import os
import torch
import numpy as np
from pathlib import Path
from abc import ABC, abstractmethod
from xuance.common import Optional, List, Union
from argparse import Namespace
from operator import itemgetter
from xuance.torch import Tensor, Module
MAX_GPUs = torch.cuda.device_count()
[docs]
class Learner(ABC):
def __init__(self,
config: Namespace,
policy: Module,
callback):
self.value_normalizer = None
self.config = config
self.distributed_training = config.distributed_training
self.episode_length = config.episode_length
self.learning_rate = config.learning_rate if hasattr(config, 'learning_rate') else None
self.use_linear_lr_decay = config.use_linear_lr_decay if hasattr(config, 'use_linear_lr_decay') else False
self.end_factor_lr_decay = config.end_factor_lr_decay if hasattr(config, 'end_factor_lr_decay') else 1.0
self.gamma = config.gamma if hasattr(config, 'gamma') else 0.99
self.use_rnn = config.use_rnn if hasattr(config, 'use_rnn') else False
self.use_actions_mask = config.use_actions_mask if hasattr(config, 'use_actions_mask') else False
self.policy = policy
self.optimizer: Union[dict, list, Optional[torch.optim.Optimizer]] = None
self.scheduler: Union[dict, list, Optional[torch.optim.lr_scheduler.LinearLR]] = None
self.callback = callback
if self.distributed_training:
self.world_size = int(os.environ['WORLD_SIZE'])
self.rank = self.device = int(os.environ['RANK'])
self.snapshot_path = os.path.join(os.getcwd(), config.model_dir, "DDP_Snapshot")
if os.path.exists(self.snapshot_path):
if os.path.exists(os.path.join(self.snapshot_path, "snapshot.pt")):
print("Loading Snapshot...")
self.load_snapshot(self.snapshot_path)
else:
if self.device == 0:
os.makedirs(self.snapshot_path)
else:
self.world_size = 1
self.rank = 0
self.device = config.device
self.use_grad_clip = config.use_grad_clip
self.grad_clip_norm = config.grad_clip_norm
self.device = config.device
self.model_dir = config.model_dir
self.total_iters = self.estimate_total_iterations()
self.iterations = 0
[docs]
def estimate_total_iterations(self):
"""Estimated total number of training iterations"""
start_training = getattr(self.config, "start_training", 0)
training_frequency = getattr(self.config, "training_frequency", 1)
total_iters = (self.config.running_steps - start_training) // (training_frequency * self.config.parallels)
return total_iters
[docs]
def save_model(self, model_path):
if type(self.optimizer) is dict:
torch.save(
{
'policy': self.policy.state_dict(),
'optimizer': {k: v.state_dict() for k, v in self.optimizer.items()},
'rng_state': torch.get_rng_state(),
'cuda_rng_state': torch.cuda.get_rng_state_all(),
},
model_path)
else:
torch.save(
{
'policy': self.policy.state_dict(),
'optimizer': self.optimizer.state_dict(),
'rng_state': torch.get_rng_state(),
'cuda_rng_state': torch.cuda.get_rng_state_all(),
},
model_path)
if self.distributed_training:
self.save_snapshot()
[docs]
def load_model(self, path, model=None):
target_path = os.path.join(path, model) if model is not None else path
if os.path.isfile(target_path): # load the specified model file
model_path = target_path
dir_name = os.path.dirname(model_path)
else:
if not os.path.isdir(path):
raise RuntimeError(f"The path '{path}' is not a valid directory or file!")
folder_names = [f for f in os.listdir(path) if "seed_" in f]
folder_names.sort()
if not folder_names:
raise RuntimeError(f"No model files with 'seed_' found in '{path}'!")
