hakuyume / chainer-ssd Goto Github PK

seems the file in the URL is no longer available in the server (http://www.cs.unc.edu/~wliu/projects/SSD/)

http://www.cs.unc.edu/%7Ewliu/projects/SSD/models_VGGNet_VOC0712_SSD_300x300.tar.gz

In Windows, if not explicitly cast to float, it will be 0 and 'demo.py' will not work. work by this fix.
your code is very helpful. thank you!

class SSD300(chainer.Chain):
---------------------
#steps = [s / 300 for s in (8, 16, 32, 64, 100, 300)] ----- original code
steps = [np.float32(s) / 300 for s in (8, 16, 32, 64, 100, 300)]
#sizes = [np.float32(s) / 300 for s in (30, 60, 111, 162, 213, 264, 315)] ----- original code
sizes = [np.float32(s) / 300 for s in (30, 60, 111, 162, 213, 264, 315)]

Where is your code to calculate mAP?

The pascal provided code is in matlab, and it is frustrating that there are no good python implementation available anywhere. I have been struggling to get a proper mAP for my pytorch implementation and am looking to check with someone else's.

Thanks

Hello! @Hakuyume Great jobs. Your work makes the world a better place. I have seen your project carefully. But, now I have some with the file voc.py.

Just as the picture shows: os.path.join(root, 'ImageSets', 'Main', name + '.txt'), after unzip VOCtrainval_06-Nov-2007.zip ,it's folder "Main" include many *.txt files, but in file voc.py, I guess 'name' should be in ['train', 'val','trainval'] or similar format. I'm a little confused here. It would be great if you can help me out. Thanks in advance..

Hello Hakuyume-san,
On Windows and Python 3.6.0 :: Anaconda 4.3.0 (64-bit), This line occurs error.

multibox.py:56: 
hard_neg = rank < (np.count_nonzero(t_conf, axis=1) * 3)[:, np.newaxis]
TypeError: 'int' object is not subscriptable

I thought count_nonzero with axis does not work on windows.
Output type is 'int' expected 'numpy'.
So I replaced source code like bellow and it seems working well.

        non_zero = np.zeros((n,1))
        for i in range(n):
            non_zero[i] = (np.count_nonzero(t_conf[i])) 
        non_zero *= 3
        hard_neg = rank < non_zero

Honestly saying, I have just started python for 2 month, so it is not beautiful.
But I like this chainer and SSD. I would like to contribute you. Please tell me how to do.
Thanks for your excellent always working.

Hi, when I try to convert VGG_ILSVRC_16_layers_fc_reduced.caffemodel to vgg16.npz, some error occured:
on Win10 OS(Python 3.5.2, protobuf 3.2.0, chainer 1.21.0):

and on Ubuntu 14.04 (python 3.4.3, protobuf 3.1.0, chainer 1.21.0):
.Now, I can not come up with the reason, so I ask for a hand from you. Thanks in advance .

I try VGG_VOC0712Plus_SSD_512 model from here.

When I do demo.py, it displays wrong coordinates for dog.

For example:

$ ./demo.py --arch 512 VGG_VOC0712Plus_SSD_512.npz VOCdevkit/VOC2007/JPEGImages/000001.jpg

result:

Is this expected behavior?
I use python2.7, cuda-8.0 and chainer 1.24.

Thanks in advance.