DL-Art-School/codes/trainer/steps.py

374 lines
20 KiB
Python
Raw Normal View History

from torch.cuda.amp import GradScaler
2021-12-26 04:32:01 +00:00
from torch.distributed.optim import ZeroRedundancyOptimizer
2020-08-22 14:24:34 +00:00
from utils.loss_accumulator import LossAccumulator
from torch.nn import Module
import logging
2020-12-18 16:18:34 +00:00
from trainer.losses import create_loss
2020-08-22 14:24:34 +00:00
import torch
from collections import OrderedDict
2020-12-30 03:58:02 +00:00
from trainer.inject import create_injector
from utils.util import recursively_detach, opt_get, clip_grad_norm
2020-08-12 14:45:23 +00:00
2020-08-22 14:24:34 +00:00
logger = logging.getLogger('base')
2020-08-12 14:45:23 +00:00
2023-02-22 23:07:05 +00:00
import bitsandbytes as bnb
2020-08-12 14:45:23 +00:00
2020-08-22 14:24:34 +00:00
# Defines the expected API for a single training step
class ConfigurableStep(Module):
def __init__(self, opt_step, env):
super(ConfigurableStep, self).__init__()
self.step_opt = opt_step
self.env = env
self.opt = env['opt']
self.gen_outputs = opt_step['generator_outputs']
self.loss_accumulator = LossAccumulator(buffer_sz=opt_get(opt_step, ['loss_log_buffer'], 50))
self.optimizers = None
self.scaler = GradScaler(enabled=self.opt['fp16'] or opt_get(self.opt, ['grad_scaler_enabled'], False))
self.grads_generated = False
2021-07-09 04:13:44 +00:00
self.clip_grad_eps = opt_get(opt_step, ['clip_grad_eps'], None)
2020-08-22 14:24:34 +00:00
# This is a half-measure that can be used between anomaly_detection and running a potentially problematic
# trainer bare. With this turned on, the optimizer will not step() if a nan grad is detected. If a model trips
# this warning 10 times in a row, the training session is aborted and the model state is saved. This has a
# noticeable affect on training speed, but nowhere near as bad as anomaly_detection.
self.check_grads_for_nan = opt_get(opt_step, ['check_grads_for_nan'], False)
self.nan_counter = 0
# This is a similar mechanism plugged into the forward() pass. It cannot be turned off.
self.nan_loss_counter = 0
2020-08-22 14:24:34 +00:00
self.injectors = []
if 'injectors' in self.step_opt.keys():
injector_names = []
2020-08-22 14:24:34 +00:00
for inj_name, injector in self.step_opt['injectors'].items():
assert inj_name not in injector_names # Repeated names are always an error case.
injector_names.append(inj_name)
2020-08-22 14:24:34 +00:00
self.injectors.append(create_injector(injector, env))
losses = []
self.weights = {}
if 'losses' in self.step_opt.keys():
for loss_name, loss in self.step_opt['losses'].items():
assert loss_name not in self.weights.keys() # Repeated names are always an error case.
2020-09-29 04:09:42 +00:00
losses.append((loss_name, create_loss(loss, env)))
self.weights[loss_name] = loss['weight']
2020-08-22 14:24:34 +00:00
self.losses = OrderedDict(losses)
def get_network_for_name(self, name):
return self.env['generators'][name] if name in self.env['generators'].keys() \
else self.env['discriminators'][name]
