wcfzl

> 我从https://github.com/tensorflow/tensorflow/blob/r1.12/tensorflow/contrib/lite/g3doc/models.md > 下载了Mobilenet_v2_1.0_224_quant，并用mace master转换成功。 > 你用的哪个模型，mace是哪个版本？ mace是latest版本，tensorflow1.14。我又试了一下，tensorflow官方的模型可以正常转换，但我自己训练的模型mobilenetv2+deeplabv3+ 无法转换（可以正常转tflite） 报错如下 `library_name: mobilenet-v2-quantize-retrain target_abis: [arm64-v8a,armeabi-v7a] model_graph_format: file model_data_format: file models: mobilenet_v2_quantize_retrain: platform: tensorflow model_file_path: ./model.pb model_sha256_checksum: c90e0a6ac00ef6969723bb90d27068df2734792e4eb6c65c24a9a2ee3d693042 subgraphs: - input_tensors: -...

@lee-bin 我打印了出错的地方 `File "/home/wcf/mace-master/bazel-bin/mace/python/tools/converter.runfiles/mace/mace/python/tools/quantization/quantize_util.py", line 143, in quantize output = np.clip((np.round(zero + data / scale).astype(np.int32)), 0, 255)` 发现data scale 都是全0,似乎是没有读到相应的值

> 可以提供一个能复现该问题的模型吗 https://github.com/wcfzl/deeplab-mobilenetv2-Quantization-aware-training-model

I have the same problem,that cause the net doesn't work. "ValueError: Shape of a new variable (conv6_1/weights) must be fully defined, but instead was (3, 3, ?, 512)." do you...