[FRONTEND][TENSORFLOW] Add Pad and PadV2 support #1545
Conversation
| @@ -669,6 +669,22 @@ def _impl(inputs, in_state_c, in_state_h, attr, params): | |||
| return _impl | |||
|
|
|||
|
|
|||
| def _pad(name): | |||
| def _impl(inputs, attr, params): | |||
| padlist = params.pop(inputs[1].list_output_names()[0]).asnumpy() | |||
There was a problem hiding this comment.
Suggest to assert if padlist is not available in params.
There was a problem hiding this comment.
Please assert for mode not equal to CONSTANT
And the mode string is case-insensitive.
There was a problem hiding this comment.
Thank you for the review.
For the second pointed out, I think that _pad is called from only CONSTANT in this PR, because Pad is generated by only tf.pad with CONSTANT mode without constant_value. In more detail, tf.pad generates different operators depending on the mode and the option:
- mode is
SYMMETRICorREFLECT=>MirrorPad - mode is
CONSTANTwithoutconstant_value=>Pad - mode is
CONSTANTwithconstant_value=>PadV2
I confirmed these little strange behavior by following script:
import tensorflow as tf
import nnvm
def main(_):
def check_pad_graph(**kwargs):
print("Pad: args = ", kwargs)
with tf.Graph().as_default() as graph:
sess = tf.InteractiveSession()
a = tf.constant([[1,2,3], [4,5,6]], name='a')
padding = tf.constant([[1, 1], [2,2]])
c = tf.pad(a, padding, **kwargs)
sess.run(tf.global_variables_initializer())
print(sess.run(c))
graph_def = graph.as_graph_def(add_shapes=True)
pad_mode = kwargs['mode']
if pad_mode == 'CONSTANT':
name = "constant"
if 'constant_values' in kwargs:
name += "_with_values"
elif pad_mode == 'REFLECT':
name = "reflect"
elif pad_mode == 'SYMMETRIC':
name = "symmetric"
name += "_pad.pbtxt"
tf.train.write_graph(graph_def, '.', name, as_text=True)
check_pad_graph(mode='CONSTANT')
check_pad_graph(mode='CONSTANT', constant_values=1)
check_pad_graph(mode='SYMMETRIC', constant_values=1)
check_pad_graph(mode='REFLECT')
if __name__ == '__main__':
tf.app.run()I put the script's result on my gist:
https://gist.github.com/nishi-t/a72c477c5f07812d327a9cb7f06b923d
There was a problem hiding this comment.
@nishi-t thanks for the details.
I think this PR is good to go with the other cosmetic change in testcase.
Any data about demand for MirrorPad ? we could propose additional attribute to pad operator to support mode.
There was a problem hiding this comment.
@srkreddy1238 Sorry for the delay. In terms of wether I have use cases to use MirrorPad (use tf.pad with symmetric or reflect mode), I don't have any data at the moment. But, since Pad is required for resnet, I want to add Pad support for now.
To support MirrorPad, we will have to add mode option(attribute?) to pad operator (in nnvm and perhaps tvm) to distinguish behavior of padding(CONSTANT, REFLECT, SYMMETRIC) . Since MirrorPad has mode as attribute to distinguish between REFLECT and SYMMETRIC, we will be able to pass it to the pad operator.
I might have a misunderstand so please let me if we can give you any other information.
There was a problem hiding this comment.
@nishi-t don't worry about MirrorPad, I just asked about it to know the requirement of it.
| padlist = params.pop(inputs[1].list_output_names()[0]).asnumpy() | ||
| paddings = tuple([tuple(l) for l in padlist]) | ||
| attr['pad_width'] = paddings | ||
| attr['pad_value'] = 0 |
There was a problem hiding this comment.
For Pad constant_value should be taken from attributes (should be zero only if not available in attributes.).
There was a problem hiding this comment.
Pad doesn't have constant_value as attribute and pad_value always should be 0 for Pad. Please check my previous comment.
| @@ -704,6 +704,44 @@ def test_forward_resize_bilinear(): | |||
| _test_resize_bilinear((4, 16, 32, 32), [50, 50], False) | |||
| _test_resize_bilinear((6, 32, 64, 64), [20, 20], True) | |||
|
|
|||
| ####################################################################### | |||
| # Pad | |||
| # ------- | |||
There was a problem hiding this comment.
Underline up to the text size !
| pad_values = constant_op.constant(paddings) | ||
| pad = tf.pad(in_data, paddings=pad_values, **kwargs) | ||
|
|
||
| if 'constant_values' in kwargs: |
There was a problem hiding this comment.
constant_values could be there for both Pad and PadV2
| @@ -669,6 +669,22 @@ def _impl(inputs, in_state_c, in_state_h, attr, params): | |||
| return _impl | |||
|
|
|||
|
|
|||
| def _pad(name): | |||
| def _impl(inputs, attr, params): | |||
| padlist = params.pop(inputs[1].list_output_names()[0]).asnumpy() | |||
There was a problem hiding this comment.
@nishi-t thanks for the details.
I think this PR is good to go with the other cosmetic change in testcase.
Any data about demand for MirrorPad ? we could propose additional attribute to pad operator to support mode.
|
@srkreddy1238 I addressed. Please review again. |
|
Thanks for contribution and reviewing. This is merged. |
* [FRONTEND][TENSORFLOW] Add Pad and PadV2 support * Add assettion to _pad, and fix testcase for pad.
This PR adds Pad and PadV2 support for tensorflow.
@PariksheetPinjari909 @srkreddy1238 Could you please review?