2020-08-22 14:24:34 +00:00
# Subclasses should override this to define individual optimizers. They should all go into self.optimizers.
# This default implementation defines a single optimizer for all Generator parameters.
# Must be called after networks are initialized and wrapped.
2020-08-22 14:24:34 +00:00
def define_optimizers(self):
opt_configs = [opt_get(self.step_opt, ['optimizer_params'], None)]
2021-06-11 21:31:10 +00:00
self.optimizers = []
if opt_configs[0] is None:
2021-06-11 21:31:10 +00:00
return
training = self.step_opt['training']
training_net = self.get_network_for_name(training)
nets = [training_net]
training = [training]
for net_name, net, opt_config in zip(training, nets, opt_configs):
# Configs can organize parameters by-group and specify different learning rates for each group. This only
# works in the model specifically annotates which parameters belong in which group using PARAM_GROUP.
optim_params = {'default': {'params': [], 'lr': opt_config['lr']}}
2021-06-11 21:31:10 +00:00
if opt_config is not None and 'param_groups' in opt_config.keys():
for k, pg in opt_config['param_groups'].items():
optim_params[k] = {'params': [], 'lr': pg['lr']}
2021-07-31 21:58:41 +00:00
import torch.nn as nn
norm_modules = (nn.BatchNorm2d, nn.InstanceNorm2d, nn.BatchNorm1d, nn.InstanceNorm1d,
nn.BatchNorm3d, nn.InstanceNorm3d, nn.GroupNorm, nn.LayerNorm)
2023-02-22 23:07:05 +00:00
# nn.Embedding
emb_modules = (bnb.nn.StableEmbedding, nn.EmbeddingBag)
2021-08-04 03:08:27 +00:00
param_names_notweights = set()
all_param_names = set()
param_map = {}
2021-07-31 21:58:41 +00:00
for mn, m in net.named_modules():
for k, v in m.named_parameters():
v.is_bias = k.endswith(".bias")
v.is_weight = k.endswith(".weight")
v.is_norm = isinstance(m, norm_modules)
v.is_emb = isinstance(m, emb_modules)
2021-08-04 03:08:27 +00:00
fpn = '%s.%s' % (mn, k) if mn else k # full param name
all_param_names.add(fpn)
param_map[fpn] = v
2021-07-31 21:58:41 +00:00
if v.is_bias or v.is_norm or v.is_emb:
2021-08-04 03:08:27 +00:00
param_names_notweights.add(fpn)
2021-07-31 21:58:41 +00:00
# Some models can specify some parameters to be in different groups.
param_group = "default"
if hasattr(v, 'PARAM_GROUP'):
if v.PARAM_GROUP in optim_params.keys():
param_group = v.PARAM_GROUP
else:
logger.warning(f'Model specifies a custom param group {v.PARAM_GROUP} which is not configured. '
f'The same LR will be used for all parameters.')
if v.requires_grad:
optim_params[param_group]['params'].append(v)
else:
if self.env['rank'] <= 0:
logger.warning('Params [{:s}] will not optimize.'.format(k))
params_names_notweights = sorted(list(param_names_notweights))
params_notweights = [param_map[k] for k in params_names_notweights]
params_names_weights = sorted(list(all_param_names ^ param_names_notweights))
params_weights = [param_map[k] for k in params_names_weights]
2022-02-16 03:53:07 +00:00
if 'optimizer' not in self.step_opt.keys() or self.step_opt['optimizer'] == 'adamw':
2021-07-31 21:58:41 +00:00
groups = [
2021-08-04 03:08:27 +00:00
{ 'params': params_weights, 'weight_decay': opt_get(opt_config, ['weight_decay'], 0) },
{ 'params': params_notweights, 'weight_decay': 0 }
2021-07-31 21:58:41 +00:00
]
2023-02-22 23:07:05 +00:00
# torch.optim.AdamW
opt = bnb.optim.AdamW8bit(groups, lr=opt_config['lr'],
weight_decay=opt_get(opt_config, ['weight_decay'], 1e-2),
betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
opt._group_names = [params_names_weights, params_names_notweights]
2022-06-10 03:15:09 +00:00
elif self.step_opt['optimizer'] == 'mu_adamw':
groups = [
{ 'params': params_weights, 'weight_decay': opt_get(opt_config, ['weight_decay'], 0) },
{ 'params': params_notweights, 'weight_decay': 0 }
]
from mup.optim import MuAdamW
opt = MuAdamW(groups, lr=opt_config['lr'],
weight_decay=opt_get(opt_config, ['weight_decay'], 1e-2),
betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
opt._group_names = [params_names_weights, params_names_notweights]
2021-12-26 04:32:01 +00:00
elif self.step_opt['optimizer'] == 'adamw_zero':
