dynamics_model_output(state_pred,
K).squeeze()
pts_pred = pts_pred.detach().cpu().numpy()
name = "pred_3D/%d" % (i)
meshcat_utils.visualize_points(
vis=vis,
name=name,
pts=pts_pred,
color=color,
size=0.01,
)
# input("finished visualizing GT rollout\npress Enter to continue")
index_dict = get_object_and_robot_state_indices(config)
object_indices = index_dict['object_indices']
# reset the environment and use the MPC controller to stabilize this
# now setup the MPC to try to stabilize this . . . .
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
episode.clear()
# add just some large number of these
for i in range(n_history):
action_zero = np.zeros(2)
obs_tmp = env.get_observation()
episode.add_observation_action(obs_tmp, action_zero)
input("press Enter to continue")
# make a planner config
planner_config = copy.copy(config)
config_tmp = load_yaml(
os.path.join(get_project_root(),
'experiments/drake_pusher_slider/eval_config.yaml'))
planner_config['mpc'] = config_tmp['mpc']
planner_config['mpc']['mppi']['terminal_cost_only'] = TERMINAL_COST_ONLY
planner_config['mpc']['random_shooting'][
'terminal_cost_only'] = TERMINAL_COST_ONLY
planner = None
if PLANNER_TYPE == "random_shooting":
planner = RandomShootingPlanner(planner_config)
elif PLANNER_TYPE == "mppi":
planner = PlannerMPPI(planner_config)
else:
raise ValueError("unknown planner type: %s" % (PLANNER_TYPE))
mpc_out = None
action_seq_mpc = None
state_pred_mpc = None
counter = -1
while True:
counter += 1
print("\n\n-----Running MPC Optimization: Counter (%d)-------" %
(counter))
obs_cur = env.get_observation()
episode.add_observation_only(obs_cur)
if counter == 0 or REPLAN:
print("replanning")
####### Run the MPC ##########
# [1, state_dim]
n_look_ahead = N - counter
if USE_FIXED_MPC_HORIZON:
n_look_ahead = MPC_HORIZON
elif USE_SHORT_HORIZON_MPC:
n_look_ahead = min(MPC_HORIZON, N - counter)
if n_look_ahead == 0:
break
start_idx = counter
end_idx = counter + n_look_ahead
print("start_idx:", start_idx)
print("end_idx:", end_idx)
# start_time = time.time()
# idx of current observation
idx = episode.get_latest_idx()
mpc_start_time = time.time()
mpc_input_data = mpc_input_builder.get_dynamics_model_input(
idx, n_history=n_history)
state_cur = mpc_input_data['states']
action_his = mpc_input_data['actions']
if SEED_WITH_NOMINAL_ACTIONS:
action_seq_rollout_init = action_seq_gt_torch[
start_idx:end_idx]
else:
if mpc_out is not None:
action_seq_rollout_init = mpc_out['action_seq'][1:]
print("action_seq_rollout_init.shape",
action_seq_rollout_init.shape)
if action_seq_rollout_init.shape[0] < n_look_ahead:
num_steps = n_look_ahead - action_seq_rollout_init.shape[
0]
action_seq_zero = torch.zeros([num_steps, 2])
action_seq_rollout_init = torch.cat(
(action_seq_rollout_init, action_seq_zero), dim=0)
print("action_seq_rollout_init.shape",
action_seq_rollout_init.shape)
else:
action_seq_rollout_init = None
# run MPPI
z_cur = None
with torch.no_grad():
z_cur = model.compute_z_state(
state_cur.unsqueeze(0).cuda())['z'].squeeze(0)
if action_seq_rollout_init is not None:
n_look_ahead = action_seq_rollout_init.shape[0]
obs_goal = None
print("z_object_rollout_gt.shape", z_object_rollout_gt.shape)
if TRAJECTORY_GOAL:
obs_goal = z_object_rollout_gt[start_idx:end_idx]
print("n_look_ahead", n_look_ahead)
print("obs_goal.shape", obs_goal.shape)
# add the final goal state on as needed
if obs_goal.shape[0] < n_look_ahead:
obs_goal_final = z_object_rollout_gt[-1].unsqueeze(0)
num_repeat = n_look_ahead - obs_goal.shape[0]
obs_goal_final_expand = obs_goal_final.expand(
[num_repeat, -1])
obs_goal = torch.cat((obs_goal, obs_goal_final_expand),
dim=0)
else:
obs_goal = z_object_rollout_gt[-1]
obs_goal = torch_utils.cast_to_numpy(obs_goal)
print("obs_goal.shape", obs_goal.shape)
seed = 1
set_seed(seed)
mpc_out = planner.trajectory_optimization(
state_cur=z_cur,
action_his=action_his,
obs_goal=obs_goal,
model_dy=model,
action_seq_rollout_init=action_seq_rollout_init,
n_look_ahead=n_look_ahead,
eval_indices=object_indices,
rollout_best_action_sequence=True,
verbose=True,
add_grid_action_samples=True,
)
print("MPC step took %.4f seconds" %
(time.time() - mpc_start_time))
action_seq_mpc = mpc_out['action_seq'].cpu().numpy()
# Rollout with ground truth simulator dynamics
action_seq_mpc = torch_utils.cast_to_numpy(mpc_out['action_seq'])
env2.set_simulator_state_from_observation_dict(
env2.get_mutable_context(), obs_cur)
obs_mpc_gt = env_utils.rollout_action_sequence(
env2, action_seq_mpc)['observations']
state_pred_mpc = torch_utils.cast_to_numpy(mpc_out['state_pred'])
vis['mpc_3D'].delete()
vis['mpc_GT_3D'].delete()
L = len(obs_mpc_gt)
print("L", L)
if L == 0:
break
for i in range(L):
# red
color = np.array([1, 0, 0]) * get_color_intensity(i, L)
state_pred = state_pred_mpc[i, :]
state_pred = np.expand_dims(state_pred,
0) # may need to expand dims here
pts_pred = keypoints_3D_from_dynamics_model_output(state_pred,
K).squeeze()
name = "mpc_3D/%d" % (i)
meshcat_utils.visualize_points(
vis=vis,
name=name,
pts=pts_pred,
color=color,
size=0.01,
)
# ground truth rollout of the MPC action_seq
name = "mpc_GT_3D/%d" % (i)
T_W_obj = slider_pose_from_observation(obs_mpc_gt[i])
# green
color = np.array([1, 1, 0]) * get_color_intensity(i, L)
meshcat_utils.visualize_points(vis=vis,
name=name,
pts=z_keypoints_obj,
color=color,
size=0.01,
T=T_W_obj)
action_cur = action_seq_mpc[0]
print("action_cur", action_cur)
print("action_GT", initial_cond['action'])
input("press Enter to continue")
# add observation actions to the episode
obs_cur = env.get_observation()
episode.replace_observation_action(obs_cur, action_cur)
# step the simulator
env.step(action_cur)
# visualize current keypoint positions
obs_cur = env.get_observation()
T_W_obj = slider_pose_from_observation(obs_cur)
# yellow
color = np.array([1, 1, 0])
meshcat_utils.visualize_points(vis=vis,
name="keypoint_cur",
pts=z_keypoints_obj,
color=color,
size=0.02,
T=T_W_obj)
action_seq_mpc = action_seq_mpc[1:]
state_pred_mpc = state_pred_mpc[1:]
pose_error = compute_pose_error(env_obs_rollout_gt[-1], obs_cur)
print("position_error: %.3f" % (pose_error['position_error']))
print("angle error degrees: %.3f" %
(pose_error['angle_error_degrees']))
obs_final = env.get_observation()
pose_error = compute_pose_error(env_obs_rollout_gt[-1], obs_final)
print("position_error: %.3f" % (pose_error['position_error']))
print("angle error degrees: %.3f" % (pose_error['angle_error_degrees']))
def eval_dynamics(config,
train_dir,
eval_dir,
state_dict_path=None,
keypoint_observation=False,
debug=False,
render_human=False):
