jaggeryoung / c3d-mxnet Goto Github PK

Thanks for sharing C3D implementation. I met this error:

mxnet.base.MXNetError: InferShape Error in pooling4: [16:56:37] src/operator/./pooling-inl.h:232: Check failed: (param_.kernel[0]) < (dshape[2] + 2 * param_.pad[0]) kernel size exceeds input

This can be seen whenever I want to collapse the last two frames into one.
I think it is related to cudnn as 3D conv is only supported by cudnn(??). Is your mxnet the latest version? What is your cudnn version?
Thanks

If you have 160frames in video. Following the code,It extracted the 1 frame, 10frame ...160frame. It is not continuous, but the paper said extracted continuous frame.In my understanding, it should be 116frames,1732frames,....Am I wrong? Hope replying!

` pic = []
#print dirName
for filename in glob.glob(dirName+'/*.jpg'):
pic.append(filename)
pic.sort()
#print len(pic)
ret = []
len_pic = len(pic)
tmp = len_pic/num
for i in range(num):
ret.append(pic[i * tmp])
r_1 = []
g_1 = []
b_1 = []
mat = []

for i in range(len(ret)):
    img = cv2.imread(ret[i], cv2.IMREAD_COLOR)
#img = img.resize(data_shape[2],data_shape[3])
    b,g,r = cv2.split(img)
r = cv2.resize(r, (data_shape[3], data_shape[2]))
g = cv2.resize(g, (data_shape[3], data_shape[2]))
b = cv2.resize(b, (data_shape[3], data_shape[2]))
r = np.multiply(r, 1/255.0)
g = np.multiply(g, 1/255.0)
b = np.multiply(b, 1/255.0)
r_1.append(r)
g_1.append(g)
b_1.append(b)
#mat.append(img)
mat.append(r_1)
mat.append(g_1)
mat.append(b_1)
#print len(mat),len(mat[0][0])
return mat`

I use the latest mxnet and c3d-mxnet and follow the instruction to train the ucf101 dataset,
but the Validation-accuracy remains the same
does anyone have the same problem?

~/C3D-mxnet$ sudo python train_ucf101.py
9537
3783
[('data', (10, 3, 28, 122, 122))] [('label', (10,))]
train_ucf101.py:164: DeprecationWarning: mxnet.model.FeedForward has been deprecated. Please use mxnet.mod.Module instead.
initializer = mx.init.Xavier(factor_type="in", magnitude=2.34))
begin fit
/usr/local/lib/python2.7/dist-packages/mxnet-0.9.4-py2.7.egg/mxnet/model.py:516: DeprecationWarning: Calling initializer with init(str, NDArray) has been deprecated.please use init(mx.init.InitDesc(...), NDArray) instead.
self.initializer(k, v)
2017-02-20 13:06:16,231 Start training with [gpu(0)]
[13:06:24] src/operator/./cudnn_convolution-inl.h:55: Running performance tests to find the best convolution algorithm, this can take a while... (setting env variable MXNET_CUDNN_AUTOTUNE_DEFAULT to 0 to disable)
2017-02-20 13:36:05,022 Epoch[0] Resetting Data Iterator
2017-02-20 13:36:05,022 Epoch[0] Time cost=1777.772
2017-02-20 13:47:20,723 Epoch[0] Validation-accuracy=0.011640
2017-02-20 14:11:57,614 Epoch[1] Resetting Data Iterator
2017-02-20 14:11:57,614 Epoch[1] Time cost=1476.891
2017-02-20 14:20:42,997 Epoch[1] Validation-accuracy=0.011640
2017-02-20 14:43:38,866 Epoch[2] Resetting Data Iterator
2017-02-20 14:43:38,867 Epoch[2] Time cost=1375.870
2017-02-20 14:52:21,879 Epoch[2] Validation-accuracy=0.011640