machrisaa / tensorflow-vgg Goto Github PK

View Code? Open in Web Editor NEW

2.2K 2.2K 1.1K 57 KB

VGG19 and VGG16 on Tensorflow

Python 100.00%

tensorflow-vgg's People

Contributors

Stargazers

Watchers

Forkers

fooway iamaaditya dionysio boscong88 wkentaro marvinteichmann zbxzc35 alanguo001 dominicgwak apo-j butterflyaichinese aecker twitwi ruotianluo marc-moreaux suhangpro libphy kklemon linzhineng gongxijun wangqi1919 vonclites zerolocker alexzhou1995 nicebam yatuzhang freedom079215 skycache waleedgondal leiup garfield2005 patriciasr92 tae-hyun codeashu v1vekkumar samxiaozy jogo23 supermancc lszhuhaichao mingyue312 soledad89 caomw kamransial iflier zhangxinnan lyuha yenchi-hsu xhygh ozgurgundogan rioyokotalab huyng qinpersevere hhappy06 amylizzle scorpiodoctor chrisying nsokhand yunfanz insikk dkollias jamoque wujiahongpku malhagi liupeng89 saitamandd calleamaris jaypy-h wiibrew antlerros ryomaniihara mengfanyang irfanicmll frhrdr olegjakushkin schaber valeman flrngel a514514772 hujiajia0401 andersenhuang yrj90 ywpkwon priteshgudge kanul wlau0721 xiaoyangzz yheno jjpallisgaard andrewszwec dcarlyle bowen03 tfy1028 ichae engp0958 raghavendranpm davidgengenbach h312h liu-yangyang sebsacuna suprespark

tensorflow-vgg's Issues

AssertionError

When I try thetransf_learning.py,some wrong happens:
Traceback (most recent call last):
File "transfer_learning.py", line 19, in
vgg.build(input_)
File "F:\VGG16-1\tensorflow_vgg\vgg16.py", line 36, in build
assert red.get_shape().as_list()[1:] == [224, 224, 1]
AssertionError.
Could you give me some advice?

I wonder if my vgg npy file broken

The range of numbers in the output tensor generated by conv5_4 is very big, can up to 1e5, is this the correct output?

code for training the net

Great work. Do you have code that train the vgg net and save the model weights to the npy file ?

Thanks,

Which dataset is VGG_MEAN calculated from?

I want to use existing pre-trained weights to estimate top1/5 error of this model. Could you kindly tell me which dataset VGG_MEAN is calculated from?

Wondering about data input format

Hi! My group and I are trying to modify this model to accomplish a model proposed in the paper "Deep3D"
We are just wondering if using the ndarray input is required. Tensorflow has a decode_jpeg method. Would that work with these pre trained weights? we are training on around 1 million 3D movie frames and it would save a lot of memory scaling problems if jpegs are used instead.

Please let us know!

Thanks

Training does not work

I use test_vgg19_trainable to train the vgg19 net.
I feed a photo of a face (not tiger), telling the net it is a tiger (#292=1).
But in every run the system says it is a "cup" (#968) with p=1.

Wget links for VGG*.npy.

It's a pain to create an account etc to just download that file. Also a wget link will be better since I am guessing folks are downloading this to a server instead of to local machine.

Odd classification results

Hi,

i just cloned your rep and downloaded the *.npy files provided. I didn't change any file and run the test_vgg16.py example, which works but leads to odd classification results:

[...]
npy file loaded
[[  3.54046060e-05   6.26692083e-03   2.32575258e-04 ...,   1.31252409e-05
    4.17124102e-04   5.32930251e-03]
 [  2.74504941e-07   1.49452433e-04   5.41723421e-06 ...,   1.50292351e-06
    1.41892961e-04   7.05321319e-03]
 [  2.73421301e-05   2.04334548e-03   5.73722122e-04 ...,   1.37338357e-05
    5.19106514e-04   4.67342697e-03]]
('prob shape', (1000,))
('Top1: ', 'n04589890 window screen', 0.17731868)
('Top5: ', [('n04589890 window screen', 0.17731868), ('n03788365 mosquito net', 0.050591175), ('n03000134 chainlink fence', 0.038932227), ('n04476259 tray', 0.036342748), ('n03485794 handkerchief, hankie, hanky, hankey', 0.033405658)])
('prob shape', (1000,))
('Top1: ', 'n04209239 shower curtain', 0.3111819)
('Top5: ', [('n04209239 shower curtain', 0.3111819), ('n02971356 carton', 0.061327737), ('n04548280 wall clock', 0.036779501), ('n03131574 crib, cot', 0.032009162), ('n03291819 envelope', 0.031746123)])

Any thoughts on what might be the problem? I'm using tensorflow 0.8.0 and tensorflow-vgg16 actually works fine.

