hizhangp / yolo_tensorflow Goto Github PK

Hi,

I spot a minor typo in train.py line 110, 111. I think the closing bracket is misplaced.

Best

Hi, deeply thanks for sharing your work!

I found that there is a transpose in your code:
net = tf.transpose(net, [0, 3, 1, 2], name='trans_31')

This is where a conv layer gets flattened before fc layers. May I ask why transpose is there? Is this a standard way to do so? The YOLO paper recommends doing so?

I have trained a model with your coded and the total loss was about 8 after 30000 iterations. But when I test the trained weights, it cannot dectect objects correctly. For example , the confidents it given are lower than 0.01.

I took 18 hours for training, but this training haven't end.
When will the work which is training end?
Thanks!

When I trained the network ,there was a value error -"uknown code 'f' for object of type 'str'" in train.py, and the program stoped in line of 'train_timer.remain(step, self.max_iter))'. Why?

Can you put on a license?

Hi, im interested in your amazing project.
I've trained a model followed by your steps. But when i tried to test the model.
It seems that nothing was detected in the picture.
However, run the test.py with YOLO_small.ckpt is OK. But I cannot detect anything with save.ckpt-15000 etc.
i changed nothing in your train.py. could you tell me what's possibly going wrong?

When i restored the YOLO_small.ckpt，there was an error: "Data loss: not an sstable (bad magic number): perhaps your file is in a different file format and you need to use a different restore operator?"

As we know, tensorflow saves model into four files including checkpoint, index, data and meta, why this checkpoint only has one file, how can i use it?
Thanks for anyone's help.

By the way, when i trained the model in Pascal VOC 2007 without using the pre-trained model, the loss couldn't decrease and changed between 8 and 12.

Whatever photos I put , the net_ouput will be the same.And the final image showed last have the same frame.I am upset about this phenomenon.
Hope someone could tell me any advice.
Thank you.

def detect_from_cvmat(self, inputs):
net_output = self.sess.run(self.net.logits,
feed_dict={self.net.images: inputs})

AttributeError: 'YOLONet' object has no attribute 'loss'?
I check YONONet class and there is no 'loss' actually? is it wrong?

Dear all:
I have been training this network thousands of iterations, but I nearly cannot find any change about the loss value. So what's wrong? Does anyone have validated its validity?

InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [539] rhs shape= [1470]
[[Node: save/Assign_52 = Assign[T=DT_FLOAT, _class=["loc:@yolo/fc_36/biases"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/gpu:0"](yolo/fc_36/biases, save/RestoreV2_52/_5)]]

I trained my own data set and only one class has this problem

hi again!
recently I've spent time reading YOLO papers and trying to understand your codes.
now i think it's time to fine-tune the yolo and then make some modification on the basic structure of yolo.

but soon after I managed to prepare my own datasets from IMAGENET, I got another issue regarding python version.. I'm using python3.x and i have absolutely no idea about python2 or compatibility issue..

now i decided to first convert your codes to the codes that can be used in python3.x and then use them.
this weekend i'm just gonna convert all the python2-type print to python3-type print function and then i will execute the codes.
ex) print 'Hello world!' ----> print('Hello World')

so.. can you tell me what kind of other issues i have to consider except for print function issues when py2.x -> py3.x ?

thanks in advance!

What are steps in training our own dataset

NotFoundError (see above for traceback): Tensor name "yolo/conv_2/biases" not found in checkpoint files data/weights/YOLO_small.ckpt

@hizhangp
If I just want to test people, how should I modify the code?

I have successfully trained the model with my own dataset(which have two classes)and saved the weight file. But when I loaded the saved model to conduct testing, something wrong with the weight file

/core/util/tensor_slice_reader.cc:95] Could not open ./data/weights/save.ckpt: Data loss: not an sstable (bad magic number): perhaps your file is in a different file format and you need to use a different restore operator?

The error appears in the line:

self.saver.restore(self.sess, self.weights_file)

The version of tensorflow I am currently using is 1.2. I'm not sure if it's the version problem(I have searched on the Internet for this error and found the same problems related to the version of tensorflow ). Could you help me with this issue and how to restore weight files which have been trained?

I‘ve tried to use test.py to test the model trained by train.py. I rename the model to save.ckpt and use it as the weights to run the test.py. But it seems that test.py cannot open that file. Am i wrong in testing methods?

Hello, why hasn't my loss changed from the beginning to the end of training? Do you know what the reason is? Thank you very much!

• Did you implement yolo v2 in this project? If it is not, do you have a plan to do it?

