| from torch import optim |
| import logging |
|
|
|
|
| class EmptyOptimizer: |
| def __init__(self): |
| self.param_groups = [] |
|
|
| def step(self, *args, **kwargs): |
| pass |
|
|
| def state_dict(self): |
| return dict() |
|
|
| def load_state_dict(self, *args, **kwargs): |
| pass |
|
|
| def zero_grad(self): |
| pass |
|
|
|
|
| def build_optimizer(args, model): |
| def exclude( |
| n, p): return p.ndim < 2 or "bn" in n or "ln" in n or "bias" in n or 'logit_scale' in n |
|
|
| def include(n, p): return not exclude(n, p) |
|
|
| named_parameters = list(model.named_parameters()) |
| |
| model_parts = [ |
| list(model.image_named_params()), |
| list(model.text_named_params()), |
| list(model.joint_named_params()), |
| ] |
|
|
| cnt1 = sum(v.numel() for k, v in named_parameters if v.requires_grad) |
| cnt2 = sum(sum(v.numel() for k, v in part if v.requires_grad) |
| for part in model_parts) |
| assert cnt1 == cnt2, f"cnt1 {cnt1} != cnt2 {cnt2}" |
|
|
| optimizer = [] |
| part_names = ['image', 'text', 'joint'] |
| assert len(model_parts) == len(part_names) |
| for name, named_parameters in zip(part_names, model_parts): |
| gain_or_bias_params = [p for n, p in named_parameters if exclude( |
| n, p) and p.requires_grad and "l0_module" not in n] |
| rest_params = [p for n, p in named_parameters if include( |
| n, p) and p.requires_grad and "l0_module" not in n] |
| params_groups = [ |
| {"params": gain_or_bias_params, "weight_decay": 0.}, |
| {"params": rest_params, "weight_decay": args.wd}, |
| ] |
|
|
| num_opt_params = 0 |
| for pg in params_groups: |
| num_opt_params += sum(p.numel() for p in pg['params']) |
|
|
| logging.info(f'number of optimizer ({name}) params: {num_opt_params}') |
|
|
| if num_opt_params > 0: |
| optimizer_i = optim.AdamW( |
| params_groups, |
| lr=args.lr, |
| betas=(args.beta1, args.beta2), |
| eps=args.eps, |
| ) |
| else: |
| optimizer_i = EmptyOptimizer() |
| optimizer.append(optimizer_i) |
|
|
| if args.prune_image or args.prune_text: |
| lr_l0 = 0.02 |
| lr_lamda = args.l0lr |
| l0_params = [] |
| |
| if args.prune_image: |
| l0_params.extend([ |
| { |
| "params": [p for n, p in model.image_named_params() if p.requires_grad and "lambda" not in n and "l0_module" in n], |
| "weight_decay": 0.0, |
| "lr": lr_l0 |
| }, { |
| "params": [p for n, p in model.image_named_params() if p.requires_grad and "lambda" in n and "l0_module" in n], |
| "weight_decay": 0.0, |
| "lr": lr_lamda |
| }]) |
| if args.prune_text: |
| l0_params.extend([ |
| { |
| "params": [p for n, p in model.text_named_params() if p.requires_grad and "lambda" not in n and "l0_module" in n], |
| "weight_decay": 0.0, |
| "lr": lr_l0 |
| }, { |
| "params": [p for n, p in model.text_named_params() if p.requires_grad and "lambda" in n and "l0_module" in n], |
| "weight_decay": 0.0, |
| "lr": lr_lamda |
| }]) |
| l0_optimizer = optim.AdamW(l0_params) |
| optimizer.append(l0_optimizer) |
|
|
| return optimizer |
|
|
