我对所有这些工具都不熟悉。我正试图开始使用Tensorflow Lite，最终在Coral Edge TPU上运行我自己的深度学习模型。
我用Keras API构建了一个玩具XOR网络，写出了tensorflow图，并将其冻结。现在我试图使用TOCO将冻结的模型转换为tflite格式。我得到以下错误：
值错误：节点dense_1/weights_quant/AssignMinLast的输入0从dense_1/weights_quant/min传递浮点值：0与预期的浮点值不兼容。
我看到其他人在github上谈论类似的错误，但是我没有找到解决方案。
完整代码如下：
training_data = np.array([[0,0],[0,1],[1,0],[1,1]], "uint8")
target_data = np.array([[0],[1],[1],[0]], "uint8")
model = Sequential()
model.add(Dense(16, input_dim=2, use_bias=False, activation='relu'))
model.add(Dense(1, use_bias=False, activation='sigmoid'))
session = tf.keras.backend.get_session()
tf.contrib.quantize.create_training_graph(session.graph)
session.run(tf.global_variables_initializer())
model.compile(loss='mean_squared_error',
optimizer='adam',
metrics=['binary_accuracy'])
model.fit(training_data, target_data, nb_epoch=1000, verbose=2)
print model.predict(training_data).round()
model.summary()
saver = tf.train.Saver()
saver.save(keras.backend.get_session(), 'xor-keras.ckpt')
tf.io.write_graph(session.graph, '.', 'xor-keras.pb')
然后冻结模型：
python freeze_graph.py \
--input_graph='xor-keras.pb' \
--input_checkpoint='xor-keras.ckpt' \
--output_graph='xor-keras-frozen.pb' \
--output_node_name='dense_2/Sigmoid'
然后像这样给toco打电话：
toco \
--graph_def_file=xor-keras-frozen.pb \
--output_file=xor-keras-frozen.tflite \
--input_shapes=1,2 \
--input_arrays='dense_1_input' \
--output_arrays='dense_2/Sigmoid' \
--inference_type=QUANTIZED_UINT8
以下是TOCO的完整输出：
2019-06-26 15:31:17.374904: I tensorflow/core/platform/cpu_feature_guard.cc:142] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 AVX512F FMA
2019-06-26 15:31:17.404237: I tensorflow/core/platform/profile_utils/cpu_utils.cc:94] CPU Frequency: 2600000000 Hz
2019-06-26 15:31:17.407613: I tensorflow/compiler/xla/service/service.cc:168] XLA service 0x55bbcf9a5ed0 executing computations on platform Host. Devices:
2019-06-26 15:31:17.407741: I tensorflow/compiler/xla/service/service.cc:175] StreamExecutor device (0): <undefined>, <undefined>
Traceback (most recent call last):
File "/home/redacted/.local/bin/toco", line 11, in <module>
sys.exit(main())
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/lite/python/tflite_convert.py", line 503, in main
app.run(main=run_main, argv=sys.argv[:1])
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/python/platform/app.py", line 40, in run
_run(main=main, argv=argv, flags_parser=_parse_flags_tolerate_undef)
File "/home/redacted/.local/lib/python2.7/site-packages/absl/app.py", line 300, in run
_run_main(main, args)
File "/home/redacted/.local/lib/python2.7/site-packages/absl/app.py", line 251, in _run_main
sys.exit(main(argv))
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/lite/python/tflite_convert.py", line 499, in run_main
_convert_tf1_model(tflite_flags)
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/lite/python/tflite_convert.py", line 124, in _convert_tf1_model
converter = _get_toco_converter(flags)
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/lite/python/tflite_convert.py", line 111, in _get_toco_converter
return converter_fn(**converter_kwargs)
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/lite/python/lite.py", line 628, in from_frozen_graph
_import_graph_def(graph_def, name="")
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/python/util/deprecation.py", line 507, in new_func
return func(*args, **kwargs)
File "/home/redacted/.local/lib/python2.7/site-packages/tensorflow/python/framework/importer.py", line 431, in import_graph_def
raise ValueError(str(e))
ValueError: Input 0 of node dense_1/weights_quant/AssignMinLast was passed float from dense_1/weights_quant/min:0 incompatible with expected float_ref.
最佳答案
我自己解决了问题。原来“训练图”不能转换为TFLite，但“eval图”是。从训练会话中保存图形会产生错误的输入。
在我看来，freeze_graph脚本应该足够聪明来处理这个问题，但唉，它不是。
下面是生成TOCO正确输入的代码。
# <Load the model into a new session>
session = tf.keras.backend.get_session()
saver = tf.train.Saver()
saver.restore(session, 'xor-keras.ckpt')
tf.contrib.quantize.create_eval_graph(session.graph)
tf.io.write_graph(session.graph, '.', 'xor-keras-eval.pb', as_text=False)