Thank you very much.

if i want to train this network on my own pictures, should i make my data the same format as VOC?

Hi I'm a beginner in this deep-learning field and now I'm trying to fine-tune some of the well-known object detectors such as Yolo. first I want to train Yolo to classify a single class(person or none) and also to find bounding boxes of multiple objects(people) in an image or in videos

I think I can use just the PASCAL VOC data sets for fine-tuning. and I have a laptop with i5-6200U, Geforce 940M. I know its not such a good environment and condition especially for training CNNs.

so.. how do you expect the duration time for fine-tuning this Yolo, if I use roughly 1000 images?

thanks in advance

p.s
sorry for my bad english and i hope you understood my intention!

predict_boxes_tran = tf.stack(
[(predict_boxes[..., 0] + offset) / self.cell_size,
(predict_boxes[..., 1] + offset_tran) / self.cell_size,
tf.square(predict_boxes[..., 2]),
tf.square(predict_boxes[..., 3])], axis=-1)

Why do you take the square of predicted h and w here?

And in
boxes_delta = coord_mask * (predict_boxes - boxes_tran)

You don't take the square root of predicted h and w.

So the output W and H of the network has already been taken root?

HOW to train my own dataset?I need to use my image to detect the defects,so I test the VOC2007, it works.But when I change the images the Keyerror happens.In addition, CAN the model in the program be changed?or use a simpler model?please connect with me,[email protected]

Hi,
I'm try train with another number class (2,3,21...)
What parameter should I change to train with difference number classes?

I have run test.py successfully, and got a visualization picture of person.jpg. And I also want to get the iou on VOC2007 test dataset. How could I do? Did anyone get the test iou on VOC2007 test dataset? Can you tell me ? Thank you.

Hi, anyone can help

when I try python test.py I get the following error

abdulrahman@abdulrahman-ThinkPad-X230-Tablet:~/yolo_tensorflow$ python test.py
Restoring weights from: data/weights/YOLO_small.ckpt
Average detecting time: 0.608s
Traceback (most recent call last):
File "test.py", line 190, in
main()
File "test.py", line 186, in main
detector.image_detector(imname)
File "test.py", line 161, in image_detector
self.draw_result(image, result)
File "test.py", line 43, in draw_result
cv2.putText(img, result[i][0] + ' : %.2f' % result[i][5], (x - w + 5, y - h - 7), cv2.FONT_HERSHEY_SIMPLEX, 0.5, (0, 0, 0), 1, cv2.CV_AA)
AttributeError: 'module' object has no attribute 'CV_AA'

Hi,
I'm trying to run train.py with gpu options with
$ python train.py --gpu 0
it runs without errors but it takes too much time that it shows no improvements in speed of computation compared to the one with no gpu option
I've installed cuda without problems and other cuda-based applications run fine.
I printed out device placements with log_device_placement option and found out everything were run
on CPU
can anyone tell me why this happens?

Hi, I trained my own dataset with the yolo_tensorflow, but get the following error:
InvalidArgumentError (see above for traceback): Assign requires shapes of both tensors to match. lhs shape= [539] rhs shape= [1470] [[Node: save/Assign_49 = Assign[T=DT_FLOAT, _class=["loc:@Variable_53"], use_locking=true, validate_shape=true, _device="/job:localhost/replica:0/task:0/gpu:0"](Variable_53, save/RestoreV2_49/_1)]]
I don't know how to do with this error

I try to train the yolo networks, and don't use pre-train weight, but I get some error.
that is my process:

1. changed '\yolo\config' WEIGHTS_FILE = os.path.join(DATA_PATH, 'weights', 'YOLO_small.ckpt')
   to WEIGHTS_FILE = None
2. 'parser.add_argument('--weights', default="YOLO_small.ckpt", type=str)' was comment out, because I want train yolo network without pre-train weights
3. 'cfg.WEIGHTS_FILE = os.path.join(cfg.WEIGHTS_DIR, args.weights)' was comment out, the reason is same as above

Please tell how, wo know my way is stupid, but I still want to try. Thx.

python test.py
Restoring weights from: data/weights/YOLO_small.ckpt
Traceback (most recent call last):
File "test.py", line 183, in
main()
File "test.py", line 175, in main
detector.camera_detector(cap)
File "test.py", line 140, in camera_detector
result = self.detect(frame)
UnboundLocalError: local variable 'frame' referenced before assignment

OS: Ubuntu 16.04.1 (64 Bits)
Python 2.7.12
Python 2.7.12 (default, Nov 19 2016, 06:48:10)
[GCC 5.4.0 20160609] on linux2
Type "help", "copyright", "credits" or "license" for more information.

