Question

我试图让Mobilenetv2模型（将最后一层重新训练到我的数据中）运行在Google边缘TPU珊瑚上。

我遵循这个教程https://www.tensorflow.org/lite/performance/post_training_quantization?hl=en进行训练后量化。相关代码为：

...
train = tf.convert_to_tensor(np.array(train, dtype='float32'))
my_ds = tf.data.Dataset.from_tensor_slices(train).batch(1)


# POST TRAINING QUANTIZATION
def representative_dataset_gen():
    for input_value in my_ds.take(30):
        yield [input_value]

converter = tf.lite.TFLiteConverter.from_keras_model_file(saved_model_dir)
converter.optimizations = [tf.lite.Optimize.DEFAULT]
converter.representative_dataset = representative_dataset_gen
converter.target_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
tflite_quant_model = converter.convert()

我已经成功生成了tflite量化模型，但是当我运行edgetpu_compiler时（此页面https://coral.withgoogle.com/docs/edgetpu/compiler/#usage之后），我得到了以下输出：

edgetpu_compiler  Notebooks/MobileNetv2_3class_visit_split_best-val- 
acc.h5.quant.tflite

Edge TPU Compiler version 2.0.258810407
INFO: Initialized TensorFlow Lite runtime.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
ERROR: quantized_dimension must be in range [0, 1). Was 3.
Invalid model: Notebooks/MobileNetv2_3class_visit_split_best-val-        
acc.h5.quant.tflite
Model could not be parsed

我对量化过程还很陌生，所以可能缺少一些简单的东西。有什么想法吗？

模型的输入形状是3通道rgb图像。是否需要仅1通道图像作为输入？可以对3个通道的图像进行全整数量化吗？

Answer 1

我有相同的问题和相同的错误消息。我使用tensorflow.keras.applications mobilenetv2重新训练了MobilenetV2。我发现我的模型与Coral的示例模型（https://coral.withgoogle.com/models/）之间的TFLite张量存在较大差异。

首先，输入和输出的类型不同。当我将tf.keras模型转换为tflite时，它包含浮点类型的输入和输出张量，而示例模型具有整数类型。如果我使用tensorflow-lite（https://www.tensorflow.org/lite/convert/）的命令行转换和python转换，则这是不同的。命令行转换输出整数类型io，而python转换输出浮点类型io。（这真的很奇怪。）

第二，示例模型中没有批处理标准化（BN）层，但是Keras MobilenetV2中有一些BN。我认为'ERROR：quantized_dimension'的数量必须在[0，1）范围内。是3。与BN的数量有关，因为Keras模型中有17个BN层。

我仍在努力解决这个问题。我将按照Coral的再培训示例进行解决。（https://coral.withgoogle.com/docs/edgetpu/retrain-detection/）

Answer 2

此问题已在tensorflow1.15-rc中修复。在新的tf版本中将模型转换为TFLite。然后TFLite模型将在TPU编译器中工作。

并放置这些行，使TFlite模型的输入和输出成为uint8类型。（我认为应该是tf.int8。）

converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
converter.inference_input_type = tf.uint8
converter.inference_output_type = tf.uint8

检查下面的链接。 https://www.tensorflow.org/lite/performance/post_training_quantization

Answer 3

我有类似的错误，使用tf-nightly build 1.15和使用.tflite文件进行后期训练全整数量化，然后使用应该可以工作的edge TPU编译器进行编译。这种方法解决了我的错误。

在github中提出了相同的问题，您可以看到-here

Answer 4

更新到最新的编译器版本后，您还会遇到此问题吗？

Edge TPU Compiler version 2.0.267685300

Edge TPU编译器|错误：quantized_dimension必须在[0，1）范围内。是3

4 个答案: