From addfc7c1ac719a312f5cd6260bfa79bddc6c4dc5 Mon Sep 17 00:00:00 2001 From: juntang Date: Sun, 4 Apr 2021 23:48:15 -0400 Subject: [PATCH 1/3] adabelief --- distributed_train_adabelief.sh | 3 + .../args.yaml | 101 ++++++++ timm/optim/__init__.py | 4 +- timm/optim/adabelief.py | 244 ++++++++++++++++++ timm/optim/optim_factory.py | 5 +- 5 files changed, 354 insertions(+), 3 deletions(-) create mode 100755 distributed_train_adabelief.sh create mode 100644 output/train/20210404-234631-efficientnet_b0-224/args.yaml create mode 100644 timm/optim/adabelief.py diff --git a/distributed_train_adabelief.sh b/distributed_train_adabelief.sh new file mode 100755 index 00000000..964e6803 --- /dev/null +++ b/distributed_train_adabelief.sh @@ -0,0 +1,3 @@ +#!/bin/bash +NUM_PROC=2 +CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=$NUM_PROC train.py "$@" --model efficientnet_b0 --weight-decay 2.5e-2 --drop 0.2 --drop-path 0.2 --lr 0.002 --batch-size 192 --epochs 400 --sched cosine --opt adabelief --workers 8 --warmup-lr 1e-4 --aa rand-m9-mstd0.5 --remode pixel --reprob 0.2 --amp --bn-momentum 0.1 --mixup 0.2 --mixup-off-epoch 400 --min-lr 1e-5 diff --git a/output/train/20210404-234631-efficientnet_b0-224/args.yaml b/output/train/20210404-234631-efficientnet_b0-224/args.yaml new file mode 100644 index 00000000..a796dbbb --- /dev/null +++ b/output/train/20210404-234631-efficientnet_b0-224/args.yaml @@ -0,0 +1,101 @@ +aa: rand-m9-mstd0.5 +amp: true +apex_amp: false +aug_splits: 0 +batch_size: 192 +bn_eps: null +bn_momentum: 0.1 +bn_tf: false +channels_last: false +checkpoint_hist: 10 +clip_grad: null +clip_mode: norm +color_jitter: 0.4 +cooldown_epochs: 10 +crop_pct: null +cutmix: 0.0 +cutmix_minmax: null +data_dir: /media/juntang/Samsung_T5/ImageNet +dataset: '' +decay_epochs: 30 +decay_rate: 0.1 +dist_bn: '' +drop: 0.2 +drop_block: null +drop_connect: null +drop_path: 0.2 +epoch_repeats: 0.0 +epochs: 400 +eval_metric: top1 +experiment: '' +gp: null +hflip: 0.5 +img_size: null +initial_checkpoint: '' +input_size: null +interpolation: '' +jsd: false +local_rank: 0 +log_interval: 50 +lr: 0.002 +lr_cycle_limit: 1 +lr_cycle_mul: 1.0 +lr_noise: null +lr_noise_pct: 0.67 +lr_noise_std: 1.0 +mean: null +min_lr: 1.0e-05 +mixup: 0.2 +mixup_mode: batch +mixup_off_epoch: 400 +mixup_prob: 1.0 +mixup_switch_prob: 0.5 +model: efficientnet_b0 +model_ema: false +model_ema_decay: 0.9998 +model_ema_force_cpu: false +momentum: 0.9 +native_amp: false +no_aug: false +no_prefetcher: false +no_resume_opt: false +num_classes: null +opt: adabelief +opt_betas: null +opt_eps: null +output: '' +patience_epochs: 10 +pin_mem: false +pretrained: false +ratio: +- 0.75 +- 1.3333333333333333 +recount: 1 +recovery_interval: 0 +remode: pixel +reprob: 0.2 +resplit: false +resume: '' +save_images: false +scale: +- 0.08 +- 1.0 +sched: cosine +seed: 42 +smoothing: 0.1 +split_bn: false +start_epoch: null +std: null +sync_bn: false +torchscript: false +train_interpolation: random +train_split: train +tta: 0 +use_multi_epochs_loader: false +val_split: validation +validation_batch_size_multiplier: 1 +vflip: 0.0 +warmup_epochs: 3 +warmup_lr: 0.0001 +weight_decay: 0.025 +workers: 8 diff --git a/timm/optim/__init__.py b/timm/optim/__init__.py index 7c4f4d36..e2bf46aa 100644 --- a/timm/optim/__init__.py +++ b/timm/optim/__init__.py @@ -9,5 +9,5 @@ from .nvnovograd import NvNovoGrad from .radam import RAdam from .rmsprop_tf import RMSpropTF from .sgdp import SGDP - -from .optim_factory import