# set random seed for reproduction
set_seed(config['train']['random_seed'])
tee = Tee(os.path.join(eval_dir, 'eval.log'), 'w')
print(config)
use_gpu = torch.cuda.is_available()
'''
model
'''
model_dy = DynaNetMLP(config)
# print model #params
print("model #params: %d" % count_trainable_parameters(model_dy))
if state_dict_path is None:
if config['eval']['eval_dy_epoch'] == -1:
state_dict_path = os.path.join(train_dir, 'net_best_dy.pth')
else:
state_dict_path = os.path.join(
train_dir, 'net_dy_epoch_%d_iter_%d.pth' % \
(config['eval']['eval_dy_epoch'], config['eval']['eval_dy_iter']))
print("Loading saved ckp from %s" % state_dict_path)
model_dy.load_state_dict(torch.load(state_dict_path))
model_dy.eval()
if use_gpu:
model_dy.cuda()
criterionMSE = nn.MSELoss()
bar = ProgressBar()
st_idx = config['eval']['eval_st_idx']
ed_idx = config['eval']['eval_ed_idx']
# load the data
episodes = load_episodes_from_config(config)
# generate action/observation functions
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
dataset = MultiEpisodeDataset(config,
action_function=action_function,
observation_function=observation_function,
episodes=episodes,
phase="valid")
episode_names = dataset.get_episode_names()
episode_names.sort()
num_episodes = None
# for backwards compatibility
if "num_episodes" in config["eval"]:
num_episodes = config["eval"]["num_episodes"]
else:
num_episodes = 10
episode_list = []
if debug:
episode_list = [episode_names[0]]
else:
episode_list = episode_names[:num_episodes]
for roll_idx, episode_name in enumerate(episode_list):
print("episode_name", episode_name)
if keypoint_observation:
eval_episode_keypoint_observations(config,
dataset,
episode_name,
roll_idx,
model_dy,
eval_dir,
start_idx=9,
n_prediction=30,
render_human=render_human)
else:
eval_episode(config,
dataset,
episode_name,
roll_idx,
model_dy,
eval_dir,
start_idx=9,
n_prediction=30,
render_human=render_human)
def train_autoencoder(config,
train_dir,
ckp_dir=None,
multi_episode_dict=None,
type=None, # ["SpatialAutoencoder", . . .]
):
assert multi_episode_dict is not None
if ckp_dir is None:
ckp_dir = os.path.join(train_dir, 'checkpoints')
if not os.path.exists(ckp_dir):
os.makedirs(ckp_dir)
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
images_dir = os.path.join(train_dir, 'images')
if not os.path.exists(images_dir):
os.makedirs(images_dir)
# save the config
save_yaml(config, os.path.join(train_dir, 'config.yaml'))
# set random seed for reproduction
set_seed(config['train_autoencoder']['random_seed'])
camera_names = [config['perception']['camera_name']]
model = None
image_preprocess_func = None
if type == "SpatialAutoencoder":
model = SpatialAutoencoder.from_global_config(config)
image_preprocess_func = functools.partial(spatial_autoencoder_image_preprocessing,
H_in=model.input_image_shape[0],
W_in=model.input_image_shape[1],
H_out=model.output_image_shape[0],
W_out=model.output_image_shape[1])
elif type == "ConvolutionalAutoencoder":
model = ConvolutionalAutoencoder.from_global_config(config)
image_preprocess_func = AutoencoderImagePreprocessFunctionFactory.convolutional_autoencoder(config)
else:
raise ValueError("unknown model type: %s" % (type))
writer = SummaryWriter(log_dir=tensorboard_dir)
# only use images from this specific config
### data
datasets = {}
dataloaders = {}
for phase in ['train', 'valid']:
datasets[phase] = AutoencoderImageDataset(config,
phase=phase,
episodes=multi_episode_dict,
camera_names=camera_names,
image_preprocess_func=image_preprocess_func)
dataloaders[phase] = DataLoader(
datasets[phase], batch_size=config['train_autoencoder']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train_autoencoder']['batch_size'])
use_gpu = torch.cuda.is_available()
params = model.parameters()
optimizer = optim.Adam(
params, lr=float(config['train_autoencoder']['lr']),
betas=(config['train_autoencoder']['adam_beta1'], 0.999))
scheduler = None
if config['train_autoencoder']['lr_scheduler']['enabled']:
scheduler = ReduceLROnPlateau(
optimizer, 'min', factor=0.6, patience=2, verbose=True)
if use_gpu:
model = model.cuda()
best_valid_loss = np.inf
global_iteration = 0
log_fout = open(os.path.join(ckp_dir, 'log.txt'), 'w')
# criterion
criterionMSE = nn.MSELoss()
# a little test
if False:
data = datasets['train'][0]
print(data.keys())
print("data['target_tensor'].shape", data['target_tensor'].shape)
print("data['target_mask'].shape", data['target_mask'].shape)
fig = plt.figure()
ax = fig.subplots(2)
target_img = data['target']
print("target_img.dtype", target_img.dtype)
ax[0].imshow(data['input'])
ax[1].imshow(data['target'], cmap='gray', vmin=0, vmax=255)
plt.show()
quit()
# a little test
if False:
data = datasets['train'][0]
print(data.keys())
print("data['target_tensor'].shape", data['target_tensor'].shape)
print("data['target_mask'].shape", data['target_mask'].shape)
fig = plt.figure()
ax = fig.subplots(2)
target_img = data['target']
target_tensor = data['target_tensor'].unsqueeze(0)
target_tensor_np = torch_utils.convert_torch_image_to_numpy(target_tensor).squeeze()
print("target_img.dtype", target_img.dtype)
ax[0].imshow(target_img)
ax[1].imshow(target_tensor_np)
plt.show()
quit()
counters = {'train': 0, 'valid': 0}
n_epoch = config['train_autoencoder']['n_epoch']
for epoch in range(n_epoch):
phases = ['train', 'valid']
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
model.train(phase == 'train')
meter_loss = AverageMeter()
loader = dataloaders[phase]
bar = ProgressBar(max_value=len(loader))
step_duration_meter = AverageMeter()
epoch_start_time = time.time()
prev_time = time.time()
print("\n\n")
for i, data in bar(enumerate(loader)):
loss_container = dict() # store the losses for this step
counters[phase] += 1
with torch.set_grad_enabled(phase == 'train'):
input = data['input_tensor']
target = data['target_tensor']
if use_gpu:
input = input.cuda()
target = target.cuda()
out = model(input)
target_pred = out['output']
# print("target.shape", target.shape)
# print("target_pred.shape", target_pred.shape)
# reconstruction loss
l2_recon = criterionMSE(target, target_pred)
loss_container['l2_recon'] = l2_recon
# loss_masked
# [B, H', W']
mask = data['target_mask'].to(target.device)
mask_idx = mask > 0
# convert to BHWC ordering so we can directly index
target_masked = target.permute(0, 2, 3, 1)[mask_idx]
target_pred_masked = target_pred.permute(0, 2, 3, 1)[mask_idx]
# print('target_masked.shape', target_masked.shape)
# print("target_pred_masked.shape", target_pred_masked.shape)
l2_recon_masked = criterionMSE(target_masked, target_pred_masked)
loss_container['l2_recon_masked'] = l2_recon_masked
# compute the loss
loss = 0
