[Relay][Op] Adaptive pooling #3085
Conversation
include/tvm/relay/attrs/nn.h
Outdated
| @@ -332,6 +332,22 @@ struct GlobalPool2DAttrs : public tvm::AttrsNode<GlobalPool2DAttrs> { | |||
| } | |||
| }; | |||
|
|
|||
| /*! \brief Attributes for global pool operator */ | |||
python/tvm/relay/op/nn/nn.py
Outdated
| layout="NCHW"): | ||
| r"""2D adaptive max pooling operator. | ||
|
|
||
| This operator takes data as input and does 2D average value calculation |
| auto i_end_h = end_index(output[height_axis], out_height, height); | ||
| auto i_start_w = start_index(output[width_axis], out_width, width); | ||
| auto i_end_w = end_index(output[width_axis], out_width, width); | ||
| auto dheight = tvm::reduce_axis(Range(0, i_end_h - i_start_h), "rv1"); |
There was a problem hiding this comment.
Out of curiosity, is naming the IterVar rv1 necessary?
There was a problem hiding this comment.
This is optional.
| const Expr& odim, | ||
| const Expr& idim) { | ||
| Expr tmp = (out_index + 1) * idim / odim; | ||
| return tvm::ir::Select::make((out_index + 1) * idim % odim == 0, |
There was a problem hiding this comment.
Any reason not to explicitly take the ceil here?
There was a problem hiding this comment.
ceil will involve several casts and causing result inaccurate.
topi/python/topi/cuda/pooling.py
Outdated
|
|
||
| Parameters | ||
| ---------- | ||
| outs: Array of Tensor | ||
| The computation graph description of global_pool | ||
| The computation graph description of adaptive_poo |
There was a problem hiding this comment.
adaptive_poo->adaptive_pool
src/relay/op/nn/pooling.cc
Outdated
| const auto* data = types[0].as<TensorTypeNode>(); | ||
| if (data == nullptr) { return false; } | ||
| const auto dshape = data->shape; | ||
| CHECK_NE(dshape.size(), 0); |
There was a problem hiding this comment.
Remove? since dshape.size() >= 2
src/relay/op/nn/pooling.cc
Outdated
| CHECK_NE(dshape.size(), 0); | ||
| CHECK_GE(dshape.size(), 2U) | ||
| << "Pool2D only support input >= 2-D: input must have height and width"; | ||
| const auto param = attrs.as<AdaptivePool2DAttrs>(); |
There was a problem hiding this comment.
const auto* to reduce confusion since as returns a pointer?
kevinthesun
force-pushed
the
AdaptivePooling
branch
from
c23b218 to
060309b
Compare
|
Please take some time to discuss the API naming and docs, as adaptive_pooling seems to be something that is not conventional. |
|
@tqchen This naming comes from pytorch pooling family： https://pytorch.org/docs/stable/_modules/torch/nn/modules/pooling.html MXNet has a similar API for adaptive avg pooing: https://mxnet.apache.org/api/python/ndarray/contrib.html?highlight=adap#mxnet.ndarray.contrib.AdaptiveAvgPooling2D For now I don't see a more appropriate name in mainsteam DL frameworks. Also adaptive describes the behavior of this kind of pooling. |
|
let us make it contrib.adaptive_pooling then and mark as experimental |
kevinthesun
force-pushed
the
AdaptivePooling
branch
from
c357607 to
7fdb3f2
Compare
kevinthesun
force-pushed
the
AdaptivePooling
branch
from
abbd9dd to
2de44ca
Compare
|
@tqchen This is ready to be reviewed. |
|
@kevinthesun Could you also update the mxnet converter for adaptive pooling? |
Already updated. |
|
Sorry, didn't notice that... |
There was a problem hiding this comment.
Notice Tianqi's comment. Could you change the nn.contrib_adaptive_avg_pool2d and nn.contrib_adaptive_max_pool2d to contrib.adaptive_avg_pool2d and contrib.adaptive_avg_pool2d?
Please also update the doc at docs/langref/relay_op.rst and docs/api/python/topi.rst.
|
I will leave the PR management to @icemelon9 |
|
Thanks everyone. This is now merged. |
* Add topi adaptive_pool * Use adaptive_pool to compute global_pool * Add relay adaptive pool2d * Fix lint * Fix typo * Minor change * Change support level to 10 * Add contrib * Remove global pool schedule * Add contrib module * Fix lint * Update doc * Update doc
* Add topi adaptive_pool * Use adaptive_pool to compute global_pool * Add relay adaptive pool2d * Fix lint * Fix typo * Minor change * Change support level to 10 * Add contrib * Remove global pool schedule * Add contrib module * Fix lint * Update doc * Update doc
Adaptive pooling operator.