# The torch ZeRO implementation does not seem to support parameter groups, so do not shard the non-weighted
# parameters and just use a normal AdamW implementation. In a large network, these weights will normally
# be a tiny fraction of the total weights.
2023-02-22 23:07:05 +00:00
# torch.optim.AdamW
opt_unweighted = bnb.optim.AdamW8bit(params_notweights, lr=opt_config['lr'], weight_decay=0,
2021-12-26 04:32:01 +00:00
betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
opt_unweighted._config = opt_config
opt_unweighted._config['network'] = net_name
opt_unweighted._group_names = []
2022-02-15 13:58:11 +00:00
self.optimizers.append(opt_unweighted)
2022-02-15 14:08:17 +00:00
2023-02-22 23:07:05 +00:00
# torch.optim.AdamW
opt = ZeroRedundancyOptimizer(params_weights, optimizer_class=bnb.optim.AdamW8bit, lr=opt_config['lr'],
2021-12-26 04:32:01 +00:00
weight_decay=opt_get(opt_config, ['weight_decay'], 1e-2),
betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
2022-01-30 23:14:06 +00:00
opt.param_groups[0]['initial_lr'] = opt_config['lr']
2021-12-26 04:32:01 +00:00
opt._group_names = []
elif self.step_opt['optimizer'] == 'lars':
from trainer.optimizers.larc import LARC
from trainer.optimizers.sgd import SGDNoBiasMomentum
optSGD = SGDNoBiasMomentum(list(optim_params.values()), lr=opt_config['lr'], momentum=opt_config['momentum'],
weight_decay=opt_config['weight_decay'])
opt = LARC(optSGD, trust_coefficient=opt_config['lars_coefficient'])
opt._group_names = sorted(list(all_param_names))
2022-02-15 13:48:13 +00:00
elif self.step_opt['optimizer'] == 'lamb':
from trainer.optimizers.lamb import Lamb
2023-02-22 23:07:05 +00:00
# torch.optim.AdamW
opt_unweighted = bnb.optim.AdamW8bit(params_notweights, lr=opt_config['lr'], weight_decay=0,
2022-02-15 13:48:13 +00:00
betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
opt_unweighted._config = opt_config
opt_unweighted._config['network'] = net_name
2022-02-15 13:58:11 +00:00
opt_unweighted._group_names = []
2022-02-15 13:48:13 +00:00
self.optimizers.append(opt_unweighted)
opt = Lamb(params_weights, lr=opt_config['lr'],
weight_decay=opt_get(opt_config, ['weight_decay'], 1e-2),
betas=(opt_get(opt_config, ['beta1'], .9), opt_get(opt_config, ['beta2'], .999)))
2022-02-15 14:08:17 +00:00
opt._group_names = []
2021-01-12 03:09:16 +00:00
elif self.step_opt['optimizer'] == 'sgd':
from torch.optim import SGD
opt = SGD(list(optim_params.values()), lr=opt_config['lr'], momentum=opt_config['momentum'], weight_decay=opt_config['weight_decay'])
opt._group_names = sorted(list(all_param_names))
opt._config = opt_config # This is a bit seedy, but we will need these configs later.
opt._config['network'] = net_name
self.optimizers.append(opt)
2020-08-12 14:45:23 +00:00
# Returns all optimizers used in this step.
def get_optimizers(self):
2020-08-22 14:24:34 +00:00
assert self.optimizers is not None
return self.optimizers
2020-08-12 14:45:23 +00:00
# Returns optimizers which are opting in for default LR scheduling.
def get_optimizers_with_default_scheduler(self):