for key, val in config['train_autoencoder']['loss_function'].items():
if val['enabled']:
loss += loss_container[key] * val['weight']
meter_loss.update(loss.item())
step_duration_meter.update(time.time() - prev_time)
prev_time = time.time()
if phase == 'train':
optimizer.zero_grad()
loss.backward()
nn.utils.clip_grad_norm_(params, 1)
optimizer.step()
if global_iteration > 100:
writer.add_scalar("Params/learning rate", get_lr(optimizer), global_iteration)
# writer.add_scalar("Loss/%s" % (phase), loss.item(), global_iteration)
writer.add_scalar("Loss_train/%s" % (phase), loss.item(), counters[phase])
for loss_type, loss_obj in loss_container.items():
plot_name = "Loss/%s/%s" % (loss_type, phase)
writer.add_scalar(plot_name, loss_obj.item(), counters[phase])
if i % config['train_autoencoder']['log_per_iter'] == 0:
log = '%s [%d/%d][%d/%d] LR: %.6f, Loss: %.6f (%.6f)' % (
phase, epoch, n_epoch, i, len(loader), get_lr(optimizer),
loss.item(), meter_loss.avg)
log += ', step time %.6f' % (step_duration_meter.avg)
step_duration_meter.reset()
print(log)
log_fout.write(log + '\n')
log_fout.flush()
if phase == 'train' and i % config['train_autoencoder']['ckp_per_iter'] == 0:
torch.save(
model.state_dict(),
'%s/net_kp_epoch_%d_iter_%d.pth' % (ckp_dir, epoch, i))
if i % config['train_autoencoder']['img_save_per_iter'] == 0:
nrows = 4
ncols = 2
fig_width = 5
B, _, H, W = target.shape
fig_height = fig_width * ((nrows * H) / (ncols * W))
figsize = (fig_width, fig_height)
fig = plt.figure(figsize=figsize)
ax = fig.subplots(nrows=nrows, ncols=ncols)
target_np = torch_utils.convert_torch_image_to_numpy(target)
target_pred_np = torch_utils.convert_torch_image_to_numpy(target_pred)
for n in range(nrows):
ax[n, 0].imshow(target_np[n])
ax[n, 1].imshow(target_pred_np[n])
save_file = os.path.join(images_dir,
'%s_epoch_%d_iter_%d.png' %(phase, epoch, i))
fig.savefig(save_file)
plt.close(fig)
writer.flush() # flush SummaryWriter events to disk
global_iteration += 1
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, n_epoch, meter_loss.avg, best_valid_loss)
print(log)
print("Epoch Duration:", time.time() - epoch_start_time)
log_fout.write(log + '\n')
log_fout.flush()
if phase == 'valid':
if scheduler is not None:
scheduler.step(meter_loss.avg)
if meter_loss.avg < best_valid_loss:
best_valid_loss = meter_loss.avg
torch.save(model.state_dict(), '%s/net_best.pth' % ckp_dir)
log_fout.close()
def train_transporter(
config,
train_dir,
ckp_dir=None,
multi_episode_dict=None,
):
assert multi_episode_dict is not None
if ckp_dir is None:
ckp_dir = os.path.join(train_dir)
if not os.path.exists(ckp_dir):
os.makedirs(ckp_dir)
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
images_dir = os.path.join(train_dir, 'images')
if not os.path.exists(images_dir):
os.makedirs(images_dir)
# save the config
save_yaml(config, os.path.join(train_dir, 'config.yaml'))
# set random seed for reproduction
set_seed(config['train_transporter']['random_seed'])
writer = SummaryWriter(log_dir=tensorboard_dir)
# only use images from this specific config
camera_names = [config['perception']['camera_name']]
### data
datasets = {}
dataloaders = {}
for phase in ['train', 'valid']:
datasets[phase] = ImageTupleDataset(config,
phase=phase,
episodes=multi_episode_dict,
tuple_size=2,
camera_names=camera_names)
dataloaders[phase] = DataLoader(
datasets[phase],
batch_size=config['train_transporter']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train_transporter']['batch_size'])
use_gpu = torch.cuda.is_available()
crop_enabled = datasets['train'].crop_enabled
rgb_tensor_key = None
if crop_enabled:
rgb_image_key = "rgb_crop"
rgb_tensor_key = "rgb_crop_tensor"
else:
rgb_image_key = "rgb_masked_scaled"
rgb_tensor_key = "rgb_masked_scaled_tensor"
if False:
dataset = datasets["train"]
dataset_size = len(dataset)
print("len(dataset)", len(dataset))
print("len(dataset._image_dataset)", len(dataset._image_dataset))
print("len(dataset['valid'])", len(datasets['valid']))
print("len(dataset['train'])", len(datasets['train']))
print("dataset.crop_enabled", dataset.crop_enabled)
data = dataset[0]
print("data.keys()", data.keys())
print("data[0].keys()", data[0].keys())
# rgb_crop_tensor = data[0]['rgb_crop_tensor']
# print("rgb_crop_tensor.max()", rgb_crop_tensor.max())
# print("rgb_crop_tensor.min()", rgb_crop_tensor.min())
#
# rgb_image = data[0]['rgb_masked_scaled']
# print("rgb_crop.dtype", rgb_image.dtype)
# print("rgb_image.shape", rgb_image.shape)
rgb_image = data[0][rgb_image_key]
rgb_tensor = data[0][rgb_tensor_key]
print("rgb_image.shape", rgb_image.shape)
print("rgb_tensor.shape", rgb_tensor.shape)
plt.figure()
# plt.imshow(rgb_image)
plt.imshow(data[0][rgb_image_key])
plt.show()
quit()
#
### model
model_kp = Transporter(config, use_gpu=use_gpu)
print("model_kp #params: %d" % count_parameters(model_kp))
if config['train_transporter']['resume_epoch'] >= 0:
model_kp_path = os.path.join(
ckp_dir, 'net_kp_epoch_%d_iter_%d.pth' %
(config['train_transporter']['resume_epoch'],
config['train_transporter']['resume_iter']))
print("Loading saved ckp from %s" % model_kp_path)
model_kp.load_state_dict(torch.load(model_kp_path))
# criterion
criterionMSE = nn.MSELoss()
# optimizer
params = model_kp.parameters()
optimizer = optim.Adam(params,
lr=float(config['train_transporter']['lr']),
betas=(config['train_transporter']['adam_beta1'],
0.999))
scheduler = ReduceLROnPlateau(optimizer,
'min',
factor=0.6,
patience=2,
verbose=True)
if use_gpu:
model_kp = model_kp.cuda()
best_valid_loss = np.inf
global_iteration = 0
log_fout = open(os.path.join(ckp_dir, 'log.txt'), 'w')
n_epoch = config['train_transporter']['n_epoch']
for epoch in range(n_epoch):
phases = ['train', 'valid']
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
model_kp.train(phase == 'train')
meter_loss = AverageMeter()
loader = dataloaders[phase]
bar = ProgressBar(max_value=len(loader))
for i, data in bar(enumerate(loader)):
with torch.set_grad_enabled(phase == 'train'):
src = data[0][rgb_tensor_key]
des = data[1][rgb_tensor_key]
if use_gpu:
src = src.cuda()
des = des.cuda()
des_pred = model_kp(src, des)
# reconstruction loss
loss = criterionMSE(des_pred, des)
meter_loss.update(loss.item(), src.size(0))
if phase == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
if global_iteration > 100:
writer.add_scalar("Params/learning rate",
get_lr(optimizer), global_iteration)
writer.add_scalar("Loss/%s" % (phase), loss.item(),
global_iteration)
if i % config['train_transporter']['log_per_iter'] == 0:
log = '%s [%d/%d][%d/%d] LR: %.6f, Loss: %.6f (%.6f)' % (
phase, epoch, n_epoch, i, len(loader),
