| import io |
| import os |
| import math |
| import time |
| import json |
| import shutil |
| from collections import defaultdict, deque |
| import datetime |
| import numpy as np |
| from timm.utils import get_state_dict |
| from torch.utils.data._utils.collate import default_collate |
| from pathlib import Path |
| import subprocess |
| import torch |
| import torch.distributed as dist |
| from torch import inf |
| import random |
|
|
| from tensorboardX import SummaryWriter |
| from aurora.model.fish import create_mask_gradient_dynamic |
|
|
| class SmoothedValue(object): |
| """Track a series of values and provide access to smoothed values over a |
| window or the global series average. |
| """ |
|
|
| def __init__(self, window_size=20, fmt=None): |
| if fmt is None: |
| fmt = "{global_avg:.3f}" |
| self.deque = deque(maxlen=window_size) |
| self.total = 0.0 |
| self.count = 0 |
| self.fmt = fmt |
|
|
| def update(self, value, n=1): |
| self.deque.append(value) |
| self.count += n |
| self.total += value * n |
|
|
| def synchronize_between_processes(self): |
| """ |
| Warning: does not synchronize the deque! |
| """ |
| if not is_dist_avail_and_initialized(): |
| return |
| t = torch.tensor([self.count, self.total], dtype=torch.float64, device='cuda') |
| dist.all_reduce(t) |
| t = t.tolist() |
| self.count = int(t[0]) |
| self.total = t[1] |
|
|
| @property |
| def median(self): |
| if len(self.deque) == 0: |
| return 0.0 |
| d = torch.tensor(list(self.deque)) |
| return d.median().item() |
|
|
| @property |
| def avg(self): |
| if len(self.deque) == 0: |
| return 0.0 |
| d = torch.tensor(list(self.deque), dtype=torch.float32) |
| return d.mean().item() |
|
|
| @property |
| def global_avg(self): |
| if self.count == 0: |
| return 0.0 |
| return self.total / self.count |
|
|
| @property |
| def max(self): |
| if len(self.deque) == 0: |
| return 0.0 |
| return max(self.deque) |
|
|
| @property |
| def value(self): |
| if len(self.deque) == 0: |
| return 0.0 |
| return self.deque[-1] |
|
|
| def __str__(self): |
| return self.fmt.format( |
| median=self.median, |
| avg=self.avg, |
| global_avg=self.global_avg, |
| max=self.max, |
| value=self.value) |
|
|
|
|
| class MetricLogger(object): |
| def __init__(self, delimiter="\t"): |
| self.meters = defaultdict(SmoothedValue) |
| self.delimiter = delimiter |
|
|
| def update(self, **kwargs): |
| for k, v in kwargs.items(): |
| if v is None: |
| continue |
| if isinstance(v, torch.Tensor): |
| v = v.item() |
| assert isinstance(v, (float, int)) |
| self.meters[k].update(v) |
|
|
| def __getattr__(self, attr): |
| if attr in self.meters: |
| return self.meters[attr] |
| if attr in self.__dict__: |
| return self.__dict__[attr] |
| raise AttributeError("'{}' object has no attribute '{}'".format( |
| type(self).__name__, attr)) |
|
|
| def __str__(self): |
| loss_str = [] |
| for name, meter in self.meters.items(): |
| loss_str.append( |
| "{}: {}".format(name, str(meter)) |
| ) |
| return self.delimiter.join(loss_str) |
|
|
| def synchronize_between_processes(self): |
| if not is_dist_avail_and_initialized(): |
| return |
| |
| local_keys = sorted(self.meters.keys()) |
| gathered_keys = [None for _ in range(dist.get_world_size())] |
| dist.all_gather_object(gathered_keys, local_keys) |
| union_keys = sorted({k for keys in gathered_keys for k in keys}) |
| for k in union_keys: |
| if k not in self.meters: |
| self.meters[k] = SmoothedValue() |
| for k in union_keys: |
