when i caffe convert tensorflow,the error accur

[la735681253]

Platform (like ubuntu 16.04/win10): 16.04

Python version: 3.6

Source framework with version (like Tensorflow 1.4.1 with GPU):caffe 1.0

Destination framework with version (like CNTK 2.3 with GPU): tensorflow 1.4

Pre-trained model path (webpath or webdisk path):

Running scripts: mmconvert -sf caffe -in deploy_vgg16_places365.prototxt -iw vgg16_places365.caffemodel -df tensorflow -om places365_tf.npy

Traceback (most recent call last): File "/home/liang/anaconda3/envs/caffe/lib/python3.6/site-packages/mmdnn/conversion/common/DataStructure/emitter.py", line 36, in _load_weights self.weights_dict = np.load(file_name).item() File "/home/liang/anaconda3/lib/python3.6/site-packages/numpy/lib/npyio.py", line 447, in load pickle_kwargs=pickle_kwargs) File "/home/liang/anaconda3/lib/python3.6/site-packages/numpy/lib/format.py", line 696, in read_array raise ValueError("Object arrays cannot be loaded when " ValueError: Object arrays cannot be loaded when allow_pickle=False

During handling of the above exception, another exception occurred:

Traceback (most recent call last): File "/home/liang/anaconda3/envs/caffe/bin/mmconvert", line 11, in sys.exit(_main()) File "/home/liang/anaconda3/envs/caffe/lib/python3.6/site-packages/mmdnn/conversion/_script/convert.py", line 108, in _main ret = IRToCode._convert(code_args) File "/home/liang/anaconda3/envs/caffe/lib/python3.6/site-packages/mmdnn/conversion/_script/IRToCode.py", line 26, in _convert emitter = TensorflowEmitter((args.IRModelPath, args.IRWeightPath)) File "/home/liang/anaconda3/envs/caffe/lib/python3.6/site-packages/mmdnn/conversion/tensorflow/tensorflow_emitter.py", line 66, in init self._load_weights(model[1]) File "/home/liang/anaconda3/envs/caffe/lib/python3.6/site-packages/mmdnn/conversion/common/DataStructure/emitter.py", line 38, in _load_weights self.weights_dict = np.load(file_name, encoding='bytes').item() File "/home/liang/anaconda3/lib/python3.6/site-packages/numpy/lib/npyio.py", line 447, in load pickle_kwargs=pickle_kwargs) File "/home/liang/anaconda3/lib/python3.6/site-packages/numpy/lib/format.py", line 696, in read_array raise ValueError("Object arrays cannot be loaded when " ValueError: Object arrays cannot be loaded when allow_pickle=False

[juliagusak]

Hi @la735681253 , @rainLiuplus

the cause of this problem is overviewed here. In your case, if you did nothing with numpy version

Firstly, navigate to "/home/liang/anaconda3/envs/caffe/lib/python3.6/site-packages/mmdnn/conversion/common/DataStructure/emitter.py" and replace line 36 self.weights_dict = np.load(file_name).item() with self.weights_dict = np.load(file_name, allow_pickle = True).item()

Secondly, perform 3-step conversion instead 1-step:

Let IR_DIR will be a repo to store IR representation of the model and CONVERTED_MODEL_DIR - a repo to store TensorFlow model. MODEL_NAME is a name you use for the model ( my_vgg16, for example)

• Convert the model from Caffe to IR format mkdir -p $IR_DIR mmtoir -f caffe -n $PATH_PROTOTXT -w $PATH_CAFFEMODEL --inputShape $INPUT_SHAPE -d $IR_DIR/$MODEL_NAME

• Convert the model from IR to TensorFlow code snippet mkdir -p $CONVERTED_MODEL_DIR mmtocode -f tensorflow -n "$IR_DIR/$MODEL_NAME.pb" -w "$IR_DIR/$MODEL_NAME.npy" -d "$CONVERTED_MODEL_DIR/$MODEL_NAME.py"

  After this step you probably will get the similar error Object arrays cannot be loaded when allow_pickle=False . Now it happens because of np.load(file_name).item() in generated "$CONVERTED_MODEL_DIR/$MODEL_NAME.py" file. Navigate to the error string and replace it with np.load(file_name, allow_pickle = True).item()

• Convert the model from IR to TensorFlow model mmtomodel -f tensorflow -in "$CONVERTED_MODEL_DIR/$MODEL_NAME.py" -iw "$IR_DIR/$MODEL_NAME.npy" -o "$CONVERTED_MODEL_DIR/$MODEL_NAME.ckpt" --dump_tag SERVING

P.S.: you can use --dump_tag TRAINING instead --dump_tag SERVING

[rainLiuplus]

@juliagusak, Thank you very much for the detailed explanation!

[bersbersbers]

I suggest to rename this issue to "Object arrays cannot be loaded when allow_pickle=False", which will occur for many people who upgraded to numpy>=1.16.3 (where the default allow_pickle=True has been changed to allow_pickle=False). By the way, @rainLiuplus's commit above just fails in two Python 2.7 tests - if that could be fixed, that issue would be gone. Until then, installing numpy==1.16.2 is a workaround.