hma / train_multi_diffusion.py
LeroyWaa's picture
draft
246c106
import logging
import math
import os
import mup
import numpy as np
import torch
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import set_seed
from torch.utils.data import DataLoader
from tqdm.auto import tqdm
import transformers
from transformers import (
default_data_collator,
get_scheduler,
)
import wandb
from cont_data import RawFeatureDataset, get_maskgit_collator_feature
from genie.config import DiffusionGenieConfig
from genie.st_mar import STMAR
from datetime import datetime
from accelerate import DistributedDataParallelKwargs
from common import data_sampler
import yaml
from train_diffusion import parse_args, train
# Get current date and time
now = datetime.now()
# Format the datetime object as a string
formatted_date = now.strftime("%Y-%m-%d %H:%M:%S")
torch.set_float32_matmul_precision("medium")
logger = get_logger(__name__)
torch.autograd.set_detect_anomaly(True)
def parse_args_multi():
# parser = argparse.ArgumentParser(description="Train a MaskGIT or Llama-style LLM on video generation.")
parser = parse_args()
# Data
parser.add_argument(
"--train_split", type=str, default="experiments/datasplit/dataset2.yaml",
help="Config files for using multiple datasets."
)
parser.add_argument(
"--num_episodes_per_dataset",
type=int,
default=1000000,
help="Maximum number of trajectories per dataset",
)
parser.add_argument(
"--image_maskgit_path",
type=str,
default=None,
help="Optional path to the official MaskGIT checkpoint. "
"If specified, will copy relevant weights from the checkpoint. "
"These weights will have a different (hard-coded) warmup schedule.",
)
parser.add_argument(
"--action_network",
type=str,
default=None,
choices=["concat", "cross_attention"], # TODO: add other methods (resampler_concat, modulate, etc)
help="If specified, will override the action in the config. Helps reduce the number of config jsons."
)
args = parser.parse_args()
return args
def main():
args = parse_args_multi()
assert (args.llama_config is not None) ^ (args.genie_config is not None), \
"Exactly one of `llama_config` and `genie_config` should be set."
# Manual gradient accumulation
ddp_kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
accelerator = Accelerator(gradient_accumulation_steps=1, log_with=args.report_to,
even_batches=False, project_dir=args.output_dir, kwargs_handlers=[ddp_kwargs])
accelerator.init_trackers("video")
if accelerator.is_main_process:
accelerator.trackers[0].run.name = formatted_date + "_" + args.run_name
# Make one log on every process with the configuration for debugging.
logging.basicConfig(
format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
datefmt="%m/%d/%Y %H:%M:%S",
level=logging.INFO,
)
logger.info(accelerator.state, main_process_only=False)
if accelerator.is_local_main_process:
transformers.utils.logging.set_verbosity_info()
else:
transformers.utils.logging.set_verbosity_error()
if args.seed is not None:
set_seed(args.seed)
if accelerator.is_main_process:
os.makedirs(args.output_dir, exist_ok=True)
accelerator.wait_for_everyone()
# create multiple datasets
with open(args.train_split, 'r') as file:
datasplit = yaml.safe_load(file)
config = DiffusionGenieConfig.from_pretrained(args.genie_config)
# Extract the 'domains' value and split it into a list
domains_list = [domain.strip() for domain in datasplit['domains'].split(',')]
train_datasets = []
val_datasets = []
dataset_num_samples = []
val_dataset_num_samples = []
action_dimensions = []
action_stats = []
shared_keys = ("s", "h", "w", "vocab_size", "latent_channels",
"encoder_type", "encoder_name_or_path", "quantized") # TODO: check train/val hz per dataset?
for domain in domains_list:
try:
# train_data_dir = f"data/{domain}_vae_traj500_train" # {args.num_episodes_per_dataset}
# val_data_dir = f"data/{domain}_vae_traj500_val"
train_data_dir = f"data/{domain}_noquant_temporalvae_shard0_of_1_train" # {args.num_episodes_per_dataset}
val_data_dir = f"data/{domain}_noquant_temporalvae_shard0_of_1_val"
# train_data_dir = f"data/{domain}_vae_traj{args.num_episodes_per_dataset}_train" # {args.num_episodes_per_dataset}
# val_data_dir = f"data/{domain}_vae_traj{args.num_episodes_per_dataset}_val"
if config.drop_action_ratio > 0:
raise NotImplementedError
train_dataset = RawFeatureDataset(train_data_dir, window_size=args.window_size,
stride=args.stride, filter_overlaps=args.filter_overlaps,
max_traj_num=args.num_episodes_per_dataset,
use_actions=config.use_actions, domain=domain)
dataset_num_samples.append(len(train_dataset))
action_dimensions.append(train_dataset.n_action)
if config.use_actions:
action_stats.append(train_dataset.action_stat)
if not args.overfit_first_batch:
eval_dataset = RawFeatureDataset(val_data_dir, window_size=args.window_size,
stride=args.stride, filter_overlaps=True,
use_actions=config.use_actions, domain=domain)
else:
train_dataset.valid_start_inds = train_dataset.valid_start_inds[:args.per_device_train_batch_size
* args.gradient_accumulation_steps
* accelerator.num_processes]