Thanks!

Not able to download the pretrained models VGG16 and VGG19 from the links given

The browser is not able to connect as the link is insecure. Is there a mirror or any other link that I can use to download?

Training from scratch

Hello, i am tf beginner. When i look at your code, i didnt see any instruction of how to training vgg from scratch or fine tuning from specified weights. Would you mind showing me how to do with your code ? Thanks a lot

KeyError Message after loading the npy file

Hi, Thank you for sharing the code.

I have been trying to running the test_vgg16.py, and it gives the error message as follows. Would you like to let me know the possible reason for this?

npy file loaded build model started Traceback (most recent call last): File "test_vgg16.py", line 23, in <module> vgg.build(images) File "/home/DL-Phase3/tensorflow-vgg-master/vgg16.py", line 47, in build self.conv1_1 = self.conv_layer(bgr, "conv1_1") File "/home/DL-Phase3/tensorflow-vgg-master/vgg16.py", line 92, in conv_layer filt = self.get_conv_filter(name) File "/home/DL-Phase3/tensorflow-vgg-master/vgg16.py", line 120, in get_conv_filter return tf.constant(self.data_dict[name][0], name="filter") KeyError: 0

preprocessing image for VGG16, BGR or RGB ?

VGG16 were trained using Caffe, and Caffe uses OpenCV to load images which uses BGR by default, so VGG models are expecting BGR images. In the example, RGB images are used, but I'm not sure whether is it right because I couldn't find the op of swapping the first-layer filters of the CNN when convert caffe model to tensorflow model.

what is VGG_MEAN ?

when convert rgb to bgr ,why need to minus VGG_MEAN?

What database did this vgg implementation trained on?

Hi,

I'm using this library to do some research, and it will be a great help if I can know which database was the model trained on. If it is Imagenet, what's the version of the data from Imagenet, is it ILSVRC2012 or ILSVRC2014?
Thanks.

Wanquan,

Order of image channels?

Hi, I wonder which order is used in your implementation of VGG19? As far as I know, BGR is used for Caffe, and since the trained model you provide here is adapted from Caffe's model, so by default BGR is also used here?

please add a vgg-16 trainable

I would like to ask for extra feature of trainable vgg-16. thx

Mismatch in the number of units of fc6

ImportError: undefined symbol: PyClass_Type

when I try to load module vgg19.py,it comes to this error:
import tensorflowvgg.vgg19

Traceback (most recent call last):

File "", line 1, in
import tensorflowvgg.vgg19

File "/home/yan/下载/tensorflowvgg/vgg19.py", line 2, in
import tensorflow as tf

File "/usr/local/lib/python2.7/dist-packages/tensorflow/init.py", line 23, in
from tensorflow.python import *

File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/init.py", line 45, in
from tensorflow.python import pywrap_tensorflow

File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/pywrap_tensorflow.py", line 28, in
_pywrap_tensorflow = swig_import_helper()

File "/usr/local/lib/python2.7/dist-packages/tensorflow/python/pywrap_tensorflow.py", line 24, in swig_import_helper
_mod = imp.load_module('_pywrap_tensorflow', fp, pathname, description)

File "/home/yan/anaconda3/lib/python3.6/imp.py", line 242, in load_module
return load_dynamic(name, filename, file)

File "/home/yan/anaconda3/lib/python3.6/imp.py", line 342, in load_dynamic
return _load(spec)

ImportError: /usr/local/lib/python2.7/dist-packages/tensorflow/python/_pywrap_tensorflow.so: undefined symbol: PyClass_Type

my python version is: Python 3.6.1 |Anaconda custom (64-bit)| (default, May 11 2017, 13:09:58)

Memory Usage

I am looking to fine tune the VGG19 model for an image classification problem, and am running into memory issues on my GPU.

