如何将 Lenet model h5 转换为 .tflite

Question

How do I correctly convert Lenet Model (input 32x32, 5 layers, 10 classes) to Tensorflow Lite? I used this lines of codes but it gives me really bad confidences in android, like this image . The confidences are all around 0.1, or 10%.

This is the code I used

model = tf.keras.models.load_model('model.h5')
converter = tf.lite.TFLiteConverter.from_keras_model(model)
converter.post_training_quantize = True
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

The.h5 file above can predict images with good confidences and accuracy, like this image . Or should I ask, does Tensorflow Lite not support custom model (Lenet)? why is tflite file does so much worse than.h5?

Answer 1

If.tflite file is generated with no mistakes it doesn't matter if the model is called Lenet or anything else. Also quantization will have a small decrease in accuracy but no major difference like you are stating. I would see how u are making bytebuffer to insert it inside interpreter. If u are using gray scale images u have to divide with 3/255... for colored images is only /255. If during your training u haven't used pixel normalization then do not use /255 during bitmap to bytebuffer. So your code would be like:

private ByteBuffer convertBitmapToByteBuffer(Bitmap bitmap) {
    ByteBuffer byteBuffer = ByteBuffer.allocateDirect(ModelConfig.MODEL_INPUT_SIZE);
    byteBuffer.order(ByteOrder.nativeOrder());
    int[] pixels = new int[ModelConfig.INPUT_WIDTH * ModelConfig.INPUT_HEIGHT];
    bitmap.getPixels(pixels, 0, bitmap.getWidth(), 0, 0, bitmap.getWidth(), bitmap.getHeight());
    for (int pixel : pixels) {
        float rChannel = (pixel >> 16) & 0xFF;
        float gChannel = (pixel >> 8) & 0xFF;
        float bChannel = (pixel) & 0xFF;
        float pixelValue = (rChannel + gChannel + bChannel);
        byteBuffer.putFloat(pixelValue);
    }
    return byteBuffer;
}

and not:

private ByteBuffer convertBitmapToByteBuffer(Bitmap bitmap) {
    ByteBuffer byteBuffer = ByteBuffer.allocateDirect(ModelConfig.MODEL_INPUT_SIZE);
    byteBuffer.order(ByteOrder.nativeOrder());
    int[] pixels = new int[ModelConfig.INPUT_WIDTH * ModelConfig.INPUT_HEIGHT];
    bitmap.getPixels(pixels, 0, bitmap.getWidth(), 0, 0, bitmap.getWidth(), bitmap.getHeight());
    for (int pixel : pixels) {
        float rChannel = (pixel >> 16) & 0xFF;
        float gChannel = (pixel >> 8) & 0xFF;
        float bChannel = (pixel) & 0xFF;
        float pixelValue = (rChannel + gChannel + bChannel) / 255.f;
        byteBuffer.putFloat(pixelValue);
    }
    return byteBuffer;
}

Answer 2

Its because of quantization.
It reduces the size of the model so does the accuracy. Try not to quantizatise the model.
Try this.

import tensorflow as tf

model = tf.keras.models.load_model('model.h5')
converter = tf.lite.TFLiteConverter.from_keras_model(model)
tflite_model = converter.convert()
open("converted_model.tflite", "wb").write(tflite_model)

It might increase the size the tflite model but won't degrade the accuracy to that extent.