| self.meters[k].synchronize_between_processes() |
|
|
| def add_meter(self, name, meter): |
| self.meters[name] = meter |
|
|
| def log_every(self, iterable, print_freq, header=None): |
| i = 0 |
| if not header: |
| header = '' |
| start_time = time.time() |
| end = time.time() |
| iter_time = SmoothedValue(fmt='{avg:.4f}') |
| data_time = SmoothedValue(fmt='{avg:.4f}') |
| space_fmt = ':' + str(len(str(len(iterable)))) + 'd' |
| log_msg = [ |
| header, |
| '[{0' + space_fmt + '}/{1}]', |
| 'eta: {eta}', |
| '{meters}', |
| 'time: {time}', |
| 'data: {data}' |
| ] |
| if torch.cuda.is_available(): |
| log_msg.append('max mem: {memory:.0f}') |
| log_msg = self.delimiter.join(log_msg) |
| MB = 1024.0 * 1024.0 |
| for obj in iterable: |
| data_time.update(time.time() - end) |
| yield obj |
| iter_time.update(time.time() - end) |
| if i % print_freq == 0 or i == len(iterable) - 1: |
| eta_seconds = iter_time.global_avg * (len(iterable) - i) |
| eta_string = str(datetime.timedelta(seconds=int(eta_seconds))) |
| if torch.cuda.is_available(): |
| print(log_msg.format( |
| i, len(iterable), eta=eta_string, |
| meters=str(self), |
| time=str(iter_time), data=str(data_time), |
| memory=torch.cuda.max_memory_allocated() / MB)) |
| else: |
| print(log_msg.format( |
| i, len(iterable), eta=eta_string, |
| meters=str(self), |
| time=str(iter_time), data=str(data_time))) |
| i += 1 |
| end = time.time() |
| total_time = time.time() - start_time |
| total_time_str = str(datetime.timedelta(seconds=int(total_time))) |
| print('{} Total time: {} ({:.4f} s / it)'.format( |
| header, total_time_str, total_time / len(iterable))) |
|
|
|
|
| class TensorboardLogger(object): |
| def __init__(self, log_dir): |
| self.writer = SummaryWriter(logdir=log_dir) |
| self.step = 0 |
|
|
| def set_step(self, step=None): |
| if step is not None: |
| self.step = step |
| else: |
| self.step += 1 |
|
|
| def update(self, head='scalar', step=None, **kwargs): |
| for k, v in kwargs.items(): |
| if v is None: |
| continue |
| if isinstance(v, torch.Tensor): |
| v = v.item() |
| assert isinstance(v, (float, int)) |
| self.writer.add_scalar(head + "/" + k, v, self.step if step is None else step) |
|
|
| def flush(self): |
| self.writer.flush() |
|
|
|
|
| class WandbLogger(object): |
| def __init__(self, project, run_name=None, config=None, entity=None, tags=None, mode="online", dir=None): |
| try: |
| import wandb |
| except Exception as e: |
| raise RuntimeError( |
| "Weights & Biases logging is enabled, but `wandb` is not installed. " |
| "Install it with `pip install wandb`." |
| ) from e |
| self._wandb = wandb |
| self.step = 0 |
| self._pending = {} |
| init_kwargs = { |
| "project": project, |
| "name": run_name if run_name else None, |
| "config": config if config is not None else None, |
| "entity": entity if entity else None, |
| "tags": tags if tags else None, |
| "mode": mode if mode else "online", |
| "dir": dir if dir else None, |
| } |
| self._run = self._wandb.init(**{k: v for k, v in init_kwargs.items() if v is not None}) |
|
|
| def set_step(self, step=None): |
| |
| if len(self._pending) > 0: |
| self._wandb.log(self._pending, step=self.step, commit=True) |
| self._pending = {} |
| if step is not None: |
| self.step = int(step) |
| else: |
| self.step += 1 |
|
|
| def update(self, head='scalar', step=None, **kwargs): |
| target_step = self.step if step is None else int(step) |
| if target_step != self.step: |
| |
