tensorflow-模型训练后tflite量化（tensorflow1.15)

 1.方式一： converter.post_training_quantize=True

def convert_model_qtv1():
    converter = tf.lite.TFLiteConverter.from_frozen_graph('model_convert/ckpt_new/densenet_ocr.pb', input_arrays=["Placeholder"],
                                                              input_shapes={"Placeholder": [1, 32, 280, 1]}, output_arrays=[
            "sequence_rnn_module/transpose_time_major"])
    converter.post_training_quantize=True
    tflite_model = converter.convert()
    open("model_convert/ckpt_new/densenet_ocr_metal_v1.tflite", "wb").write(tflite_model)

2.方式二：Dynamic range quantization

The simplest form of post-training quantization statically quantizes only the weights from floating point to integer, which has 8-bits of precision:

def convert_model_qtv1():
    converter = tf.lite.TFLiteConverter.from_frozen_graph('model_convert/ckpt_new/densenet_ocr.pb', input_arrays=["Placeholder"],
                                                              input_shapes={"Placeholder": [1, 32, 280, 1]}, output_arrays=[
            "sequence_rnn_module/transpose_time_major"])
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    tflite_model = converter.convert()
    open("model_convert/ckpt_new/densenet_ocr_metal-DY_v1.tflite", "wb").write(tflite_model)

3.方式三：Full integer quantization

Integer with float fallback (using default float input/output):

def representative_dataset():
    for _ in range(100):
      #这里一定记得要修改data的输入shape与模型的输入shape吻合
      data = np.random.rand(1, 32, 280, 1)
      yield [data.astype(np.float32)]

def convert_model_qtv1():
    converter = tf.lite.TFLiteConverter.from_frozen_graph('model_convert/ckpt_new/densenet_ocr.pb', input_arrays=["Placeholder"],
                                                              input_shapes={"Placeholder": [1, 32, 280, 1]}, output_arrays=[
            "sequence_rnn_module/transpose_time_major"])
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    converter.representative_dataset = representative_dataset
    tflite_model = converter.convert()
    open("model_convert/ckpt_new/densenet_ocr_metal-int_v1.tflite", "wb").write(tflite_model)

4.方式四：Float16 quantization

To enable float16 quantization of weights, use the following steps:

def convert_model_qtv1():
    converter = tf.lite.TFLiteConverter.from_frozen_graph('model_convert/ckpt_new/densenet_ocr.pb', input_arrays=["Placeholder"],
                                                              input_shapes={"Placeholder": [1, 32, 280, 1]}, output_arrays=[
            "sequence_rnn_module/transpose_time_major"])
    converter.optimizations = [tf.lite.Optimize.DEFAULT]
    converter.target_spec.supported_types = [tf.float16]
    tflite_model = converter.convert()
    open("model_convert/ckpt_new/densenet_ocr_metal-fl16_v1.tflite", "wb").write(tflite_model)

转换之后的tflite大小对比：

 ---没量化前

 ---方式一

 ---方式二

---方式三

---方式四