create_optimizer, create_optimizer_v2, optimizer_kwargs \ No newline at end of file +from .adabelief import AdaBelief +from .optim_factory import create_optimizer, create_optimizer_v2, optimizer_kwargs diff --git a/timm/optim/adabelief.py b/timm/optim/adabelief.py new file mode 100644 index 00000000..71075524 --- /dev/null +++ b/timm/optim/adabelief.py @@ -0,0 +1,244 @@ +import math +import torch +from torch.optim.optimizer import Optimizer +from tabulate import tabulate +from colorama import Fore, Back, Style + +version_higher = ( torch.__version__ >= "1.5.0" ) + +class AdaBelief(Optimizer): + r"""Implements AdaBelief algorithm. Modified from Adam in PyTorch + Arguments: + params (iterable): iterable of parameters to optimize or dicts defining + parameter groups + lr (float, optional): learning rate (default: 1e-3) + betas (Tuple[float, float], optional): coefficients used for computing + running averages of gradient and its square (default: (0.9, 0.999)) + eps (float, optional): term added to the denominator to improve + numerical stability (default: 1e-16) + weight_decay (float, optional): weight decay (L2 penalty) (default: 0) + amsgrad (boolean, optional): whether to use the AMSGrad variant of this + algorithm from the paper `On the Convergence of Adam and Beyond`_ + (default: False) + weight_decouple (boolean, optional): ( default: True) If set as True, then + the optimizer uses decoupled weight decay as in AdamW + fixed_decay (boolean, optional): (default: False) This is used when weight_decouple + is set as True. + When fixed_decay == True, the weight decay is performed as + $W_{new} = W_{old} - W_{old} \times decay$. + When fixed_decay == False, the weight decay is performed as + $W_{new} = W_{old} - W_{old} \times decay \times lr$. Note that in this case, the + weight decay ratio decreases with learning rate (lr). + rectify (boolean, optional): (default: True) If set as True, then perform the rectified + update similar to RAdam + degenerated_to_sgd (boolean, optional) (default:True) If set as True, then perform SGD update + when variance of gradient is high + print_change_log (boolean, optional) (default: True) If set as True, print the modifcation to + default hyper-parameters + reference: AdaBelief Optimizer, adapting stepsizes by the belief in observed gradients, NeurIPS 2020 + """ + + def __init__(self, params, lr=1e-3, betas=(0.9, 0.999), eps=1e-16, + weight_decay=0, amsgrad=False, weight_decouple=True, fixed_decay=False, rectify=True, + degenerated_to_sgd=True, print_change_log = True): + + # ------------------------------------------------------------------------------ + # Print modifications to default arguments + if print_change_log: + print(Fore.RED + 'Please check your arguments if you have upgraded adabelief-pytorch from version 0.0.5.') + print(Fore.RED + 'Modifications to default arguments:') + default_table = tabulate([ + ['adabelief-pytorch=0.0.5','1e-8','False','False'], + ['>=0.1.0 (Current 0.2.0)','1e-16','True','True']], + headers=['eps','weight_decouple','rectify']) + print(Fore.RED + default_table) + + recommend_table = tabulate([ + ['Recommended eps = 1e-8', 'Recommended eps = 1e-16'], + ], + headers=['SGD better than Adam (e.g. CNN for Image Classification)','Adam better than SGD (e.g. Transformer, GAN)']) + print(Fore.BLUE + recommend_table) + + print(Fore.BLUE +'For a complete table of recommended hyperparameters, see') + print(Fore.BLUE + 'https://github.com/juntang-zhuang/Adabelief-Optimizer') + + print(Fore.GREEN + 'You can disable the log message by setting "print_change_log = False", though it is recommended to keep as a reminder.') + + print(Style.RESET_ALL) + # ------------------------------------------------------------------------------ + + if not 0.0 <= lr: + raise ValueError("Invalid learning rate: {}".format(lr)) + if not 0.0 <= eps: + raise ValueError("Invalid epsilon value: {}".format(eps)) + if not 0.0 <= betas[0] < 1.0: + raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0])) + if not 0.0 <= betas[1] < 1.0: + raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1])) + + self.degenerated_to_sgd = degenerated_to_sgd + if isinstance(params, (list, tuple)) and len(params) > 0 and isinstance(params[0], dict): + for param in params: + if 'betas' in param and (param['betas'][0] != betas[0] or param['betas'][1] != betas[1]): + param['buffer'] = [[None, None, None] for _ in range(10)] + + defaults = dict(lr=lr, betas=betas, eps=eps, + weight_decay=weight_decay, amsgrad=amsgrad, buffer=[[None, None, None] for _ in range(10)]) + super(AdaBelief, self).__init__(params, defaults) + + self.degenerated_to_sgd = degenerated_to_sgd + self.weight_decouple = weight_decouple + self.rectify = rectify + self.fixed_decay = fixed_decay + if self.weight_decouple: + print('Weight decoupling enabled in AdaBelief') + if self.fixed_decay: + print('Weight decay fixed') + if self.rectify: + print('Rectification enabled in AdaBelief') + if amsgrad: + print('AMSGrad enabled in AdaBelief') + + def __setstate__(self, state): + super(AdaBelief, self).__setstate__(state) + for group in self.param_groups: + group.setdefault('amsgrad', False) + + def reset(self): + for group in self.param_groups: + for p in group['params']: + state = self.state[p] + amsgrad = group['amsgrad'] + + # State initialization + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p.data,memory_format=torch.preserve_format) \ + if version_higher else torch.zeros_like(p.data) + + # Exponential moving average of squared gradient values + state['exp_avg_var'] = torch.zeros_like(p.data,memory_format=torch.preserve_format) \ + if version_higher else torch.zeros_like(p.data) + + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state['max_exp_avg_var'] = torch.zeros_like(p.data,memory_format=torch.preserve_format) \ + if version_higher else torch.zeros_like(p.data) + + def step(self, closure=None): + """Performs a single optimization step. + Arguments: + closure (callable, optional): A closure that reevaluates the model + and returns the loss. + """ + loss = None + if closure is not None: + loss = closure() + + for group in self.param_groups: + for p in group['params']: + if p.grad is None: + continue + + # cast data type + half_precision = False + if p.data.dtype == torch.float16: + half_precision = True + p.data = p.data.float() + p.grad = p.grad.float() + + grad = p.grad.data + if grad.is_sparse: + raise RuntimeError( + 'AdaBelief does not support sparse gradients, please consider SparseAdam instead') + amsgrad = group['amsgrad'] + + state = self.state[p] + + beta1, beta2 = group['betas'] + + # State initialization + if len(state) == 0: + state['step'] = 0 + # Exponential moving average of gradient values + state['exp_avg'] = torch.zeros_like(p.data,memory_format=torch.preserve_format) \ + if version_higher else torch.zeros_like(p.data) + # Exponential moving average of squared gradient values + state['exp_avg_var'] = torch.zeros_like(p.data,memory_format=torch.preserve_format) \ + if version_higher else torch.zeros_like(p.data) + if amsgrad: + # Maintains max of all exp. moving avg. of sq. grad. values + state['max_exp_avg_var'] = torch.zeros_like(p.data,memory_format=torch.preserve_format) \ + if version_higher else torch.zeros_like(p.data) + + # perform weight decay, check if decoupled weight decay + if self.weight_decouple: + if not self.fixed_decay: + p.data.mul_(1.0 - group['lr'] * group['weight_decay']) + else: + p.data.mul_(1.0 - group['weight_decay']) + else: + if group['weight_decay'] != 0: + grad.add_(p.data, alpha=group['weight_decay']) + + # get current state variable + exp_avg, exp_avg_var = state['exp_avg'], state['exp_avg_var'] + + state['step'] += 1 + bias_correction1 = 1 - beta1 ** state['step'] + bias_correction2 = 1 - beta2 ** state['step'] + + # Update first and second moment running average + exp_avg.mul_(beta1).add_(grad, alpha=1 - beta1) + grad_residual = grad - exp_avg + exp_avg_var.mul_(beta2).addcmul_( grad_residual, grad_residual, value=1 - beta2) + + if amsgrad: + max_exp_avg_var = state['max_exp_avg_var'] + # Maintains the maximum of all 2nd moment running avg. till now + torch.max(max_exp_avg_var, exp_avg_var.add_(group['eps']), out=max_exp_avg_var) + + # Use the max. for normalizing running avg. of gradient + denom = (max_exp_avg_var.sqrt() / math.sqrt(bias_correction2)).add_(group['eps']) + else: + denom = (exp_avg_var.add_(group['eps']).sqrt() / math.sqrt(bias_correction2)).add_(group['eps']) + + # update + if not self.rectify: + # Default update + step_size = group['lr'] / bias_correction1 + p.data.addcdiv_( exp_avg, denom, value=-step_size) + + else: # Rectified update, forked from RAdam + buffered = group['buffer'][int(state['step'] % 10)] + if state['step'] == buffered[0]: + N_sma, step_size = buffered[1], buffered[2] + else: + buffered[0] = state['step'] + beta2_t = beta2 ** state['step'] + N_sma_max = 2 / (1 - beta2) - 1 + N_sma = N_sma_max - 2 * state['step'] * beta2_t / (1 - beta2_t) + buffered[1] = N_sma + + # more conservative since it's an approximated value + if N_sma >= 5: + step_size = math.sqrt( + (1 - beta2_t) * (N_sma - 4) / (N_sma_max - 4) * (N_sma - 2) / N_sma * N_sma_max / ( + N_sma_max - 2)) / (1 - beta1 ** state['step']) + elif self.degenerated_to_sgd: + step_size = 1.0 / (1 - beta1 ** state['step']) + else: + step_size = -1 + buffered[2] = step_size + + if N_sma >= 5: + denom = exp_avg_var.sqrt().add_(group['eps']) + p.data.addcdiv_(exp_avg, denom, value=-step_size * group['lr']) + elif step_size > 0: + p.data.add_( exp_avg, alpha=-step_size * group['lr']) + + if half_precision: + p.data = p.data.half() + p.grad = p.grad.half() + + return loss diff --git a/timm/optim/optim_factory.py b/timm/optim/optim_factory.py index a10607cb..c9b3b6df 100644 --- a/timm/optim/optim_factory.py +++ b/timm/optim/optim_factory.py @@ -17,6 +17,7 @@ from .nvnovograd import NvNovoGrad from .radam import RAdam from .rmsprop_tf import RMSpropTF from .sgdp import SGDP +from .adabelief import AdaBelief try: from apex.optimizers import FusedNovoGrad, FusedAdam, FusedLAMB, FusedSGD @@ -118,7 +119,9 @@ def create_optimizer_v2( opt_args.pop('eps', None) optimizer = optim.SGD(parameters, momentum=momentum, nesterov=False, **opt_args) elif opt_lower == 'adam': - optimizer = optim.Adam(parameters, **opt_args) + optimizer = optim.Adam(parameters, **opt_args) + elif opt_lower == 'adabelief': + optimizer = AdaBelief(parameters, rectify = False, print_change_log = False,**opt_args) elif opt_lower == 'adamw': optimizer = optim.AdamW(parameters, **opt_args) elif opt_lower == 'nadam': From 1d848f409a8a2887a5ebbad6b84621c51c14baf5 Mon Sep 17 00:00:00 2001 From: Juntang Zhuang <44451229+juntang-zhuang@users.noreply.github.com> Date: Mon, 5 Apr 2021 15:11:35 -0400 Subject: [PATCH 2/3] Delete args.yaml --- .../args.yaml | 101 ------------------ 1 file changed, 101 deletions(-) delete mode 100644 output/train/20210404-234631-efficientnet_b0-224/args.yaml diff --git a/output/train/20210404-234631-efficientnet_b0-224/args.yaml b/output/train/20210404-234631-efficientnet_b0-224/args.yaml deleted file mode 100644 index a796dbbb..00000000 --- a/output/train/20210404-234631-efficientnet_b0-224/args.yaml +++ /dev/null @@ -1,101 +0,0 @@ -aa: rand-m9-mstd0.5 -amp: true -apex_amp: false -aug_splits: 0 -batch_size: 192 -bn_eps: null -bn_momentum: 0.1 -bn_tf: false -channels_last: false -checkpoint_hist: 10 -clip_grad: null -clip_mode: norm -color_jitter: 0.4 -cooldown_epochs: 10 -crop_pct: null -cutmix: 0.0 -cutmix_minmax: null -data_dir: /media/juntang/Samsung_T5/ImageNet -dataset: '' -decay_epochs: 30 -decay_rate: 0.1 -dist_bn: '' -drop: 0.2 -drop_block: null -drop_connect: null -drop_path: 0.2 -epoch_repeats: 0.0 -epochs: 400 -eval_metric: top1 -experiment: '' -gp: null -hflip: 0.5 -img_size: null -initial_checkpoint: '' -input_size: null -interpolation: '' -jsd: false -local_rank: 0 -log_interval: 50 -lr: 0.002 -lr_cycle_limit: 1 -lr_cycle_mul: 1.0 -lr_noise: null -lr_noise_pct: 0.67 -lr_noise_std: 1.0 -mean: null -min_lr: 1.0e-05 -mixup: 0.2 -mixup_mode: batch -mixup_off_epoch: 400 -mixup_prob: 1.0 -mixup_switch_prob: 0.5 -model: efficientnet_b0 -model_ema: false -model_ema_decay: 0.9998 -model_ema_force_cpu: false -momentum: 0.9 -native_amp: false -no_aug: false -no_prefetcher: false -no_resume_opt: false -num_classes: null -opt: adabelief -opt_betas: null -opt_eps: null -output: '' -patience_epochs: 10 -pin_mem: false -pretrained: false -ratio: -- 0.75 -- 1.3333333333333333 -recount: 1 -recovery_interval: 0 -remode: pixel -reprob: 0.2 -resplit: false -resume: '' -save_images: false -scale: -- 0.08 -- 1.0 -sched: cosine -seed: 42 -smoothing: 0.1 -split_bn: false -start_epoch: null -std: null -sync_bn: false -torchscript: false -train_interpolation: random -train_split: train -tta: 0 -use_multi_epochs_loader: false -val_split: validation -validation_batch_size_multiplier: 1 -vflip: 0.0 -warmup_epochs: 3 -warmup_lr: 0.0001 -weight_decay: 0.025 -workers: 8 From 74366f733cd14461dca3ee537cbe275c333da2da Mon Sep 17 00:00:00 2001 From: Juntang Zhuang <44451229+juntang-zhuang@users.noreply.github.com> Date: Mon, 5 Apr 2021 15:11:45 -0400 Subject: [PATCH 3/3] Delete distributed_train_adabelief.sh --- distributed_train_adabelief.sh | 3 --- 1 file changed, 3 deletions(-) delete mode 100755 distributed_train_adabelief.sh diff --git a/distributed_train_adabelief.sh b/distributed_train_adabelief.sh deleted file mode 100755 index 964e6803..00000000 --- a/distributed_train_adabelief.sh +++ /dev/null @@ -1,3 +0,0 @@ -#!/bin/bash -NUM_PROC=2 -CUDA_VISIBLE_DEVICES=0,1 python -m torch.distributed.launch --nproc_per_node=$NUM_PROC train.py "$@" --model efficientnet_b0 --weight-decay 2.5e-2 --drop 0.2 --drop-path 0.2 --lr 0.002 --batch-size 192 --epochs 400 --sched cosine --opt adabelief --workers 8 --warmup-lr 1e-4 --aa rand-m9-mstd0.5 --remode pixel --reprob 0.2 --amp --bn-momentum 0.1 --mixup 0.2 --mixup-off-epoch 400 --min-lr 1e-5