path = Path(os.path.join(path, folder_names[-1]))
dir_name = str(path)
model_names = list(path.glob("*.pth"))
model_path = None
if len(model_names) == 0:
raise FileNotFoundError(f"No .pth file found in {path}")
else:
for f in model_names:
if "final_train_model.pth" in str(f):
model_path = f
break
model_path = str(model_names)
checkpoint = torch.load(str(model_path), map_location={f"cuda:{i}": self.device
for i in range(MAX_GPUs)}, weights_only=True)
self.policy.load_state_dict(checkpoint['policy'], strict=False)
if 'optimizer' in checkpoint and self.optimizer is not None:
if type(self.optimizer) is dict:
for k, v in self.optimizer.items():
v.load_state_dict(checkpoint['optimizer'][k])
else:
self.optimizer.load_state_dict(checkpoint['optimizer'])
current_lr = self.optimizer.param_groups[0]['lr']
self.learning_rate = current_lr
if 'rng_state' in checkpoint:
rng_state = checkpoint['rng_state']
rng_state = rng_state.cpu().to(dtype=torch.uint8)
torch.set_rng_state(rng_state)
if 'cuda_rng_state' in checkpoint and torch.cuda.is_available():
cuda_states = checkpoint['cuda_rng_state']
if isinstance(cuda_states, list):
num_available_gpus = torch.cuda.device_count()
cuda_states = cuda_states[:num_available_gpus]
for i, state in enumerate(cuda_states):
state = state.cpu().to(dtype=torch.uint8)
torch.cuda.set_rng_state(state, device=i)
self._safe_scheduler_step()
print(f"Successfully load model from '{model_path}'.")
return dir_name
[docs]
def load_snapshot(self, snapshot_path):
loc = f"cuda:{self.device}" if torch.cuda.is_available() else "cpu"
snapshot = torch.load(snapshot_path, map_location=loc)
if "MODEL_STATE" in snapshot:
self.policy.load_state_dict(snapshot["MODEL_STATE"])
elif "policy" in snapshot:
self.policy.load_state_dict(snapshot["policy"])
if "optimizer" in snapshot and self.optimizer is not None:
self.optimizer.load_state_dict(snapshot["optimizer"])
if "rng_state" in snapshot:
torch.set_rng_state(snapshot["rng_state"].to('cpu'))
if "cuda_rng_state" in snapshot and torch.cuda.is_available():
cuda_states = snapshot["cuda_rng_state"]
if isinstance(cuda_states, list):
for i, state in enumerate(cuda_states):
torch.cuda.set_rng_state(state.to(f'cuda:{i}'), device=i)
print("Resuming training from snapshot (including optimizer/rng state).")
[docs]
def save_snapshot(self):
snapshot = {
"policy": self.policy.state_dict(),
"optimizer": self.optimizer.state_dict(),
"rng_state": torch.get_rng_state(),
"cuda_rng_state": torch.cuda.get_rng_state_all() if torch.cuda.is_available() else None,
}
snapshot_pt = os.path.join(self.snapshot_path, "snapshot.pt")
os.makedirs(self.snapshot_path, exist_ok=True)
torch.save(snapshot, snapshot_pt)
def _safe_scheduler_step(self):
if not hasattr(self, 'scheduler'):
return
if not hasattr(self.config, 'rt_epoch'):
return
try:
train_steps = self.config.running_steps // self.config.parallels
eval_interval = self.config.eval_interval // self.config.parallels
num_epoch = int(train_steps / eval_interval)
current_iters = int(self.total_iters * self.config.rt_epoch / num_epoch)
self.scheduler.step(current_iters)
print(f"scheduler.step success，rt_epoch={self.config.rt_epoch}")
except TypeError as e:
if "positional argument" in str(e) or "takes 1 positional argument" in str(e):
self.scheduler.step()
print(f"scheduler.step success, rt_epoch={self.config.rt_epoch}")
except Exception as e:
print(f"scheduler.step failure：{e}")
[docs]
@abstractmethod
def update(self, *args):
raise NotImplementedError
[docs]
class LearnerMAS(Learner):
def __init__(self,
config: Namespace,
model_keys: List[str],
agent_keys: List[str],
policy: Module,
callback):
self.value_normalizer = None
self.config = config
self.distributed_training = config.distributed_training
self.n_agents = config.n_agents
self.dim_id = self.n_agents
