topic Re: Re:INT8 optimization for YOLOv4 in Intel® Distribution of OpenVINO™ Toolkit

INT8 optimization for YOLOv4

Rahila_T_Intel — Mon, 07 Dec 2020 13:14:34 GMT

I was trying to do POT optimization with custom trained yolov4 weights.

Initially converted the yolov4 weights to .pb file and then created IR files using openvino. Also I am able to optimize the model using POT.

I have checked the benchmark app with both IR FP32 model and INT8 model.

BENCHMARK APP RESULT :

------------------Using FP32 IR model----------------

Count: 860 iterations

Duration: 61260.79 ms

Latency: 695.84 ms

Throughput: 14.04 FPS

----------------Using INT8 model----------------------

Command tried :

python3 /openvino_2021.1.110/deployment_tools/tools/benchmark_tool/benchmark_app.py -m /results/TFYOLOV4_DefaultQuantization/2020-12-02_22-57-00/optimized/TFYOLOV4.xml -i cam --api_type async --number_iterations 10

[Step 1/11] Parsing and validating input arguments

/openvino_2021.1.110/python/python3.6/openvino/tools/benchmark/main.py:29: DeprecationWarning: The 'warn' method is deprecated, use 'warning' instead

logger.warn(" -nstreams default value is determined automatically for a device. "

[ WARNING ] -nstreams default value is determined automatically for a device. Although the automatic selection usually provides a reasonable performance, but it still may be non-optimal for some cases, for more information look at README.

[Step 2/11] Loading Inference Engine

[ INFO ] InferenceEngine:

API version............. 2.1.2021.1.0-1237-bece22ac675-releases/2021/1

[ INFO ] Device info

CPU

MKLDNNPlugin............ version 2.1

Build................... 2021.1.0-1237-bece22ac675-releases/2021/1

[Step 3/11] Setting device configuration

[ WARNING ] -nstreams default value is determined automatically for CPU device. Although the automatic selection usually provides a reasonable performance,but it still may be non-optimal for some cases, for more information look at README.

[Step 4/11] Reading network files

[ INFO ] Read network took 377.16 ms

[Step 5/11] Resizing network to match image sizes and given batch

[ INFO ] Network batch size: 1

[Step 6/11] Configuring input of the model

[Step 7/11] Loading the model to the device

[ ERROR ] Quantize layer StatefulPartitionedCall/functional_7/up_sampling2d_2/resize/ResizeNearestNeighbor/fq_input_0 supports only parameter levels > 1

Traceback (most recent call last):

File "/openvino_2021.1.110/python/python3.6/openvino/tools/benchmark/main.py", line 218, in run

exe_network = benchmark.load_network(ie_network)

File "/openvino_2021.1.110/python/python3.6/openvino/tools/benchmark/benchmark.py", line 73, in load_network

num_requests=1 if self.api_type == 'sync' else self.nireq or 0)

File "ie_api.pyx", line 311, in openvino.inference_engine.ie_api.IECore.load_network

File "ie_api.pyx", line 320, in openvino.inference_engine.ie_api.IECore.load_network

RuntimeError: Quantize layer StatefulPartitionedCall/functional_7/up_sampling2d_2/resize/ResizeNearestNeighbor/fq_input_0 supports only parameter levels > 1

I tried out the same procedure for publicaly available Yolov4 weight file, where I am able to run the benchmark app without any errors.

Re:INT8 optimization for YOLOv4

Adli — Wed, 09 Dec 2020 02:57:11 GMT

Hi Rahali,

Thank you for reaching out to us. The current release of the OpenVINO toolkit does not officially support the YOLOv4. We encourage our users to try and explore the YOLOv4 on the OpenVINO toolkit.

If possible, could you set a single image as the benchmark app input and then run the benchmark app? Please share the outcome here.

Regards,

Adli

Re: Re:INT8 optimization for YOLOv4

Rahila_T_Intel — Wed, 09 Dec 2020 04:33:38 GMT

Hi Adli,

There is no change with a single image as the benchmark app input.

Command : python3 /openvino_2021.1.110/deployment_tools/tools/benchmark_tool/benchmark_app.py -m /results1/TFYOLOV4_DefaultQuantization/2020-12-02_22-57-00/optimized//TFYOLOV4.xml -i /models/6dogs.jpg

Output :

[Step 1/11] Parsing and validating input arguments
/openvino_2021.1.110/python/python3.6/openvino/tools/benchmark/main.py:29: DeprecationWarning: The 'warn' method is deprecated, use 'warning' instead
logger.warn(" -nstreams default value is determined automatically for a device. "
[ WARNING ] -nstreams default value is determined automatically for a device. Although the automatic selection usually provides a reasonable performance, but it still may be non-optimal for some cases, for more information look at README.
[Step 2/11] Loading Inference Engine
[ INFO ] InferenceEngine:
API version............. 2.1.2021.1.0-1237-bece22ac675-releases/2021/1
[ INFO ] Device info
CPU
MKLDNNPlugin............ version 2.1
Build................... 2021.1.0-1237-bece22ac675-releases/2021/1

[Step 3/11] Setting device configuration
[ WARNING ] -nstreams default value is determined automatically for CPU device. Although the automatic selection usually provides a reasonable performance,but it still may be non-optimal for some cases, for more information look at README.
[Step 4/11] Reading network files
[ INFO ] Read network took 392.63 ms
[Step 5/11] Resizing network to match image sizes and given batch
[ INFO ] Network batch size: 1
[Step 6/11] Configuring input of the model
[Step 7/11] Loading the model to the device
[ ERROR ] Quantize layer StatefulPartitionedCall/functional_7/up_sampling2d_2/resize/ResizeNearestNeighbor/fq_input_0 supports only parameter levels > 1
Traceback (most recent call last):
File "/openvino_2021.1.110/python/python3.6/openvino/tools/benchmark/main.py", line 218, in run
exe_network = benchmark.load_network(ie_network)
File "/openvino_2021.1.110/python/python3.6/openvino/tools/benchmark/benchmark.py", line 73, in load_network
num_requests=1 if self.api_type == 'sync' else self.nireq or 0)
File "ie_api.pyx", line 311, in openvino.inference_engine.ie_api.IECore.load_network
File "ie_api.pyx", line 320, in openvino.inference_engine.ie_api.IECore.load_network
RuntimeError: Quantize layer StatefulPartitionedCall/functional_7/up_sampling2d_2/resize/ResizeNearestNeighbor/fq_input_0 supports only parameter levels > 1

Thanks,

Rahila

Re: Re:INT8 optimization for YOLOv4

Rahila_T_Intel — Wed, 09 Dec 2020 05:13:50 GMT

Hi Adli,

I have fixed the issue

I created a new int8 model by referring the below link.

https://community.intel.com/t5/Intel-Distribution-of-OpenVINO/How-to-enable-the-Post-Training-Optimization-Tool/m-p/1180921?profile.language=ko

Now I am able to run the bench mark app without any error.

Regards,

Rahila T

Re:INT8 optimization for YOLOv4

Adli — Wed, 09 Dec 2020 06:45:27 GMT

Hi Rahila,

This thread will no longer be monitored since this issue has been resolved. If you need any additional information from Intel, please submit a new question.

Regards,

Adli