Unable to run yolov3_tiny with different input_size

[cfdcfc]

Hi,

First of all, I have successfully implement openvino_tiny-yolov3_test.py.
Then, I would like to trade the accuracy for speed by reducing the input_size (416 -> 320), I have successfully achieve this in yolov3. However, I am not able to achieve this in yolov3_tiny, please see the below details:

get model
wget https://pjreddie.com/media/files/yolov3-tiny.weights

convert yolo to tensorflow
python3 /opt/OpenVINO-YoloV3/convert_weights_pb.py --class_names /opt/darknet/data/coco.names --weights_file /opt/darknet/yolov3-tiny.weights --data_format NHWC --output_graph /data/train/frozen_darknet_yolov3_tiny_model.pb --size 320 --tiny

convert tensorflow to IR
python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py --input_model /data/train/frozen_darknet_yolov3_tiny_model.pb --tensorflow_use_custom_operations_config /opt/intel/openvino/deployment_tools/model_optimizer/extensions/front/tf/yolo_v3_tiny.json --data_type FP16 --batch 1 --output_dir /data/train/

try to test with python(On Raspberry Pi Stretch with NCS2)

import os

try:
    from armv7l.openvino.inference_engine import IENetwork, IEPlugin
except:
    from openvino.inference_engine import IENetwork, IEPlugin
    
model_xml = "/home/pi/models/frozen_darknet_yolov3_tiny_model.xml"
model_bin = os.path.splitext(model_xml)[0] + ".bin"

plugin = IEPlugin(device="MYRIAD")
net = IENetwork(model=model_xml, weights=model_bin)
input_blob = next(iter(net.inputs))
exec_net = plugin.load(network=net)

Then error occurs:

Traceback (most recent call last):
File "yolov3_tiny_test.py", line 14, in
exec_net = plugin.load(network=net)
File "ie_api.pyx", line 395, in openvino.inference_engine.ie_api.IEPlugin.load
File "ie_api.pyx", line 406, in openvino.inference_engine.ie_api.IEPlugin.load
RuntimeError: [VPU] Internal error: Output in detector/yolo-v3-tiny/pool2_5/MaxPool has incorrect width dimension. Expected: 9 or 9 Actual: 10

[PINTO0309]

1. Install Tensorflow v1.12.0

$ sudo -H pip3 install tensorflow==1.12.0 --upgrade

2. Execute below.

$ python3 /opt/intel/openvino/deployment_tools/model_optimizer/mo_tf.py \
--input_model /data/train/frozen_darknet_yolov3_tiny_model.pb \
--tensorflow_use_custom_operations_config /opt/intel/openvino/deployment_tools/model_optimizer/extensions/front/tf/yolo_v3_tiny.json \
--data_type FP16 \
--batch 1 \
--output_dir /data/train/ \
--input_shape [1,320,320,3]

[cfdcfc]

Hi PINTO0309,

Thanks for the response, I followed your instruction but got this error:

[ ERROR ] Both --input_shape and --batch were provided. Please provide only one of them.
For more information please refer to Model Optimizer FAQ (<INSTALL_DIR>/deployment_tools/documentation/docs/MO_FAQ.html), question #56.

I removed the --batch 1 and successfully convert the tensorflow to IR, but eventually I end up with the same error like before

Traceback (most recent call last):
File "yolov3_tiny_test.py", line 14, in
exec_net = plugin.load(network=net)
File "ie_api.pyx", line 395, in openvino.inference_engine.ie_api.IEPlugin.load
File "ie_api.pyx", line 406, in openvino.inference_engine.ie_api.IEPlugin.load
RuntimeError: [VPU] Internal error: Output in detector/yolo-v3-tiny/pool2_5/MaxPool has incorrect width dimension. Expected: 9 or 9 Actual: 10

[PINTO0309]

Did you do it at 320 when training? I think it needs to be adjusted when training on Darknet.

[cfdcfc]

I did not train at 320.
However, I was able to apply the same trick on normal yolo v3.