Merge pull request #32 from rwightman/opt

More optimizer work

timm/models/helpers.py

@@ -29,7 +29,7 @@ def load_checkpoint(model, checkpoint_path, use_ema=False):
 
 
 def resume_checkpoint(model, checkpoint_path):
-    optimizer_state = None
+    other_state = {}
     resume_epoch = None
     if os.path.isfile(checkpoint_path):
         checkpoint = torch.load(checkpoint_path, map_location='cpu')
@@ -40,7 +40,9 @@ def resume_checkpoint(model, checkpoint_path):
                 new_state_dict[name] = v
             model.load_state_dict(new_state_dict)
             if 'optimizer' in checkpoint:
-                optimizer_state = checkpoint['optimizer']
+                other_state['optimizer'] = checkpoint['optimizer']
+            if 'amp' in checkpoint:
+                other_state['amp'] = checkpoint['amp']
             if 'epoch' in checkpoint:
                 resume_epoch = checkpoint['epoch']
                 if 'version' in checkpoint and checkpoint['version'] > 1:
@@ -49,7 +51,7 @@ def resume_checkpoint(model, checkpoint_path):
         else:
             model.load_state_dict(checkpoint)
             logging.info("Loaded checkpoint '{}'".format(checkpoint_path))
-        return optimizer_state, resume_epoch
+        return other_state, resume_epoch
     else:
         logging.error("No checkpoint found at '{}'".format(checkpoint_path))
         raise FileNotFoundError()

timm/optim/__init__.py

@@ -3,5 +3,6 @@ from .rmsprop_tf import RMSpropTF
 from .adamw import AdamW
 from .radam import RAdam
 from .novograd import NovoGrad
+from .nvnovograd import NvNovoGrad
 from .lookahead import Lookahead
 from .optim_factory import create_optimizer

timm/optim/lookahead.py

@@ -13,37 +13,40 @@ class Lookahead(Optimizer):
             raise ValueError(f'Invalid slow update rate: {alpha}')
         if not 1 <= k:
             raise ValueError(f'Invalid lookahead steps: {k}')
-        self.alpha = alpha
-        self.k = k
+        defaults = dict(lookahead_alpha=alpha, lookahead_k=k, lookahead_step=0)
         self.base_optimizer = base_optimizer
         self.param_groups = self.base_optimizer.param_groups
         self.defaults = base_optimizer.defaults
+        self.defaults.update(defaults)
         self.state = defaultdict(dict)
-        for group in self.param_groups:
-            group["step_counter"] = 0
+        # manually add our defaults to the param groups
+        for name, default in defaults.items():
+            for group in self.param_groups:
+                group.setdefault(name, default)
 
-    def update_slow_weights(self, group):
+    def update_slow(self, group):
         for fast_p in group["params"]:
             if fast_p.grad is None:
                 continue
             param_state = self.state[fast_p]
-            if "slow_buffer" not in param_state:
-                param_state["slow_buffer"] = torch.empty_like(fast_p.data)
-                param_state["slow_buffer"].copy_(fast_p.data)
-            slow = param_state["slow_buffer"]
-            slow.add_(self.alpha, fast_p.data - slow)
+            if 'slow_buffer' not in param_state:
+                param_state['slow_buffer'] = torch.empty_like(fast_p.data)
+                param_state['slow_buffer'].copy_(fast_p.data)
+            slow = param_state['slow_buffer']
+            slow.add_(group['lookahead_alpha'], fast_p.data - slow)
             fast_p.data.copy_(slow)
 
     def sync_lookahead(self):
         for group in self.param_groups:
-            self.update_slow_weights(group)
+            self.update_slow(group)
 
     def step(self, closure=None):
+        #assert id(self.param_groups) == id(self.base_optimizer.param_groups)
         loss = self.base_optimizer.step(closure)
         for group in self.param_groups:
-            group['step_counter'] += 1
-            if group['step_counter'] % self.k == 0:
-                self.update_slow_weights(group)
+            group['lookahead_step'] += 1
+            if group['lookahead_step'] % group['lookahead_k'] == 0:
+                self.update_slow(group)
         return loss
 
