CNN-LSTM model performs poor in OpenVINO environment when pre-processing is applied

[nat98]

Hello,

I have two deep learning models, that is a CNN model and a hybrid CNN-LSTM model, both are used for image classification tasks. Note that CNN has no pre-processing required and CNN-LSTM requires pre-processing.

TensorFlow environment: The training performance of the CNN model is about 98% accuracy; the CNN-LSTM model is about 96% accuracy with pre-processing applied.

TensorFlow environment: The inference performance of the CNN model is about 100% accuracy; the CNN-LSTM model is about 75% accuracy without pre-processing, and 100% accuracy with pre-processing applied.

OpenVINO environment: The inference performance of the CNN model is about 100% accuracy; the CNN-LSTM model is about 75% accuracy without pre-processing, and 6% accuracy with pre-processing applied.

The pre-processing is a simple code as shown below:

if(np.max(image)>1):
image = image/255.0

Why does the CNN-LSTM model accuracy in the OpenVINO environment drops significantly once the pre-processing is applied?

Regards,

nat98

[Wan_Intel]

Hi Nat98,

Thanks for reaching out to us.

Please share the following information with us for investigation purposes:

• Steps to set up TensorFlow environment
• Steps to set up OpenVINO environment
• CNN and CNN-LSTM model in TensorFlow and Intermediate Representation
• Model Optimizer command used to convert TensorFlow model into Intermediate Representation.
• Scripts to check accuracy for both TensorFlow and Intermediate Representation model

Regards,

Wan

[nat98 · Accepted solution]

Hi Wan,

Thanks for your reply.

I found out that the problem is the pre-processing code:

if(np.max(image)>1):
image = image/255.0

The pre-processing code is simply to make the input images in the range of 0 to 1 instead of 0 to 255. Not sure why I can't use it.

The solution is to implement this pre-processing code during conversion, which uses the "Use Scales" of the Model Optimizer command. (Screenshot attached).

Now I'm able to get 100% accuracy in the OpenVINO environment.

Regards,

nat98

[Wan_Intel]

Hi Nat98,

Thanks for sharing your solution in OpenVINO™ toolkit community!

This thread will no longer be monitored since the issue has been resolved.

If additional information is required, please submit a new question.

Regards,

Wan