-
Notifications
You must be signed in to change notification settings - Fork 411
/
xdecoder_trainer.py
182 lines (153 loc) · 7.9 KB
/
xdecoder_trainer.py
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
181
182
# --------------------------------------------------------
# X-Decoder -- Generalized Decoding for Pixel, Image, and Language
# Copyright (c) 2022 Microsoft
# Licensed under The MIT License [see LICENSE for details]
# Modified by Xueyan Zou ([email protected])
# --------------------------------------------------------
import logging
import os
import json
import random
import copy
import itertools
from typing import Any, Dict, List, Set, Union
from datetime import datetime
from mpi4py import MPI
import numpy as np
import torch
from torch.utils.data import DataLoader
from detectron2.projects.deeplab import build_lr_scheduler
from fvcore.common.config import CfgNode
from infinibatch import iterators
from utils.distributed import is_main_process, get_world_size
from .default_trainer import DefaultTrainer
from .utils.serialization import JSONEncoder, filter_jsonable
logger = logging.getLogger(__name__)
class XDecoder_Trainer(DefaultTrainer):
"""
Construct Mask2Former_Trainer for optimizer and lr_scheduler
"""
def create_optimizer_and_scheduler(self):
"""
Set up self.optimizers and self.lr_schedulers
This method initializes self.optimizers and self.lr_schedulers as dictionaries of
instances of the classes that OPTIMIZER and LR_SCHEDULER in the config file points to.
One optimizer and lr scheduler for each model in self.raw_models. They have the same keys
as self.raw_models.
"""
self.opt['init_optimizer_in_deepspeed'] = False
self.opt['init_lr_scheduler_in_deepspeed'] = False
self.optimizers = {module_name: None for module_name in self.model_names}
self.lr_schedulers = {module_name: None for module_name in self.model_names}
cfg_solver = self.opt['SOLVER']
weight_decay_norm = cfg_solver['WEIGHT_DECAY_NORM']
weight_decay_embed = cfg_solver['WEIGHT_DECAY_EMBED']
weight_decay_bias = cfg_solver.get('WEIGHT_DECAY_BIAS', 0.0)
defaults = {}
defaults["lr"] = cfg_solver['BASE_LR']
defaults["weight_decay"] = cfg_solver['WEIGHT_DECAY']
norm_module_types = (
torch.nn.BatchNorm1d,
torch.nn.BatchNorm2d,
torch.nn.BatchNorm3d,
torch.nn.SyncBatchNorm,
# NaiveSyncBatchNorm inherits from BatchNorm2d
torch.nn.GroupNorm,
torch.nn.InstanceNorm1d,
torch.nn.InstanceNorm2d,
torch.nn.InstanceNorm3d,
torch.nn.LayerNorm,
torch.nn.LocalResponseNorm,
)
fix_param = self.opt['SOLVER'].get('FIX_PARAM',{})
ignore_fix = self.opt['SOLVER'].get('IGNORE_FIX',[])
for _module_name in self.model_names:
flag_continue = False
for name, param in self.raw_models[_module_name].named_parameters():
for ig in ignore_fix:
if ig in name:
flag_continue = True
break
if flag_continue:
flag_continue = False
continue
for key, value in fix_param.items():
if key in name and value == True:
param.requires_grad = False
lr_multiplier = self.opt['SOLVER']['LR_MULTIPLIER']
for _module_name in self.model_names:
# parameters = self.raw_models[module_name].get_training_parameters()
# self.optimizers[module_name] = optimizer_class(parameters, **optimizer_parameters)
# params = []
# for module_param_name, value in self.raw_models[module_name].named_parameters(recurse=True):
params: List[Dict[str, Any]] = []
memo: Set[torch.nn.parameter.Parameter] = set()
for module_name, module in self.raw_models[_module_name].named_modules():
for module_param_name, value in module.named_parameters(recurse=False):
if not value.requires_grad:
continue
# Avoid duplicating parameters
if value in memo:
continue
memo.add(value)
hyperparams = copy.copy(defaults)
for key, lr_mul in lr_multiplier.items():
if key in "{}.{}".format(module_name, module_param_name):
hyperparams["lr"] = hyperparams["lr"] * lr_mul
if is_main_process():
logger.info("Modify Learning rate of {}: {}".format("{}.{}".format(module_name, module_param_name), lr_mul))
if (
"relative_position_bias_table" in module_param_name
or "absolute_pos_embed" in module_param_name
):
hyperparams["weight_decay"] = 0.0
if isinstance(module, norm_module_types):
hyperparams["weight_decay"] = weight_decay_norm
if isinstance(module, torch.nn.Embedding):
hyperparams["weight_decay"] = weight_decay_embed
if "bias" in module_name:
hyperparams["weight_decay"] = weight_decay_bias
params.append({"params": [value], **hyperparams})
def maybe_add_full_model_gradient_clipping(optim):
# detectron2 doesn't have full model gradient clipping now
clip_norm_val = cfg_solver['CLIP_GRADIENTS']['CLIP_VALUE']
enable = (
cfg_solver['CLIP_GRADIENTS']['ENABLED']
and cfg_solver['CLIP_GRADIENTS']['CLIP_TYPE'] == "full_model"
and clip_norm_val > 0.0
)
class FullModelGradientClippingOptimizer(optim):
def step(self, closure=None):
all_params = itertools.chain(*[x["params"] for x in self.param_groups])
torch.nn.utils.clip_grad_norm_(all_params, clip_norm_val)
super().step(closure=closure)
return FullModelGradientClippingOptimizer if enable else optim
optimizer_type = cfg_solver['OPTIMIZER']
if optimizer_type == "SGD":
optimizer = maybe_add_full_model_gradient_clipping(torch.optim.SGD)(
params, cfg_solver['BASE_LR'], momentum=cfg_solver['MOMENTUM']
)
elif optimizer_type == "ADAMW":
optimizer = maybe_add_full_model_gradient_clipping(torch.optim.AdamW)(
params, cfg_solver['BASE_LR']
)
else:
raise NotImplementedError(f"no optimizer type {optimizer_type}")
self.optimizers[_module_name] = optimizer
self.optimizers[_module_name].zero_grad()
num_epoch = self.opt['SOLVER']['MAX_NUM_EPOCHS']
cfg_solver['MAX_ITER'] = num_epoch * self.train_params['updates_per_epoch']
cfg_solver['STEPS'] = [int(x*cfg_solver['MAX_ITER']) for x in cfg_solver['STEPS']]
logger.info(f"Calculate MAX_ITER @ {cfg_solver['MAX_ITER']} and STEPS @ {cfg_solver['STEPS']}")
for module_name in self.model_names:
scheduler_cfg = CfgNode({'SOLVER': cfg_solver})
self.lr_schedulers[module_name] = build_lr_scheduler(scheduler_cfg, self.optimizers[module_name])
for module_name in self.model_names:
num_params = 0
num_trainable_params = 0
for name, param in self.raw_models[module_name].named_parameters():
num_params += param.numel()
if param.requires_grad:
num_trainable_params += param.numel()
logger.info(f"Total number of parameters in {module_name} module (on each GPU): {num_params}")
logger.info(f"Number of trainable parameters in {module_name} module (on each GPU): {num_trainable_params}")