self.use_parameter_sharing = config.use_parameter_sharing
self.model_keys = model_keys
self.agent_keys = agent_keys
self.episode_length = config.episode_length
self.learning_rate = getattr(config, 'learning_rate', None)
self.use_linear_lr_decay = getattr(config, 'use_linear_lr_decay', False)
self.end_factor_lr_decay = getattr(config, 'end_factor_lr_decay', 0.5)
self.gamma = getattr(config, 'gamma', 0.99)
self.use_cnn = getattr(config, "use_cnn", False)
self.use_rnn = getattr(config, 'use_rnn', False)
self.use_actions_mask = getattr(config, 'use_actions_mask', False)
self.policy = policy
self.optimizer: Union[dict, list, Optional[torch.optim.Optimizer]] = None
self.scheduler: Union[dict, list, Optional[torch.optim.lr_scheduler.LinearLR]] = None
self.callback = callback
self.use_grad_clip = config.use_grad_clip
self.grad_clip_norm = config.grad_clip_norm
self.device = config.device
self.model_dir = config.model_dir
self.total_iters = self.estimate_total_iterations()
self.iterations = 0
[docs]
def estimate_total_iterations(self):
"""Estimated total number of training iterations"""
start_training = getattr(self.config, "start_training", 0)
training_frequency = getattr(self.config, "training_frequency", 1)
n_epochs = getattr(self.config, "n_epochs", 1)
episode_length = self.episode_length
if self.use_rnn:
total_iters = (self.config.running_steps - start_training) // (episode_length * self.config.parallels)
else:
total_iters = (self.config.running_steps - start_training) // (training_frequency * self.config.parallels)
total_iters *= n_epochs
return total_iters
[docs]
def build_training_data(self, sample: Optional[dict],
use_parameter_sharing: Optional[bool] = False,
use_actions_mask: Optional[bool] = False,
use_global_state: Optional[bool] = False):
"""
Prepare the training data.
Parameters:
sample (dict): The raw sampled data.
use_parameter_sharing (bool): Whether to use parameter sharing for individual agent models.
use_actions_mask (bool): Whether to use actions mask for unavailable actions.
use_global_state (bool): Whether to use global state.
Returns:
sample_Tensor (dict): The formatted sampled data.
"""
batch_size = sample['batch_size']
seq_length = sample['sequence_length'] if self.use_rnn else 1
state, avail_actions, filled = None, None, None
obs_next, state_next, avail_actions_next = None, None, None
IDs = None
if use_parameter_sharing:
k = self.model_keys[0]
bs = batch_size * self.n_agents
if self.n_agents == 1:
obs_tensor = Tensor(sample['obs'][k]).to(self.device).unsqueeze(1)
actions_tensor = Tensor(sample['actions'][k]).to(self.device).unsqueeze(1)
rewards_tensor = Tensor(sample['rewards'][k]).to(self.device).unsqueeze(1)
ter_tensor = Tensor(sample['terminals'][k]).float().to(self.device).unsqueeze(1)
msk_tensor = Tensor(sample['agent_mask'][k]).float().to(self.device).unsqueeze(1)
else:
obs_tensor = Tensor(np.stack(itemgetter(*self.agent_keys)(sample['obs']),
axis=1)).to(self.device)
actions_tensor = Tensor(np.stack(itemgetter(*self.agent_keys)(sample['actions']),
axis=1)).to(self.device)
rewards_tensor = Tensor(np.stack(itemgetter(*self.agent_keys)(sample['rewards']),
axis=1)).to(self.device)
ter_tensor = Tensor(np.stack(itemgetter(*self.agent_keys)(sample['terminals']),
axis=1)).float().to(self.device)
msk_tensor = Tensor(np.stack(itemgetter(*self.agent_keys)(sample['agent_mask']),
axis=1)).float().to(self.device)
if self.use_cnn and len(obs_tensor.shape) > 3: # obs_array consists of images
obs_shape_item = obs_tensor.shape[2:]
else:
obs_shape_item = (-1,)
if self.use_rnn:
obs = {k: obs_tensor.reshape(bs, seq_length + 1, -1)}
if len(actions_tensor.shape) == 3:
actions = {k: actions_tensor.reshape(bs, seq_length)}
elif len(actions_tensor.shape) == 4:
actions = {k: actions_tensor.reshape(bs, seq_length, -1)}
else:
raise AttributeError("Wrong actions shape.")
