Forums - dlc convert

Please also provide the error messages produced by the command above.

Hi Enrico,

I have also tried a model trained with code: https://github.com/mrharicot/monodepth

When converting checkpoint or frozed model to dlc model, the convert tool is just hang in session run:

snpe-1.17.0/lib/python/converters/tensorflow/util.py:

Quote:

232         if len(requiring_evaluation) > 0:
233             try:
234                 outputs = self._session.run(fetches=requiring_evaluation, feed_dict=input_tensors)
235                 outputs = dict(zip(requiring_evaluation, outputs))
236                 for t, o in outputs.iteritems():
237                     self._tensor_value_cache[t.name] = o
238                 outputs_map.update(outputs)
239                 requiring_evaluation = []
240             except InvalidArgumentError:
241                 pass

So why would it hang in session run?

My convert command line is :

Quote:

./snpe-tensorflow-to-dlc --graph /home/damon/work/monodepth/my_model_resnet/frozen_model.pb -i split 1,256,512,3 --out_node model/decoder/mul_21 --dlc ~/work/monodepth/my_model_resnet/frozen_model.dlc

Any comments would be appreciated! Thanks.

I'm trying to do the same, converting a tensorflow ResNet-50_v1 model to dlc and facing errors . Model itself is not getting loaded . Can someone  suggest how to fix this.

Im using SNPE 1.19.2 and tensorflow 1.6 ( since the documentation says that SNPE is tested with this version of tensorflow).

Upgrading tensorflow version to 1.11 didnt help either.

snpe-tensorflow-to-dlc --graph /home/test/Desktop/ResNet-50_v1/Frozen_graphs/resnet50_v1.pb --input_dim "input" 1,224,224,3 --out_node resnet_v1_50/SpatialSqueeze --dlc /home/test/snpe-sdk/models/tf_resnet50v1/ --allow_unconsumed_nodes --verbose

ERROR

2018-10-08 18:46:28.372973: I tensorflow/core/platform/cpu_feature_guard.cc:140] Your CPU supports instructions that this TensorFlow binary was not compiled to use: AVX2 FMA
2018-10-08 18:46:28,374 - 109 - ERROR - Encountered Error: Error parsing message
Traceback (most recent call last):
  File "/home/test/snpe-sdk/bin/x86_64-linux-clang/snpe-tensorflow-to-dlc", line 99, in main
    model = loader.load(args.graph, in_nodes, in_dims, args.in_type, args.out_node, session)
  File "/home/test/snpe-sdk/lib/python/converters/tensorflow/loader.py", line 50, in load
    graph_def = self.__import_graph(graph_pb_or_meta_path, session, out_node_names)
  File "/home/test/snpe-sdk/lib/python/converters/tensorflow/loader.py", line 102, in __import_graph
    graph_def = cls.__import_from_frozen_graph(graph_path)
  File "/home/test/snpe-sdk/lib/python/converters/tensorflow/loader.py", line 115, in __import_from_frozen_graph
    graph_def.ParseFromString(f.read())
DecodeError: Error parsing message

question is how you are selecting the  --input_dim [shape is [?,?,?,3]] as input layer and why resnet_v2_101/predictions/Reshape_1 is a output layer [are you able to get boxes label and score from there] ????

please confirm

Thankyou