I was wondering how much memory is expected to be used when training this network? When running it on a K80 with a 11 gb available, it was using 6.4 gb while training on just the two demo images.

I have heard of people running vgg19 on graphics cards with less than 4 gb of memory, so I was wondering if there room for improvement reducing the memory usage of this implementation?

python3 compability: load() got an unexpected keyword argument 'encoding'

I got the following error when I run it on python3:

  File "/home/kang/Documents/work_code_PC1/tensorflow-vgg/vgg16.py", line 20, in __init__
    self.data_dict = np.load(vgg16_npy_path, encoding='latin1').item()

TypeError: load() got an unexpected keyword argument 'encoding'

How to fix it?
Thank you very much.

The vgg16.npy can't be downed.

VGG Numpy files distribution.

Hi Chris,
I've used your TF implementation of VGGnet in my project and linked your dropbox storage to download the numpy file. However, it looks like the link was disabled due to too much traffic. Would it be alright if I reuploaded it somewhere and linked there?

Thanks,
Jan

why we devide image by 255.0?

How was the npy files created?

Accuracy of the trained model file.

Is the model file VGG19.NPY trained on imagenet? How about the accuracy?

I am sorry to open this issue in a hurry. I just found that I should convert image to BGR.

Cannot download the pretrained vgg19.npz

Hi, the address of VGG19 NPY is unreachable, can you check the address' correctness, or would you please send it me to my email address, [email protected] is in use, thanks a lot. Looking forward to your early reply.

Organize the project as Python package

It would be really nice if the repo/project would be organized as an ordinary Python package so it could be installed and used globally more easily. Currently this is a bit difficult, especially when you work in environments when it's not so easy to setup custom modules likes it's often applies in ML.

Update: check my fork which I have organized as package. If you satisfied with that, I could open a pull request.

download vgg16.npy

你好，我发现vgg16.npy的下载链接打不开，是链接过期了吗？

How to use it to train my own data

I am having a program to classify pollen in 2 types: sugi and hinoki. The teaching data are 100x100 images of only 1 pollen, and the real data are bigger images with lots of pollen inside (which I used opencv to mark circle in each pollen, with suitable color according to its type).

Teacher asked me to change the program into a new version using vgg (16 or 19), but I really dont know how. I have looked up in the internet, and most of versions use image_net weight, not trainable one. There is only this version seem to be useful but I dont know how to change it to fit my program . I am really new to neuron network and tensorflow, so please help me.

Here is my program.

``# coding: UTF-8
#opencv bgr
#rgb
import numpy as np
import tensorflow as tf
import os
import sys
import csv
import random
from datetime import datetime
from PIL import Image
from matplotlib import pylab as plt
from time import sleep
import cv2
#import cv2.cv as cv

DataShape = (100,100,3)
Ratio = 1#2000.0
TestNum = 100

def show_img(dataDir):

    data = np.load(dataDir)#.astype(float)
    #label = int(row[LabelIdx])
    #data += np.random.normal(0,0.1*np.max(data),(512,512))
    #data -= np.min(data)
    #data = data*255/np.max(data)
    #print name
    #print data.shape
    #if name=="os10-020_cancer_full.npy":
    for i in range(data.shape[0]):
        print (i)
        #lt.imshow(data[i])
        #plt.show()
    #plt.imshow(data[data.shape[0]-ZIdx])
    #plt.show()
    #pilImg = Image.fromarray(np.uint8(data[15]))
    #pilImg.show()
    #sleep(1)
    
def show_npy(dataList):

    #data = np.load(dataDir)#.astype(float)
    #label = int(row[LabelIdx])
    #data += np.random.normal(0,0.1*np.max(data),(512,512))
    #data -= np.min(data)
    #data = data*255/np.max(data)
    #print name
    #print data.shape
    #if name=="os10-020_cancer_full.npy":
    for i in range(len(dataList[0])):
        #print i
        #plt.imshow(dataList[2][i])
        #plt.show()
        pilImg = Image.fromarray(np.uint8(dataList[2][i]*255.0))
        pilImg.show()
        sleep(1)
    #plt.imshow(data[data.shape[0]-ZIdx])
    #plt.show()
    #pilImg = Image.fromarray(np.uint8(data[15]))
    #pilImg.show()
    #sleep(1)
 
# nameList  : list of each data name
# labelList : list of teaching data (written by 0 or 1 and 3 calams)
# dataList  : 4 dimensionary data
def GetData():
    sugi_data = np.load('sugi.npy')/255.0
    hinoki_data = np.load('hinoki.npy')/255.0
    nameList = []
    labelList = []
    dataList = []
    print (sugi_data.shape)
    print (hinoki_data.shape)
    #sys.exit()
    
    repeat = np.min((len(sugi_data),len(hinoki_data)))%100
    for i_repeat in range(repeat):
        for i,data in enumerate(sugi_data[i_repeat*100:(i_repeat+1)*100]):
            name = 'sugi_'+str(i)
            data = np.reshape(data,[DataShape[0],DataShape[1],DataShape[2]])
            label = 0
            l = np.zeros(2)
            l[label] += 1

            labelList.append(l)
            nameList.append([name])
            dataList.append(data)

        print (len(sugi_data))

        for i,data in enumerate(hinoki_data[i_repeat*100:(i_repeat+1)*100]):
            name = 'hinoki_'+str(i)
            data = np.reshape(data,[DataShape[0],DataShape[1],DataShape[2]])
            label = 1
            l = np.zeros(2)
            l[label] += 1

            labelList.append(l)
            nameList.append([name])
            dataList.append(data)
        #print len(hinoki_data)
        #print np.array(dataList).shape
        #print np.max(dataList)
    
    #data_shuffle = zip(nameList,labelList,dataList)
    #random.shuffle(data_shuffle)
    #return zip(*data_shuffle)

    return nameList,labelList,dataList

# dataList : 4 dimensionary data
# testIdxList : 1=test data , 0 = training data
def GetDataset(dataList,testIdxList):
    teDataList = []
    trDataList = []
    for i in range(len(dataList)):
        if i in testIdxList:
            teDataList.append(dataList[i])
        else:
            trDataList.append(dataList[i])
    return np.vstack([teDataList]),np.vstack([trDataList])


def Train(alldata,testIdxList,resultName):
    print ("start train initialize")
    os.mkdir("output/" + resultName)
    print ("get dataset")
    ten,trn = GetDataset(alldata[0], testIdxList)
    tel,trl = GetDataset(alldata[1], testIdxList)
    ted,trd = GetDataset(alldata[2], testIdxList)
    #tef,trf = GetDataset(alldata[3], testIdxList)

    #print trd.shape
    #sys.exit()


    #mizumasi 
    #data_num = len(trd)
    #nlist = []
    #llist = []
    #dlist = []
    #flist = []
    #nlist.extend(trn)
    #llist.extend(trl)
    #dlist.extend(trd)
    #flist.extend(trf)
    
    #print data_num
    #label_num = np.zeros(3).astype(np.int32)
    #for i in range(data_num):
    #    for n in range(3):
    #        if trl[i][n]==1 :
    #            label_num[n]+=1
    #print label_num

    #for n in range(3):
    #    for j in range(100-label_num[n]):
    #        flag=1
    #        while(flag==1):
    #          i = random.randint(0,data_num-1)
    #          if trl[i][n]==1:
    #            flag=0  #
    #        nlist.append(trn[i])
    #        llist.append(trl[i])
    #        data = trd[i] + np.random.normal(0,0.1,(DataShape[0],DataShape[1],DataShape[2],1))
    #       dlist.append(data)
            #flist.append(trf[i])

    #for i in range(data_num):
    #    for j in range(10): 
    #        nlist.append(trn[i])
    #        llist.append(trl[i])
    #        data = trd[i] + np.random.normal(0,0.1,(DataShape[0],DataShape[1],DataShape[2],1))
    #        dlist.append(data)
            #flist.append(trf[i])