get_lr(optimizer), loss.item(), meter_loss.avg)
print()
print(log)
log_fout.write(log + '\n')
log_fout.flush()
if phase == 'train' and i % config['train_transporter'][
'ckp_per_iter'] == 0:
torch.save(
model_kp.state_dict(),
'%s/net_kp_epoch_%d_iter_%d.pth' % (ckp_dir, epoch, i))
# compute some images and draw them
if global_iteration % config['train_transporter'][
'image_per_iter'] == 0:
with torch.no_grad():
kp = model_kp.predict_keypoint(des)
heatmap = model_kp.keypoint_to_heatmap(
kp, inv_std=config['perception']['inv_std'])
images = visualize_transporter_output(
des=des, des_pred=des_pred, heatmap=heatmap, kp=kp)
print("images[0].shape", images[0].shape)
save_img = np.concatenate(images[:4], axis=0)
print("save_img.dtype", save_img.dtype)
print("save_img.shape", save_img.shape)
save_file = os.path.join(
images_dir,
'%s_epoch_%d_iter_%d.png' % (phase, epoch, i))
cv2.imwrite(save_file, save_img)
pass
writer.flush() # flush SummaryWriter events to disk
global_iteration += 1
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, n_epoch, meter_loss.avg, best_valid_loss)
print(log)
log_fout.write(log + '\n')
log_fout.flush()
if phase == 'valid':
scheduler.step(meter_loss.avg)
if meter_loss.avg < best_valid_loss:
best_valid_loss = meter_loss.avg
torch.save(model_kp.state_dict(),
'%s/net_best.pth' % ckp_dir)
log_fout.close()
def train_dynamics(config, data_path, train_dir):
# access dict values as attributes
config = edict(config)
# set random seed for reproduction
set_seed(config.train.random_seed)
st_epoch = config.train.resume_epoch if config.train.resume_epoch > 0 else 0
tee = Tee(os.path.join(train_dir, 'train_st_epoch_%d.log' % st_epoch), 'w')
print(config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(config, data_path, phase=phase)
dataloaders[phase] = DataLoader(
datasets[phase], batch_size=config.train.batch_size,
shuffle=True if phase == 'train' else False,
num_workers=config.train.num_workers)
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
'''
define model for dynamics prediction
'''
model_dy = DynaNetMLP(config)
print("model_dy #params: %d" % count_trainable_parameters(model_dy))
if config.train.resume_epoch >= 0:
# if resume from a pretrained checkpoint
model_dy_path = os.path.join(
train_dir, 'net_dy_epoch_%d_iter_%d.pth' % (
config.train.resume_epoch, config.train.resume_iter))
print("Loading saved ckp from %s" % model_dy_path)
model_dy.load_state_dict(torch.load(model_dy_path))
# criterion
criterionMSE = nn.MSELoss()
# optimizer
params = model_dy.parameters()
optimizer = optim.Adam(params, lr=config.train.lr, betas=(config.train.adam_beta1, 0.999))
scheduler = ReduceLROnPlateau(optimizer, 'min', factor=0.9, patience=10, verbose=True)
if use_gpu:
model_dy = model_dy.cuda()
best_valid_loss = np.inf
for epoch in range(st_epoch, config.train.n_epoch):
phases = ['train', 'valid']
for phase in phases:
model_dy.train(phase == 'train')
meter_loss_rmse = AverageMeter()
bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in bar(enumerate(loader)):
if use_gpu:
if isinstance(data, list):
data = [d.cuda() for d in data]
else:
data = data.cuda()
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config.train.n_history, config.train.n_rollout
n_samples = n_his + n_roll
if config.env.type in ['PusherSlider']:
states, actions = data
assert states.size(1) == n_samples
B = states.size(0)
loss_mse = 0.
# state_cur: B x n_his x state_dim
state_cur = states[:, :n_his]
for j in range(n_roll):
state_des = states[:, n_his + j]
# action_cur: B x n_his x action_dim
action_cur = actions[:, j : j + n_his] if actions is not None else None
# state_pred: B x state_dim
state_pred = model_dy(state_cur, action_cur)
loss_mse_cur = criterionMSE(state_pred, state_des)
loss_mse += loss_mse_cur / config.train.n_rollout
# update state_cur
state_cur = torch.cat([state_cur[:, 1:], state_pred.unsqueeze(1)], 1)
meter_loss_rmse.update(np.sqrt(loss_mse.item()), B)
if phase == 'train':
optimizer.zero_grad()
loss_mse.backward()
optimizer.step()
if i % config.train.log_per_iter == 0:
log = '%s [%d/%d][%d/%d] LR: %.6f' % (
phase, epoch, config.train.n_epoch, i, data_n_batches[phase],
get_lr(optimizer))
log += ', rmse: %.6f (%.6f)' % (
np.sqrt(loss_mse.item()), meter_loss_rmse.avg)
print(log)
if phase == 'train' and i % config.train.ckp_per_iter == 0:
torch.save(model_dy.state_dict(), '%s/net_dy_epoch_%d_iter_%d.pth' % (train_dir, epoch, i))
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, config.train.n_epoch, meter_loss_rmse.avg, best_valid_loss)
print(log)
if phase == 'valid':
scheduler.step(meter_loss_rmse.avg)
if meter_loss_rmse.avg < best_valid_loss:
best_valid_loss = meter_loss_rmse.avg
torch.save(model_dy.state_dict(), '%s/net_best_dy.pth' % (train_dir))
def f():
seed = np.random.RandomState().randint(0, 10000)
# print(np.random.RandomState().randint(0, 10000))
set_seed(seed)
print(np.random.rand())
def train_dynamics(
config,
train_dir, # str: directory to save output
multi_episode_dict=None,
spatial_descriptors_idx=None,
metadata=None,
spatial_descriptors_data=None,
):
assert multi_episode_dict is not None
# assert spatial_descriptors_idx is not None
# set random seed for reproduction
set_seed(config['train']['random_seed'])
st_epoch = config['train'][
'resume_epoch'] if config['train']['resume_epoch'] > 0 else 0
tee = Tee(os.path.join(train_dir, 'train_st_epoch_%d.log' % st_epoch), 'w')
tensorboard_dir = os.path.join(train_dir, "tensorboard")
if not os.path.exists(tensorboard_dir):
os.makedirs(tensorboard_dir)
writer = SummaryWriter(log_dir=tensorboard_dir)
# save the config
save_yaml(config, os.path.join(train_dir, "config.yaml"))
if metadata is not None:
save_pickle(metadata, os.path.join(train_dir, 'metadata.p'))
if spatial_descriptors_data is not None:
save_pickle(spatial_descriptors_data,
os.path.join(train_dir, 'spatial_descriptors.p'))
training_stats = dict()
training_stats_file = os.path.join(train_dir, 'training_stats.yaml')
# load the data
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.function_from_config(
config)
datasets = {}
dataloaders = {}
data_n_batches = {}
for phase in ['train', 'valid']:
print("Loading data for %s" % phase)
datasets[phase] = MultiEpisodeDataset(
config,
action_function=action_function,
observation_function=observation_function,
episodes=multi_episode_dict,
phase=phase)
dataloaders[phase] = DataLoader(
datasets[phase],
batch_size=config['train']['batch_size'],
shuffle=True if phase == 'train' else False,
num_workers=config['train']['num_workers'],
drop_last=True)
data_n_batches[phase] = len(dataloaders[phase])
use_gpu = torch.cuda.is_available()