2020-08-22 14:24:34 +00:00
assert self.optimizers is not None
return self.optimizers
2020-08-12 14:45:23 +00:00
# Returns the names of the networks this step will train. Other networks will be frozen.
def get_networks_trained(self):
if isinstance(self.step_opt['training'], list):
return self.step_opt['training']
else:
return [self.step_opt['training']]
2020-08-22 14:24:34 +00:00
2020-09-29 04:09:42 +00:00
def get_training_network_name(self):
if isinstance(self.step_opt['training'], list):
return self.step_opt['training'][0]
else:
return self.step_opt['training']
2020-08-22 14:24:34 +00:00
# Performs all forward and backward passes for this step given an input state. All input states are lists of
# chunked tensors. Use grad_accum_step to dereference these steps. Should return a dict of tensors that later
# steps might use. These tensors are automatically detached and accumulated into chunks.
def do_forward_backward(self, state, grad_accum_step, amp_loss_id, train=True, no_ddp_sync=False, loss_accumulator=None):
2020-12-30 16:51:59 +00:00
local_state = {} # <-- Will store the entire local state to be passed to injectors & losses.
new_state = {} # <-- Will store state values created by this step for returning to ExtensibleTrainer.
2020-08-22 14:24:34 +00:00
for k, v in state.items():
local_state[k] = v[grad_accum_step]
local_state['train_nets'] = str(self.get_networks_trained())
loss_accumulator = self.loss_accumulator if loss_accumulator is None else loss_accumulator
2020-08-12 14:45:23 +00:00
# Some losses compute backward() internally. Accommodate this by stashing the amp_loss_id in env.
2020-09-25 22:38:23 +00:00
self.env['amp_loss_id'] = amp_loss_id
self.env['current_step_optimizers'] = self.optimizers
self.env['training'] = train
2020-09-25 22:38:23 +00:00
# Inject in any extra dependencies.
for inj in self.injectors:
# Don't do injections tagged with eval unless we are not in train mode.
if train and 'eval' in inj.opt.keys() and inj.opt['eval']:
continue
# Likewise, don't do injections tagged with train unless we are not in eval.
if not train and 'train' in inj.opt.keys() and inj.opt['train']:
continue
# Don't do injections tagged with 'after' or 'before' when we are out of spec.
if 'after' in inj.opt.keys() and self.env['step'] < inj.opt['after'] or \
2020-11-20 06:47:24 +00:00
'before' in inj.opt.keys() and self.env['step'] > inj.opt['before'] or \
'every' in inj.opt.keys() and self.env['step'] % inj.opt['every'] != 0:
continue
2021-06-06 19:57:22 +00:00
if 'no_accum' in inj.opt.keys() and grad_accum_step > 0:
continue
training_net = self.get_network_for_name(self.step_opt['training'])
if no_ddp_sync and hasattr(training_net, 'no_sync'):
with training_net.no_sync():
injected = inj(local_state)
2021-08-04 03:08:27 +00:00
elif opt_get(inj.opt, ['no_grad'], False):
with torch.no_grad():
injected = inj(local_state)
else:
injected = inj(local_state)
local_state.update(injected)
new_state.update(injected)
2022-05-16 03:50:38 +00:00
if hasattr(inj, 'extra_metrics'):
for n, v in inj.extra_metrics().items():
# Doesn't really work for training setups where multiple of the same injector are used.
loss_accumulator.add_loss(n, v)
if len(self.losses) > 0:
# Finally, compute the losses.
total_loss = 0
for loss_name, loss in self.losses.items():
multiplier = 1
# Some losses only activate after a set number of steps. For example, proto-discriminator losses can
# be very disruptive to a generator.
if 'after' in loss.opt.keys() and loss.opt['after'] > self.env['step'] or \
2020-11-12 22:42:05 +00:00
'before' in loss.opt.keys() and self.env['step'] > loss.opt['before'] or \
'every' in loss.opt.keys() and self.env['step'] % loss.opt['every'] != 0:
multiplier = 0 # Multiply by 0 so gradients still flow and DDP works. Effectively this means the loss is unused.