rewards = {k: rewards_tensor.reshape(batch_size, self.n_agents, seq_length)}
terminals = {k: ter_tensor.reshape(batch_size, self.n_agents, seq_length)}
agent_mask = {k: msk_tensor.reshape(bs, seq_length)}
IDs = torch.eye(self.n_agents).unsqueeze(1).unsqueeze(0).expand(
batch_size, -1, seq_length + 1, -1).reshape(bs, seq_length + 1, self.n_agents).to(self.device)
else:
obs = {k: obs_tensor.reshape(bs, *obs_shape_item)}
if len(actions_tensor.shape) == 2:
actions = {k: actions_tensor.reshape(bs)}
elif len(actions_tensor.shape) == 3:
actions = {k: actions_tensor.reshape(bs, -1)}
else:
raise AttributeError("Wrong actions shape.")
rewards = {k: rewards_tensor.reshape(batch_size, self.n_agents)}
terminals = {k: ter_tensor.reshape(batch_size, self.n_agents)}
agent_mask = {k: msk_tensor.reshape(bs)}
obs_next = {k: Tensor(np.stack(itemgetter(*self.agent_keys)(sample['obs_next']),
axis=1)).to(self.device).reshape(bs, *obs_shape_item)}
IDs = torch.eye(self.n_agents).unsqueeze(0).expand(
batch_size, -1, -1).reshape(bs, self.n_agents).to(self.device)
if use_actions_mask:
avail_a = np.stack(itemgetter(*self.agent_keys)(sample['avail_actions']), axis=1)
if self.use_rnn:
avail_actions = {k: Tensor(avail_a.reshape([bs, seq_length + 1, -1])).float().to(self.device)}
else:
avail_actions = {k: Tensor(avail_a.reshape([bs, -1])).float().to(self.device)}
avail_a_next = np.stack(itemgetter(*self.agent_keys)(sample['avail_actions_next']), axis=1)
avail_actions_next = {k: Tensor(avail_a_next.reshape([bs, -1])).float().to(self.device)}
else:
obs = {k: Tensor(sample['obs'][k]).to(self.device) for k in self.agent_keys}
actions = {k: Tensor(sample['actions'][k]).to(self.device) for k in self.agent_keys}
rewards = {k: Tensor(sample['rewards'][k]).to(self.device) for k in self.agent_keys}
terminals = {k: Tensor(sample['terminals'][k]).float().to(self.device) for k in self.agent_keys}
agent_mask = {k: Tensor(sample['agent_mask'][k]).float().to(self.device) for k in self.agent_keys}
if not self.use_rnn:
obs_next = {k: Tensor(sample['obs_next'][k]).to(self.device) for k in self.agent_keys}
if use_actions_mask:
avail_actions = {k: Tensor(sample['avail_actions'][k]).float().to(self.device) for k in self.agent_keys}
if not self.use_rnn:
avail_actions_next = {k: Tensor(sample['avail_actions_next'][k]).float().to(self.device) for k in
self.model_keys}
if use_global_state:
state = Tensor(sample['state']).to(self.device)
if not self.use_rnn:
state_next = Tensor(sample['state_next']).to(self.device)
if self.use_rnn:
filled = Tensor(sample['filled']).float().to(self.device)
sample_Tensor = {
'batch_size': batch_size,
'state': state,
'state_next': state_next,
'obs': obs,
'actions': actions,
'obs_next': obs_next,
'rewards': rewards,
'terminals': terminals,
'agent_mask': agent_mask,
'avail_actions': avail_actions,
'avail_actions_next': avail_actions_next,
'agent_ids': IDs,
'filled': filled,
'seq_length': seq_length,
}
return sample_Tensor
[docs]
@abstractmethod
def update(self, *args):
raise NotImplementedError
[docs]
