Я следую примеру микро речи здесь:
https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/experimental/micro/examples/micro_speech
Поезд завершается с высокой (> 95%) валидацией и итоговой точностью теста. Замораживание завершается без каких-либо предупреждений или ошибок.
Toco сообщает о неподдерживаемом типе данных и неподдерживаемом AudioMicrofrontend op. Это проблема?
2019-07-10 13:56:39.687589: I tensorflow/lite/toco/import_tensorflow.cc:1336] Converting unsupported operation: DecodeWav
2019-07-10 13:56:39.694101: I tensorflow/lite/toco/import_tensorflow.cc:193] Unsupported data type in placeholder op: 5
2019-07-10 13:56:39.694142: I tensorflow/lite/toco/import_tensorflow.cc:1336] Converting unsupported operation: AudioMicrofrontend
2019-07-10 13:56:39.697231: I tensorflow/lite/toco/import_tensorflow.cc:1385] Unable to determine output type for op: AudioMicrofrontend
2019-07-10 13:56:39.698124: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] Before Removing unused ops: 44 operators, 70 arrays (0 quantized)
2019-07-10 13:56:39.698328: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] After Removing unused ops pass 1: 37 operators, 58 arrays (0 quantized)
2019-07-10 13:56:39.698557: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] Before general graph transformations: 37 operators, 58 arrays (0 quantized)
2019-07-10 13:56:39.698919: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 1: 10 operators, 18 arrays (1 quantized)
2019-07-10 13:56:39.699006: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 2: 9 operators, 17 arrays (1 quantized)
2019-07-10 13:56:39.699068: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] After general graph transformations pass 3: 8 operators, 15 arrays (1 quantized)
2019-07-10 13:56:39.699116: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] Before pre-quantization graph transformations: 8 operators, 15 arrays (1 quantized)
2019-07-10 13:56:39.699138: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] After pre-quantization graph transformations pass 1: 5 operators, 12 arrays (1 quantized)
2019-07-10 13:56:39.699162: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] Before Group bidirectional sequence lstm/rnn: 5 operators, 12 arrays (1 quantized)
2019-07-10 13:56:39.699184: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] Before quantization graph transformations: 5 operators, 12 arrays (1 quantized)
2019-07-10 13:56:39.699315: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] After quantization graph transformations pass 1: 4 operators, 11 arrays (11 quantized)
2019-07-10 13:56:39.699344: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] After quantization graph transformations pass 2: 4 operators, 11 arrays (11 quantized)
2019-07-10 13:56:39.699367: I tensorflow/lite/toco/graph_transformations/graph_transformations.cc:39] Before shuffling of FC weights: 4 operators, 11 arrays (11 quantized)
2019-07-10 13:56:39.699401: I tensorflow/lite/toco/allocate_transient_arrays.cc:345] Total transient array allocated size: 2496 bytes, theoretical optimal value: 2496 bytes.
2019-07-10 13:56:39.699422: I tensorflow/lite/toco/toco_tooling.cc:436] Estimated count of arithmetic ops: 0.000448371 billion (note that a multiply-add is counted as 2 ops).
Я подозреваю, что неподдерживаемая операция AudioMicrofrontend является проблемой, потому что развернутая модель имеет мрачную производительность по сравнению с точностью, сообщенной при обучении. Как я могу подтвердить, действительно ли это проблема?
Я использую последний код от tenorflow master с коммитами от Пита Уордена до 8 июля 2019 года.
https://github.com/tensorflow/tensorflow/commits?author=petewarden
Я пробовал модели: model_architecture = tiny_embedding_conv и model_architecture = tiny_conv
Оба дают одинаковый выходной ток.