    #trn = np.array(nlist)
    #trl = np.array(llist)
    #trd = np.array(dlist)
    print ("generate cnn model")
    x = tf.placeholder(tf.float32, [None,DataShape[0],DataShape[1],DataShape[2]])
    y = tf.placeholder(tf.float32, [None,2])

    conv_W0 = tf.Variable(tf.truncated_normal([5,5,3,32],stddev=0.01), name="conv_W0")
    conv_b0 = tf.Variable(tf.zeros([32]), name="conv_b0")
    conv_z0 = tf.nn.conv2d(x, conv_W0, [1,1,1,1], "SAME") + conv_b0
    conv_u0 = tf.nn.relu(conv_z0)
    pool0 = tf.nn.max_pool(conv_u0, [1,2,2,1], [1,2,2,1], "SAME")
    norm0 = pool0#tf.nn.lrn(pool0)   

    conv_W1 = tf.Variable(tf.truncated_normal([5,5,32,64],stddev=0.01), name="conv_W1")
    conv_b1 = tf.Variable(tf.zeros([64]), name="conv_b1")
    conv_z1 = tf.nn.conv2d(norm0, conv_W1, [1,1,1,1], "SAME") + conv_b1
    conv_u1 = tf.nn.relu(conv_z1)
    pool1 = tf.nn.max_pool(conv_u1, [1,2,2,1], [1,2,2,1], "SAME")
    norm1 = pool1#tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta = 0.75, name='norm1')

    reshapeVal = 25*25*64
    flat  = tf.nn.dropout(tf.reshape(norm1,[-1,reshapeVal]),0.5)

    full_W0 = tf.Variable(tf.truncated_normal([reshapeVal,200],stddev=0.01), name="full_W0")
    full_b0 = tf.Variable(tf.zeros([200]), name="full_b0")
    full_z0 = tf.matmul(flat, full_W0)  + full_b0
    full_u0 = tf.nn.relu(full_z0)

    full_W = tf.Variable(tf.truncated_normal([200,2],stddev=0.01), name="full_W")
    full_b = tf.Variable(tf.zeros([2]), name="full_b")
    full_z = tf.matmul(full_u0, full_W)  + full_b
    full_u = tf.nn.softmax(full_z)

    #conv_L2 = tf.nn.l2_loss(conv_W0) + tf.nn.l2_loss(full_W)

    loss = -tf.reduce_mean(y*tf.log(tf.clip_by_value(full_u,1e-10,1.0))) #+ 0.001*conv_L2
    trainstep = tf.train.AdamOptimizer(1e-5).minimize(loss)
    #trainstep = tf.train.GradientDescentOptimizer(0.0001).minimize(loss)
    #trainstep = tf.train.FtrlOptimizer(0.0001).minimize(loss)

    collect_prediction = tf.equal(tf.argmax(full_u, 1), tf.argmax(y, 1))
    accuracy = tf.reduce_mean(tf.cast(collect_prediction,tf.float32))
    
    print ("server define")
    sess = tf.Session()
    sess.run(tf.global_variables_initializer())
    saver = tf.train.Saver()
    #saver.restore(sess, "/home/miyake/Desktop/kahun_ohira/finalmodel.ckpt")

    f = open("output/" + resultName+"/LearningResult.csv","w")
    writer = csv.writer(f)
    writer.writerow(["loop","train loss","test accuracy"])
    print ("start learning")

    for i in range(1000001):
    #for i in range(101):
        idxList = []
        for j in range(10):
            idxList.append(random.randint(0,trd.shape[0]-1))

        if i % 10 == 0:
            l = sess.run(loss,feed_dict={x:trd[idxList], y:trl[idxList]})
            acc = sess.run(accuracy,feed_dict={x:ted,y:tel})
            writer.writerow([i,l,acc])
            print (str(i) + ": loss = " + str(l) + ", acc = " + str(acc))
            if l<0.0001 :
                saver.save(sess, datetime.now().strftime('%s')+"finalmodel.ckpt")
                break
        if i % 1000 == 0:
            saver.save(sess, "model.ckpt")

        sess.run(trainstep,feed_dict={x:trd[idxList], y:trl[idxList]})

    f.close()
    print ("output result")
    f = open("output/"+resultName+"/TestResult.csv","w")
    writer = csv.writer(f)
    writer.writerow(["Name","IsCollect"])
    isCollect,answer,output,output_data = sess.run((collect_prediction,tf.argmax(y, 1),tf.argmax(full_u, 1),full_u),feed_dict={x:ted,y:tel})
    for i in range(len(output)):
        writer.writerow([ten[i][0],int(isCollect[i]),answer[i],output[i],output_data[i][0],output_data[i][1]])
    tf.reset_default_graph()

def Test(image_path,label,filename,file_path):


    limit_data = []
    limit_label = []
    limit_circles = []
    img = cv2.imread(image_path,cv2.IMREAD_COLOR)
    img_array = np.array(img)
    cimg = cv2.cvtColor(img,cv2.COLOR_BGR2GRAY)
    cimg = cv2.medianBlur(cimg,5)
    #_,cimg = cv2.threshold(cimg,0,255,cv2.THRESH_BINARY| cv2.THRESH_OTSU)
    #cv2.imwrite(datetime.now().strftime('%s')+"binary.jpg",cimg)
    #sys.exit()