# compute normalization parameters if not starting from pre-trained network . . .
'''
Build model for dynamics prediction
'''
model_dy = build_dynamics_model(config)
camera_name = config['vision_net']['camera_name']
# criterion
criterionMSE = nn.MSELoss()
l1Loss = nn.L1Loss()
smoothL1 = nn.SmoothL1Loss()
# optimizer
params = model_dy.parameters()
lr = float(config['train']['lr'])
optimizer = optim.Adam(params,
lr=lr,
betas=(config['train']['adam_beta1'], 0.999))
# setup scheduler
sc = config['train']['lr_scheduler']
scheduler = None
if config['train']['lr_scheduler']['enabled']:
if config['train']['lr_scheduler']['type'] == "ReduceLROnPlateau":
scheduler = ReduceLROnPlateau(optimizer,
mode='min',
factor=sc['factor'],
patience=sc['patience'],
threshold_mode=sc['threshold_mode'],
cooldown=sc['cooldown'],
verbose=True)
elif config['train']['lr_scheduler']['type'] == "StepLR":
step_size = config['train']['lr_scheduler']['step_size']
gamma = config['train']['lr_scheduler']['gamma']
scheduler = StepLR(optimizer, step_size=step_size, gamma=gamma)
else:
raise ValueError("unknown scheduler type: %s" %
(config['train']['lr_scheduler']['type']))
if use_gpu:
print("using gpu")
model_dy = model_dy.cuda()
# print("model_dy.vision_net._ref_descriptors.device", model_dy.vision_net._ref_descriptors.device)
# print("model_dy.vision_net #params: %d" %(count_trainable_parameters(model_dy.vision_net)))
best_valid_loss = np.inf
valid_loss_type = config['train']['valid_loss_type']
global_iteration = 0
counters = {'train': 0, 'valid': 0}
epoch_counter_external = 0
loss = 0
index_map = get_object_and_robot_state_indices(config)
object_state_indices = torch.LongTensor(index_map['object_indices'])
robot_state_indices = torch.LongTensor(index_map['robot_indices'])
object_state_shape = config['dataset']['object_state_shape']
try:
for epoch in range(st_epoch, config['train']['n_epoch']):
phases = ['train', 'valid']
epoch_counter_external = epoch
writer.add_scalar("Training Params/epoch", epoch, global_iteration)
for phase in phases:
# only validate at a certain frequency
if (phase == "valid") and (
(epoch % config['train']['valid_frequency']) != 0):
continue
model_dy.train(phase == 'train')
average_meter_container = dict()
step_duration_meter = AverageMeter()
# bar = ProgressBar(max_value=data_n_batches[phase])
loader = dataloaders[phase]
for i, data in enumerate(loader):
loss_container = dict() # store the losses for this step
step_start_time = time.time()
global_iteration += 1
counters[phase] += 1
with torch.set_grad_enabled(phase == 'train'):
n_his, n_roll = config['train']['n_history'], config[
'train']['n_rollout']
n_samples = n_his + n_roll
if DEBUG:
print("global iteration: %d" % (global_iteration))
print("n_samples", n_samples)
# [B, n_samples, obs_dim]
observations = data['observations']
visual_observations_list = data[
'visual_observations_list']
# [B, n_samples, action_dim]
actions = data['actions']
B = actions.shape[0]
if use_gpu:
observations = observations.cuda()
actions = actions.cuda()
# compile the visual observations
# compute the output of the visual model for all timesteps
visual_model_output_list = []
for visual_obs in visual_observations_list:
# visual_obs is a dict containing observation for a single
# time step (of course across a batch however)
# visual_obs[<camera_name>]['rgb_tensor'] has shape [B, 3, H, W]
# probably need to cast input to cuda
# [B, -1, 3]
keypoints = visual_obs[camera_name][
'descriptor_keypoints_3d_world_frame']
# [B, K, 3] where K = len(spatial_descriptors_idx)
keypoints = keypoints[:, spatial_descriptors_idx]
B, K, _ = keypoints.shape
# [B, K*3]
keypoints_reshape = keypoints.reshape([B, K * 3])
if DEBUG:
print("keypoints.shape", keypoints.shape)
print("keypoints_reshape.shape",
keypoints_reshape.shape)
visual_model_output_list.append(keypoints_reshape)
visual_model_output = None
if len(visual_model_output_list) > 0:
# concatenate this into a tensor
# [B, n_samples, vision_model_out_dim]
visual_model_output = torch.stack(
visual_model_output_list, dim=1)
else:
visual_model_output = torch.Tensor(
) # empty tensor
# states, actions = data
assert actions.shape[1] == n_samples
# cast this to float so it can be concatenated below
visual_model_output = visual_model_output.type_as(
observations)
# we don't have any visual observations, so states are observations
# states is gotten by concatenating visual_observations and observations
# [B, n_samples, vision_model_out_dim + obs_dim]
states = torch.cat((visual_model_output, observations),
dim=-1)
# state_cur: B x n_his x state_dim
# state_cur = states[:, :n_his]
# [B, n_his, state_dim]
state_init = states[:, :n_his]
# We want to rollout n_roll steps
# actions = [B, n_his + n_roll, -1]
# so we want action_seq.shape = [B, n_roll, -1]
action_start_idx = 0
action_end_idx = n_his + n_roll - 1
action_seq = actions[:, action_start_idx:
action_end_idx, :]
if DEBUG:
print("states.shape", states.shape)
print("state_init.shape", state_init.shape)
print("actions.shape", actions.shape)
print("action_seq.shape", action_seq.shape)
# try using models_dy.rollout_model instead of doing this manually
rollout_data = rollout_model(state_init=state_init,
action_seq=action_seq,
dynamics_net=model_dy,
compute_debug_data=False)
# [B, n_roll, state_dim]
state_rollout_pred = rollout_data['state_pred']
# [B, n_roll, state_dim]
state_rollout_gt = states[:, n_his:]
if DEBUG:
print("state_rollout_gt.shape",
state_rollout_gt.shape)
print("state_rollout_pred.shape",
state_rollout_pred.shape)
# the loss function is between
# [B, n_roll, state_dim]
state_pred_err = state_rollout_pred - state_rollout_gt
# [B, n_roll, object_state_dim]
object_state_err = state_pred_err[:, :,
object_state_indices]
B, n_roll, object_state_dim = object_state_err.shape
# [B, n_roll, *object_state_shape]
object_state_err_reshape = object_state_err.reshape(
[B, n_roll, *object_state_shape])
# num weights
J = object_state_err_reshape.shape[2]
weights = model_dy.weight_matrix
assert len(
weights) == J, "len(weights) = %d, but J = %d" % (
len(weights), J)
# loss mse object, note the use of broadcasting semantics
# [B, n_roll]
object_state_loss_mse = weights * torch.pow(
object_state_err_reshape, 2).sum(dim=-1)
object_state_loss_mse = object_state_loss_mse.mean()
l2_object = (weights * torch.norm(
object_state_err_reshape, dim=-1)).mean()
l2_object_final_step = (weights * torch.norm(
object_state_err_reshape[:, -1], dim=-1)).mean()
# [B, n_roll, robot_state_dim]
robot_state_err = state_pred_err[:, :,
robot_state_indices]
robot_state_loss_mse = torch.pow(robot_state_err,
2).sum(dim=-1).mean()
loss_container[
'object_state_loss_mse'] = object_state_loss_mse
loss_container[
'robot_state_loss_mse'] = robot_state_loss_mse
loss_container['l2_object'] = l2_object
loss_container[
'l2_object_final_step'] = l2_object_final_step
# total loss
loss = object_state_loss_mse + robot_state_loss_mse
loss_container['loss'] = loss
for key, val in loss_container.items():
if not key in average_meter_container:
average_meter_container[key] = AverageMeter()
average_meter_container[key].update(val.item(), B)
step_duration_meter.update(time.time() - step_start_time)
if phase == 'train':
optimizer.zero_grad()
loss.backward()
optimizer.step()
if (i % config['train']['log_per_iter']
== 0) or (global_iteration %
config['train']['log_per_iter'] == 0):
log = '%s [%d/%d][%d/%d] LR: %.6f' % (
phase, epoch, config['train']['n_epoch'], i,
data_n_batches[phase], get_lr(optimizer))
# log += ', l2: %.6f' % (loss_container['l2'].item())
# log += ', l2_final_step: %.6f' %(loss_container['l2_final_step'].item())
log += ', step time %.6f' % (step_duration_meter.avg)
step_duration_meter.reset()
print(log)
# log data to tensorboard
# only do it once we have reached 100 iterations
if global_iteration > 100:
writer.add_scalar("Params/learning rate",
get_lr(optimizer),
global_iteration)
writer.add_scalar("Loss_train/%s" % (phase),
loss.item(), global_iteration)
for loss_type, loss_obj in loss_container.items():
plot_name = "Loss/%s/%s" % (loss_type, phase)
writer.add_scalar(plot_name, loss_obj.item(),
counters[phase])
