mariolew / deep-alignment-network-tensorflow Goto Github PK

Hi
How is the performance of mobileNet on iphone 7?
Thanks

stupid misstake.0.0

想请问一下在300-w数据集上，stage1和stage2差不多要训练几个epoch呢。
另外训练好之后的误差在多少说明模型效果比较好呢？我目前只训练了stage1，batch_error在0.8左右一直下不去，在300W-test上做测试的时候mean_error在0.07。
最好还想请问一下原代码中的validationSet是不是有问题，和trainSet的数据重复了？
非常希望得到您的解答，谢谢。

1、has this code reached the same accuracy as the theano ?
2、 i saw in theano document , bn(conv(relu)) means conv + bn +relu
like this: >>> from lasagne.layers import InputLayer, DenseLayer, batch_norm
>>> from lasagne.nonlinearities import tanh
>>> l1 = InputLayer((64, 768))
>>> l2 = batch_norm(DenseLayer(l1, num_units=500, nonlinearity=tanh))
This introduces batch normalization right before its nonlinearity:
>>> from lasagne.layers import get_all_layers
>>> [l.class.name for l in get_all_layers(l2)]
['InputLayer', 'DenseLayer', 'BatchNormLayer', 'NonlinearityLayer']

 so , which is right about bn. 

你好，
我用你的mobilenet版本来从头训练DAN，使用工程提供的脚本生成的数据集，只训练stage1，训了两天，loss已经不在下降，但是测试效果很差。对于mobilenet的训练，您有没有什么训练经验可以分享一下吗？之前用mobilenet从头训练其他的任务，效果也不好。
谢谢！

Hi,
When I run trainingsetpreparation .py,
Creating perturbations of 0 shapes
E:\anaconda\lib\site-packages\numpy\core\fromnumeric.py:2957: RuntimeWarning: Mean of empty slice.
out=out, **kwargs)
E:\anaconda\lib\site-packages\numpy\core_methods.py:80: RuntimeWarning: invalid value encountered in true_divide
ret = ret.dtype.type(ret / rcount)
E:\anaconda\lib\site-packages\numpy\core_methods.py:135: RuntimeWarning: Degrees of freedom <= 0 for slice
keepdims=keepdims)
E:\anaconda\lib\site-packages\numpy\core_methods.py:105: RuntimeWarning: invalid value encountered in true_divide
arrmean, rcount, out=arrmean, casting='unsafe', subok=False)
E:\anaconda\lib\site-packages\numpy\core_methods.py:127: RuntimeWarning: invalid value encountered in true_divide
ret = ret.dtype.type(ret / rcount)
Traceback (most recent call last):
File "TrainingSetPreparation.py", line 17, in
trainSet.NormalizeImages()#去均值，除以标准差
File "F:\code\Deep-Alignment-Network-tensorflow-master\DAN-TF\ImageServer.py", line 219, in NormalizeImages
plt.imshow(meanImg[:,:,0], cmap=plt.cm.gray)
IndexError: invalid index to scalar variable.
Can you help me ?
Thank you !

不知道可以不可以用resnet换掉第一层的vgg16

Do you know how to save this model to pb? the output tensor name is needed ,but i don't know the name ?

您好，请问训练的batch_size怎么设置，还有当我运行trainDAN.py时候，提示 FailedPreconditionError(see above for traceback):Attemping to use uninitialized value Stage2/beta1_power[[{{node Stage2/beta1_power/read}}=Identity[T=DT_FLoat,_CLASS={'"LOC:@STAGE2/adam/assign_1"},-device="job:localhost/replace:0/task:0/device:GPU:0"(stage2/beta1_power)]]]，请问这个错误怎么避免

没法训练，一直在爆内存。

Hi,
What size of model when you train Deep alignment network using mobile net? And if I have GPU so I can run your code on GPU?

您好，我在训练模型的时候发现loss下降的速度很慢，想修改模型训练的学习率，但是整个代码都找了，没发现学习率在哪设置的，您这边能告诉学习率的参数到底是哪一个么?

您好，我在做实验的时候发现一个问题，我现在只训练了第一级网络，第一级网络其实能预测关键点了，在第一级网络中加了meanshape的效果比没有加meanshape的效果要差一下，您有遇到这个问题吗？

RT
i do not found weight decay in all three implementation. Did i miss something?

您好：
我想实际的看下视频流的Landmark效果，源码只有测试集的Error评估，能指点一下，怎么实现测试视频流吗？万分感谢~

你好，我在运行python trainDAN.py，会出现以下错误：
Traceback (most recent call last): │
File "trainDAN.py", line 12, in │
trainSet = ImageServer.Load(datasetDir + "dataset_nimgs=40_perturbations=[0.2, 0.2, 20,│
0.25]_size=[112, 112].npz") │
File "/home/zhanggl/dantf/DAN-TF/ImageServer.py", line 31, in Load │
arrays = np.load(filename) │
File "/home/zhanggl/anaconda3/envs/dantf/lib/python3.6/site-packages/numpy/lib/npyio.py",│
line 415, in load │
fid = open(os_fspath(file), "rb") │
FileNotFoundError: [Errno 2] No such file or directory: '../data/dataset_nimgs=40_perturbat│
ions=[0.2, 0.2, 20, 0.25]_size=[112, 112].npz'

