训练结果acc 为什么一直这么低，但是lfw face verification accuracy: 0.9095 threshold: 0.37753257 结果是这样的 #115
Description
total time is 254.24768543243408, average time is 0.9892906047954634
lfw face verification accuracy: 0.9095 threshold: 0.37753257
Sat Dec 23 08:19:27 2023 train epoch 2 iter 10 0.08301005559180279 iters/s loss 12.163952827453613 acc 0.0234375
Sat Dec 23 08:35:10 2023 train epoch 2 iter 110 0.10600735235136766 iters/s loss 11.601116180419922 acc 0.02734375
Sat Dec 23 08:51:06 2023 train epoch 2 iter 210 0.10467295813176283 iters/s loss 11.698826789855957 acc 0.02734375
Sat Dec 23 09:07:12 2023 train epoch 2 iter 310 0.10353925427018165 iters/s loss 11.612323760986328 acc 0.0390625
Sat Dec 23 09:23:40 2023 train epoch 2 iter 410 0.10121389428942351 iters/s loss 11.697545051574707 acc 0.04296875
Sat Dec 23 09:39:57 2023 train epoch 2 iter 510 0.10231825089000418 iters/s loss 11.737585067749023 acc 0.046875
Sat Dec 23 09:57:14 2023 train epoch 2 iter 610 0.09639180206759532 iters/s loss 11.6640625 acc 0.03515625
训练数据是CASIA-WebFace 10575的那份数据
以下是配置文件：
class Config(object):
env = 'default'
backbone = 'resnet18'
classify = 'softmax'
num_classes = 13938 #10575
metric = 'arc_margin'
easy_margin = False
use_se = False
loss = 'focal_loss'
display = False
finetune = False
train_root = /facedata/CASIA-WebFace_align/'
train_list = '/img_train.txt'
#val_list = '/data/Datasets/webface/val_data_13938.txt'
#test_root = '/data1/Datasets/anti-spoofing/test/data_align_256'
#test_list = 'test.txt'
lfw_root = ‘/facedata/lfw-align-128'
lfw_test_list = './lfw_test_pair.txt'
checkpoints_path = 'checkpoint'
load_model_path = 'models/resnet18.pth'
test_model_path = 'checkpoint/resnet18_110.pth'
save_interval = 5
train_batch_size = 256 # batch size
test_batch_size = 60
input_shape = (1, 128, 128)
optimizer = 'sgd'
use_gpu = True # use GPU or not
gpu_id = '0, 1'
num_workers = 4 # how many workers for loading data
print_freq = 100 # print info every N batch
debug_file = '/tmp/debug' # if os.path.exists(debug_file): enter ipdb
result_file = 'result.csv'
max_epoch = 50
lr = 1e-1 # initial learning rate
lr_step = 10
lr_decay = 0.95 # when val_loss increase, lr = lr*lr_decay
weight_decay = 5e-4