| if len(self._pending) > 0: |
| self._wandb.log(self._pending, step=self.step, commit=True) |
| self._pending = {} |
| self.step = target_step |
| for k, v in kwargs.items(): |
| if v is None: |
| continue |
| if isinstance(v, torch.Tensor): |
| v = v.item() |
| assert isinstance(v, (float, int)) |
| self._pending[f"{head}/{k}"] = float(v) |
|
|
| def flush(self): |
| if len(self._pending) > 0: |
| self._wandb.log(self._pending, step=self.step, commit=True) |
| self._pending = {} |
|
|
| def close(self): |
| self.flush() |
| try: |
| self._wandb.finish() |
| except Exception: |
| pass |
|
|
|
|
| class CompositeLogger(object): |
| """Fan-out logger that mirrors TensorboardLogger API.""" |
|
|
| def __init__(self, loggers): |
| self.loggers = [lg for lg in loggers if lg is not None] |
| self.step = 0 |
|
|
| def set_step(self, step=None): |
| for lg in self.loggers: |
| lg.set_step(step) |
| if step is not None: |
| self.step = int(step) |
| else: |
| self.step += 1 |
|
|
| def update(self, head='scalar', step=None, **kwargs): |
| for lg in self.loggers: |
| lg.update(head=head, step=step, **kwargs) |
|
|
| def flush(self): |
| for lg in self.loggers: |
| if hasattr(lg, "flush"): |
| lg.flush() |
|
|
| def close(self): |
| for lg in self.loggers: |
| if hasattr(lg, "close"): |
| lg.close() |
|
|
| def seed_worker(worker_id): |
| worker_seed = torch.initial_seed() % 2**32 |
| np.random.seed(worker_seed) |
| random.seed(worker_seed) |
| |
| def _load_checkpoint_for_ema(model_ema, checkpoint): |
| """ |
| Workaround for ModelEma._load_checkpoint to accept an already-loaded object |
| """ |
| mem_file = io.BytesIO() |
| torch.save(checkpoint, mem_file) |
| mem_file.seek(0) |
| model_ema._load_checkpoint(mem_file) |
|
|
|
|
| def setup_for_distributed(is_master): |
| """ |
| This function disables printing when not in master process |
| """ |
| import builtins as __builtin__ |
| builtin_print = __builtin__.print |
|
|
| def print(*args, **kwargs): |
| force = kwargs.pop('force', False) |
| if is_master or force: |
| builtin_print(*args, **kwargs) |
|
|
| __builtin__.print = print |
|
|
|
|
| def is_dist_avail_and_initialized(): |
| if not dist.is_available(): |
| return False |
| if not dist.is_initialized(): |
| return False |
| return True |
|
|
|
|
| def get_world_size(): |
| if not is_dist_avail_and_initialized(): |
| return 1 |
| return dist.get_world_size() |
|
|
|
|
| def get_rank(): |
| if not is_dist_avail_and_initialized(): |
| return 0 |
| return dist.get_rank() |
|
|
|
|
| def is_main_process(): |
| return get_rank() == 0 |
|
|
|
|
| def save_on_master(*args, **kwargs): |
| if is_main_process(): |
| torch.save(*args, **kwargs) |
|
|
|
|
| def init_distributed_mode(args): |
| if args.dist_on_itp: |
| args.rank = int(os.environ['OMPI_COMM_WORLD_RANK']) |
| args.world_size = int(os.environ['OMPI_COMM_WORLD_SIZE']) |
| args.gpu = int(os.environ['OMPI_COMM_WORLD_LOCAL_RANK']) |
| args.dist_url = "tcp://%s:%s" % (os.environ['MASTER_ADDR'], os.environ['MASTER_PORT']) |
| os.environ['LOCAL_RANK'] = str(args.gpu) |
| os.environ['RANK'] = str(args.rank) |
| os.environ['WORLD_SIZE'] = str(args.world_size) |
| elif 'SLURM_PROCID' in os.environ: |
| args.rank = int(os.environ['SLURM_PROCID']) |
| args.gpu = int(os.environ['SLURM_LOCALID']) |
| args.world_size = int(os.environ['SLURM_NTASKS']) |
| os.environ['RANK'] = str(args.rank) |
| os.environ['LOCAL_RANK'] = str(args.gpu) |
| os.environ['WORLD_SIZE'] = str(args.world_size) |