请问这是我哪里出来问题？（我已经过运行TestSetPreparation.py 和 TrainingSetPreparation.py）
非常感谢。

Could you please put the pre-trained model to this project ? or do you know how to use the pre-trained model in the origin theano project

def LandmarkImageLayer(Landmarks):

def draw_landmarks(L):
    def draw_landmarks_helper(Point):
        intLandmark = tf.to_int32(Point)
        locations = Offsets + intLandmark
        dxdy = Point - tf.to_float(intLandmark)
        offsetsSubPix = tf.to_float(Offsets) - dxdy
        vals = 1 / (1 + tf.norm(offsetsSubPix, axis=2))
        img = tf.scatter_nd(locations, vals, shape=(IMGSIZE, IMGSIZE))
        return img

这个函数中，Offsets和intLandmark维度不匹配呀？
intLandmark的维度是（图片的数量，136,2）儿Offsets的维度是（16,16,2）

训练集：300w indoor
stage0和stage1的时候表现良好
stage2的时候testerror爆发式增长，有人遇到类似情况吗？

There are two lines in testDan which uses format like:
.format{np.mean(errs)})
that should of course be
.format(np.mean(errs)))

如题，非人脸数据，想进行关键点检测的训练，该如何进行呢

I can not see any explanation about Face Normalization. It this project can normalize the faces? Could you provide some details please?

时隔快一年了,还是没弄懂为什么在替换了MobilentV2后,没法训练STAGE2.是因为mobilenet本身的特性么?

有人知道stage1 和 stage2 的训练周期多少比较合适吗？感谢感谢

小姐姐，你好^^.
我在看代码的实现细节，但有些涉及到矩阵运算的代码，我不太理解为什么要那样做。
比如TransformParamsLayer和AffineTransformLayer中的矩阵运算，
你能提供一些资料吗，多谢。^^

我训练完 s1 50次 s2 50次之后发现效果 和theano版还有一定差距，又训练了40次s2,结果没什么改变，我想再继续训练一下s1 看看有没有提升，可以直接训练吗？ 还是必须要从头开始训练s1

challengingSet model 140 0.08934
commonSet.npz model 140 0.04839

theano:
commonSet 0.04287
challengeSet 0.07040

评价指标是 normalization = 'centers'

In TrainingSetPreparation.py,

I wonder whats the format of LFPW or HELEN and how to ensure that images corresponding to data?

my stage 1 training failed with this message:
ptxas fatal : Memory allocation failure

my gpu memory is 12GB, memory is 64GB, python3.6, CUDA9.2, python3.6, tensorflow-gpu 1.10.0
running
trainDAN.py with stage=1 and using npz file generated

您好，我在进行trainDAN训练时，STAGE=2，执行sess.sun()；获取数据训练时，运行到models.py中的S2_InputImage = AffineTransformLayer(InputImage, S2_AffineParam)会报错；

错误如下：

Caused by op 'Stage2/MatrixInverse', defined at:
File "D:/softmares/Pycharm_workplace/Deep-Alignment-Network-tensorflow-master/DAN-TF/python_test1.py", line 73, in
dan = DAN(initLandmarks)
File "D:\softmares\Pycharm_workplace\Deep-Alignment-Network-tensorflow-master\DAN-TF\python_test2.py", line 102, in DAN
S2_InputImage = AffineTransformLayer(InputImage, S2_AffineParam) ## 通过变换矩阵对原图进行矫正，得到新的图片
File "D:\softmares\Pycharm_workplace\Deep-Alignment-Network-tensorflow-master\DAN-TF\layers.py", line 60, in AffineTransformLayer
A = tf.matrix_inverse(A)
File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\ops\gen_linalg_ops.py", line 330, in matrix_inverse
name=name)
File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\op_def_library.py", line 768, in apply_op
op_def=op_def)
File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 2336, in create_op
original_op=self._default_original_op, op_def=op_def)
File "C:\Program Files\Anaconda3\lib\site-packages\tensorflow\python\framework\ops.py", line 1228, in init
self._traceback = _extract_stack()

InvalidArgumentError (see above for traceback): Input is not invertible.
[[Node: Stage2/MatrixInverse = MatrixInverseT=DT_FLOAT, adjoint=false, _device="/job:localhost/replica:0/task:0/cpu:0"]]

请问您遇到过吗？是我开始数据变换出了问题吗？能请教您下吗？

I was wondering what npz-files do I have to load in line 15 and 16 in order to train the model? At the moment those files aren't in the directory "data".

(14)>>>datasetDir = "../data/"
(15)>>>trainSet = ImageServer.Load(datasetDir + "dataset_nimgs=40_perturbations=[0.2, 0.2, 20, 0.25]_size=[112, 112].npz")
(16)>>>validationSet = ImageServer.Load(datasetDir + "dataset_nimgs=9_perturbations=[]_size=[112, 112].npz")

when run the testDAN.py,there some error

nChannels = testSet.imgs.shape[1]

error：tuple index out of range

Hi sir, i appreciate your work, and I would like to know which values should be in TestErr and BatchErr on training phase of stage 1 and 2 to obtain a good model?. How many epochs you used?
Recently i started training phase stage 1. Best Regards...


Hi,
I have an error when run TestsetPreparation.py file:
Traceback (most recent call last):
File "TestSetPreparation.py", line 19, in
commonSet.PrepareData(commonSetImageDirs, commonSetBoundingBoxFiles, meanShape, 0, 1000, False)
File "/home/diffdeep/Documents/train_dan/Deep-Alignment-Network-tensorflow-master-2/DAN-TF/ImageServer.py", line 60, in PrepareData
boundingBoxDict = pickle.load(open(boundingBoxFiles[i], 'rb'))
UnicodeDecodeError: 'ascii' codec can't decode byte 0xd3 in position 0: ordinal not in range(128)
Please suggest me some advice. Thanks