使用lightened_cnn.py训练 CASIA-WebFace ,accuracy 一直都小0.1,是为什么呢？

[adeagle]

No description provided.

[tornadomeet]

i don't know, because no detail about your experiment.

[adeagle]

一、数据集CASIA-webface

二、数据清洗方法:
https://github.com/happynear/FaceVerification
Good News: @潘泳苹果皮 and his colleagues have washed the CASIA-webface database manually. After washing, 27703 wrong images are deleted. The washed list can be downloaded from http://pan.baidu.com/s/1hrKpbm8

三、制作数据集
raw_data_path=/opt/mxnet/mxnet-face-master/verification/FullData
align_data_path=/opt/mxnet/mxnet-face-master/verification/FullData-align/
makelist_path=/opt/mxnet/mxnet/tools/make_list.py

the number threads used for align data, you shold change this depend on your environment

num_process=4
landmarks=innerEyesAndBottomLip
face_size=144
ts=0.1
list_name=casia
rec_name=casia
im2rec_path=/opt/mxnet/mxnet/bin/im2rec

step1:align the face iamge

if ! [ -e $align_data_path ];then
mkdir -p $align_data_path
for N in $(seq $num_process);do
echo "the sub-process is : $N"
python ../util/align_face.py $raw_data_path align $landmarks $align_data_path --ts $ts --size $face_size &
done
else
echo "$align_data_path already exist."
fi
wait
echo "Align face image done"

step2: generate .lst for im2rec

if ! [ -e ${list_name}image_train.lst ];then
python -u $makelist_path $align_data_path $list_name --train_ratio 0.95 --recursive True
else
echo ".lst file for training already exist."
fi
echo "generated .lst file done"

step3: use img2rec to generate .rec file for training

if ! [ -e ${rec_name}_train.rec ]; then
$im2rec_path ${list_name}image_train.lst $align_data_path ${rec_name}_train.rec color=0 encoding='.png' &
$im2rec_path ${list_name}image_val.lst $align_data_path ${rec_name}_val.rec color=0 encoding='.png' &
else
echo "$rec_name already exist."
fi
wait
echo "generate .rec done"
echo "trining done!"

四、数据集类数和训练样本数
ls FullData-align/ | wc
10575
cat casiaimage_train.lst | wc
413914

五、训练源码
lightened_cnn.py:

import argparse,logging
import mxnet as mx
import matplotlib.pyplot as plt
import numpy as np
import time
import threading
import Queue

logger = logging.getLogger()
logger.setLevel(logging.INFO)
ctx = mx.gpu(0)
msg=Queue.Queue(maxsize = 1000)

def group(data, num_r, num, kernel, stride, pad, layer):
if num_r > 0:
conv_r = mx.symbol.Convolution(data=data, num_filter=num_r, kernel=(1,1), name=('conv%s_r' % layer))
slice_r = mx.symbol.SliceChannel(data=conv_r, num_outputs=2, name=('slice%s_r' % layer))
mfm_r = mx.symbol.maximum(slice_r[0], slice_r[1])
conv = mx.symbol.Convolution(data=mfm_r, kernel=kernel, stride=stride, pad=pad, num_filter=num, name=('conv%s' % layer))
else:
conv = mx.symbol.Convolution(data=data, kernel=kernel, stride=stride, pad=pad, num_filter=num, name=('conv%s' % layer))
slice = mx.symbol.SliceChannel(data=conv, num_outputs=2, name=('slice%s' % layer))
mfm = mx.symbol.maximum(slice[0], slice[1])
pool = mx.symbol.Pooling(data=mfm, pool_type="max", kernel=(2, 2), stride=(2,2), name=('pool%s' % layer))
return pool

def lightened_cnn_a_feature():
data = mx.symbol.Variable(name="data")
pool1 = group(data, 0, 96, (9,9), (1,1), (0,0), str(1))
pool2 = group(pool1, 0, 192, (5,5), (1,1), (0,0), str(2))
pool3 = group(pool2, 0, 256, (5,5), (1,1), (0,0), str(3))
pool4 = group(pool3, 0, 384, (4,4), (1,1), (0,0), str(4))
flatten = mx.symbol.Flatten(data=pool4)
fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=512, name="fc1")
slice_fc1 = mx.symbol.SliceChannel(data=fc1, num_outputs=2, name="slice_fc1")
mfm_fc1 = mx.symbol.maximum(slice_fc1[0], slice_fc1[1])
drop1 = mx.symbol.Dropout(data=mfm_fc1, p=0.7, name="drop1")
return drop1

def lightened_cnn_a(num_classes=10575):
drop1 = lightened_cnn_a_feature()
fc2 = mx.symbol.FullyConnected(data=drop1, num_hidden=num_classes, name="fc2")
softmax = mx.symbol.SoftmaxOutput(data=fc2, name='softmax')
return softmax

def lightened_cnn_b_feature():
data = mx.symbol.Variable(name="data")
pool1 = group(data, 0, 96, (5,5), (1,1), (2,2), str(1))
pool2 = group(pool1, 96, 192, (3,3), (1,1), (1,1), str(2))
pool3 = group(pool2, 192, 384, (3,3), (1,1), (1,1), str(3))
pool4 = group(pool3, 384, 256, (3,3), (1,1), (1,1), str(4))
pool5 = group(pool4, 256, 256, (3,3), (1,1), (1,1), str(5))
flatten = mx.symbol.Flatten(data=pool5)
fc1 = mx.symbol.FullyConnected(data=flatten, num_hidden=512, name="fc1")
slice_fc1 = mx.symbol.SliceChannel(data=fc1, num_outputs=2, name="slice_fc1")
mfm_fc1 = mx.symbol.maximum(slice_fc1[0], slice_fc1[1])
drop1 = mx.symbol.Dropout(data=mfm_fc1, p=0.7, name="drop1")
return drop1

