I met some bugs when convert MobileNetV2（tensorflow）to tvm

block98k · September 7, 2018, 8:04am

everytime I executed the code below,

sym, params = nnvm.frontend.from_tensorflow(gd)

I would get this error:

NotImplementedError: Please freeze the graph with add_shapes=True

It was given by line 994 of ~/tvm/nnvm/python/nnvm/frontend/tensorflow.py

Is it the problem of .pb file of tensorflow? How should i freeze the graph?
the function freeze_graph.freeze_graph() does not have the add_shapes attr. Could someone help me?

srkreddy1238 · September 7, 2018, 8:39am

Tensorflow frontend expect the graph to be frozen with add_shapes=True.

You may refer to below util

github.com

srkreddy1238/dmlc_data/blob/master/work/tf-to-nnvm.py

"""
Sample Utility to convert tensorflow model compatible to TVM
============================================================

Args:

<model_file> <model_out> <out_node>

Ex: tf-to-nnvm.py mobilenet_v1_1.0_224_frozen.pb mobilenet_v1_1.0_224_frozen-shapes.pb 'MobilenetV1/Predictions/Reshape_1'

"""

# Tensorflow imports
import tensorflow as tf
from tensorflow.core.framework import graph_pb2
from tensorflow.python.framework import dtypes
from tensorflow.python.framework import tensor_util
import sys

model_file=sys.argv[1]

This file has been truncated. show original

lixiaoquan · September 7, 2018, 10:23am

Very useful script ! Do you think this conversion can be merged to
nnvm.testing.tf.ProcessGraphDefParam() or other similar utility function?

srkreddy1238 · September 7, 2018, 10:42am

I proposed to embed into frontend and the decision was to keep this outside front end and warn the user.
nnvm.testing.tf may be a better place. Thanks for the suggestion and I will try to enhance it.

block98k · September 7, 2018, 11:44am

thank you ! this problem has been solved with your help. but if it could only inference with CPU? there are some bugs when i change tvm.cpu(0) to tvm.gpu(0). because the TVM has not surpported the DW convolution?

srkreddy1238 · September 7, 2018, 11:56am

Majority of Tensorflow models are NHWC and the GPU is good with NCHW. I am working on some changes to frontend to enable GPU by converting operators to NCHW.

This is the related discussion for the same.