|
|
| node_list = os.environ['SLURM_NODELIST'] |
| addr = subprocess.getoutput( |
| f'scontrol show hostname {node_list} | head -n1') |
| if 'MASTER_ADDR' not in os.environ: |
| os.environ['MASTER_ADDR'] = addr |
| elif 'RANK' in os.environ and 'WORLD_SIZE' in os.environ: |
| args.rank = int(os.environ["RANK"]) |
| args.world_size = int(os.environ['WORLD_SIZE']) |
| args.gpu = int(os.environ['LOCAL_RANK']) |
| else: |
| print('Not using distributed mode') |
| args.distributed = False |
| return |
|
|
| args.distributed = True |
|
|
| torch.cuda.set_device(args.gpu) |
| args.dist_backend = 'nccl' |
| print('| distributed init (rank {}): {}, gpu {}'.format( |
| args.rank, args.dist_url, args.gpu), flush=True) |
| torch.distributed.init_process_group(backend=args.dist_backend, init_method=args.dist_url, |
| world_size=args.world_size, rank=args.rank, timeout=datetime.timedelta(seconds=3600)) |
| torch.distributed.barrier(device_ids=[args.gpu]) |
| |
| setup_for_distributed(args.rank == 0) |
|
|
|
|
| def load_state_dict(model, state_dict, prefix='', ignore_missing="relative_position_index"): |
| missing_keys = [] |
| unexpected_keys = [] |
| error_msgs = [] |
| metadata = getattr(state_dict, '_metadata', None) |
| state_dict = state_dict.copy() |
| if metadata is not None: |
| state_dict._metadata = metadata |
|
|
| def load(module, prefix=''): |
| local_metadata = {} if metadata is None else metadata.get( |
| prefix[:-1], {}) |
| module._load_from_state_dict( |
| state_dict, prefix, local_metadata, True, missing_keys, unexpected_keys, error_msgs) |
| for name, child in module._modules.items(): |
| if child is not None: |
| load(child, prefix + name + '.') |
|
|
| load(model, prefix=prefix) |
|
|
| warn_missing_keys = [] |
| ignore_missing_keys = [] |
| for key in missing_keys: |
| keep_flag = True |
| for ignore_key in ignore_missing.split('|'): |
| if ignore_key in key: |
| keep_flag = False |
| break |
| if keep_flag: |
| warn_missing_keys.append(key) |
| else: |
| ignore_missing_keys.append(key) |
|
|
| missing_keys = warn_missing_keys |
|
|
| if len(missing_keys) > 0: |
| print( |
| "Weights of {} not initialized from pretrained model (count={}).".format( |
| model.__class__.__name__, |
| len(missing_keys), |
| ) |
| ) |
| if len(unexpected_keys) > 0: |
| print( |
| "Weights from pretrained model not used in {} (count={}).".format( |
| model.__class__.__name__, |
| len(unexpected_keys), |
| ) |
| ) |
| if len(ignore_missing_keys) > 0: |
| print( |
| "Ignored weights of {} not initialized from pretrained model (count={}).".format( |
| model.__class__.__name__, |
| len(ignore_missing_keys), |
| ) |
| ) |
| if len(error_msgs) > 0: |
| print('\n'.join(error_msgs)) |
|
|
|
|
| class NativeScalerWithGradNormCount: |
| state_dict_key = "amp_scaler" |
|
|
| def __init__(self): |
| self._scaler = torch.GradScaler('cuda') |
|
|
| def __call__( |
| self, |
| loss, |
| optimizer, |
| clip_grad=None, |
| parameters=None, |
| create_graph=False, |
| update_grad=True, |
| use_ours=False, |
| weight=None, |
| k_value=0.001, |
| component_masker=None, |
| ): |
| self._scaler.scale(loss).backward(create_graph=create_graph) |
| if update_grad: |
| self._scaler.unscale_(optimizer) |
|
|
| if component_masker is not None: |
| if hasattr(component_masker, "prepare"): |
| component_masker.prepare() |
| elif hasattr(component_masker, "update_and_apply"): |
| component_masker.update_and_apply() |
| elif use_ours: |
