forked from deepcam-cn/FaceQuality
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain_quality.py
executable file
·177 lines (159 loc) · 7.69 KB
/
train_quality.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
import torch
import torch.nn as nn
import torch.optim as optim
import math
import cv2
import torchvision.transforms as transforms
from dataset.dataset import ImageFolder
from config import config
from models.model_resnet import ResNet, FaceQuality
from models.metrics import GaussianFace
from models.focal import FocalLoss
from util.utils import *
import torch.distributed as dist
import torch.multiprocessing as mp
from tensorboardX import SummaryWriter
from util.cosine_lr_scheduler import CosineDecayLR
from tqdm import tqdm
import os
import random
import numbers
import shutil
import argparse
import numpy as np
from ptflops import get_model_complexity_info
def load_state_dict(model, state_dict):
all_keys = {k for k in state_dict.keys()}
for k in all_keys:
if k.startswith('module.'):
state_dict[k[7:]] = state_dict.pop(k)
model_dict = model.state_dict()
pretrained_dict = {k:v for k, v in state_dict.items() if k in model_dict and v.size() == model_dict[k].size()}
if len(pretrained_dict) == len(model_dict):
print("all params loaded")
else:
not_loaded_keys = {k for k in pretrained_dict.keys() if k not in model_dict.keys()}
print("not loaded keys:", not_loaded_keys)
model_dict.update(pretrained_dict)
model.load_state_dict(model_dict)
def train():
DEVICE = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
writer = SummaryWriter(config.LOG_ROOT)
train_transform = transforms.Compose([
transforms.RandomApply([transforms.RandomResizedCrop(112, scale=(0.95, 1), ratio=(1, 1))]),
transforms.Resize(112),
transforms.RandomHorizontalFlip(),
transforms.RandomGrayscale(0.01),
transforms.ToTensor(),
transforms.Normalize(mean = config.RGB_MEAN, std = config.RGB_STD),
])
dataset_train = ImageFolder(config.TRAIN_FILES, train_transform)
train_loader = torch.utils.data.DataLoader(
dataset_train, batch_size = config.BATCH_SIZE, pin_memory = True, shuffle=True,
num_workers = 8, drop_last = True
)
NUM_CLASS = train_loader.dataset.classes
print("Number of Training Classes: {}".format(NUM_CLASS))
QUALITY = FaceQuality(512 * 7 * 7)
BACKBONE = ResNet(num_layers=100, feature_dim=512)
flops, params = get_model_complexity_info(BACKBONE, (3, 112, 112), as_strings=True, print_per_layer_stat=False)
print('BACKBONE FLOPs:', flops)
print('BACKBONE PARAMS:', params)
HEAD = GaussianFace(in_features = config.EMBEDDING_SIZE, out_features = NUM_CLASS)
LOSS = FocalLoss()
if config.BACKBONE_RESUME_ROOT and config.HEAD_RESUME_ROOT:
print("=" * 60)
if os.path.isfile(config.BACKBONE_RESUME_ROOT):
print("Loading Backbone Checkpoint '{}'".format(config.BACKBONE_RESUME_ROOT))
checkpoint = torch.load(config.BACKBONE_RESUME_ROOT)
load_state_dict(BACKBONE, checkpoint)
else:
print("No Checkpoint Found at '{}' Please Have a Check or Continue to Train from Scratch".format(config.BACKBONE_RESUME_ROOT))
if os.path.isfile(config.HEAD_RESUME_ROOT):
print("Loading Head Checkpoint '{}'".format(config.HEAD_RESUME_ROOT))
checkpoint = torch.load(config.HEAD_RESUME_ROOT)
load_state_dict(HEAD, checkpoint)
else:
print("No Checkpoint Found at '{}' Please Have a Check or Continue to Train from Scratch".format(config.HEAD_RESUME_ROOT))
print("=" * 60)
else:
print('Error: Pretrained backbone and head are necessary for quality training')
return