import cv2
cv2.version
'2.4.13'
OpenCV version 2.4.13
pip 9.0.1

pip list:

appdirs (1.4.2)
apturl (0.5.2)
asgi-redis (1.0.0)
asgiref (1.0.0)
autobahn (0.17.1)
beautifulsoup4 (4.4.1)
bleach (1.5.0)
blinker (1.3)
Brlapi (0.6.4)
channels (1.0.3)
chardet (2.3.0)
checkbox-support (0.22)
command-not-found (0.3)
constantly (15.1.0)
cryptography (1.2.3)
daphne (1.0.2)
decorator (4.0.11)
defer (1.0.6)
Django (1.10.5)
entrypoints (0.2.2)
feedparser (5.1.3)
guacamole (0.9.2)
html5lib (0.9999999)
httplib2 (0.9.1)
idna (2.0)
incremental (16.10.1)
ipykernel (4.5.2)
ipython (5.2.2)
ipython-genutils (0.1.0)
ipywidgets (5.2.2)
Jinja2 (2.9.5)
jsonschema (2.6.0)
jupyter (1.0.0)
jupyter-client (4.4.0)
jupyter-console (5.1.0)
jupyter-core (4.2.1)
language-selector (0.1)
louis (2.6.4)
lxml (3.5.0)
Mako (1.0.3)
MarkupSafe (0.23)
mistune (0.7.3)
msgpack-python (0.4.8)
nbconvert (5.1.1)
nbformat (4.2.0)
nltk (3.2.2)
notebook (4.3.2)
numpy (1.12.0)
oauthlib (1.0.3)
onboard (1.2.0)
packaging (16.8)
padme (1.1.1)
pandas (0.19.2)
pandocfilters (1.4.1)
pexpect (4.2.1)
pickleshare (0.7.4)
Pillow (3.1.2)
pip (9.0.1)
plainbox (0.25)
prompt-toolkit (1.0.13)
protobuf (3.2.0)
ptyprocess (0.5.1)
pyasn1 (0.1.9)
pycups (1.9.73)
pycurl (7.43.0)
Pygments (2.2.0)
pygobject (3.20.0)
PyJWT (1.3.0)
pyparsing (2.1.10)
python-apt (1.1.0b1)
python-dateutil (2.6.0)
python-debian (0.1.27)
python-systemd (231)
pytz (2016.10)
pyxdg (0.25)
pyzmq (16.0.2)
qtconsole (4.2.1)
redis (2.10.5)
reportlab (3.3.0)
requests (2.9.1)
scikit-learn (0.18.1)
scipy (0.18.1)
screen-resolution-extra (0.0.0)
sessioninstaller (0.0.0)
setuptools (34.3.0)
simplegeneric (0.8.1)
six (1.10.0)
system-service (0.3)
tensorflow (1.0.0)
terminado (0.6)
testpath (0.3)
tornado (4.4.2)
tqdm (4.11.2)
traitlets (4.3.1)
Twisted (16.6.0)
txaio (2.6.0)
ubuntu-drivers-common (0.0.0)
ufw (0.35)
unattended-upgrades (0.1)
unity-scope-calculator (0.1)
unity-scope-chromiumbookmarks (0.1)
unity-scope-colourlovers (0.1)
unity-scope-devhelp (0.1)
unity-scope-firefoxbookmarks (0.1)
unity-scope-gdrive (0.7)
unity-scope-manpages (0.1)
unity-scope-openclipart (0.1)
unity-scope-texdoc (0.1)
unity-scope-tomboy (0.1)
unity-scope-virtualbox (0.1)
unity-scope-yelp (0.1)
unity-scope-zotero (0.1)
urllib3 (1.13.1)
usb-creator (0.3.0)
wcwidth (0.1.7)
wheel (0.29.0)
widgetsnbextension (1.2.6)
xdiagnose (3.8.4.1)
xkit (0.0.0)
XlsxWriter (0.7.3)
zope.interface (4.3.3)

2018-03-31 21:29:02.437666: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:247] PoolAllocator: After 10024 get requests, put_count=10022 evicted_count=1000 eviction_rate=0.0997805 and unsatisfied allocation rate=0.109936
2018-03-31 21:29:02.437717: I tensorflow/core/common_runtime/gpu/pool_allocator.cc:259] Raising pool_size_limit_ from 100 to 110
Traceback (most recent call last):
File "train.py", line 165, in
main()
File "train.py", line 158, in main
solver.train()
File "train.py", line 90, in train
train_timer.remain(step, self.max_iter))
ValueError: Unknown format code 'f' for object of type 'str'

thanks for your code! and i find this error when i training yolo-tensorflow in GPU device! i want to know how to solve it? if you have better ideal please answer me ! my GPU device is GT 1080!