| |
| fish_mask = create_mask_gradient_dynamic( |
| keep_ratio=k_value, |
| grad_type='absolute', |
| optimizer=optimizer, |
| weight=weight, |
| ) |
| for fish_params in optimizer.param_groups: |
| for fish_param, fish_name in zip(fish_params['params'], fish_params['names']): |
| if 'ours' not in fish_name and 'encoder.' not in fish_name and 'decoder.' not in fish_name: |
| device = fish_param.device |
| fish_mask[fish_name] = fish_mask[fish_name].to(device) |
| fish_param.grad.data.copy_(fish_param.grad.data * fish_mask[fish_name].data) |
|
|
| if clip_grad is not None: |
| assert parameters is not None |
| norm = torch.nn.utils.clip_grad_norm_(parameters, clip_grad) |
| else: |
| norm = get_grad_norm_(parameters) |
| if component_masker is not None and hasattr(component_masker, "capture_pre_step"): |
| component_masker.capture_pre_step() |
| self._scaler.step(optimizer) |
| if component_masker is not None and hasattr(component_masker, "apply_post_step"): |
| component_masker.apply_post_step() |
| self._scaler.update() |
| else: |
| norm = None |
| return norm |
|
|
| def state_dict(self): |
| return self._scaler.state_dict() |
|
|
| def load_state_dict(self, state_dict): |
| self._scaler.load_state_dict(state_dict) |
|
|
|
|
| def get_grad_norm_(parameters, norm_type: float = 2.0) -> torch.Tensor: |
| if isinstance(parameters, torch.Tensor): |
| parameters = [parameters] |
| parameters = [p for p in parameters if p.grad is not None] |
| norm_type = float(norm_type) |
| if len(parameters) == 0: |
| return torch.tensor(0.) |
| device = parameters[0].grad.device |
| if norm_type == inf: |
| total_norm = max(p.grad.detach().abs().max().to(device) for p in parameters) |
| else: |
| total_norm = torch.norm(torch.stack([torch.norm(p.grad.detach(), norm_type).to(device) for p in parameters]), norm_type) |
| return total_norm |
|
|
|
|
| def cosine_scheduler(base_value, final_value, epochs, niter_per_ep, warmup_epochs=0, |
| start_warmup_value=0, warmup_steps=-1): |
| warmup_schedule = np.array([]) |
| total_iters = epochs * niter_per_ep |
| warmup_iters = warmup_epochs * niter_per_ep |
| if warmup_steps > 0: |
| warmup_iters = warmup_steps |
| warmup_iters = max(0, min(int(warmup_iters), int(total_iters))) |
| print("Set warmup steps = %d" % warmup_iters) |
| if warmup_iters > 0: |
| warmup_schedule = np.linspace(start_warmup_value, base_value, warmup_iters) |
|
|
| iters = np.arange(total_iters - warmup_iters) |
| if len(iters) > 0: |
| schedule = np.array( |
| [ |
| final_value |
| + 0.5 * (base_value - final_value) * (1 + math.cos(math.pi * i / len(iters))) |
| for i in iters |
| ] |
| ) |
| else: |
| schedule = np.array([]) |
|
|
| schedule = np.concatenate((warmup_schedule, schedule)) |
|
|
| assert len(schedule) == total_iters |
| return schedule |
|
|
|
|
| def save_model(args, epoch, model, model_without_ddp, optimizer, loss_scaler, model_ema=None, keep_last_n=None): |
| output_dir = Path(args.output_dir) |
| epoch_name = str(epoch) |
| if loss_scaler is not None: |
| checkpoint_paths = [output_dir / ('checkpoint-%s.pth' % epoch_name)] |
| for checkpoint_path in checkpoint_paths: |
| to_save = { |
| 'model': model_without_ddp.state_dict(), |
| 'optimizer': optimizer.state_dict(), |
| 'epoch': epoch, |
| 'scaler': loss_scaler.state_dict(), |
| 'args': args, |
| } |
|
|
| if model_ema is not None: |
| to_save['model_ema'] = get_state_dict(model_ema) |
|
|
| save_on_master(to_save, checkpoint_path) |