# only plot the weights if we are in the train phase . . . .
if phase == "train":
for i in range(len(weights)):
plot_name = "Weights/%d" % (i)
writer.add_scalar(plot_name,
weights[i].item(),
counters[phase])
if phase == 'train' and global_iteration % config['train'][
'ckp_per_iter'] == 0:
save_model(
model_dy, '%s/net_dy_epoch_%d_iter_%d' %
(train_dir, epoch, i))
log = '%s [%d/%d] Loss: %.6f, Best valid: %.6f' % (
phase, epoch, config['train']['n_epoch'],
average_meter_container[valid_loss_type].avg,
best_valid_loss)
print(log)
# record all average_meter losses
for key, meter in average_meter_container.items():
writer.add_scalar("AvgMeter/%s/%s" % (key, phase),
meter.avg, epoch)
if phase == "train":
if (scheduler is not None) and (
config['train']['lr_scheduler']['type']
== "StepLR"):
scheduler.step()
if phase == 'valid':
if (scheduler is not None) and (
config['train']['lr_scheduler']['type']
== "ReduceLROnPlateau"):
scheduler.step(
average_meter_container[valid_loss_type].avg)
if average_meter_container[
valid_loss_type].avg < best_valid_loss:
best_valid_loss = average_meter_container[
valid_loss_type].avg
training_stats['epoch'] = epoch
training_stats['global_iteration'] = counters['valid']
save_yaml(training_stats, training_stats_file)
save_model(model_dy, '%s/net_best_dy' % (train_dir))
writer.flush() # flush SummaryWriter events to disk
except KeyboardInterrupt:
# save network if we have a keyboard interrupt
save_model(
model_dy, '%s/net_dy_epoch_%d_keyboard_interrupt' %
(train_dir, epoch_counter_external))
writer.flush() # flush SummaryWriter events to disk
from key_dynam.dataset.vision_function_factory import VisualObservationFunctionFactory from key_dynam.dataset.online_episode_reader import OnlineEpisodeReader from key_dynam.dataset.mpc_dataset import DynamicsModelInputBuilder from key_dynam.utils import meshcat_utils from key_dynam.planner.planners import RandomShootingPlanner, PlannerMPPI from key_dynam.utils import transform_utils, torch_utils from key_dynam.experiments.exp_09 import utils as exp_utils from key_dynam.utils import drake_image_utils from key_dynam.dynamics.utils import keypoints_3D_from_dynamics_model_output, set_seed from key_dynam.eval.utils import compute_pose_error from key_dynam.models import model_builder from key_dynam.models.model_builder import build_dynamics_model from key_dynam.autoencoder.autoencoder_models import ConvolutionalAutoencoder # constants set_seed(0) USE_FIXED_MPC_HORIZON = False USE_SHORT_HORIZON_MPC = False TERMINAL_COST_ONLY = True TRAJECTORY_GOAL = False REPLAN = True SEED_WITH_NOMINAL_ACTIONS = False N = 2 * 10 # horizon length for ground truth push N = 10 MPC_HORIZON = 10 PUSHER_VELOCITY = 0.20 PUSHER_ANGLE = np.deg2rad(20)
def main():
# load dynamics model
model_dict = load_autoencoder_model()
model = model_dict['model_dy']
model_ae = model_dict['model_ae']
visual_observation_function = model_dict['visual_observation_function']
config = model.config
env_config = load_yaml(
os.path.join(get_project_root(),
'experiments/exp_18_box_on_side/config.yaml'))
env_config['env']['observation']['depth_int16'] = True
n_history = config['train']['n_history']
# create the environment
# create the environment
env = DrakePusherSliderEnv(env_config)
env.reset()
# create another environment for doing rollouts
env2 = DrakePusherSliderEnv(env_config, visualize=False)
env2.reset()
action_function = ActionFunctionFactory.function_from_config(config)
observation_function = ObservationFunctionFactory.drake_pusher_position_3D(
config)
episode = OnlineEpisodeReader()
mpc_input_builder = DynamicsModelInputBuilder(
observation_function=observation_function,
visual_observation_function=visual_observation_function,
action_function=action_function,
episode=episode)
vis = meshcat_utils.make_default_visualizer_object()
vis.delete()
initial_cond = get_initial_state()
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
obs_init = env.get_observation()
# visualize starting position of the object
print("obs_init.keys()", obs_init.keys())
print("obs_init['slider']['position']", obs_init['slider']['position'])
T = DrakePusherSliderEnv.object_position_from_observation(obs_init)
vis['start_pose'].set_object(triad(scale=0.1))
vis['state_pose'].set_transform(T)
#### ROLLOUT USING LEARNED MODEL + GROUND TRUTH ACTIONS ############
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
# add just some large number of these
episode.clear()
for i in range(n_history):
action_zero = np.zeros(2)
obs_tmp = env.get_observation()
episode.add_observation_action(obs_tmp, action_zero)
#### ROLLOUT THE ACTION SEQUENCE USING THE SIMULATOR ##########
# rollout single action sequence using the simulator
gt_rollout_data = env_utils.rollout_action_sequence(
env, initial_cond['action_sequence'].cpu().numpy())
env_obs_rollout_gt = gt_rollout_data['observations']
gt_rollout_episode = gt_rollout_data['episode_reader']
for i, env_obs in enumerate(gt_rollout_data['observations']):
T = DrakePusherSliderEnv.object_position_from_observation(env_obs)
vis_name = "GT_trajectory/%d" % (i)
vis[vis_name].set_object(triad(scale=0.1))
vis[vis_name].set_transform(T)
action_state_gt = mpc_input_builder.get_action_state_tensors(
start_idx=0, num_timesteps=N, episode=gt_rollout_episode)
state_rollout_gt = action_state_gt['states']
action_rollout_gt = action_state_gt['actions']
z_object_rollout_gt = model.compute_z_state(
state_rollout_gt)['z_object_flat']
print('state_rollout_gt.shape', state_rollout_gt.shape)
print("z_object_rollout_gt.shape", z_object_rollout_gt.shape)
def goal_func(obs_tmp):
state_tmp = mpc_input_builder.get_state_input_single_timestep(
{'observation': obs_tmp})['state']
return model.compute_z_state(
state_tmp.unsqueeze(0))['z_object'].flatten()
# using the vision model to get "goal" keypoints
z_object_goal = goal_func(env_obs_rollout_gt[-1])
z_object_goal_np = torch_utils.cast_to_numpy(z_object_goal)
# input("press Enter to continue")
#### ROLLOUT USING LEARNED MODEL + GROUND TRUTH ACTIONS ############
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
# add just some large number of these
episode.clear()
for i in range(n_history):
action_zero = np.zeros(2)
