[NNVM] Support argmax/argmin in tensorflow frontend #1514
Conversation
sergei-mironov
force-pushed
the
add-argmax-tf
branch
2 times, most recently
from
f2ea7ee to
0f3d7c4
Compare
|
|
||
| try: | ||
| from tensorflow.python.framework import tensor_util | ||
| except ImportError as e: | ||
| raise ImportError( | ||
| "Unable to import tensorflow which is required {}".format(e)) | ||
|
|
||
| # Parse the nodes briefly to determine the preconditions: |
There was a problem hiding this comment.
Brief parsing is a good addition, Thanks. It helps to list all missing ops at one place.
Please make it a separate function instead of inline.
There was a problem hiding this comment.
OK, I'll move it to a separate function.
| @@ -940,7 +976,7 @@ def from_tensorflow(self, graph): | |||
| "Please freeze the graph with add_shapes=True") | |||
| elif node.op == "Const": | |||
| # Assuming first Const node as Graph Input node | |||
| if self._input_node == '': | |||
| if self._input_node == '' and not has_placeholders: | |||
There was a problem hiding this comment.
Technically there is nothing does with _input_node. Hence _input_node has no significance here.
As you touched this area I suggest to change the below or ignore this part in this PR.
A graph can be executed as ConstA+ConstB=C.
where there is no place holders and we don't need to convert an Const to placeholder and expect input from user (Example test cases.).
Hence I suggest to simplify as "All placeholders are inputs" and "all Constants are Params".
There was a problem hiding this comment.
Treating constants as inputs seemed strange to me too, so I am glad to see you suggestion.
I attempted to make this simplification and it seems that at least some operations implicitly depend on Consts being an input. For example, _conv operation uses _input_shape here . Input shape in its turn is set only for input nodes (i.e. nodes which aren't listed in params attribute, here is the key point) Thus, removing constants from inputs unfortunately breaks some tests involving convolution, since _input_shapes is no longer available for input[0].
I think it is a bug in _conv since any operation should AFAIK work with any type of input nodes. Let me suggest that I fix it in a separate commit. As for this commit, I'd better remove the placeholder tweak completely to let other changes go.
There was a problem hiding this comment.
Yes, you could hold the placeholder change from this PR and let this go first.
We could open a discussion @ discuss.tvm.ai and discuss it further.
BTW, _input_shapes is a conclusion after a long discussion around SAME padding implementation of tensorflow convolution. This is primarily due to NNVM compilation process expects to build based on a fixed input shapes (which is different for tensorflow or other frameworks which accepts variable shapes at runtime).
_input_shape for constant is not a problem as it's a constant and shape is available from Tensor.
There was a problem hiding this comment.
Done, postponing the placeholders part. I've started https://discuss.tvm.ai/t/importing-tensorflow-models-placeholder-vs-const-issues/561 topic for further discussion.
sergei-mironov
force-pushed
the
add-argmax-tf
branch
2 times, most recently
from
d50bf27 to
fa69b93
Compare
|
Thanks @grwlf @srkreddy1238 this is merged |
This PR aims to improve Tensorflow frontend by: