2

我尝试按照tensorflow 量化中的说明生成量化的 tensorflow lite 模型。

首先，我在训练过程中使用 tf.contrib.quantize.create_training_graph() 和 tf.contrib.quantize.create_eval_graph() 将假量化节点插入到图中，最后生成一个冻结的 pb 文件（model.pb）。

其次，我使用以下命令将我的假量化 tensorflow 模型转换为量化 tensorflow lite 模型。

bazel-bin/tensorflow/contrib/lite/toco/toco \
--input_file=model.pb \
--input_format=TENSORFLOW_GRAPHDEF \
--output_format=TFLITE \
--output_file=model.tflite \
--inference_type=QUANTIZED_UINT8 --input_shapes=1,1:1,5002 \
--input_arrays=Test/Model/input,Test/Model/apps \
--output_arrays=Test/Model/output_probs,Test/Model/final_state  \
--mean_values=127.5,127.5 --std_values=127.5,127.5 --allow_custom_ops

隐蔽过程失败，日志如下：

2018-03-28 18:00:38.348403: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before Removing unused ops: 118 operators, 193 arrays (0 quantized)
2018-03-28 18:00:38.349394: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before general graph transformations: 118 operators, 193 arrays (0 quantized)
2018-03-28 18:00:38.382854: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 1: 57 operators, 103 arrays (1 quantized)
2018-03-28 18:00:38.384327: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 2: 56 operators, 101 arrays (1 quantized)
2018-03-28 18:00:38.385235: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 3: 55 operators, 100 arrays (1 quantized)
2018-03-28 18:00:38.385995: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] Before pre-quantization graph transformations: 55 operators, 100 arrays (1 quantized)
2018-03-28 18:00:38.386047: W tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc:131] Skipping min-max setting for {TensorFlowSplit operator with output Test/Model/RNN/RNN/multi_rnn_cell/cell_0/basic_lstm_cell/split} because output Test/Model/RNN/RNN/multi_rnn_cell/cell_0/basic_lstm_cell/split already has min-max.
2018-03-28 18:00:38.386076: W tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc:131] Skipping min-max setting for {TensorFlowSplit operator with output Test/Model/RNN/RNN/multi_rnn_cell/cell_1/basic_lstm_cell/split} because output Test/Model/RNN/RNN/multi_rnn_cell/cell_1/basic_lstm_cell/split already has min-max.
2018-03-28 18:00:38.386328: I tensorflow/contrib/lite/toco/graph_transformations/graph_transformations.cc:39] After pre-quantization graph transformations pass 1: 48 operators, 93 arrays (1 quantized)
2018-03-28 18:00:38.386484: W tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc:131] Skipping min-max setting for {TensorFlowSplit operator with output Test/Model/RNN/RNN/multi_rnn_cell/cell_1/basic_lstm_cell/split} because output Test/Model/RNN/RNN/multi_rnn_cell/cell_1/basic_lstm_cell/split already has min-max.
2018-03-28 18:00:38.386502: W tensorflow/contrib/lite/toco/graph_transformations/hardcode_min_max.cc:131] Skipping min-max setting for {TensorFlowSplit operator with output Test/Model/RNN/RNN/multi_rnn_cell/cell_0/basic_lstm_cell/split} because output Test/Model/RNN/RNN/multi_rnn_cell/cell_0/basic_lstm_cell/split already has min-max.
2018-03-28 18:00:38.386778: F tensorflow/contrib/lite/toco/tooling_util.cc:1432] Array Test/Model/embedding_lookup, which is an input to the TensorFlowReshape operator producing the output array Test/Model/Reshape_1, is lacking min/max data, which is necessary for quantization. Either target a non-quantized output format, or change the input graph to contain min/max information, or pass --default_ranges_min= and --default_ranges_max= if you do not care about the accuracy of results.
Aborted

有什么问题，我错在哪里？

4

1 回答 1

1

你没有做错什么。

目前 create_training_graph 和 create_eval_graph 在各种模型架构中并不是最健壮的。我们让他们在大多数 CNN 上工作，但 RNN 仍在进行中，并带来了一系列不同的挑战。

根据 RNN 的细节，目前的量化方法将涉及更多，并且可能需要手动将 FakeQuantization 操作放在正确的位置。特别是在您的错误消息中，您似乎需要在 embedding_lookup 中添加一个 FakeQuantization 操作。话虽如此，最终量化的 RNN 可能会运行，但我不知道准确度会如何。它真的最终依赖于模型和数据集:)

当自动重写正确支持 RNN 时，我将更新此答案。

于 2018-03-29T02:10:52.057 回答