Spaces:
Running
Running
File size: 6,176 Bytes
6a62ffb |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 |
# Copyright (c) Facebook, Inc. and its affiliates.
#
# This source code is licensed under the MIT license found in the
# LICENSE file in the root directory of this source tree.
import torch
from fairseq import utils
from fairseq.dataclass.utils import gen_parser_from_dataclass
class FairseqOptimizer(object):
def __init__(self, cfg):
super().__init__()
self.cfg = cfg
@classmethod
def add_args(cls, parser):
"""Add optimizer-specific arguments to the parser."""
dc = getattr(cls, "__dataclass", None)
if dc is not None:
gen_parser_from_dataclass(parser, dc())
@property
def optimizer(self):
"""Return a torch.optim.optimizer.Optimizer instance."""
if not hasattr(self, "_optimizer"):
raise NotImplementedError
if not isinstance(self._optimizer, torch.optim.Optimizer):
raise ValueError("_optimizer must be an instance of torch.optim.Optimizer")
return self._optimizer
@optimizer.setter
def optimizer(self, optimizer):
"""Reset optimizer instance."""
if not hasattr(self, "_optimizer"):
raise NotImplementedError
if not isinstance(self._optimizer, torch.optim.Optimizer):
raise ValueError("_optimizer must be an instance of torch.optim.Optimizer")
self._optimizer = optimizer
@property
def optimizer_config(self):
"""
Return a kwarg dictionary that will be used to override optimizer
args stored in checkpoints. This allows us to load a checkpoint and
resume training using a different set of optimizer args, e.g., with a
different learning rate.
"""
raise NotImplementedError
@property
def params(self):
"""Return an iterable of the parameters held by the optimizer."""
for param_group in self.param_groups:
for p in param_group["params"]:
yield p
@property
def param_groups(self):
return self.optimizer.param_groups
def __getstate__(self):
return self._optimizer.__getstate__()
def get_lr(self):
"""Return the current learning rate."""
return self.param_groups[0]["lr"]
def set_lr(self, lr):
"""Set the learning rate."""
for param_group in self.param_groups:
param_group["lr"] = lr
def state_dict(self):
"""Return the optimizer's state dict."""
return self.optimizer.state_dict()
def load_state_dict(self, state_dict, optimizer_overrides=None):
"""Load an optimizer state dict.
In general we should prefer the configuration of the existing optimizer
instance (e.g., learning rate) over that found in the state_dict. This
allows us to resume training from a checkpoint using a new set of
optimizer args.
"""
self.optimizer.load_state_dict(state_dict)
if optimizer_overrides is not None and len(optimizer_overrides) > 0:
# override learning rate, momentum, etc. with latest values
for group in self.param_groups:
group.update(optimizer_overrides)
def backward(self, loss):
"""Computes the sum of gradients of the given tensor w.r.t. graph leaves."""
loss.backward()
def all_reduce_grads(self, module):
"""Manually all-reduce gradients (if required)."""
if hasattr(module, "all_reduce_grads"):
module.all_reduce_grads()
def multiply_grads(self, c):
"""Multiplies grads by a constant *c*."""
for p in self.params:
if p.grad is not None:
if torch.is_tensor(c):
c = c.to(p.grad.device)
p.grad.data.mul_(c)
def clip_grad_norm(self, max_norm, aggregate_norm_fn=None):
"""Clips gradient norm."""
return utils.clip_grad_norm_(self.params, max_norm, aggregate_norm_fn)
def step(self, closure=None, scale=1.0, groups=None):
"""Performs a single optimization step."""
if self.supports_step_with_scale:
if self.supports_groups:
self.optimizer.step(closure, scale=scale, groups=groups)
else:
self.optimizer.step(closure, scale=scale)
else:
if scale != 1.0:
self.multiply_grads(1.0 / scale)
if self.supports_groups:
self.optimizer.step(closure, groups=groups)
else:
self.optimizer.step(closure)
def zero_grad(self):
"""Clears the gradients of all optimized parameters."""
for p in self.params:
p.grad = None
self.optimizer.zero_grad()
@property
def supports_memory_efficient_fp16(self):
if hasattr(self.optimizer, "supports_memory_efficient_fp16"):
return self.optimizer.supports_memory_efficient_fp16
return False
@property
def supports_step_with_scale(self):
if hasattr(self.optimizer, "supports_step_with_scale"):
return self.optimizer.supports_step_with_scale
return False
@property
def supports_groups(self):
if hasattr(self.optimizer, "supports_groups"):
return self.optimizer.supports_groups
return False
@property
def supports_flat_params(self):
"""
Whether the optimizer supports collapsing of the model
parameters/gradients into a single contiguous Tensor.
"""
if hasattr(self.optimizer, "supports_flat_params"):
return self.optimizer.supports_flat_params
return False
def average_params(self):
pass
def broadcast_global_state_dict(self, state_dict):
"""
Broadcasts a global state dict to all ranks.
Useful for optimizers that shard state between ranks.
"""
if hasattr(self.optimizer, "broadcast_global_state_dict"):
return self.optimizer.broadcast_global_state_dict(state_dict)
else:
return state_dict
class LegacyFairseqOptimizer(FairseqOptimizer):
def __init__(self, args):
self.args = args
|