Input shape changed after converting model to tflite

[VanBap]

Input shape changed after converting model to tflite

[doxuanhieu185]

Hey friend, please wait, I believe good things will come to you.

[pkgoogle]

Hi @VanBap, in your notebook you do not specify a specific input_shape in the API: https://keras.io/api/applications/resnet/#resnet50-function, Resizing the converted input shape is actually expected: https://www.tensorflow.org/lite/guide/inference#run_inference_with_dynamic_shape_model, so given your parameters, I believe resizing is the correct action. Does that answer your concerns? Thanks.

[VanBap]

anyone? :(

[pkgoogle]

Hi VanBap, can you upload your .h5 and .tflite file? That will help me test this faster.

[VanBap]

Hi VanBap, can you upload your .h5 and .tflite file? That will help me test this faster.

I cant upload model files here because they are >25MB even though I zipped those.
But they are on my Github:

• model .tflite: https://github.com/VanBap/Adverse_Weather/tree/master/saved_model
• training model .h5: https://github.com/VanBap/Adverse_Weather/blob/master/Detection_model_usingResNet50.ipynb

By the way when I try to run model .tflite, I just put a new line code:
" # Resize Tensor Shape
interpreter.resize_tensor_input(input_details[0]['index'], (1, 224, 224, 3)) "
then it finally ran but I dont really know (how) does it affect the way my model .tflite processing ?!

[VanBap]

Hi @VanBap, in your notebook you do not specify a specific input_shape in the API: https://keras.io/api/applications/resnet/#resnet50-function, Resizing the converted input shape is actually expected: https://www.tensorflow.org/lite/guide/inference#run_inference_with_dynamic_shape_model, so given your parameters, I believe resizing is the correct action. Does that answer your concerns? Thanks.

I really appreciate your help! I am trying your answer and will feed back when its done

[VanBap]

Hi @VanBap, in your notebook you do not specify a specific input_shape in the API: https://keras.io/api/applications/resnet/#resnet50-function, Resizing the converted input shape is actually expected: https://www.tensorflow.org/lite/guide/inference#run_inference_with_dynamic_shape_model, so given your parameters, I believe resizing is the correct action. Does that answer your concerns? Thanks.

Hi pkgoogle, I just want to feedback that I have tried your answer and my model ran properly. Thank u so much. I really appreciate it!

[pkgoogle]

Hi @VanBap, no worries, if you have no more open items, please feel free to close as completed. Thanks.