File size: 6,418 Bytes
da7bf91 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 | import os
import sys
from pathlib import Path
import torch
from torch.utils.data import DataLoader, Subset
from torch.utils.tensorboard import SummaryWriter
from accelerate import Accelerator
from accelerate.logging import get_logger
from tqdm.auto import tqdm
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from graphwm.config_graph import GraphWMArgs
from graphwm.dataset.collate_graph_wm import collate_graph_wm
from graphwm.dataset.dataset_graph_wm import GraphWorldModelDataset, SampledDataGraphWorldModelDataset
from graphwm.models.ctrl_world_graph import CtrlWorldGraph
def build_datasets(args: GraphWMArgs):
if args.use_sampled_data_loader:
full_dataset = SampledDataGraphWorldModelDataset(
sample_root=args.sampled_data_root,
type_vocab=args.graph_type_vocab,
session_id=args.sampled_session_id,
episode_id=args.sampled_episode_id,
num_history=args.num_history,
num_frames=args.num_frames,
resize_hw=args.sampled_resize_hw,
include_depth=args.include_depth,
)
else:
full_dataset = GraphWorldModelDataset(args.graph_manifest_path, args.graph_type_vocab)
if not args.use_eval_split:
return full_dataset, None
dataset_len = len(full_dataset)
val_len = max(1, int(dataset_len * args.val_ratio))
if dataset_len - val_len < 1:
val_len = max(1, dataset_len - 1)
train_len = dataset_len - val_len
train_indices = list(range(0, train_len))
val_indices = list(range(train_len, dataset_len))
return Subset(full_dataset, train_indices), Subset(full_dataset, val_indices)
def evaluate(model, loader, accelerator):
model.eval()
total = 0.0
count = 0
with torch.no_grad():
for batch in loader:
with accelerator.autocast():
loss_gen, _ = model(batch)
avg_loss = accelerator.gather(loss_gen.detach().reshape(1)).mean()
total += float(avg_loss.item())
count += 1
model.train()
return total / max(count, 1)
def main(args: GraphWMArgs):
logger = get_logger(__name__, log_level="INFO")
accelerator = Accelerator(
gradient_accumulation_steps=args.gradient_accumulation_steps,
mixed_precision=args.mixed_precision,
)
model = CtrlWorldGraph(args)
if args.ckpt_path:
state_dict = torch.load(args.ckpt_path, map_location="cpu")
model.load_state_dict(state_dict, strict=False)
train_dataset, val_dataset = build_datasets(args)
train_loader = DataLoader(
train_dataset,
batch_size=args.train_batch_size,
shuffle=args.shuffle,
num_workers=args.num_workers,
collate_fn=collate_graph_wm,
)
val_loader = None
if val_dataset is not None:
val_loader = DataLoader(
val_dataset,
batch_size=args.eval_batch_size,
shuffle=False,
num_workers=args.num_workers,
collate_fn=collate_graph_wm,
)
optimizer = torch.optim.AdamW(model.parameters(), lr=args.learning_rate)
if val_loader is not None:
model, optimizer, train_loader, val_loader = accelerator.prepare(model, optimizer, train_loader, val_loader)
else:
model, optimizer, train_loader = accelerator.prepare(model, optimizer, train_loader)
writer = None
if accelerator.is_main_process and args.use_tensorboard:
os.makedirs(args.tensorboard_log_dir, exist_ok=True)
writer = SummaryWriter(log_dir=args.tensorboard_log_dir)
model.train()
global_step = 0
running_loss = 0.0
running_count = 0
progress_bar = tqdm(range(args.max_train_steps), disable=not accelerator.is_local_main_process)
progress_bar.set_description("Graph WM Steps")
if accelerator.is_main_process:
logger.info("Train samples: %s", len(train_dataset))
if val_dataset is not None:
logger.info("Val samples: %s", len(val_dataset))
while global_step < args.max_train_steps:
for batch in train_loader:
with accelerator.accumulate(model):
with accelerator.autocast():
loss_gen, _ = model(batch)
avg_loss = accelerator.gather(loss_gen.detach().reshape(1)).mean()
running_loss += float(avg_loss.item())
running_count += 1
accelerator.backward(loss_gen)
if accelerator.sync_gradients:
accelerator.clip_grad_norm_(model.parameters(), args.max_grad_norm)
optimizer.step()
optimizer.zero_grad()
if accelerator.sync_gradients:
global_step += 1
progress_bar.update(1)
progress_bar.set_postfix({"loss": float(avg_loss.item())})
if global_step % args.log_every_steps == 0:
train_loss = running_loss / max(running_count, 1)
if accelerator.is_main_process:
logger.info("step=%s train_loss=%.6f", global_step, train_loss)
if writer is not None:
writer.add_scalar("loss/train", train_loss, global_step)
running_loss = 0.0
running_count = 0
if val_loader is not None and global_step % args.validation_steps == 0:
val_loss = evaluate(model, val_loader, accelerator)
if accelerator.is_main_process:
logger.info("step=%s val_loss=%.6f", global_step, val_loss)
if writer is not None:
writer.add_scalar("loss/val", val_loss, global_step)
if global_step % args.checkpointing_steps == 0 and accelerator.is_main_process:
os.makedirs(args.output_dir, exist_ok=True)
save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}.pt")
torch.save(accelerator.unwrap_model(model).state_dict(), save_path)
logger.info("Saved checkpoint to %s", save_path)
if global_step >= args.max_train_steps:
break
if writer is not None:
writer.close()
if __name__ == "__main__":
args = GraphWMArgs()
main(args)
|