eval_dataset = train_dataset
# Shuffle eval dataset and then set shuffle=False on the dataloader.
# Shuffling in the dataloader results in reshuffling with each iteration.
eval_dataset.valid_start_inds = torch.tensor(eval_dataset.valid_start_inds)[
torch.randperm(len(eval_dataset), generator=torch.Generator().manual_seed(0))
].tolist()
val_dataset_num_samples.append(len(eval_dataset))
except Exception as e:
import traceback
print(traceback.format_exc())
train_datasets.append(train_dataset)
val_datasets.append(eval_dataset)
assert all(train_dataset.metadata.get(shared_key) == eval_dataset.metadata.get(shared_key)
for shared_key in shared_keys) # TODO: check this across all datasets
print("dataset_num_samples:", dataset_num_samples)
# Will not store key in metadata if it's missing, so that defaults can be filled by functions later? # TODO: handle missing keys
shared_metadata = {shared_key: train_dataset.metadata[shared_key]
for shared_key in shared_keys if shared_key in train_dataset.metadata}
config.use_mup = args.mu_transfer # Note: changing this may affect pre-trained model due to attn scaling
config.image_vocab_size = None
config.T = args.window_size
config.S = shared_metadata["h"] * shared_metadata["w"] # TODO: make STMaskGIT use h and w instead of S
config.vae_embed_dim = shared_metadata["latent_channels"]
if args.action_network is not None:
print("Using action network", args.action_network)
config.action_network = args.action_network
model = STMAR(config)
if config.use_actions:
# TODO: use new list instead of domains_list, in case domain fails
model.init_action_projectors(domains_list, action_dimensions, action_stats, config.action_network)
if args.image_maskgit_path is not None:
model.init_weights()
model.load_pretrained_image_weights(args.image_maskgit_path)
if args.mu_transfer:
model.set_mup_shapes(rescale_params=False)
elif args.mu_transfer:
model.set_mup_shapes(rescale_params=True)
# model.init_weights() # might be unnecessary if `rescale_params` is True
# Optimizer. Split weights in two groups, one with weight decay and the other not.
opt_class = mup.MuAdamW if args.mu_transfer else torch.optim.AdamW
# scale base learning rate
effective_batch_size = args.per_device_train_batch_size * args.gradient_accumulation_steps \
* accelerator.num_processes
args.learning_rate = args.learning_rate * min(max(1, effective_batch_size / 64), 8)
no_decay = ["bias", "layer_norm.weight"]
pretrained_params = { # more accurately the params we want lower lr for, some weights like pos_embed_TSC are pre-trained but not treated as lower lr
param_name
for param_name, _ in model.named_parameters()
if any(term in param_name for term in ("spatial_attn.qkv", "spatial_attn.proj", "mlp"))
} if args.image_maskgit_path is not None else set()
# Give pre-trained weights 10x lower learning rate
optimizer_grouped_parameters = [
{
"params": [p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay) and n not in pretrained_params],
"weight_decay": args.weight_decay,
"lr": args.learning_rate,
},
{
"params": [p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay) and n not in pretrained_params],
"weight_decay": 0.0,
"lr": args.learning_rate,
},
{
"params": [p for n, p in model.named_parameters()
if not any(nd in n for nd in no_decay) and n in pretrained_params],
"weight_decay": args.weight_decay,
"lr": args.learning_rate * 0.1,
},
{
"params": [p for n, p in model.named_parameters()
if any(nd in n for nd in no_decay) and n in pretrained_params],
"weight_decay": 0.0,
"lr": args.learning_rate * 0.1,
},
]
optimizer = opt_class(optimizer_grouped_parameters, lr=args.learning_rate,
betas=(args.adam_beta_1, args.adam_beta_2), eps=args.adam_eps)
# DataLoaders creation:
collate_fn = default_data_collator if args.llama_config is not None else get_maskgit_collator_feature(config)
combined_dataset = torch.utils.data.ConcatDataset(train_datasets)
batch_sampler = data_sampler.MultiTaskBatchSampler(
dataset_num_samples,
batch_size=args.per_device_train_batch_size,
temperature=3. # the higher the more flat the distribution
)
dataset_traj_image = data_sampler.make_dataset_pie_plot(domains_list, dataset_num_samples)
accelerator.log(({"dataset_mixture": wandb.Image(dataset_traj_image)}), log_kwargs={"wandb": {"commit": False}})
dataset_weights = batch_sampler.generate_tasks_distribution().cpu().numpy()
dataset_weight_image = data_sampler.make_dataset_pie_plot(domains_list, dataset_weights)
accelerator.log(({"dataset_mixture_weight": wandb.Image(dataset_weight_image)}), log_kwargs={"wandb": {"commit": False}})
train_dataloader = DataLoader(combined_dataset, batch_sampler=batch_sampler, collate_fn=collate_fn,
num_workers=24, pin_memory=False)
batch_val_sampler = data_sampler.MultiTaskBatchSampler(
val_dataset_num_samples,
batch_size=args.per_device_train_batch_size,
temperature=4. # the higher the more flat the distribution
)
combined_val_dataset = torch.utils.data.ConcatDataset(val_datasets)
eval_dataloader = DataLoader(combined_val_dataset, batch_sampler=batch_val_sampler, collate_fn=collate_fn,
num_workers=24, pin_memory=False)