| |
| if is_main_process() and keep_last_n is not None and int(keep_last_n) > 0: |
| all_ckpts = [] |
| for p in output_dir.glob("checkpoint-*.pth"): |
| stem = p.stem |
| suffix = stem.split("-")[-1] |
| if suffix.isdigit(): |
| all_ckpts.append((int(suffix), p)) |
| all_ckpts.sort(key=lambda x: x[0]) |
| for _, old_path in all_ckpts[:-int(keep_last_n)]: |
| try: |
| old_path.unlink(missing_ok=True) |
| except Exception as e: |
| print(f"[checkpoint] warning: failed to remove old checkpoint {old_path}: {e}") |
| else: |
| client_state = {'epoch': epoch} |
| if model_ema is not None: |
| client_state['model_ema'] = get_state_dict(model_ema) |
| model.save_checkpoint(save_dir=args.output_dir, tag="checkpoint-%s" % epoch_name, client_state=client_state) |
| |
| |
| |
| if is_main_process() and keep_last_n is not None and int(keep_last_n) > 0: |
| all_ckpts = [] |
| for p in output_dir.glob("checkpoint-*"): |
| if not p.is_dir(): |
| continue |
| suffix = p.name.split("-")[-1] |
| if suffix.isdigit(): |
| all_ckpts.append((int(suffix), p)) |
| all_ckpts.sort(key=lambda x: x[0]) |
| for _, old_path in all_ckpts[:-int(keep_last_n)]: |
| try: |
| shutil.rmtree(old_path) |
| except Exception as e: |
| print(f"[checkpoint] warning: failed to remove old DeepSpeed checkpoint {old_path}: {e}") |
|
|
|
|
| def _load_model_state_with_shape_fallback(model, checkpoint_model): |
| """Load checkpoint model weights, allowing shape-mismatch only for changed heads/prompt blocks.""" |
| try: |
| incompatible = model.load_state_dict(checkpoint_model) |
| return incompatible, [] |
| except RuntimeError as strict_error: |
| model_state = model.state_dict() |
| allowed_prefixes = ("temporal_corrector.", "spatial_corrector.") |
| allowed_fragments = ("ours_", "prompt") |
|
|
| filtered = {} |
| dropped = [] |
| blocked = [] |
| for key, value in checkpoint_model.items(): |
| if key not in model_state: |
| continue |
|
|
| model_value = model_state[key] |
| if hasattr(value, "shape") and hasattr(model_value, "shape"): |
| ckpt_shape = tuple(value.shape) |
| model_shape = tuple(model_value.shape) |
| if ckpt_shape != model_shape: |
| info = (key, ckpt_shape, model_shape) |
| allow = key.startswith(allowed_prefixes) or any(frag in key for frag in allowed_fragments) |
| if allow: |
| dropped.append(info) |
| continue |
| blocked.append(info) |
| continue |
|
|
| filtered[key] = value |
|
|
| if blocked: |
| head = blocked[0] |
| raise RuntimeError( |
| "Checkpoint contains non-head shape mismatches. " |
| f"Example: {head[0]} ckpt{head[1]} != model{head[2]}" |
| ) from strict_error |
|
|
| incompatible = model.load_state_dict(filtered, strict=False) |
| return incompatible, dropped |
|
|
|
|
| def auto_load_model(args, model, model_without_ddp, optimizer, loss_scaler, model_ema=None): |
| output_dir = Path(args.output_dir) |
| if loss_scaler is not None: |
| |
| if args.auto_resume and len(args.resume) == 0: |
| import glob |
| all_checkpoints = glob.glob(os.path.join(output_dir, 'checkpoint-*.pth')) |
| latest_ckpt = -1 |
| for ckpt in all_checkpoints: |
| t = ckpt.split('-')[-1].split('.')[0] |
| if t.isdigit(): |
| latest_ckpt = max(int(t), latest_ckpt) |
| if latest_ckpt >= 0: |
| args.resume = os.path.join(output_dir, 'checkpoint-%d.pth' % latest_ckpt) |
| print("Auto resume checkpoint: %s" % args.resume) |
|
|
| if args.resume: |
| if args.resume.startswith('https'): |
| checkpoint = torch.hub.load_state_dict_from_url( |
