TensorFlow 重新训练的 inception v3 模型在 Android 上崩溃

Question

我已经使用我自己的数据集重新训练了 TensorFlow Inception v3 模型，如 tutorial 中所述。

现在我正在尝试使用我重新训练的模型构建和运行 TensorFlow Android example。我按原样从示例中构建了本机代码，将模型 (.pb) 和标签 (.txt) 文件复制到资产目录并更改了 TensorFlowImageListener.java 中的模型参数：

private static final int NUM_CLASSES = 5; // number of categories
private static final int INPUT_SIZE = 299;
private static final int IMAGE_MEAN = 128;
private static final float IMAGE_STD = 128;
private static final String INPUT_NAME = "Mul:0";
private static final String OUTPUT_NAME = "final_result:0";

但应用在解析从资产加载的模型文件时崩溃：

08-12 16:02:08.258 25253-25253/com.iliakplv.tensorflow I/native: tensorflow_jni.cc:115 Loading TensorFlow.
08-12 16:02:08.258 25253-25253/com.iliakplv.tensorflow I/native: tensorflow_jni.cc:117 Making new SessionOptions.
08-12 16:02:08.259 25253-25253/com.iliakplv.tensorflow I/native: tensorflow_jni.cc:120 Got config, 0 devices
08-12 16:02:08.264 25253-25253/com.iliakplv.tensorflow I/native: tensorflow_jni.cc:123 Session created.
08-12 16:02:08.264 25253-25253/com.iliakplv.tensorflow I/native: tensorflow_jni.cc:126 Graph created.
08-12 16:02:08.265 25253-25253/com.iliakplv.tensorflow I/native: tensorflow_jni.cc:130 Acquired AssetManager.
08-12 16:02:08.265 25253-25253/com.iliakplv.tensorflow I/native: tensorflow_jni.cc:132 Reading file to proto: file:///android_asset/tensorflow_inception_graph.pb
08-12 16:02:08.265 25253-25253/com.iliakplv.tensorflow I/native: jni_utils.cc:120 Opening asset tensorflow_inception_graph.pb from disk with copy.
08-12 16:02:09.382 25253-25253/com.iliakplv.tensorflow A/native: jni_utils.cc:123 Check failed: message->ParseFromArray(memory, data_size) 
08-12 16:02:09.382 25253-25253/com.iliakplv.tensorflow A/libc: Fatal signal 6 (SIGABRT), code -6 in tid 25253 (tech.tensorflow)

附：我也尝试过像这样使用tensorflow/python/tools/strip_unused.py（正如它所建议的here）：

bazel build tensorflow/python/tools:strip_unused && \
bazel-bin/tensorflow/python/tools/strip_unused \
--input_graph=some_graph_def.pb \
--output_graph=/tmp/stripped_graph.pb \
--input_node_names=Mul
--output_node_names=final_result

没有帮助。

Answer 1

问题在于 gradle 压缩资产文件导致模型解析失败。我禁用了 .pb 文件的压缩：

aaptOptions { noCompress 'pb' }