obs_tmp = env.get_observation()
episode.add_observation_action(obs_tmp, action_zero)
# [n_history, state_dim]
idx = episode.get_latest_idx()
dyna_net_input = mpc_input_builder.get_dynamics_model_input(
idx, n_history=n_history)
state_init = dyna_net_input['states'].cuda() # [n_history, state_dim]
action_init = dyna_net_input['actions'] # [n_history, action_dim]
print("state_init.shape", state_init.shape)
print("action_init.shape", action_init.shape)
print("n_history", n_history)
action_seq_gt_torch = initial_cond['action_sequence']
action_input = torch.cat(
(action_init[:(n_history - 1)], action_seq_gt_torch), dim=0).cuda()
print("action_input.shape", action_input.shape)
# rollout using the ground truth actions and learned model
# need to add the batch dim to do that
z_init = model.compute_z_state(state_init)['z']
rollout_pred = rollout_model(state_init=z_init.unsqueeze(0),
action_seq=action_input.unsqueeze(0),
dynamics_net=model,
compute_debug_data=True)
state_pred_rollout = rollout_pred['state_pred'].squeeze(0)
print("state_pred_rollout.shape", state_pred_rollout.shape)
# input("press Enter to continue")
# check L2 distance between predicted and actual
# basically comparing state_pred_rollout and state_rollout_gt
print("state_rollout_gt[-1]\n", state_rollout_gt[-1])
print("state_pred_rollout[-1]\n", state_pred_rollout[-1])
index_dict = get_object_and_robot_state_indices(config)
object_indices = index_dict['object_indices']
# reset the environment and use the MPC controller to stabilize this
# now setup the MPC to try to stabilize this . . . .
reset_environment(env, initial_cond['q_pusher'], initial_cond['q_slider'])
episode.clear()
# add just some large number of these
for i in range(n_history):
action_zero = np.zeros(2)
obs_tmp = env.get_observation()
episode.add_observation_action(obs_tmp, action_zero)
# input("press Enter to continue")
# make a planner config
planner_config = copy.copy(config)
config_tmp = load_yaml(
os.path.join(get_project_root(),
'experiments/drake_pusher_slider/eval_config.yaml'))
planner_config['mpc'] = config_tmp['mpc']
planner_config['mpc']['mppi']['terminal_cost_only'] = TERMINAL_COST_ONLY
planner_config['mpc']['random_shooting'][
'terminal_cost_only'] = TERMINAL_COST_ONLY
planner = None
if PLANNER_TYPE == "random_shooting":
planner = RandomShootingPlanner(planner_config)
elif PLANNER_TYPE == "mppi":
planner = PlannerMPPI(planner_config)
else:
raise ValueError("unknown planner type: %s" % (PLANNER_TYPE))
mpc_out = None
action_seq_mpc = None
state_pred_mpc = None
counter = -1
while True:
counter += 1
print("\n\n-----Running MPC Optimization: Counter (%d)-------" %
(counter))
obs_cur = env.get_observation()
episode.add_observation_only(obs_cur)
if counter == 0 or REPLAN:
print("replanning")
####### Run the MPC ##########
# [1, state_dim]
n_look_ahead = N - counter
if USE_FIXED_MPC_HORIZON:
n_look_ahead = MPC_HORIZON
elif USE_SHORT_HORIZON_MPC:
n_look_ahead = min(MPC_HORIZON, N - counter)
if n_look_ahead == 0:
break
start_idx = counter
end_idx = counter + n_look_ahead
print("start_idx:", start_idx)
print("end_idx:", end_idx)
print("n_look_ahead", n_look_ahead)
# start_time = time.time()
# idx of current observation
idx = episode.get_latest_idx()
mpc_start_time = time.time()
mpc_input_data = mpc_input_builder.get_dynamics_model_input(
idx, n_history=n_history)
state_cur = mpc_input_data['states']
action_his = mpc_input_data['actions']
if SEED_WITH_NOMINAL_ACTIONS:
action_seq_rollout_init = action_seq_gt_torch[
start_idx:end_idx]
else:
if mpc_out is not None:
action_seq_rollout_init = mpc_out['action_seq'][1:]
print("action_seq_rollout_init.shape",
action_seq_rollout_init.shape)
if action_seq_rollout_init.shape[0] < n_look_ahead:
num_steps = n_look_ahead - action_seq_rollout_init.shape[
0]
action_seq_zero = torch.zeros([num_steps, 2])
action_seq_rollout_init = torch.cat(
(action_seq_rollout_init, action_seq_zero), dim=0)
print("action_seq_rollout_init.shape",
action_seq_rollout_init.shape)
else:
action_seq_rollout_init = None
# run MPPI
z_cur = None
with torch.no_grad():
z_cur = model.compute_z_state(
state_cur.unsqueeze(0).cuda())['z'].squeeze(0)
if action_seq_rollout_init is not None:
n_look_ahead = action_seq_rollout_init.shape[0]
obs_goal = None
print("z_object_rollout_gt.shape", z_object_rollout_gt.shape)
if TRAJECTORY_GOAL:
obs_goal = z_object_rollout_gt[start_idx:end_idx]
print("n_look_ahead", n_look_ahead)
print("obs_goal.shape", obs_goal.shape)
# add the final goal state on as needed
if obs_goal.shape[0] < n_look_ahead:
obs_goal_final = z_object_rollout_gt[-1].unsqueeze(0)
num_repeat = n_look_ahead - obs_goal.shape[0]
obs_goal_final_expand = obs_goal_final.expand(
[num_repeat, -1])
obs_goal = torch.cat((obs_goal, obs_goal_final_expand),
dim=0)
else:
obs_goal = z_object_rollout_gt[-1]
obs_goal = torch_utils.cast_to_numpy(obs_goal)
print("obs_goal.shape", obs_goal.shape)
seed = 1
set_seed(seed)
mpc_out = planner.trajectory_optimization(
state_cur=z_cur,
action_his=action_his,
obs_goal=obs_goal,
model_dy=model,
action_seq_rollout_init=action_seq_rollout_init,
n_look_ahead=n_look_ahead,
eval_indices=object_indices,
rollout_best_action_sequence=True,
verbose=True,
add_grid_action_samples=True,
)
print("MPC step took %.4f seconds" %
(time.time() - mpc_start_time))
action_seq_mpc = torch_utils.cast_to_numpy(mpc_out['action_seq'])
state_pred_mpc = torch_utils.cast_to_numpy(mpc_out['state_pred'])
# Rollout with ground truth simulator dynamics
env2.set_simulator_state_from_observation_dict(
env2.get_mutable_context(), obs_cur)
obs_mpc_gt = env_utils.rollout_action_sequence(
env2, action_seq_mpc)['observations']
vis['mpc_3D'].delete()
vis['mpc_GT_3D'].delete()
L = len(obs_mpc_gt)
print("L", L)
if L == 0:
break
for i in range(L):
# ground truth rollout of the MPC action_seq
name = "mpc_GT_3D/%d" % (i)
T_W_obj = DrakePusherSliderEnv.object_position_from_observation(
obs_mpc_gt[i])
vis[name].set_object(triad(scale=0.1))
vis[name].set_transform(T_W_obj)
action_cur = action_seq_mpc[0]
print("action_cur", action_cur)
print("action_GT", initial_cond['action'])
input("press Enter to continue")
# add observation actions to the episode
obs_cur = env.get_observation()
