Quantization of MobileNet-SSD fails

[salcanmor]

Hi guys, I've read #53 and #42 but I cannot quantize MobileNet-SSD.

Link: https://github.com/chuanqi305/MobileNet-SSD

This is what I've done: 1- Download prototxt: https://drive.google.com/file/d/0B3gersZ2cHIxWGEzbG5nSXpNQzA/view 2- Download caffe model: https://drive.google.com/open?id=0B3gersZ2cHIxRm5PMWRoTkdHdHc 3- Run XportDNN.pyc:

python XportDNN.pyc --quant_type "Xilinx" \
--deploy_model ./mobilenet/MobileNet_deploy.prototxt \
--weights ./mobilenet/MobileNetSSD_deploy.caffemodel \
--quantized_deploy_model ./mobilenet/quantized_deploy.prototxt \
--calibration_directory ./xilinx_quant_example_model/sample_calibration_dataset/ --calibration_size 10 \
--bitwidths 6,6,6 --dims 3,300,300 --transpose 2,0,1 \
--channel_swap 2,1,0 --raw_scale 255.0 \
--mean_value 127.5 --input_scale 0.007843

Please find attached both prototxt and caffemodel.

MobileNet-SSD.zip

Error:

F1205 09:58:58.985906 29455 net.cpp:813] Check failed: target_blobs.size() == source_layer.blobs_size() (1 vs. 2) Incompatible number of blobs for layer conv0
*** Check failure stack trace: ***
Abortado (`core' generado)

What am I doing wrong?

Thanks in advance!

[cuongdv1]

@salcanmor Have you modified the first layer in prototxt as below?

name: "MobileNet-SSD" layer { name: "data" type: "Input" top: "data" input_param { shape { dim: 1 dim: 3 dim: 300 dim: 300 } } }

[salcanmor]

@salcanmor Have you modified the first layer in prototxt as below?

name: "MobileNet-SSD" layer { name: "data" type: "Input" top: "data" input_param { shape { dim: 1 dim: 3 dim: 300 dim: 300 } } }

I've tried that with no luck. See prototxt attached. Did you manage to make the quantization of MobileNet-SSD work?

MobileNet_deploy_orig.prototxt.zip

[cuongdv1]

@salcanmor I made the quantization of Mobilenet-SSD and can run MobilenetSSD on CHaiDNN but the accuracy is too bad. I think depthwise conv was not supported, I am waiting for the help

[salcanmor]

@salcanmor I made the quantization of Mobilenet-SSD and can run MobilenetSSD on CHaiDNN but the accuracy is too bad. I think depthwise conv was not supported, I am waiting for the help

Can you please share the model and prototxt you used? Thanks in advance!

[cuongdv1]

Prototxt and caffemodel I cloned from here https://github.com/chuanqi305/MobileNet-SSD

[salcanmor]

Prototxt and caffemodel I cloned from here https://github.com/chuanqi305/MobileNet-SSD

If we are using the same Prototxt and caffemodel, why do I get an error and you not ?

[cuongdv1]

Can you try this?

python XportDNN.pyc --quant_type "Xilinx" --deploy_model ./models/MobilenetSSD_300_deploy.prototxt --weights ./models/MobilenetSSD_300_deploy.caffemodel --quantized_deploy_model ./models/MobilenetSSD_300_quantized_deploy.prototxt --calibration_directory ./data/VOC0712 --calibration_size 32 --bitwidths 8,8,8 --dims 3,300,300 --transpose 2,0,1 --channel_swap 2,1,0 --raw_scale 255.0 --mean_value 127.5,127.5,127.5 --input_scale 0.007843

[salcanmor]

Can you try this?

python XportDNN.pyc --quant_type "Xilinx" --deploy_model ./models/MobilenetSSD_300_deploy.prototxt --weights ./models/MobilenetSSD_300_deploy.caffemodel --quantized_deploy_model ./models/MobilenetSSD_300_quantized_deploy.prototxt --calibration_directory ./data/VOC0712 --calibration_size 32 --bitwidths 8,8,8 --dims 3,300,300 --transpose 2,0,1 --channel_swap 2,1,0 --raw_scale 255.0 --mean_value 127.5,127.5,127.5 --input_scale 0.007843

I'm gettting the same error:

F1210 09:14:50.208392 27097 net.cpp:813] Check failed: target_blobs.size() == source_layer.blobs_size() (1 vs. 2) Incompatible number of blobs for layer conv1/dw
*** Check failure stack trace: ***
Abortado (`core' generado)

[Jiaxinging]

@salcanmor I made the quantization of Mobilenet-SSD and can run MobilenetSSD on CHaiDNN but the accuracy is too bad. I think depthwise conv was not supported, I am waiting for the help

Hi, can you please share the FPS of MobilenetSSD running on ZCU102? Thanks a lot~

[Jiaxinging]

@cuongdv1 ,
Hi, can you please share the FPS of MobilenetSSD running on ZCU102? Thanks a lot~

[cuongdv1]

30fps, but the accuracy is too bad. i think it did not run full network

[Jiaxinging]

Thank you! Maybe the bad accuracy is caused by the Xilinx quantization, it is not good enough to quantize the light models like mobilenet......

[averr5]

@cuongdv1 Did you use deploy.prototxt and mobilenet_iter_73000.caffemodel as inputs ?