    #circles = cv2.HoughCircles(cimg,cv.CV_HOUGH_GRADIENT,1,10,param1=10,param2=18,minRadius=10,maxRadius=25)
    circles = cv2.HoughCircles(cimg,cv2.HOUGH_GRADIENT,1,20,param1=15,param2=20,minRadius=10,maxRadius=25)
    
    circles = np.uint16(np.around(circles))[0,:]
    print (len(circles))

    for i in circles:
        half = DataShape[0]//2
        zoom_data = img_array[i[1]-half:i[1]+half,i[0]-half:i[0]+half,:]/255.0
        if zoom_data.shape!=DataShape : continue
        limit_data.append(zoom_data)
        l = np.zeros(2)
        l[label] += 1
        limit_label.append(l)
        limit_circles.append(i)
        #cv2.circle(img,(i[0],i[1]),i[2],(0,255,0),2)
        #v2.circle(img,(i[0],i[1]),2,(0,0,255),3)
        #print img_array[i[0]-half:i[0]+half,i[1]-half:i[1]+half,:].shape
    limit_data = np.array(limit_data)
    limit_label = np.array(limit_label)
    limit_circles = np.array(limit_circles)
    label_num = limit_data.shape[0]
    #cv2.imwrite(datetime.now().strftime('%s')+"output.jpg",img)
    #sys.exit()

    x = tf.placeholder(tf.float32, [None,DataShape[0],DataShape[1],DataShape[2]])
    y = tf.placeholder(tf.float32, [None,2])

    conv_W0 = tf.Variable(tf.truncated_normal([5,5,3,32],stddev=0.01), name="conv_W0")
    conv_b0 = tf.Variable(tf.zeros([32]), name="conv_b0")
    conv_z0 = tf.nn.conv2d(x, conv_W0, [1,1,1,1], "SAME") + conv_b0
    conv_u0 = tf.nn.relu(conv_z0)
    pool0 = tf.nn.max_pool(conv_u0, [1,2,2,1], [1,2,2,1], "SAME")
    norm0 = pool0#tf.nn.lrn(pool0)   

    conv_W1 = tf.Variable(tf.truncated_normal([5,5,32,64],stddev=0.01), name="conv_W1")
    conv_b1 = tf.Variable(tf.zeros([64]), name="conv_b1")
    conv_z1 = tf.nn.conv2d(norm0, conv_W1, [1,1,1,1], "SAME") + conv_b1
    conv_u1 = tf.nn.relu(conv_z1)
    pool1 = tf.nn.max_pool(conv_u1, [1,2,2,1], [1,2,2,1], "SAME")
    norm1 = pool1#tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta = 0.75, name='norm1')

    reshapeVal = 25*25*64
    flat  = tf.nn.dropout(tf.reshape(norm1,[-1,reshapeVal]),0.5)

    full_W0 = tf.Variable(tf.truncated_normal([reshapeVal,200],stddev=0.01), name="full_W0")
    full_b0 = tf.Variable(tf.zeros([200]), name="full_b0")
    full_z0 = tf.matmul(flat, full_W0)  + full_b0
    full_u0 = tf.nn.relu(full_z0)

    full_W = tf.Variable(tf.truncated_normal([200,2],stddev=0.01), name="full_W")
    full_b = tf.Variable(tf.zeros([2]), name="full_b")
    full_z = tf.matmul(full_u0, full_W)  + full_b
    full_u = tf.nn.softmax(full_z)