episode.replace_observation_action(obs_cur, action_cur)
# step the simulator
env.step(action_cur)
# update the trajectories, in case we aren't replanning
action_seq_mpc = action_seq_mpc[1:]
state_pred_mpc = state_pred_mpc[1:]
pose_error = compute_pose_error(env_obs_rollout_gt[-1], obs_cur)
print("position_error: %.3f" % (pose_error['position_error']))
print("angle error degrees: %.3f" %
(pose_error['angle_error_degrees']))
obs_final = env.get_observation()
pose_error = compute_pose_error(env_obs_rollout_gt[-1], obs_final)
print("position_error: %.3f" % (pose_error['position_error']))
print("angle error degrees: %.3f" % (pose_error['angle_error_degrees']))
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
from progressbar import ProgressBar
from torch.optim.lr_scheduler import ReduceLROnPlateau
from torch.utils.data import Dataset, DataLoader
import torchvision.transforms as transforms
from torch.distributions.multivariate_normal import MultivariateNormal
from key_dynam.dynamics.config import gen_args
from key_dynam.dynamics.data import PhysicsDataset, load_data
from key_dynam.dynamics.models_dy import DynaNetGNN
from key_dynam.dynamics.utils import rand_int, count_trainable_parameters, Tee, AverageMeter, get_lr, to_np, set_seed
args = gen_args()
set_seed(args.random_seed)
torch.manual_seed(args.random_seed)
np.random.seed(args.random_seed)
os.system('mkdir -p ' + args.dataf)
os.system('mkdir -p ' + args.outf_dy)
tee = Tee(os.path.join(args.outf_dy, 'train.log'), 'w')
print(args)
# generate data
trans_to_tensor = transforms.Compose([
transforms.ToTensor(),
transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5))
])
# GLOBAL_TRANSLATION = np.array([-0.1, 0.0, 0]) # doesn't work with any offset
GLOBAL_TRANSLATION = np.array([0,0,0])
SEED_MPC_W_GT_ACTION_SEQUENCE = False
RANDOM_SEED = 3
OBJECT_YAW = np.deg2rad(-90)
REPLAN = True
PLANNER_TYPE = "mppi"
# PLANNER_TYPE = "random_shooting"
BOX_ON_SIDE = True
set_seed(RANDOM_SEED)
def sample_object_position(T_aug=None,
upright=False):
pos = np.array([0, 0, 0.1])
quat = None
if upright:
quat = np.array([1,0,0,0])
else:
quat = transforms3d.euler.euler2quat(np.deg2rad(90), 0, 0)
T_O_slider = transform_utils.transform_from_pose(pos, quat)
# apply a random yaw to the object
yaw = OBJECT_YAW
def run_precompute_descriptors_pipeline(multi_episode_dict, # dict
model, # dense descriptor model file
model_file=None,
output_dir=None, # str where to save data
episode_name=None, # optional for descriptor sampling
camera_name=None, # which camera to compute descriptors for
episode_idx=None, # optional for descriptor sampling
visualize=True,
K=5,
position_diff_threshold=20,
seed=0,
):
assert model_file is not None
assert camera_name is not None
#
# ## Load Model
# model_train_dir = os.path.dirname(model_file)
#
# print("model_train_dir", model_train_dir)
# print("model_file", model_file)
# model = torch.load(model_file)
# model = model.cuda()
# model = model.eval()
#
# output_dir = os.path.join(model_train_dir,
# 'precomputed_vision_data/descriptor_keypoints/dataset_%s/' % (dataset_name))
# compute descriptor confidence scores
set_seed(seed)
if True:
print("\n\n---------Computing Descriptor Confidence Scores-----------")
metadata_file = os.path.join(output_dir, 'metadata.p')
if os.path.isfile(metadata_file):
answer = input("metadata.p file already exists, do you want to overwrite it? y/n\n")
if answer == "y":
shutil.rmtree(output_dir)
print("removing existing file and continuing")
else:
print("aborting")
quit()
compute_descriptor_confidences(multi_episode_dict,
model,
output_dir,
batch_size=10,
num_workers=20,
model_file=model_file,
camera_name=camera_name,
num_ref_descriptors=50,
num_batches=10,
episode_name_arg=episode_name,
episode_idx=episode_idx,
)
if True:
confidence_score_data_file = os.path.join(output_dir, 'data.p')
confidence_score_data = load_pickle(confidence_score_data_file)
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
print("\n\n---------Selecting Spatially Separated Keypoints-----------")
score_and_select_spatially_separated_keypoints(metadata,
confidence_score_data=confidence_score_data,
K=K,
position_diff_threshold=position_diff_threshold,
output_dir=output_dir,
)
# visualize descriptors
if True:
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
episode_name = metadata['episode_name']
episode_idx = metadata['episode_idx']
camera_name = metadata['camera_name']
episode = multi_episode_dict[episode_name]
data = episode.get_image_data(camera_name, episode_idx)
rgb = data['rgb']
uv = metadata['indices']
print("uv.shape", uv.shape)
color = [0, 255, 0]
draw_reticles(rgb, uv[:, 0], uv[:, 1], label_color=color)
save_file = os.path.join(output_dir, 'sampled_descriptors.png')
plt.figure()
plt.imshow(rgb)
plt.savefig(save_file)
if visualize:
plt.show()
# visualize spatially separated descriptors
if True:
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
spatial_descriptor_file = os.path.join(output_dir, 'spatial_descriptors.p')
spatial_descriptors_data = load_pickle(spatial_descriptor_file)
des_idx = spatial_descriptors_data['spatial_descriptors_idx']
episode_name = metadata['episode_name']
episode_idx = metadata['episode_idx']
camera_name = metadata['camera_name']
episode = multi_episode_dict[episode_name]
data = episode.get_image_data(camera_name, episode_idx)
rgb = data['rgb']
uv = metadata['indices']
print("uv.shape", uv.shape)
color = [0, 255, 0]
draw_reticles(rgb, uv[des_idx, 0], uv[des_idx, 1], label_color=color)
save_file = os.path.join(output_dir, 'spatially_separated_descriptors.png')
plt.figure()
plt.imshow(rgb)
plt.savefig(save_file)
if visualize:
plt.show()
if True:
metadata_file = os.path.join(output_dir, 'metadata.p')
metadata = load_pickle(metadata_file)
print("\n\n---------Precomputing Descriptor Keypoints-----------")
descriptor_keypoints_output_dir = os.path.join(output_dir, "descriptor_keypoints")
precompute_descriptor_keypoints(multi_episode_dict,
model,
descriptor_keypoints_output_dir,
ref_descriptors_metadata=metadata,
batch_size=8,
num_workers=20,
camera_names=[camera_name]
)
print("Data saved at: ", output_dir)
print("Finished Normally")
