xuduo35 / unet_mxnet2ncnn Goto Github PK

Hi, thanks for your sharing. I want to know how about the fps and hardware configure in your test, can you share this information. Thanks!

when i run trainunet, I get train_loss and validat_loss, It doesn't feel right

New training ...
Best model at Epoch 1
Epoch 0: Moving Training Loss -0.98716, Validation Loss -0.98693
......
Epoch 23: Moving Training Loss -0.99521, Validation Loss -0.99272
Epoch 24: Moving Training Loss -0.99550, Validation Loss -0.99271
.....

i have model trained UNET segmentation (keras), then converted to NCNN (param, bin) by tool keras2ncnn. Can i use this source code for my converted model? Thanks <3

top = (INPUT_HEIGHTtop)/width;
bottom = (INPUT_HEIGHTbottom)/width;
left = (INPUT_WIDTHleft)/height;
right = (INPUT_WIDTHright)/height;
请教您一下，这四个变量这么定义有什么具体的意义吗？

when i finish the train work and get best_unet_person_segmentation-symbol.json and best_unet_person_segmentation-0000.params
the predict encounter this error:
data: (1, 3, 256, 256)
[08:41:50] c:\jenkins\workspace\mxnet-tag\mxnet\src\common\exec_utils.h:392: InferShape pass cannot decide shapes for the following arguments (0s means unknown dimensions). Please consider providing them as inputs:
softmax_label: [],

When I train ,got these errors:
Traceback (most recent call last):
File "trainunet.py", line 95, in
main()
File "trainunet.py", line 73, in main
curr_loss = F.mean(loss).asscalar()
File "/Library/Python/2.7/site-packages/mxnet/ndarray/ndarray.py", line 1990, in asscalar
return self.asnumpy()[0]
File "/Library/Python/2.7/site-packages/mxnet/ndarray/ndarray.py", line 1972, in asnumpy
ctypes.c_size_t(data.size)))
File "/Library/Python/2.7/site-packages/mxnet/base.py", line 252, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [09:22:19] src/operator/nn/./deconvolution-inl.h:224: Check failed: req[deconv::kOut] == kWriteTo (0 vs. 1)

when i train the unet i get this loss value:
Epoch 8: Moving Training Loss -0.98323, Validation Loss -0.94432

raining 0/10 ./datasets/humanparsing/JPEGImages/2500_1005.jpg
raining 1/10 ./datasets/humanparsing/JPEGImages/2500_1004.jpg
raining 2/10 ./datasets/humanparsing/JPEGImages/2500_10.jpg
raining 3/10 ./datasets/humanparsing/JPEGImages/2500_1002.jpg
raining 4/10 ./datasets/humanparsing/JPEGImages/2500_1006.jpg
raining 5/10 ./datasets/humanparsing/JPEGImages/2500_1003.jpg
raining 6/10 ./datasets/humanparsing/JPEGImages/2500_1000.jpg
raining 7/10 ./datasets/humanparsing/JPEGImages/2500_100.jpg
raining 8/10 ./datasets/humanparsing/JPEGImages/2500_1001.jpg
raining 9/10 ./datasets/humanparsing/JPEGImages/2500_1.jpg
raining 10/10
g 0/10 ./datasets/humanparsing/JPEGImages/dataset10k_6506.jpg
g 1/10 ./datasets/humanparsing/JPEGImages/dataset10k_6507.jpg
g 2/10 ./datasets/humanparsing/JPEGImages/dataset10k_6508.jpg
g 3/10 ./datasets/humanparsing/JPEGImages/dataset10k_6509.jpg
g 4/10 ./datasets/humanparsing/JPEGImages/dataset10k_651.jpg
g 5/10 ./datasets/humanparsing/JPEGImages/dataset10k_6510.jpg
g 6/10 ./datasets/humanparsing/JPEGImages/dataset10k_6511.jpg
g 7/10 ./datasets/humanparsing/JPEGImages/dataset10k_6512.jpg
g 8/10 ./datasets/humanparsing/JPEGImages/dataset10k_6513.jpg
g 9/10 ./datasets/humanparsing/JPEGImages/dataset10k_6514.jpg
g 10/10
Epoch 9: Moving Training Loss -0.98339, Validation Loss -0.94633

is this loss value right? thank you!

when run python trainunet.py.
File "/usr/local/lib/python3.4/dist-packages/mxnet/base.py", line 251, in check_call
raise MXNetError(py_str(_LIB.MXGetLastError()))
mxnet.base.MXNetError: [01:35:45] src/storage/./cpu_device_storage.h:73: Failed to allocate CPU Memory

my unet convert to ncnn(.param/.bin)，no error appear

when I inference in vs2017,

.......
ncnn::Mat mask;
ex.extract("final_layer", mask);
std::cout << "whc " << mask.w << " " << mask.h << " " << mask.c << " " << std::endl;

get : whc 2 128 2
so ,why mask.h != mask.w ? my model should get (128,128,2) output.