| args.resume, map_location='cpu', check_hash=True) |
| else: |
| checkpoint = torch.load(args.resume, map_location='cpu', weights_only=False) |
| incompatible, dropped = _load_model_state_with_shape_fallback( |
| model_without_ddp, checkpoint['model'] |
| ) |
| print("Resume checkpoint %s" % args.resume) |
| if dropped: |
| print( |
| f"[resume] dropped {len(dropped)} shape-mismatch keys " |
| "(changed factorized heads/prompt blocks)." |
| ) |
| if len(incompatible.missing_keys) > 0: |
| print(f"[resume] missing keys after load: {len(incompatible.missing_keys)}") |
| if len(incompatible.unexpected_keys) > 0: |
| print(f"[resume] unexpected keys after load: {len(incompatible.unexpected_keys)}") |
| if 'optimizer' in checkpoint and 'epoch' in checkpoint: |
| try: |
| optimizer.load_state_dict(checkpoint['optimizer']) |
| args.start_epoch = checkpoint['epoch'] + 1 |
| if hasattr(args, 'model_ema') and args.model_ema: |
| _load_checkpoint_for_ema(model_ema, checkpoint['model_ema']) |
| if 'scaler' in checkpoint: |
| loss_scaler.load_state_dict(checkpoint['scaler']) |
| print("With optim & sched!") |
| except Exception as optimizer_error: |
| print( |
| "[resume] optimizer/scaler state not restored due to parameter-group mismatch. " |
| f"Starting optimizer fresh. ({type(optimizer_error).__name__}: {optimizer_error})" |
| ) |
| else: |
| |
| |
| explicit_resume = str(getattr(args, "resume", "") or "").strip() |
| if explicit_resume: |
| resume_path = Path(explicit_resume) |
| latest_tag = None |
| load_dir = None |
| if resume_path.is_dir(): |
| latest_tag = resume_path.name |
| load_dir = str(resume_path.parent) |
| elif resume_path.is_file() and resume_path.name == "latest": |
| latest_tag = resume_path.read_text(encoding="utf-8").strip() |
| load_dir = str(resume_path.parent) |
| elif resume_path.is_file(): |
| |
| latest_tag = resume_path.parent.name |
| load_dir = str(resume_path.parent.parent) |
| if latest_tag is None or load_dir is None: |
| raise FileNotFoundError(f"Invalid DeepSpeed resume checkpoint: {explicit_resume}") |
| print(f"Resume DeepSpeed checkpoint: load_dir={load_dir}, tag={latest_tag}") |
| _, client_states = model.load_checkpoint(load_dir, tag=latest_tag) |
| if client_states is not None and 'epoch' in client_states: |
| args.start_epoch = client_states['epoch'] + 1 |
| if model_ema is not None and getattr(args, "model_ema", False) and client_states is not None: |
| if 'model_ema' in client_states: |
| _load_checkpoint_for_ema(model_ema, client_states['model_ema']) |
| elif args.auto_resume: |
| import glob |
| latest_tag = None |
| latest_file = output_dir / "latest" |
| if latest_file.is_file(): |
| try: |
| tag = latest_file.read_text(encoding="utf-8").strip() |
| if tag and (output_dir / tag).is_dir(): |
| latest_tag = tag |
| except Exception as e: |
| print(f"[resume] warning: failed to read DeepSpeed latest file {latest_file}: {e}") |
| if latest_tag is None: |
| all_checkpoints = glob.glob(os.path.join(output_dir, 'checkpoint-*')) |
| latest_ckpt = -1 |
| for ckpt in all_checkpoints: |
| t = ckpt.split('-')[-1].split('.')[0] |
| if t.isdigit(): |
| latest_ckpt = max(int(t), latest_ckpt) |
| if latest_ckpt >= 0: |
| latest_tag = 'checkpoint-%d' % latest_ckpt |
| if latest_tag is not None: |
| args.resume = os.path.join(output_dir, latest_tag) |
| print("Auto resume checkpoint: %s" % latest_tag) |
| _, client_states = model.load_checkpoint(args.output_dir, tag=latest_tag) |