# Scheduler and math around the number of training steps.
overrode_max_train_steps = False
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if args.max_train_steps is None:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
overrode_max_train_steps = True
if args.max_train_steps < 2000 and args.resume_from_checkpoint is None: # minimal number of trainng steps
args.max_train_steps = 2000
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
if args.lr_scheduler_type == "custom_cosine": # decay to `end_ratio` of the peak learning rate
def get_lr_wrapper(warmup_steps, max_steps, end_ratio=0.1):
def get_lr(step):
if step < warmup_steps:
return (step + 1) / warmup_steps
remaining_steps = max_steps - warmup_steps
return ((1 + math.cos(math.pi * (step - warmup_steps) / remaining_steps)) / 2) \
* (1 - end_ratio) + end_ratio
return get_lr
lr_scheduler = torch.optim.lr_scheduler.LambdaLR(
optimizer, get_lr_wrapper(args.num_warmup_steps * accelerator.num_processes,
args.max_train_steps if overrode_max_train_steps
else args.max_train_steps * accelerator.num_processes)
)
else:
lr_scheduler = get_scheduler(
name=args.lr_scheduler_type,
optimizer=optimizer,
num_warmup_steps=args.num_warmup_steps * accelerator.num_processes,
num_training_steps=args.max_train_steps
if overrode_max_train_steps
else args.max_train_steps * accelerator.num_processes,
)
# Enable gradient checkpointing to save memory
if args.gradient_checkpointing:
logger.info("Enabling gradient checkpointing")
model.gradient_checkpointing_enable()
model.config.use_cache = False # incompatible with grad checkpointing
# Prepare everything with our `accelerator`.
accelerator.wait_for_everyone()
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler = accelerator.prepare(
model, optimizer, train_dataloader, eval_dataloader, lr_scheduler
)
if not args.no_compile:
torch._dynamo.config.cache_size_limit = 256
torch._dynamo.config.optimize_ddp = False # https://github.com/pytorch/pytorch/issues/104674
# TODO: https://github.com/pytorch/pytorch/issues/109774#issuecomment-2046633776
model = torch.compile(model)
# We need to recalculate our total training steps as the size of the training dataloader may have changed.
num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
if overrode_max_train_steps:
args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
# Afterwards we recalculate our number of training epochs
args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
# Figure out how many steps we should save the Accelerator states
checkpointing_steps = args.checkpointing_steps
if checkpointing_steps is not None and checkpointing_steps.isdigit():
checkpointing_steps = int(checkpointing_steps)
# We need to initialize the trackers we use, and also store our configuration.
# The trackers initialize automatically on the main process.
experiment_config = vars(args) | vars(config)
seq_len = shared_metadata["h"] * shared_metadata["w"] * args.window_size
effective_batch_size = args.per_device_train_batch_size * args.gradient_accumulation_steps \
* accelerator.num_processes
args.num_datasets = len(train_datasets)
model_module = model.module if hasattr(model, "module") else model
experiment_config.update(shared_metadata | {
"model_parameters": sum(p.numel() for p in model.parameters()),
"model_parameters_M": round(sum(p.numel() for p in model.parameters()) / 1e6),
"trunk_parameters": sum(p.numel() for p in model_module.decoder.parameters()),
"trunk_parameters_M": round(sum(p.numel() for p in model_module.decoder.parameters()) / 1e6),
"seq_len": seq_len,
"train_data_tokens": len(train_dataset) * seq_len,
"effective_batch_size": effective_batch_size,
"effective_batch_size_tokens": effective_batch_size * seq_len,
"mixed_precision": accelerator.mixed_precision,
"num_datasets": args.num_datasets
})
experiment_config["FLOPs_per_update_step"] = 6 * experiment_config["model_parameters"] \
* experiment_config["effective_batch_size_tokens"]
accelerator.init_trackers(project_name="video", config=experiment_config)
# Train!
train(accelerator, model, optimizer, lr_scheduler, train_dataloader, eval_dataloader, experiment_config, config, args)
if __name__ == "__main__":
main()