     def state_dict(self):
@@ -52,37 +55,36 @@ class Lookahead(Optimizer):
             (id(k) if isinstance(k, torch.Tensor) else k): v
             for k, v in self.state.items()
         }
-        fast_state = fast_state_dict["state"]
-        param_groups = fast_state_dict["param_groups"]
+        fast_state = fast_state_dict['state']
+        param_groups = fast_state_dict['param_groups']
         return {
-            "state": fast_state,
-            "slow_state": slow_state,
-            "param_groups": param_groups,
+            'state': fast_state,
+            'slow_state': slow_state,
+            'param_groups': param_groups,
         }
 
     def load_state_dict(self, state_dict):
-        if 'slow_state' not in state_dict:
-            print('Loading state_dict from optimizer without Lookahead applied')
-            state_dict['slow_state'] = defaultdict(dict)
-        slow_state_dict = {
-            "state": state_dict["slow_state"],
-            "param_groups": state_dict["param_groups"],
-        }
         fast_state_dict = {
-            "state": state_dict["state"],
-            "param_groups": state_dict["param_groups"],
+            'state': state_dict['state'],
+            'param_groups': state_dict['param_groups'],
         }
-        super(Lookahead, self).load_state_dict(slow_state_dict)
         self.base_optimizer.load_state_dict(fast_state_dict)
 
-    def add_param_group(self, param_group):
-        r"""Add a param group to the :class:`Optimizer` s `param_groups`.
-        This can be useful when fine tuning a pre-trained network as frozen
-        layers can be made trainable and added to the :class:`Optimizer` as
-        training progresses.
-        Args:
-            param_group (dict): Specifies what Tensors should be optimized along
-            with group specific optimization options.
-        """
-        param_group['step_counter'] = 0
-        self.base_optimizer.add_param_group(param_group)
+        # We want to restore the slow state, but share param_groups reference
+        # with base_optimizer. This is a bit redundant but least code
+        slow_state_new = False
+        if 'slow_state' not in state_dict:
+            print('Loading state_dict from optimizer without Lookahead applied.')
+            state_dict['slow_state'] = defaultdict(dict)
+            slow_state_new = True
+        slow_state_dict = {
+            'state': state_dict['slow_state'],
+            'param_groups': state_dict['param_groups'],  # this is pointless but saves code
+        }
+        super(Lookahead, self).load_state_dict(slow_state_dict)
+        self.param_groups = self.base_optimizer.param_groups  # make both ref same container
+        if slow_state_new:
+            # reapply defaults to catch missing lookahead specific ones
+            for name, default in self.defaults.items():
+                for group in self.param_groups:
+                    group.setdefault(name, default)