def lightened_cnn_b(num_classes=10575):
drop1 = lightened_cnn_b_feature()
fc2 = mx.symbol.FullyConnected(data=drop1, num_hidden=num_classes, name="fc2")
softmax = mx.symbol.SoftmaxOutput(data=fc2, name='softmax')
return softmax

class MySpeedometer(object):

def __init__(self, batch_size, frequent=50):
    self.batch_size = batch_size
    self.frequent = frequent
    self.init = False
    self.tic = 0
    self.last_count = 0
    self.data=[]
    self.epoch=0
            
def __call__(self, param):
    """Callback to Show speed."""
    count = param.nbatch
    if self.last_count > count:
        self.init = False
    self.last_count = count

    if self.init:
        if count % self.frequent == 0:
            speed = self.frequent * self.batch_size / (time.time() - self.tic)
            if param.eval_metric is not None:
                name_value = param.eval_metric.get_name_value()
                param.eval_metric.reset()
                for name, value in name_value:
                    #print name,"+",value
                    if name.find("accuracy")>=0:
                        if self.epoch!=param.epoch:
                            #msg.put(value)
                            self.epoch=param.epoch                        
                                                    
                    logging.info('Epoch[%d] Batch [%d]\tSpeed: %.2f samples/sec\tTrain-%s=%f',
                                 param.epoch, count, speed, name, value)
            else:
                logging.info("Iter[%d] Batch [%d]\tSpeed: %.2f samples/sec",
                             param.epoch, count, speed)
            self.tic = time.time()
    else:
        self.init = True
        self.tic = time.time()

def main():
# lightened_cnn = lightened_cnn_a()
lightened_cnn = lightened_cnn_b()
devs = mx.cpu() if args.gpus is None else [mx.gpu(int(i)) for i in args.gpus.split(',')]
epoch_size = args.num_examples / args.batch_size
checkpoint = mx.callback.do_checkpoint(args.model_save_prefix)
kv = mx.kvstore.create(args.kv_store)
arg_params = None
aux_params = None
if args.retrain:
_, arg_params, aux_params = mx.model.load_checkpoint(args.model_load_prefix, args.model_load_epoch)

train = mx.io.ImageRecordIter(
    path_imgrec = args.data_dir + "casia_train.rec",
    data_shape  = (1, 128, 128),
    scale       = 1./255,
    batch_size  = args.batch_size,
    rand_crop   = True,
    rand_mirror = True,
    num_parts   = kv.num_workers,
    part_index  = kv.rank)
if not args.retrain:
    val = mx.io.ImageRecordIter(
        path_imgrec = args.data_dir + "casia_val.rec",
        batch_size  = args.batch_size,
        data_shape  = (1, 128, 128),
        scale       = 1./255,
        rand_crop   = True,
        rand_mirror = False,
        num_parts   = kv.num_workers,
        part_index  = kv.rank)
else:
    val = None
model = mx.model.FeedForward(
    ctx                = devs,
    symbol             = lightened_cnn,
    arg_params         = arg_params,
    aux_params         = aux_params,
    num_epoch          = 200,
    learning_rate      = args.lr,
    momentum           = 0.9,
    wd                 = 0.0005,
    lr_scheduler       = mx.lr_scheduler.FactorScheduler(step=5*max(int(epoch_size * 1), 1), factor=0.8, stop_factor_lr=5e-5),
    initializer        = mx.init.Xavier(factor_type="in", magnitude=2.34))
model.fit(
    X                  = train,
    eval_data          = val,
    kvstore            = kv,
    batch_end_callback = MySpeedometer(args.batch_size, 100),
    epoch_end_callback = checkpoint)

if name == "main":
parser = argparse.ArgumentParser(description="command for training lightened-cnn")
parser.add_argument('--gpus', type=str, help='the gpus will be used, e.g "0,1,2,3"')
parser.add_argument('--data-dir', type=str, default='./', help='the input data directory')
parser.add_argument('--model-save-prefix', type=str, default='../model/lightened_cnn/lightened_cnn',
help='the prefix of the model to save')
parser.add_argument('--lr', type=float, default=0.05, help='initialization learning reate')
parser.add_argument('--batch-size', type=int, default=100, help='the batch size')
parser.add_argument('--num-examples', type=int, default=413914, help='the number of training examples')
parser.add_argument('--kv-store', type=str, default='local', help='the kvstore type')
parser.add_argument('--model-load-prefix', type=str, default='../model/lightened_cnn', help='the prefix of the model to load')
parser.add_argument('--model-load-epoch', type=int, default=1, help='load the model on an epoch using the model-load-prefix')
parser.add_argument('--retrain', action='store_true', default=False, help='true means continue training')
args = parser.parse_args()
logging.info(args)
t1 = threading.Thread(target=main)
t1.start()
t1.join()

plt.figure(figsize=(8, 6))    
data=[]    
while True:
    if msg.empty()==True:
        time.sleep(1)
    else:
        data.append(msg.get())
        #print data
        plt.xlabel("Epoch")
        plt.ylabel("Accuracy")
        plt.plot(range(len(data)), data, 'o', linestyle='-', color="r",label="Train accuracy")        
        plt.ylim([0,1])
        plt.draw()
        plt.waitforbuttonpress(timeout=0.001)

六、现象及问题:
1、我试过改lr 从0.008 - 0.1 如果是0.1 Train-Accuracy 会收敛在0.0001左右
如果在0.008-0.05 之间，会收敛在0.75左右
2、我的问题是，这个数据集用 lightened cnn 能否收敛到90%以上呢？