[Tensorflow][CropAndResize] Potential bug or limitation -> AttributeError: '<class 'tvm.relay.expr.Call'>' object has no attribute 'name_hint'

tico · August 9, 2019, 12:11pm

Hi,

I have an isssue with a model that has a CropAndResize operator related to the processing of the boxes input. The error is as following:

Traceback (most recent call last):
  File "tvm_eval_2nd_half.py", line 247, in <module>
    mod, params = tvm_eval_non_tuned()
  File "tvm_eval_2nd_half.py", line 75, in tvm_eval_non_tuned
    shape=shape_dict)
  File "/home/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 2441, in from_tensorflow
    mod, params = g.from_tensorflow(graph, layout, shape, outputs)
  File "/home/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 2085, in from_tensorflow
    op = self._convert_operator(node.op, inputs, attr, graph)
  File "/home/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 2404, in _convert_operator
    sym = convert_map[op_name](inputs, attrs, self._params)
  File "/home/tvm/tvm/python/tvm/relay/frontend/tensorflow.py", line 504, in _impl
    boxes = params.pop(inputs[1].name_hint).asnumpy().tolist()
  File "/home/tvm/tvm/python/tvm/_ffi/_ctypes/node.py", line 75, in __getattr__
    "'%s' object has no attribute '%s'" % (str(type(self)), name))
AttributeError: '<class 'tvm.relay.expr.Call'>' object has no attribute 'name_hint'

I have tried to add AtributteError to the except but still it does not work as I get the following error:

AssertionError: All inputs to infer must be available in params.

@yongwww @srkreddy1238 maybe you can comment on this since you added this operator recently

tico · August 9, 2019, 8:25am

Hi,

It seems that only in the inputs are constants it works. If the boxes input comes from a placeholder or another operator it fails while processing the crop_and_resize operator in the Tensorflow frontend.

I am currently debugging the Tensorflow frontend but any hint would be highly appreciated!

tico · August 9, 2019, 12:10pm

After having a look at the test_forward.py script and the generated Relay IR, I see that TVM generates one image.resize op for every box. In other words, the number of boxes must be known a compile time.

Is this right? If so, how to use it then in models where the number of boxes is not known or not constant?

@vinx13 Maybe you can also give me hint about this?

Thanks

tico · August 12, 2019, 11:31am

@yongwww @srkreddy1238 @vinx13 Any ideas about this?

The only way that I was able to properly compile my model with CropAndResize was to remove the following assertion.

github.com

dmlc/tvm/blob/master/python/tvm/relay/frontend/tensorflow.py#L43


from .common import AttrCvt, get_relay_op
from .common import infer_type as _infer_type
from .common import infer_shape as _infer_shape
from .common import infer_channels as _infer_channels


__all__ = ['from_tensorflow']


def _infer_value(input_val, params):
from tvm.contrib import graph_runtime
# Check that all free variables have associated parameters.
assert all(var.name_hint in params.keys() for var in analysis.free_vars(
    input_val)), "All inputs to infer must be available in params."
func = _expr.Function(analysis.free_vars(input_val), input_val)
with tvm.relay.build_config(opt_level=0):
    graph, lib, params = tvm.relay.build(func, target="llvm", params=params)
ctx = tvm.context("llvm", 0)
m = graph_runtime.create(graph, lib, ctx)
m.set_input(**params)
m.run()
return m.get_output(0)

The test case located at test_forward.py works since all inputs given are constants. As soon as you use a Placeholder or another operator as input then you hit the error mentioned above.

github.com

dmlc/tvm/blob/master/tests/python/frontend/tensorflow/test_forward.py#L1181




def test_forward_crop():
""" Crop to bounding box """
_test_crop((1, 224, 224, 3), 20, 20, 120, 120)




#######################################################################
# CropAndResize
# -------------


def _test_forward_crop_and_resize(img_shape, boxes, box_idx, crop_size, method='bilinear', dtype="float32"):
image = np.random.uniform(0, 10, size=img_shape).astype(dtype)
tf.reset_default_graph()
in_data = tf.placeholder(dtype, image.shape, name="in_data")
tf.image.crop_and_resize(in_data, boxes=boxes, box_ind=box_idx, crop_size=crop_size,
                         method=method, name="crop_and_resize")
compare_tf_with_tvm([image], ['in_data:0'], 'crop_and_resize:0')


def test_forward_crop_and_resize():
""" CropAndResize """
_test_forward_crop_and_resize([1, 11, 11, 3], [[0, 0, 1, 1]], [0], [5, 5])

The following is a minimal test model that also fails.

    image = tf.placeholder(tf.float32, shape=(1, 24, 42, 64), name='image')
    boxes = tf.placeholder(tf.float32, shape=(1, 4), name='boxes')
    box_ind = tf.zeros(shape=tf.shape(boxes)[0], dtype=tf.int32)
    crop_size = [4, 4]
    crop = tf.image.crop_and_resize(image, boxes, box_ind=box_ind, crop_size = crop_size, name='crop')

yongwww · December 1, 2019, 6:09pm

this issue should be fixed in pr: https://github.com/apache/incubator-tvm/pull/4417