timm/optim/nvnovograd.py

@@ -0,0 +1,118 @@
+""" Nvidia NovoGrad Optimizer.
+Original impl by Nvidia from Jasper example:
+    - https://github.com/NVIDIA/DeepLearningExamples/blob/master/PyTorch/SpeechRecognition/Jasper
+Paper: `Stochastic Gradient Methods with Layer-wise Adaptive Moments for Training of Deep Networks`
+    - https://arxiv.org/abs/1905.11286
+"""
+
+import torch
+from torch.optim.optimizer import Optimizer
+import math
+
+
+class NvNovoGrad(Optimizer):
+    """
+    Implements Novograd algorithm.
+
+    Args:
+        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: (0.95, 0.98))
+        eps (float, optional): term added to the denominator to improve
+            numerical stability (default: 1e-8)
+        weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
+        grad_averaging: gradient averaging
+        amsgrad (boolean, optional): whether to use the AMSGrad variant of this
+            algorithm from the paper `On the Convergence of Adam and Beyond`_
+            (default: False)
+    """
+
+    def __init__(self, params, lr=1e-3, betas=(0.95, 0.98), eps=1e-8,
+                 weight_decay=0, grad_averaging=False, amsgrad=False):
+        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]))
+        defaults = dict(lr=lr, betas=betas, eps=eps,
+                        weight_decay=weight_decay,
+                        grad_averaging=grad_averaging,
+                        amsgrad=amsgrad)
+
+        super(NvNovoGrad, self).__init__(params, defaults)
+
+    def __setstate__(self, state):
+        super(NvNovoGrad, self).__setstate__(state)
+        for group in self.param_groups:
+            group.setdefault('amsgrad', False)
+
+    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
+                grad = p.grad.data
+                if grad.is_sparse:
+                    raise RuntimeError('Sparse gradients are not supported.')
+                amsgrad = group['amsgrad']
+
+                state = self.state[p]
+
+                # State initialization
+                if len(state) == 0:
+                    state['step'] = 0
+                    # Exponential moving average of gradient values
+                    state['exp_avg'] = torch.zeros_like(p.data)
+                    # Exponential moving average of squared gradient values
+                    state['exp_avg_sq'] = torch.zeros([]).to(state['exp_avg'].device)
+                    if amsgrad:
+                        # Maintains max of all exp. moving avg. of sq. grad. values
+                        state['max_exp_avg_sq'] = torch.zeros([]).to(state['exp_avg'].device)
+
+                exp_avg, exp_avg_sq = state['exp_avg'], state['exp_avg_sq']
+                if amsgrad:
+                    max_exp_avg_sq = state['max_exp_avg_sq']
+                beta1, beta2 = group['betas']
+
+                state['step'] += 1
+
+                norm = torch.sum(torch.pow(grad, 2))
+
+                if exp_avg_sq == 0:
+                    exp_avg_sq.copy_(norm)
+                else:
+                    exp_avg_sq.mul_(beta2).add_(1 - beta2, norm)
+
+                if amsgrad:
+                    # Maintains the maximum of all 2nd moment running avg. till now
+                    torch.max(max_exp_avg_sq, exp_avg_sq, out=max_exp_avg_sq)
+                    # Use the max. for normalizing running avg. of gradient
+                    denom = max_exp_avg_sq.sqrt().add_(group['eps'])
+                else:
+                    denom = exp_avg_sq.sqrt().add_(group['eps'])
+
+                grad.div_(denom)
+                if group['weight_decay'] != 0:
+                    grad.add_(group['weight_decay'], p.data)
+                if group['grad_averaging']:
+                    grad.mul_(1 - beta1)
+                exp_avg.mul_(beta1).add_(grad)
+
+                p.data.add_(-group['lr'], exp_avg)
+
+        return loss

timm/optim/optim_factory.py

@@ -1,5 +1,11 @@
+import torch
 from torch import optim as optim
-from timm.optim import Nadam, RMSpropTF, AdamW, RAdam, NovoGrad, Lookahead
+from timm.optim import Nadam, RMSpropTF, AdamW, RAdam, NovoGrad, NvNovoGrad, Lookahead
+try:
+    from apex.optimizers import FusedNovoGrad, FusedAdam, FusedLAMB, FusedSGD
+    has_apex = True
+except ImportError:
+    has_apex = False
 
 
 def add_weight_decay(model, weight_decay=1e-5, skip_list=()):
@@ -20,9 +26,10 @@ def add_weight_decay(model, weight_decay=1e-5, skip_list=()):
 def create_optimizer(args, model, filter_bias_and_bn=True):
     opt_lower = args.opt.lower()
     weight_decay = args.weight_decay
-    if opt_lower == 'adamw' or opt_lower == 'radam':
-        # compensate for the way current AdamW and RAdam optimizers
-        # apply the weight-decay
+    if 'adamw' in opt_lower or 'radam' in opt_lower:
+        # Compensate for the way current AdamW and RAdam optimizers apply LR to the weight-decay
+        # I don't believe they follow the paper or original Torch7 impl which schedules weight
+        # decay based on the ratio of current_lr/initial_lr
         weight_decay /= args.lr
     if weight_decay and filter_bias_and_bn:
         parameters = add_weight_decay(model, weight_decay)
@@ -30,12 +37,14 @@ def create_optimizer(args, model, filter_bias_and_bn=True):
     else:
         parameters = model.parameters()
 
