sayakpaul / adventures-in-tensorflow-lite Goto Github PK

Hi Mr. @sayakpaul
Thank you for the example convert to TFLite model.
I have successfully converted the pretrained models of DeepLabv3 to TFLite model. I have this suggestion, if you use output_arrays=['ArgMax'] instead of ['ResizeBilinear_2'] the output_shape of the TFLite model is 513 instead of 65.

I ran the local_test.sh file from the original repo, and frozen_inference_graph.pb cannot be converted to TFLite by your method (The frozen_inference_graph.pb file has run successfully with deeplab_demo.ipynb). When converting to TFLite, an error occurs: "ValueError: Invalid tensors 'sub_7' were found."

Do you have any advice? Thank you.

Hello.
Thank you for your great deeplap usage notebooks.
I would like to know if there is anyway I can color each instance separately?

Thank you in advances.
Bill

Hi @sayakpaul

Thanks for the detailed conversion examples.

I have trained a PyTorch model, which I converted to Keras using the pytorch2keras lib.

using this Keras model to convert to tflite. I want to run the tflite on coral devices.

things which i noticed :

• Keras model size (57.6MB)
• Using dynamic range quantization the generated tflite is of size(15MB)
• Using integer only quantization the generated tflite is also of size(15MB)

Ideally, we should be able to reduce the model size even further as we convert fp32 to int8

Could you point me out, if i am missing anything here?

shared my conversion notebook and Keras modelkeras_model

data

Sorry, but I have taken a look at your example. It is good to see some complete examples like this. However, the title somehow sometimes seems misleading when the 8bit integer model is not actually 8 bit. When you let the optimizer default, it means some layers will not be quantized.

Hi there,
I've checked the code and found that the output image is square photo, does it support rectangle photos like 1024x768?

Hi Paul!

Have a question for MobileDet_Conversion_TFLite.ipynb, Quantize and serialize.

Using model: model_checkpoint_name = "ssd_mobiledet_cpu_coco"

Is my understanding correct: using
original_precision = "regular"
quantization_strategy = "int8"

We expect that the converted model is a quantized model, with input as uint8 (converter.inference_input_type = tf.uint8).
However, the resulting model input is not quantized:
input_type = interpreter.get_input_details()[0]['dtype']
print('input_type: ' + str(input_type))
input_type: <class 'numpy.float32'>

Moreover, checking the resulting tflite file using netron.app shows that the model is missing the Quantize step after normalized_input_image_tensor.

Is quantization_strategy=="int8" missing converter.representative_dataset = representative_dataset_gen?