BACKBONE = nn.DataParallel(BACKBONE, device_ids = config.BACKBONE_GPUS)
BACKBONE = BACKBONE.cuda(config.BACKBONE_GPUS[0])
QUALITY = nn.DataParallel(QUALITY, device_ids = config.BACKBONE_GPUS)
QUALITY = QUALITY.cuda(config.BACKBONE_GPUS[0])
HEAD = nn.DataParallel(HEAD, device_ids = config.HEAD_GPUS, output_device=config.HEAD_GPUS[0])
HEAD = HEAD.cuda(config.HEAD_GPUS[0])
BACKBONE.eval()
OPTIMIZER = optim.SGD([{'params': QUALITY.parameters(), 'lr': config.QUALITY_LR}, {'params': HEAD.parameters(), 'lr': config.QUALITY_LR}], momentum=config.MOMENTUM)
DISP_FREQ = len(train_loader) // 100
NUM_EPOCH_WARM_UP = config.NUM_EPOCH_WARM_UP
NUM_BATCH_WARM_UP = len(train_loader) * NUM_EPOCH_WARM_UP
batch = 0
step = 0
scheduler = CosineDecayLR(OPTIMIZER, T_max=10*len(train_loader), lr_init = config.QUALITY_LR, lr_min = 1e-5, warmup = NUM_BATCH_WARM_UP)
for epoch in range(config.NUM_EPOCH):
HEAD.train()
QUALITY.train()
arcface_losses = AverageMeter()
confidences = AverageMeter()
top1 = AverageMeter()
top5 = AverageMeter()
scaler = torch.cuda.amp.GradScaler()
for inputs, labels in tqdm(iter(train_loader)):
inputs = inputs.cuda(config.BACKBONE_GPUS[0])
labels = labels.cuda(config.HEAD_GPUS[0])
with torch.no_grad():
features, fc = BACKBONE(inputs, True)
with torch.cuda.amp.autocast():
confidence = QUALITY(fc)
outputs = HEAD(confidence.cuda(config.HEAD_GPUS[0]), features.cuda(config.HEAD_GPUS[0]), labels, True)
arcface_loss = LOSS(outputs, labels)
# measure accuracy and record loss
prec1, prec5 = accuracy(outputs.data, labels, topk = (1, 5))
arcface_losses.update(arcface_loss.data.item(), inputs.size(0))
confidences.update(torch.mean(confidence).data.item(), inputs.size(0))
top1.update(prec1.data.item(), inputs.size(0))
top5.update(prec5.data.item(), inputs.size(0))
loss = arcface_loss
# compute gradient and do SGD step
OPTIMIZER.zero_grad()
#loss.backward()
#OPTIMIZER.step()
scaler.scale(loss).backward()
scaler.step(OPTIMIZER)
scaler.update()
if ((batch + 1) % DISP_FREQ == 0) and batch != 0:
print("=" * 60)
print('Epoch {}/{} Batch {}/{}\t'
'Training Loss {arcface_loss.val:.4f}({arcface_loss.avg:.4f})\t'
'Training Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Training Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
epoch + 1, config.NUM_EPOCH, batch + 1, len(train_loader) * config.NUM_EPOCH,
arcface_loss = arcface_losses, top1 = top1, top5 = top5))
print("=" * 60)
batch += 1 # batch index
scheduler.step(batch)
if batch % 1000 == 0:
print(OPTIMIZER)
# training statistics per epoch (buffer for visualization)
epoch_loss = arcface_losses.avg
epoch_acc = top1.avg
writer.add_scalar("Training_Loss", epoch_loss, epoch + 1)
writer.add_scalar("Training_Accuracy", epoch_acc, epoch + 1)
print("=" * 60)
print('Epoch: {}/{}\t'
'Training Loss {loss.val:.4f} ({loss.avg:.4f})\t'
'Training Prec@1 {top1.val:.3f} ({top1.avg:.3f})\t'
'Training Prec@5 {top5.val:.3f} ({top5.avg:.3f})'.format(
epoch + 1, config.NUM_EPOCH, loss = arcface_losses, top1 = top1, top5 = top5))
print("=" * 60)
# save checkpoints per epoch
curTime = get_time()
if not os.path.exists(config.MODEL_ROOT):
os.makedirs(config.MODEL_ROOT)
torch.save(QUALITY.state_dict(), os.path.join(config.MODEL_ROOT, "Quality_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(epoch + 1, batch, curTime)))
torch.save(HEAD.state_dict(), os.path.join(config.MODEL_ROOT, "Head_Epoch_{}_Batch_{}_Time_{}_checkpoint.pth".format(epoch + 1, batch, curTime)))
if __name__ == "__main__":
train()