2020-12-30 16:51:59 +00:00
if loss.is_stateful():
l, lstate = loss(self.get_network_for_name(self.step_opt['training']), local_state)
local_state.update(lstate)
new_state.update(lstate)
else:
l = loss(self.get_network_for_name(self.step_opt['training']), local_state)
2021-11-10 06:58:41 +00:00
if not l.isfinite():
print(f'!!Detected non-finite loss {loss_name}')
total_loss += l * self.weights[loss_name] * multiplier
# Record metrics.
if isinstance(l, torch.Tensor):
loss_accumulator.add_loss(loss_name, l)
for n, v in loss.extra_metrics():
loss_accumulator.add_loss("%s_%s" % (loss_name, n), v)
loss.clear_metrics()
# In some cases, the loss could not be set (e.g. all losses have 'after')
if train and isinstance(total_loss, torch.Tensor) and total_loss.isfinite():
loss_accumulator.add_loss("%s_total" % (self.get_training_network_name(),), total_loss)
2020-10-28 21:46:59 +00:00
# Scale the loss down by the accumulation factor.
total_loss = total_loss / self.env['mega_batch_factor']
# Get dem grads!
self.scaler.scale(total_loss).backward()
self.grads_generated = True
# Reset nan_loss_counter
self.nan_loss_counter = 0
elif not total_loss.isfinite():
print("Non-finite loss encountered. Skipping backwards step.")
self.nan_loss_counter += 1
if self.nan_loss_counter > 10:
print("Encountered 10 NaN losses in a row. Something is screwed up. Dumping model weights and exiting.")
if self.env['rank'] == 0:
torch.save(training_net.state_dict(), "nan_error_weights.pth")
exit(1)
2020-08-22 14:24:34 +00:00
# Detach all state variables. Within the step, gradients can flow. Once these variables leave the step
# we must release the gradients.
new_state = recursively_detach(new_state)
# Prune state outputs that are not actually needed.
if 'step_outputs' in self.step_opt.keys():
nst = {}
for k in self.step_opt['step_outputs']:
nst[k] = new_state[k]
new_state = nst
2020-08-22 14:24:34 +00:00
return new_state
# Performs the optimizer step after all gradient accumulation is completed. Default implementation simply steps()
# all self.optimizers.
def do_step(self, step):
if not self.grads_generated:
return
self.grads_generated = False
2020-08-22 14:24:34 +00:00
for opt in self.optimizers:
2022-03-16 16:16:34 +00:00
# self.scaler.unscale_(opt) It would be important to do this here, but ExtensibleTrainer currently does it.
2022-03-13 22:28:23 +00:00
# Optimizers can be opted out in the early stages of training.
after = opt._config['after'] if 'after' in opt._config.keys() else 0
after_network = self.opt['networks'][opt._config['network']]['after'] if 'after' in self.opt['networks'][opt._config['network']].keys() else 0
after = max(after, after_network)
if self.env['step'] < after:
continue
before = opt._config['before'] if 'before' in opt._config.keys() else -1
if before != -1 and self.env['step'] > before:
continue
nan_found = False
if self.check_grads_for_nan:
for pg in opt.param_groups:
for p in pg['params']:
if not torch.isfinite(p.grad).any():
nan_found = True
break
if nan_found:
break
if nan_found:
print("NaN found in grads. Throwing this step out.")
self.nan_counter += 1
else:
self.nan_counter = 0
2022-02-11 17:46:37 +00:00
if self.clip_grad_eps is not None and self.clip_grad_eps != 0:
for pgn, pg in zip(opt._group_names, opt.param_groups):
grad_norm = clip_grad_norm(pg['params'], pgn, self.clip_grad_eps)
2021-07-09 04:13:44 +00:00
if torch.isnan(grad_norm):
2022-02-23 06:09:37 +00:00
print("NaN found in clip_grad; zeroing grad and trying again.")
2021-07-09 04:13:44 +00:00
nan_found = True
self.nan_counter += 1
if not nan_found:
self.scaler.step(opt)
self.scaler.update()
2022-01-09 01:24:58 +00:00
else:
2022-01-09 03:31:19 +00:00
opt.zero_grad()
2020-08-22 14:24:34 +00:00
def get_metrics(self):
2022-03-08 22:52:42 +00:00
metrics = self.loss_accumulator.as_dict()
metrics['grad_scaler_scale'] = self.scaler.get_scale()
return metrics