def update_rnn(self, *args):
raise NotImplementedError
[docs]
def save_model(self, model_path):
if type(self.optimizer) is dict:
if type(list(self.optimizer.values())[0]) is dict:
torch.save(
{
'policy': self.policy.state_dict(),
'optimizer': {k_a: {k: v.state_dict() for k, v in v_a.items()}
for k_a, v_a in self.optimizer.items()}, # agent-wise
'rng_state': torch.get_rng_state(),
'cuda_rng_state': torch.cuda.get_rng_state_all(),
},
model_path)
else:
torch.save(
{
'policy': self.policy.state_dict(),
'optimizer': {k: v.state_dict() for k, v in self.optimizer.items()},
'rng_state': torch.get_rng_state(),
'cuda_rng_state': torch.cuda.get_rng_state_all(),
},
model_path)
else:
torch.save(
{
'policy': self.policy.state_dict(),
'optimizer': self.optimizer.state_dict(),
'rng_state': torch.get_rng_state(),
'cuda_rng_state': torch.cuda.get_rng_state_all(),
},
model_path)
[docs]
def load_model(self, path, model=None):
target_path = os.path.join(path, model) if model is not None else path
if os.path.isfile(target_path): # load the specified model file
model_path = target_path
dir_name = os.path.dirname(model_path)
else:
if not os.path.isdir(path):
raise RuntimeError(f"The path '{path}' is not a valid directory or file!")
folder_names = [f for f in os.listdir(path) if "seed_" in f]
folder_names.sort()
if not folder_names:
raise RuntimeError(f"No model files with 'seed_' found in '{path}'!")
path = Path(os.path.join(path, folder_names[-1]))
dir_name = str(path)
model_names = list(path.glob("*.pth"))
model_path = None
if len(model_names) == 0:
raise FileNotFoundError(f"No .pth file found in {path}")
else:
for f in model_names:
if "final_train_model.pth" in str(f):
model_path = f
break
model_path = str(model_names)
checkpoint = torch.load(str(model_path), map_location={f"cuda:{i}": self.device
for i in range(MAX_GPUs)}, weights_only=True)
self.policy.load_state_dict(checkpoint['policy'], strict=False)
if 'optimizer' in checkpoint and self.optimizer is not None:
if type(self.optimizer) is dict:
if type(list(self.optimizer.values())[0]) is dict:
for k_a, v_a in self.optimizer.items(): # agent-wise
for k, v in v_a.items():
v.load_state_dict(checkpoint['optimizer'][k_a][k])
else:
for k, v in self.optimizer.items():
v.load_state_dict(checkpoint['optimizer'][k])
else:
self.optimizer.load_state_dict(checkpoint['optimizer'])
current_lr = self.optimizer.param_groups[0]['lr']
self.learning_rate = current_lr
if 'rng_state' in checkpoint:
rng_state = checkpoint['rng_state']
rng_state = rng_state.cpu().to(dtype=torch.uint8)
torch.set_rng_state(rng_state)
if 'cuda_rng_state' in checkpoint and torch.cuda.is_available():
cuda_states = checkpoint['cuda_rng_state']
if isinstance(cuda_states, list):
num_available_gpus = torch.cuda.device_count()
cuda_states = cuda_states[:num_available_gpus]
for i, state in enumerate(cuda_states):
state = state.cpu().to(dtype=torch.uint8)
torch.cuda.set_rng_state(state, device=i)
self._safe_scheduler_step()
print(f"Successfully load model from '{model_path}'.")
return dir_name