[Frontend][Tensorflow] Report NCHW error when convert tensorflow graph into ARM Mali runtime

Development
1

I tried to freeze the graph on GPU server but still have this error warning:

only support NCHW convolution on mali

Loading
Loading Done
Compiling
Traceback (most recent call last):
  File "convert_net_1.py", line 52, in <module>
    convert(nnvm_net, nnvm_params, (1,513, 513, 3) ,'float32','llvm --system-lib -target=aarch64-linux-gnu -mattr=+neon')
  File "convert_net_1.py", line 24, in convert
    dtype=dtype, target_host=target_host)
  File "/root/.local/lib/python2.7/site-packages/nnvm-0.8.0-py2.7.egg/nnvm/compiler/build_module.py", line 294, in build
    graph = graph.apply("GraphFusePartition").apply("GraphFuseCompile")
  File "/root/.local/lib/python2.7/site-packages/nnvm-0.8.0-py2.7.egg/nnvm/graph.py", line 234, in apply
    check_call(_LIB.NNGraphApplyPasses(self.handle, npass, cpass, ctypes.byref(ghandle)))
  File "/root/.local/lib/python2.7/site-packages/nnvm-0.8.0-py2.7.egg/nnvm/_base.py", line 75, in check_call
    raise NNVMError(py_str(_LIB.NNGetLastError()))
nnvm._base.NNVMError: TVMCall CFunc Error:
Traceback (most recent call last):
  File "/root/.local/lib/python2.7/site-packages/tvm-0.4.0-py2.7-linux-x86_64.egg/tvm/_ffi/_ctypes/function.py", line 54, in cfun
    rv = local_pyfunc(*pyargs)
  File "/root/.local/lib/python2.7/site-packages/nnvm-0.8.0-py2.7.egg/nnvm/top/nn.py", line 106, in compute_conv2d
    inputs[0], kernel, strides, padding, layout, out_dtype=out_dtype)
  File "<decorator-gen-35>", line 2, in conv2d
  File "/root/.local/lib/python2.7/site-packages/tvm-0.4.0-py2.7-linux-x86_64.egg/tvm/target.py", line 345, in dispatch_func
    return dispatch_dict[k](*args, **kwargs)
  File "/root/.local/lib/python2.7/site-packages/topi-0.4.0-py2.7.egg/topi/mali/conv2d.py", line 110, in decl_conv2d
    assert layout == 'NCHW', "only support NCHW convolution on mali"
AssertionError: only support NCHW convolution on mali
2

Can you share more details of the script that you are using? Somehow it looks like you are passing another, unsupported data layout, e.g., (NHWC).

3

Sure, the code is pretty simple:

Freeze graph:

freeze_graph_with_def_protos(my_graph_def,
         input_checkpoint=checkpoint_file,
         input_saver_def=None,
         output_node_names=output_node_names,
         restore_op_name="save/restore_all",
         filename_tensor_name="save/Const:0",
         output_graph=frozen_model_file,
         clear_devices=False,
         initializer_nodes="")

Converting:

def convert(net, params, data_shape, dtype, target_host):
    # compile
    opt_level = 2 if dtype == 'float32' else 1
    with nnvm.compiler.build_config(opt_level=opt_level):
        print("Compiling")
        graph, lib, params = nnvm.compiler.build(
            net, tvm.target.mali(), shape={"image": data_shape}, params=params,
            dtype=dtype, target_host=target_host)
        print("Compiling Done")

    lib.export_library('./net1.tar')

    with open("./net1", "w") as fo:
        fo.write(graph.json())
    with open("./net1.params", "wb") as fo:
        fo.write(nnvm.compiler.save_param_dict(params))
model_name = 'frozen_model.pb'
with tf.gfile.FastGFile(os.path.join(
        "./", model_name), 'rb') as f:
    graph_def = tf.GraphDef()
    graph_def.ParseFromString(f.read())

    detection_graph = tf.Graph()
    with detection_graph.as_default():
        #graph_def = tf.GraphDef()
        graph = tf.import_graph_def(graph_def, name='')
        graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
    # Call the utility to import the graph definition into default graph.
    graph_def = nnvm.testing.tf.ProcessGraphDefParam(graph_def)

    tf.import_graph_def(graph_def, name='')
    graph_def = tf.get_default_graph().as_graph_def(add_shapes=True)
print("Loading")
nnvm_net, nnvm_params = nnvm.frontend.from_tensorflow(graph_def)
print("Loading Done")
convert(nnvm_net, nnvm_params, (1,513, 513, 3) ,'float32','llvm --system-lib -target=aarch64-linux-gnu -mattr=+neon')
4

Oh, it seems the pre-tained model conv2d layer constructed with NHWC.

https://www.tensorflow.org/api_docs/python/tf/nn/conv2d