    #conv_L2 = tf.nn.l2_loss(conv_W0) + tf.nn.l2_loss(full_W)

    loss = -tf.reduce_mean(y*tf.log(tf.clip_by_value(full_u,1e-10,1.0))) #+ 0.001*conv_L2
    trainstep = tf.train.AdamOptimizer(1e-5).minimize(loss)
    #trainstep = tf.train.GradientDescentOptimizer(0.0001).minimize(loss)
    #trainstep = tf.train.FtrlOptimizer(0.0001).minimize(loss)

    collect_prediction = tf.equal(tf.argmax(full_u, 1), tf.argmax(y, 1))
    accuracy = tf.reduce_mean(tf.cast(collect_prediction,tf.float32))
    
    sess = tf.Session()
    #sess.run(tf.global_variables_initializer())
    saver = tf.train.Saver()
    saver.restore(sess, "model.ckpt")

    #for i in range(label_num):
    output = sess.run(tf.argmax(full_u, 1),feed_dict={x:limit_data,y:limit_label})
    for i in range(label_num):
        if output[i]==label :
            cv2.circle(img,(circles[i][0],circles[i][1]),circles[i][2],(0,255,0),2)
        else:
            cv2.circle(img,(circles[i][0],circles[i][1]),circles[i][2],(0,0,255),2)
    if not os.path.exists("output/"+ file_path):
        os.mkdir("output/"+ file_path)

    
    cv2.imwrite("output/"+ file_path + "/" + datetime.now().strftime('%s') + filename,img)
    sleep(0.1)
    #cv2.waitKey(0)
    #cv2.destroyAllWindows()

    #f = open(resultName+"/TestResult.csv","w")
    #writer = csv.writer(f)
    #writer.writerow(["Name","IsCollect"])
    #isCollect,answer,output,output_data = sess.run((collect_prediction,tf.argmax(y, 1),tf.argmax(full_u, 1),full_u),feed_dict={x:data,y:label})
    #for i in range(TestNum):
    #    writer.writerow([ten[i][0],int(isCollect[i]),answer[i],output[i],output_data[i][0],output_data[i][1]])
    tf.reset_default_graph()

def learning_data_print(dataList):
    data_num = len(dataList[0])
    for i in range(0,data_num,200):
        for j in range(200):
            path = "image/"+str(i/200)+"/"+ dataList[0][i+j][0] +".png"
            if not os.path.exists(path) :
                plt.imshow(dataList[2][i+j])
                plt.savefig(path)
                plt.clf()
            #pilImg = Image.fromarray(np.uint8(dataList[2][i+j]))
            #pilImg.save("image/"+str(i/200)+"/"+ dataList[0][j][0] +".png")
            #print dataList[2][i]
            #plt.show()

if __name__ == '__main__':
    
    #show_img("/home/miyake/Desktop/kahun_ohira/sugi.npy")
    #Test('/home/miyake/Desktop/kahun_ohira/original_data/test/hinoki/IMG_3782.JPG',1)
    #Test('/home/miyake/Desktop/kahun_ohira/original_data/test/hinoki/IMG_3785.JPG',1)
    #Test('/home/miyake/Desktop/kahun_ohira/original_data/test/hinoki/IMG_3792.JPG',1)
    #Test('/home/miyake/Desktop/kahun_ohira/original_data/test/sugi/IMG_3806.JPG',0)
    #Test('/home/miyake/Desktop/kahun_ohira/original_data/test/sugi/IMG_3808.JPG',0)
    #Test('/home/miyake/Desktop/kahun_ohira/original_data/test/sugi/IMG_3806.JPG',0)

    #Test('/home/miyake/Desktop/kahun_ohira/original_data/mix/IMG_3828.JPG',0)
    #sys.exit()

    alldata = GetData()
    #learning_data_print(alldata)
    #show_npy(alldata)
   
    
    data_num = len(alldata[0])
    now = datetime.now()
    for i in range(0,data_num,200):
        testidxList = []
        for j in range(200):
            testidxList.append(i+j)
 
        resultName = now.strftime('%s')+"_Result_"+str(i)+"_"+str(i+200)
        Train(alldata,testidxList,resultName)
        for path, subdirs, files in os.walk("check/"):
            for filename in files:
                f = os.path.join(path, filename)   
                Test(str(f),0,filename,path)
            
        
        #Test('/home/miyake/Desktop/kahun_ohira/original_data/test/sugi/IMG_3802.JPG',0)
        #Test('/home/miyake/Desktop/kahun_ohira/original_data/mix/IMG_3814.JPG',0)

Training from scratch

Is it possible to release the trainer for either VGG16 or VGG 19 from scratch? Also what is a good method for initializing the weights for training from scratch?