| args.start_epoch = client_states['epoch'] + 1 |
| if model_ema is not None: |
| if args.model_ema: |
| _load_checkpoint_for_ema(model_ema, client_states['model_ema']) |
|
|
|
|
| def create_ds_config(args): |
| args.deepspeed_config = os.path.join(args.output_dir, "deepspeed_config.json") |
| with open(args.deepspeed_config, mode="w") as writer: |
| zero_stage = int(getattr(args, "ds_zero_stage", 2)) |
| offload_optimizer = bool(getattr(args, "ds_offload_optimizer", False)) |
| offload_param = bool(getattr(args, "ds_offload_param", False)) |
| use_bf16 = bool(getattr(args, "ds_bf16", True)) |
| use_torch_adam = bool(getattr(args, "ds_torch_adam", True)) |
| overlap_comm = bool(getattr(args, "ds_overlap_comm", True)) |
| allgather_bucket_size = float(getattr(args, "ds_allgather_bucket_size", 2e8)) |
| reduce_bucket_size = float(getattr(args, "ds_reduce_bucket_size", 2e8)) |
| zero_opt = { |
| "stage": zero_stage, |
| "contiguous_gradients": True, |
| } |
| if zero_stage >= 2: |
| zero_opt.update({ |
| "overlap_comm": overlap_comm, |
| "reduce_scatter": True, |
| "allgather_partitions": True, |
| "allgather_bucket_size": allgather_bucket_size, |
| "reduce_bucket_size": reduce_bucket_size, |
| }) |
| if zero_stage >= 2 and offload_optimizer: |
| zero_opt["offload_optimizer"] = { |
| "device": "cpu", |
| "pin_memory": True, |
| } |
| if zero_stage >= 3 and offload_param: |
| zero_opt["offload_param"] = { |
| "device": "cpu", |
| "pin_memory": True, |
| } |
| zero_opt["stage3_param_persistence_threshold"] = 1e5 |
| zero_opt["stage3_max_live_parameters"] = 1e8 |
| zero_opt["stage3_prefetch_bucket_size"] = 5e7 |
|
|
| opt_params = { |
| "torch_adam": use_torch_adam, |
| "lr": args.lr, |
| "weight_decay": args.weight_decay, |
| "betas": [ |
| 0.9, |
| 0.999 |
| ], |
| "eps": 1e-8 |
| } |
| |
| |
| if not use_torch_adam: |
| opt_params["bias_correction"] = True |
|
|
| ds_config = { |
| "train_batch_size": args.batch_size * args.update_freq * get_world_size(), |
| "train_micro_batch_size_per_gpu": args.batch_size, |
| "gradient_accumulation_steps": args.update_freq, |
| "steps_per_print": 1000, |
| "optimizer": { |
| "type": "adamw", |
| "params": opt_params |
| }, |
| "zero_optimization": zero_opt, |
| } |
| if getattr(args, "clip_grad", None) is not None and float(args.clip_grad) > 0: |
| ds_config["gradient_clipping"] = float(args.clip_grad) |
| if use_bf16: |
| ds_config["bf16"] = {"enabled": True} |
| else: |
| ds_config["fp16"] = { |
| "enabled": True, |
| "loss_scale": 0, |
| "initial_scale_power": 7, |
| "loss_scale_window": 128 |
| } |
|
|
| writer.write(json.dumps(ds_config, indent=2)) |
|
|
| def multiple_samples_collate(batch, fold=False): |
| """ |
| Collate function for repeated augmentation. Each instance in the batch has |
| more than one sample. |
| Args: |
| batch (tuple or list): data batch to collate. |
| Returns: |
| (tuple): collated data batch. |
| """ |
| inputs, labels, video_idx, extra_data = zip(*batch) |
| inputs = [item for sublist in inputs for item in sublist] |
| labels = [item for sublist in labels for item in sublist] |
| video_idx = [item for sublist in video_idx for item in sublist] |
| inputs, labels, video_idx, extra_data = ( |
| default_collate(inputs), |
| default_collate(labels), |
| default_collate(video_idx), |
| default_collate(extra_data), |
| ) |
| if fold: |
| return [inputs], labels, video_idx, extra_data |
| else: |
| return inputs, labels, video_idx, extra_data |
|
|
