how to export my own mobilenet1.0

[ou525]

i train my own data using train_imagenet.py. but i donot how to export the model to mxnet.
i try：

import gluoncv as gcv
from gluoncv.utils import export_block

net = gcv.model_zoo.get_model('mobilenet1.0', pretrained=True)
net.load_parameters('./0.0877-imagenet-mobilenet1.0-88-best.params')
export_block('mobilenet1.0', net, preprocess=True, layout='HWC')
print('Done.')
But there was an error, how can I solve it?

error:
Traceback (most recent call last):
File "export_class.py", line 5, in
net.load_parameters('/home/ou/workplace/tmp_model/0.0877-imagenet-mobilenet1.0-88-best.params')
File "/usr/local/lib/python2.7/dist-packages/mxnet/gluon/block.py", line 410, in load_parameters
params[name]._load_init(loaded[name], ctx, cast_dtype=cast_dtype, dtype_source=dtype_source)
File "/usr/local/lib/python2.7/dist-packages/mxnet/gluon/parameter.py", line 279, in _load_init
self.name, str(self.shape), str(data.shape))
AssertionError: Failed loading Parameter 'mobilenet0_dense0_weight' from saved params: shape incompatible expected (1000, 1024) vs saved (5, 1024)

[hetong007]

Use

net = gcv.model_zoo.get_model('mobilenet1.0', classes=5)

instead.

[ou525]

@hetong007 ,thank you for you reply, i try do it ,but error still happen, as follow:
self.name, str(self.shape), str(data.shape))
AssertionError: Failed loading Parameter 'mobilenet0_dense0_weight' from saved params: shape incompatible expected (5, 0) vs saved (1000, 1024)

[hetong007]

basically the idea is to use the same model definition in both training and test.

So, let ask: how did you define the model when training (which generated the parameters that you want to load)?

[ou525]

train_imagenet.txt
i use the cmomand as follow:
python train_imagenet.py --rec-train train.rec --rec-train-idx train.idx --rec-val test.rec --rec-val-idx test.idx --model mobilenet1.0 --mode hybrid --lr 0.4 --lr-mode cosine --num-epochs 100 --batch-size 64 --num-gpus 1 -j 8 --use-rec --dtype float32 --warmup-epochs 5 --no-wd --label-smoothing --mixup --save-dir params_mobilenet1.0_mixup --logging-file mbilenet1.0_mixup.log

how can i do it ?

[hetong007]

what is the exact command that generates this error message:

self.name, str(self.shape), str(data.shape))
AssertionError: Failed loading Parameter 'mobilenet0_dense0_weight' from saved params: shape incompatible expected (5, 0) vs saved (1000, 1024)

[ou525]

this is run export.py
import gluoncv as gcv
from gluoncv.utils import export_block

net = gcv.model_zoo.get_model('mobilenet1.0', pretrained=True,classes=5)
net.load_parameters('./0.0877-imagenet-mobilenet1.0-88-best.params')
export_block('mobilenet1.0', net, preprocess=True, layout='HWC')
print('Done.')

[hetong007]

be careful, remember to remove pretrained=True

[ou525]

thank you very much, its ok.

[ou525]

@hetong007 Sorry to bother again, my training and export are normal, but in the c program execution has been a segmentation error, other models call normal.

error:
(gdb) bt
#0 0x00007fffddc69be8 in MXPredSetInput () from /home/ou/workplace/sv_video/third-libs/mxnet/lib/libmxnet.so
#1 0x000000000042f80c in Classification::predict (this=this@entry=0x7fffffffdb30, pred_hnd=0x7fffd52dee78 <operator new(unsigned long)+24>,
image_data=std::vector of length 150528, capacity 150528 = {...}, nd_hnd=,
synset_file="/home/ou/workplace/incubator-mxnet/example/image-classification/predict-cpp/model/resnet_ou/synset5.txt") at /home/ou/workplace/sv_video/src/classification.cpp:109
#2 0x000000000042fe2f in Classification::runabi:cxx11 (this=this@entry=0x7fffffffdb30, img=...) at /home/ou/workplace/sv_video/src/classification.cpp:464
#3 0x000000000044b9ad in SpecialVehicle::run (this=this@entry=0x7fffffffda60) at /home/ou/workplace/sv_video/src/special_vehicle.cpp:278
#4 0x000000000042ac45 in main (argc=, argv=) at /home/ou/workplace/sv_video/main.cpp:13

[hetong007]

@zhreshold can we have your help here

[zhreshold]

without your c++ code it's hard to tell. @ou525 would you mind share the piece of classification.cpp ?

[ou525]

classification.txt
i call the function run(img).@zhreshold

[zhreshold]

Can't find obvious problem so far. What if you use the pretrained model from gluoncv model zoo?

[ou525]

using the pretrained model from gluoncv model zoo and the trained model by mxnet are ok, so i think the train script may have some prblem ,but i donot kown how to do?