Input 'split_dim' of 'Split' Op has type float32 that does not match expected type of int32.

I'm getting an error from this line of code :
red, green, blue = tf.split(3, 3, rgb_scaled)

I saw some tutorials that says that i need to change the parameter order, i changed it into :

red, green, blue = tf.split(rgb_scaled, 3, 3)

And i got this error instead :

Expected int32, got list containing Tensors of type '_Message' instead.

Thanks.

tensorflow version and KeyError

Thanks for your work first. I wonder which version of tensorflow did you use in this project? My tf version is 0.10.0 and when I run your code I got such an error:

Traceback (most recent call last):
File "test_vgg16.py", line 22, in <module>
   vgg.build(images)
File "/Users/yobichi/bigdata/gln/tensorflow-vgg-master/vgg16.py", line 46, in build
   self.conv1_1 = self.conv_layer(bgr, "conv1_1")
File "/Users/yobichi/bigdata/gln/tensorflow-vgg-master/vgg16.py", line 91, in conv_layer
   filt = self.get_conv_filter(name)
 File "/Users/yobichi/bigdata/gln/tensorflow-vgg-master/vgg16.py", line 120, in get_conv_filter
   return tf.constant(self.data_dict[name][0], name="filter")
 KeyError: 0

Do you have any idea? Thanks in advance!

The conv kernel is 11. It's different with the conv kernel in the paper which is 33. It's something wrong. Thanks for your answer

The size of kernel in convolutional lay is 11. However the size of kernel in convolutional lay in the paper which is 33. It's something wrong. Thanks for your answer

How I do fine-tuning using VGG19 on my Own Data in TF

For example, I want to classify 25 labels of my image data, which means there are 25 outputs. I don't know how to ignore the last FC-layer parameters. My idea is first to load all pre-trained parameters and then change the network in memory(change the last fully connected layer, randomly initialize it), then feed my data in and add the optimizer， re-train the network? I don't if it is sensible. Hope for help!!!

train model seems not correct

I want to train the model from the initial state. The images are all 224*224. But the problem is, when I train the model in a FOR loop, the probability is always closed to 50%.（ My problem is a two classification problem.）Can you provide the complete codes? Or can you fix my problem?

Fine-tune the model

Hi, I'm trying to use vgg19 on my problem which is a two-class classification, so I want to fine-tune the model on my dataset and I have modified the 'fc8' layer in ''vgg19_trainable.py''(in-size:4096, out-size:2).
I want to set the batch-size to be 32, so should I first read and reshape the 32 images and use 'np.concatenate()' to combine the images, just like the code in the 'test_vgg10.py'? And I think the label-list for each batch is a list type which contains 32 1-hot label(each label is also a list type), am I doing right?

Please specify any kind of license

Hate to be that guy. I don't really care what license, either. :)

No such file or directory: './vgg19.npy’

I can't get the vgg19.npy

NPYs are not available for downloading

The two pre-trained models of npy files are not able to be downloaded.

MEAN

why the mean of image is [103.939, 116.779, 123.68] in VGG16.py?? THANKS.

Your trainable VGG 19 finetune issue

@machrisaa I have two questions from you in your Vgg19 trainable implementation.
I am using your code and your pretrianed, but, now if want to finetune this network for some other problem using triplet loss, then, how could i implement back propogation? Is that done automatically using your code.
I have written a code, if you have time i can send you my code to view.

Error when running test_vgg19.py

I've been trying to run your test_vgg19.py, but get the following error:

ValueError: GraphDef cannot be larger than 2GB.

It seems like the graph is getting too big. So I'm wondering whether I'm doing something really stupid or whether you have the same problem?

Generate Image using this?

Is there an example showing how to use this to actually generate an image from a source image using the trainer ?

I have it running and outputting to terminal properly but it is not clear how to feed it an image and get out a different image.

Thanks

Is the accuracy same as the description paper？

-bash: syntax error near unexpected token `('

What about use this pre-trained model to train new stuffs?

What should I do if I want to add more class to that pre-trained weight file? So now the class should goes to 1001. So that it can recognize those 1000 objects have already been trained and also can recognize the thing I added in. Will that be possible?

Thanks.

Thanks,

wenouyang