[Bug] floordiv_3 (Floor) types check error: input #0 has type S32, but FP16 is expected

[LinBaoxin]

System information (version)

• OpenVINO=> 2022.1.0
• Operating System / Platform => Windows 10 (64 Bit)
• Compiler => pip install openvino-dev[tensorflow2,caffe]
• Problem classification => Model Conversion
• Framework: TensorFlow
• Model name: MoveNet: Ultra fast and accurate pose detection model

Detailed description

Hi there, When I tried running my IR model on NCS2, I saw the following error:

RuntimeError: StatefulPartitionedCall/floordiv_3 of type Floor: [ GENERAL_ERROR ] C:\j\workspace\private-ci\ie\build-windows-vs2019@3\b\repos\openvino\src\plugins\intel_myriad\graph_transformer\src\model\stage.cpp:305 Stage node StatefulPartitionedCall/floordiv_3 (Floor) types check error: input #0 has type S32, but FP16 is expected

There is a similar issue #7799.

Steps to reproduce

The model I used is downloaded from Tensorflow Hub. Tensorflow Document Page: MoveNet: Ultra fast and accurate pose detection model I converted the movenet_multipose_lightning_1.tar.gz SavedModel (TF 2. x) format to movenet_multipose_192x256.pb PB format (TF 1. x) format and then convert it to IR format.

Here is the command:

mo --input_model "movenet_multipose_192x256.pb" --input_shape "[1,192,256,3]" --data_type FP16

Here is the code:

from openvino.runtime import Core

import numpy as np
import time

ie = Core()
model = ie.read_model(model="movenet_multipose_192x256.xml", weights="movenet_multipose_192x256.bin")
compiled_model = ie.compile_model(model=model, device_name='CPU')
# compiled_model = ie.compile_model(model=model, device_name='MYRIAD')
input_key = compiled_model.input(0)
output_key = compiled_model.output(0)

image = np.random.randint(0, 256, [192, 256, 3]).astype(np.uint8)
image = np.expand_dims(image, axis=0)
tic = time.time()
result = compiled_model([image])[output_key]
toc = time.time()
print(toc - tic)

I can run it on the CPU successfully, but no NCS2.

[Iffa-Intel]

@LinBaoxin I can see similar behaviour from my end.

Conversion into FP16 IR

CPU infer (success)

NCS2 infer (failed)

We'll further investigate this and get back to you asap.

[hbalasu1]

Hi @LinBaoxin

Look like this is the bug since the issue happened on NCS2. I have escalated this issue to our development teams and will get back to you once I have the answer as soon as possible.

Regards, Hari Chand

Ref. 88871

[lsprior]

Hello, I too have exactly this problem. Can run the FP16 version on the CPU and GPU but not on myriad. Same output

RuntimeError: StatefulPartitionedCall/floordiv_3 of type Floor: [ GENERAL_ERROR ] C:\j\workspace\private-ci\ie\build-windows-vs2019@3\b\repos\openvino\src\plugins\intel_myriad\graph_transformer\src\model\stage.cpp:305 Stage node StatefulPartitionedCall/floordiv_3 (Floor) types check error: input #0 has type S32, but FP16 is expected

Any news on this issue?

Regards

[hbalasu1]

Hi @lsprior

I received a reply from the Developer that Floor operation is only supported on Myriad for FP16. Such movenet model is not supported by the Myriad plugin. A possible workaround could be to include a convert operation in the model so that it works with the Myriad plugin.

Hope this information help

Thank you

[avitial]

Closing this. Feel free to reopen and ask any questions related to this topic.