1. The loss equation describes the confidence loss of no object is " C-C' " , but in the code, the loss is "noobject_delta = noobject_mask * predict_scales"

2. λnoobj are different λclass are different from the paper.

could it be done by using cpu only.?
it has been training in my pc since more than two hours?

Hi,
I wonder how should i edit the config file.
What parameters should I make adjustment?

I want to train my own data set. Can I provide initial weights or do I need to train them?

image = cv2.resize(image, (self.image_size, self.image_size))
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB).astype(np.float32)
image = (image / 255.0) * 2.0 - 1.0)

I have two questions for these three lines of codes.

1. After 'resize' operation, why do we need to use the cvColor function?

2. What's the function of the third code?

I want to train my own data, but my sample's resolution is too low. After i do these tree operation, the result is terrible. So i just want to resize my sample without the last two operation. But after I read your code, I don't know if there are any influence if I remove them.(I am not a native English speaker, thanks for your answer.)

@hizhangp I have use the pascal_voc2007 dataset to train the network, and try to use the trained network to do some prediction. I found that whatever image feed to the network, It always generate the same output(both bonding-box and confidence in different image are the same), through the total_loss is quite low.
So I review the code, and try to understand what have happened during the training progress, and There are two place is really hard for me to get it through, they both in "loss_layer"
First of all, “boxes = tf.tile(boxes, [1, 1, 1, self.boxes_per_cell, 1]) / self.image_size”， why the "boxes" needed to be divided by "image_size" , when the predict_boxes_train only divided by cell_size
Secondly, why the predict_box should be plused with "offset" when we calculate the predict_boxes_train,?can we remove it?

I notice the description in YOLO paper that YOLO only have 2 fc layers. However, there are 3 fc layers in this code, in which the 512 fc layer don't exist in YOLO paper. I doubt whether this code do it right way.

By the way, can you tell me how do you obtain the pretrained model? I mean how is YOLO_small.ckpt generated, for YOLO source code is a C++ code and use darknet.

Could you upload a pre-trained conv layer weights (without fc layer)? Thanks.

Hi there, sorry about that to post this here, i was trying to write yolo in torch but struggled the backward function for so many days, but the gradients are always exploding slowly, could you kindly shed a light on my codes? Thank you so much.

gradInput[{ {}, {}, 1, {}, {} }] = self.mse:backward(torch.cmul(self.x_buffer, x, coord_mask), tx)
gradInput[{ {}, {}, 2, {}, {} }] = self.mse:backward(torch.cmul(self.y_buffer, y, coord_mask), ty)
gradInput[{ {}, {}, 3, {}, {} }] = self.mse:backward(torch.cmul(self.w_buffer, w, coord_mask), tw)
gradInput[{ {}, {}, 4, {}, {} }] = self.mse:backward(torch.cmul(self.h_buffer, h, coord_mask), th)
gradInput[{ {}, {}, 5, {}, {} }] = self.mse:backward(torch.cmul(self.conf_buffer, conf, coord_mask), tconf)
gradInput[{ {}, {}, { 6, 5 + nC }, {}, {} }][self.cls_mask] = self.ce:backward(torch.cmul(self.cls_buffer, cls), tcls)

@hizhangp
net = tf.pad(images, np.array([[0, 0], [3, 3], [3, 3], [0, 0]]), name='pad_1')
net = slim.conv2d(net, 64, 7, 2, padding='VALID', scope='conv_2')
I don't understand why the tf.pad was added before slim.conv2d， furthermore，it expands the size of the network。

hi,
how can train the model to recognize new type of images ?

in config.py
i modified just

WEIGHTS_FILE = None
BATCH_SIZE = 5

BUT!!
after training abt 3hours,
i did test.py, but there are no box on imname = 'test/cat.jpg'
..
what is the problem?? I just want to get my own weights file by my training

when i train with WEIGHTS_FILE = os.path.join(DATA_PATH, 'weights', 'YOLO_small.ckpt')
it worked well

Hi!
How can I changed the code if I want to input data (only train 'car' and 'cat')?
Thank you!
Shu Guo

Dear author! I want to know how can i get this number. According to the way in yolo's paper, the output tensor of yolo could be [7x7x(2x5+20)], thus [7x7x30], but i found the output of your code is 7x7x25. So i want to know is there any different?(Forgive for my grammar errors, I am not a native English speaker.）