Skip to content

Open-Sora: Democratizing Efficient Video Production for All

Python

fix dist issue

Shen-Chenhuic7a698e85bf7
diff --cc configs/vae_magvit_v2/train/16x128x128_pixabay_8_GPU_pipeline.py
index 8405bfa,e3aa372..86f8711
--- a/configs/vae_magvit_v2/train/16x128x128_pixabay_8_GPU_pipeline.py
+++ b/configs/vae_magvit_v2/train/16x128x128_pixabay_8_GPU_pipeline.py
@@@ -84,9 -84,9 +84,11 @@@ magvit uses about # samples (K) * epoch
epochs = 200
log_every = 1
- ckpt_every = 1 # 50
-ckpt_every = 5
++
++ckpt_every = 50
load = None
- batch_size = 4 # 32
-batch_size = 4
++batch_size = 32
++
lr = 1e-4
grad_clip = 1.0
diff --cc scripts/train-vae-v2.py
index b6cdd25,f386546..4967935
--- a/scripts/train-vae-v2.py
+++ b/scripts/train-vae-v2.py
@@@ -224,14 -235,13 +238,14 @@@ def main()
booster.load_lr_scheduler(lr_scheduler, os.path.join(cfg.load, "lr_scheduler"))
if disc_lr_scheduler is not None:
booster.load_lr_scheduler(disc_lr_scheduler, os.path.join(cfg.load, "disc_lr_scheduler"))
- lecam_path = os.path.join(cfg.load, "lecam_state.json")
++
+ # LeCam EMA for discriminator
+ lecam_path = os.path.join(cfg.load, "lecam_states.json")
if cfg.lecam_loss_weight is not None and os.path.exists(lecam_path):
lecam_state = load_json(lecam_path)
lecam_ema_real, lecam_ema_fake = lecam_state["lecam_ema_real"], lecam_state["lecam_ema_fake"]
lecam_ema = LeCamEMA(decay=cfg.ema_decay, ema_real=lecam_ema_real, ema_fake=lecam_ema_fake, dtype=dtype, device=device)
-- else:
- print(f"lecan not loaded, path: {lecam_path}, lecame loss weight {cfg.lecam_loss_weight}")
- lecam_ema = LeCamEMA(decay=cfg.ema_decay, dtype=dtype, device=device)
++
running_states = load_json(os.path.join(cfg.load, "running_states.json"))
dist.barrier()
start_epoch, start_step, sampler_start_idx = running_states["epoch"], running_states["step"], running_states["sample_start_index"]
@@@ -274,9 -284,10 +288,6 @@@
disc_time_padding = 0
video_contains_first_frame = cfg.video_contains_first_frame
--
-- # lecam_ema_real = torch.tensor(0.0)
-- # lecam_ema_fake = torch.tensor(0.0)
-
for epoch in range(start_epoch, cfg.epochs):
dataloader.sampler.set_epoch(epoch)
@@@ -405,10 -416,10 +416,7 @@@
real_video = real_video if cfg.gradient_penalty_loss_weight is not None else None,
)
disc_loss = weighted_d_adversarial_loss + lecam_loss + gradient_penalty_loss
- if cfg.ema_decay is not None:
- # SCH: TODO: is this written properly like this for moving average? e.g. distributed training etc.
- # lecam_ema_real = lecam_ema_real * cfg.ema_decay + (1 - cfg.ema_decay) * torch.mean(real_logits.clone().detach())
- # lecam_ema_fake = lecam_ema_fake * cfg.ema_decay + (1 - cfg.ema_decay) * torch.mean(fake_logits.clone().detach())
+ if cfg.lecam_loss_weight is not None:
- # SCH: TODO: is this written properly like this for moving average? e.g. distributed training etc.
- # lecam_ema_real = lecam_ema_real * cfg.ema_decay + (1 - cfg.ema_decay) * torch.mean(real_logits.clone().detach())
- # lecam_ema_fake = lecam_ema_fake * cfg.ema_decay + (1 - cfg.ema_decay) * torch.mean(fake_logits.clone().detach())
ema_real = torch.mean(real_logits.clone().detach()).to(device, dtype)
ema_fake = torch.mean(fake_logits.clone().detach()).to(device, dtype)
all_reduce_mean(ema_real)
@@@ -463,11 -474,15 +471,11 @@@
# Save checkpoint
if cfg.ckpt_every > 0 and (global_step + 1) % cfg.ckpt_every == 0:
save_dir = os.path.join(exp_dir, f"epoch{epoch}-global_step{global_step+1}")
- os.makedirs(os.path.join(save_dir, "model"), exist_ok=True)
+ os.makedirs(os.path.join(save_dir, "model"), exist_ok=True) # already handled in booster save_model
booster.save_model(vae, os.path.join(save_dir, "model"), shard=True)
- print("model saved")
booster.save_model(discriminator, os.path.join(save_dir, "discriminator"), shard=True)
- print("discriminator saved")
booster.save_optimizer(optimizer, os.path.join(save_dir, "optimizer"), shard=True, size_per_shard=4096)
- print("optimizer saved")
booster.save_optimizer(disc_optimizer, os.path.join(save_dir, "disc_optimizer"), shard=True, size_per_shard=4096)
- print("disc opt saved")
if lr_scheduler is not None:
booster.save_lr_scheduler(lr_scheduler, os.path.join(save_dir, "lr_scheduler"))
@@@ -480,17 -504,7 +488,18 @@@
"global_step": global_step+1,
"sample_start_index": (step+1) * cfg.batch_size,
}
- save_json(running_states, os.path.join(save_dir, "running_states.json"))
+
+ lecam_ema_real, lecam_ema_fake = lecam_ema.get()
+ lecam_state = {
+ "lecam_ema_real": lecam_ema_real.item(),
+ "lecam_ema_fake": lecam_ema_fake.item(),
+ }
+ if coordinator.is_master():
+ save_json(running_states, os.path.join(save_dir, "running_states.json"))
+ if cfg.lecam_loss_weight is not None:
+ save_json(lecam_state, os.path.join(save_dir, "lecam_states.json"))
+ dist.barrier()
++
logger.info(
f"Saved checkpoint at epoch {epoch} step {step + 1} global_step {global_step + 1} to {exp_dir}"
)