Fix for monodle failure #629
Conversation
meenakshiramanathan1
requested review from
nvukobratTT,
pilkicTT and
dgolubovicTT
as code owners
meenakshiramanathan1
force-pushed
the
mramanathan/convtranspose2d
branch
from
0a982bf
to
fd53e87
Compare
meenakshiramanathan1
changed the title
meenakshiramanathan1
force-pushed
the
mramanathan/convtranspose2d
branch
3 times, most recently
from
0aeabea
to
f9c284f
Compare
| input_nid = input_[0] | ||
| input_node = graph["nodes"][input_nid] | ||
| if input_node["op"] == "parameter" and input_node["name"].endswith("weight"): | ||
| in_channel = input_node["attrs"]["shape"][0][0][0] |
There was a problem hiding this comment.
Will this handle different shape ranks?
There was a problem hiding this comment.
Yes, inputs are either going to be 3d or 4d for convtranspose2d which is handled here, I have added different shape rank shape inputs in sanity now.
| input_nid = input_[0] | ||
| input_node = graph["nodes"][input_nid] | ||
| if input_node["op"] == "parameter" and input_node["name"].endswith("weight"): | ||
| in_channel = input_node["attrs"]["shape"][0][0][0] |
There was a problem hiding this comment.
How does it work for channel first and channel last cases?
There was a problem hiding this comment.
input_node["attrs"]["shape"][0][0][0] corresponds to the in-channels from the param being sent
In case of below example, it will be 16 and it would be the same for channelfirst and channel last cases
"in_channels, out_channels, kernel_size, stride, padding, groups, bias, dilation, padding_mode, input_shape",
[ 16, 33, (3, 3), 2, 0, 1, True, 1, "zeros", (20, 16, 50, 100)) ]
f9c284f
to
8f9f3b1
Compare
meenakshiramanathan1
force-pushed
the
mramanathan/convtranspose2d
branch
from
8f9f3b1
to
a5ef92b
Compare
| self.padding_bottom, | ||
| ] | ||
|
|
||
| assert self.padding_top == self.padding_bottom, "Padding values for top and bottom must be the same." |
There was a problem hiding this comment.
As this is a TTNN limitation, do we need to assert this here?
Let's track this as their issues, and model it further if we see that this is by design :))
There was a problem hiding this comment.
@nvukobratTT , the TTNN limitation is for a different model i.e mobilenetv2 and not for this model, these changes has been added to modify the padding handled in eval function of convtranspose2d op as a fix for monodle model.
There was a problem hiding this comment.
Is this an implementation limitation from our size regarding conv2d transpose? More precisely, my question is can we support this on FFE with different padding for each side?
If yes, let's update assert message to be a bit more cleaner about why this is limitation :))
There was a problem hiding this comment.
Sure Nikola, I will check that and make required changes.
Fix #513