+    if 'fused' in opt_lower:
+        assert has_apex and torch.cuda.is_available(), 'APEX and CUDA required for fused optimizers'
+
     opt_split = opt_lower.split('_')
     opt_lower = opt_split[-1]
     if opt_lower == 'sgd':
         optimizer = optim.SGD(
-            parameters, lr=args.lr,
-            momentum=args.momentum, weight_decay=weight_decay, nesterov=True)
+            parameters, lr=args.lr, momentum=args.momentum, weight_decay=weight_decay, nesterov=True)
     elif opt_lower == 'adam':
         optimizer = optim.Adam(
             parameters, lr=args.lr, weight_decay=weight_decay, eps=args.opt_eps)
@@ -61,6 +70,22 @@ def create_optimizer(args, model, filter_bias_and_bn=True):
             momentum=args.momentum, weight_decay=weight_decay)
     elif opt_lower == 'novograd':
         optimizer = NovoGrad(parameters, lr=args.lr, weight_decay=weight_decay, eps=args.opt_eps)
+    elif opt_lower == 'nvnovograd':
+        optimizer = NvNovoGrad(parameters, lr=args.lr, weight_decay=weight_decay, eps=args.opt_eps)
+    elif opt_lower == 'fusedsgd':
+        optimizer = FusedSGD(
+            parameters, lr=args.lr, momentum=args.momentum, weight_decay=weight_decay, nesterov=True)
+    elif opt_lower == 'fusedadam':
+        optimizer = FusedAdam(
+            parameters, lr=args.lr, adam_w_mode=False, weight_decay=weight_decay, eps=args.opt_eps)
+    elif opt_lower == 'fusedadamw':
+        optimizer = FusedAdam(
+            parameters, lr=args.lr, adam_w_mode=True, weight_decay=weight_decay, eps=args.opt_eps)
+    elif opt_lower == 'fusedlamb':
+        optimizer = FusedLAMB(parameters, lr=args.lr, weight_decay=weight_decay, eps=args.opt_eps)
+    elif opt_lower == 'fusednovograd':
+        optimizer = FusedNovoGrad(
+            parameters, lr=args.lr, betas=(0.95, 0.98), weight_decay=weight_decay, eps=args.opt_eps)
     else:
         assert False and "Invalid optimizer"
         raise ValueError

timm/utils.py

@@ -11,6 +11,12 @@ import operator
 import logging
 import numpy as np
 from collections import OrderedDict
+try:
+    from apex import amp
+    has_apex = True
+except ImportError:
+    amp = None
+    has_apex = False
 
 from torch import distributed as dist
 
@@ -50,7 +56,7 @@ class CheckpointSaver:
         self.max_history = max_history
         assert self.max_history >= 1
 
-    def save_checkpoint(self, model, optimizer, args, epoch, model_ema=None, metric=None):
+    def save_checkpoint(self, model, optimizer, args, epoch, model_ema=None, metric=None, use_amp=False):
         assert epoch >= 0
         worst_file = self.checkpoint_files[-1] if self.checkpoint_files else None
         if (len(self.checkpoint_files) < self.max_history
@@ -59,7 +65,7 @@ class CheckpointSaver:
                 self._cleanup_checkpoints(1)
             filename = '-'.join([self.save_prefix, str(epoch)]) + self.extension
             save_path = os.path.join(self.checkpoint_dir, filename)
-            self._save(save_path, model, optimizer, args, epoch, model_ema, metric)
+            self._save(save_path, model, optimizer, args, epoch, model_ema, metric, use_amp)
             self.checkpoint_files.append((save_path, metric))
             self.checkpoint_files = sorted(
                 self.checkpoint_files, key=lambda x: x[1],
@@ -77,7 +83,7 @@ class CheckpointSaver:
 
         return (None, None) if self.best_metric is None else (self.best_metric, self.best_epoch)
 
-    def _save(self, save_path, model, optimizer, args, epoch, model_ema=None, metric=None):
+    def _save(self, save_path, model, optimizer, args, epoch, model_ema=None, metric=None, use_amp=False):
         save_state = {
             'epoch': epoch,
             'arch': args.model,
@@ -86,6 +92,8 @@ class CheckpointSaver:
             'args': args,
             'version': 2,  # version < 2 increments epoch before save
         }
+        if use_amp and 'state_dict' in amp.__dict__:
+            save_state['amp'] = amp.state_dict()
         if model_ema is not None:
             save_state['state_dict_ema'] = get_state_dict(model_ema)
         if metric is not None:
@@ -106,11 +114,11 @@ class CheckpointSaver:
                 logging.error("Exception '{}' while deleting checkpoint".format(e))
         self.checkpoint_files = self.checkpoint_files[:delete_index]
 
-    def save_recovery(self, model, optimizer, args, epoch, model_ema=None, batch_idx=0):
+    def save_recovery(self, model, optimizer, args, epoch, model_ema=None, use_amp=False, batch_idx=0):
         assert epoch >= 0
         filename = '-'.join([self.recovery_prefix, str(epoch), str(batch_idx)]) + self.extension
         save_path = os.path.join(self.recovery_dir, filename)
-        self._save(save_path, model, optimizer, args, epoch, model_ema)
+        self._save(save_path, model, optimizer, args, epoch, model_ema, use_amp=use_amp)
         if os.path.exists(self.last_recovery_file):
             try:
                 logging.debug("Cleaning recovery: {}".format(self.last_recovery_file))

train.py

@@ -38,6 +38,8 @@ parser.add_argument('--initial-checkpoint', default='', type=str, metavar='PATH'
                     help='Initialize model from this checkpoint (default: none)')
 parser.add_argument('--resume', default='', type=str, metavar='PATH',
                     help='Resume full model and optimizer state from checkpoint (default: none)')
+parser.add_argument('--no-resume-opt', action='store_true', default=False,
+                    help='prevent resume of optimizer state when resuming model')
 parser.add_argument('--num-classes', type=int, default=1000, metavar='N',
                     help='number of label classes (default: 1000)')
 parser.add_argument('--gp', default='avg', type=str, metavar='POOL',
@@ -189,12 +191,6 @@ def main():
 
     data_config = resolve_data_config(vars(args), model=model, verbose=args.local_rank == 0)
 
-    # optionally resume from a checkpoint
-    optimizer_state = None
-    resume_epoch = None
-    if args.resume:
-        optimizer_state, resume_epoch = resume_checkpoint(model, args.resume)
-
     if args.num_gpu > 1:
         if args.amp:
             logging.warning(
@@ -205,8 +201,6 @@ def main():
         model.cuda()
 
     optimizer = create_optimizer(args, model)
-    if optimizer_state is not None:
-        optimizer.load_state_dict(optimizer_state)
 
     use_amp = False
     if has_apex and args.amp:
@@ -216,6 +210,22 @@ def main():
         logging.info('NVIDIA APEX {}. AMP {}.'.format(
             'installed' if has_apex else 'not installed', 'on' if use_amp else 'off'))
 
+    # optionally resume from a checkpoint
+    resume_state = {}
+    resume_epoch = None
+    if args.resume:
+        resume_state, resume_epoch = resume_checkpoint(model, args.resume)
+    if resume_state and not args.no_resume_opt:
+        if 'optimizer' in resume_state:
+            if args.local_rank == 0:
+                logging.info('Restoring Optimizer state from checkpoint')
+            optimizer.load_state_dict(resume_state['optimizer'])
+        if use_amp and 'amp' in resume_state and 'load_state_dict' in amp.__dict__:
+            if args.local_rank == 0:
+                logging.info('Restoring NVIDIA AMP state from checkpoint')
+            amp.load_state_dict(resume_state['amp'])
+    resume_state = None
+
     model_ema = None
     if args.model_ema:
         # Important to create EMA model after cuda(), DP wrapper, and AMP but before SyncBN and DDP wrapper
@@ -363,7 +373,7 @@ def main():
                 save_metric = eval_metrics[eval_metric]
                 best_metric, best_epoch = saver.save_checkpoint(
                     model, optimizer, args,
-                    epoch=epoch, model_ema=model_ema, metric=save_metric)
+                    epoch=epoch, model_ema=model_ema, metric=save_metric, use_amp=use_amp)
 
     except KeyboardInterrupt:
         pass
@@ -456,7 +466,7 @@ def train_epoch(
         if saver is not None and args.recovery_interval and (
                 last_batch or (batch_idx + 1) % args.recovery_interval == 0):
             saver.save_recovery(
-                model, optimizer, args, epoch, model_ema=model_ema, batch_idx=batch_idx)
+                model, optimizer, args, epoch, model_ema=model_ema, use_amp=use_amp, batch_idx=batch_idx)
 
         if lr_scheduler is not None:
             lr_scheduler.step_update(num_updates=num_updates, metric=losses_m.avg)