[TOPI, Relay] ROI Pool operator (apache#2811)

include/tvm/relay/attrs/vision.h

@@ -121,6 +121,26 @@ struct ROIAlignAttrs : public tvm::AttrsNode<ROIAlignAttrs> {
   }
 };
 
+/*! \brief Attributes used in roi_pool operators */
+struct ROIPoolAttrs : public tvm::AttrsNode<ROIPoolAttrs> {
+  Array<IndexExpr> pooled_size;
+  double spatial_scale;
+  std::string layout;
+  TVM_DECLARE_ATTRS(ROIPoolAttrs, "relay.attrs.ROIPoolAttrs") {
+    TVM_ATTR_FIELD(pooled_size).describe("Output size of roi pool.");
+    TVM_ATTR_FIELD(spatial_scale)
+        .describe(
+            "Ratio of input feature map height (or w) to raw image height (or w). "
+            "Equals the reciprocal of total stride in convolutional layers, which should be "
+            "in range (0.0, 1.0]");
+    TVM_ATTR_FIELD(layout).set_default("NCHW").describe(
+        "Dimension ordering of data and weight. Can be 'NCHW', 'NHWC', etc."
+        "'N', 'C', 'H', 'W' stands for batch, channel, height, and width"
+        "dimensions respectively. Convolution is applied on the 'H' and"
+        "'W' dimensions.");
+  }
+};
+
 /*! \brief Attributes used in yolo reorg operators */
 struct YoloReorgAttrs : public tvm::AttrsNode<YoloReorgAttrs> {
   Integer stride;

python/tvm/relay/frontend/mxnet.py

@@ -438,6 +438,14 @@ def _mx_roi_align(inputs, attrs):
     return _op.vision.roi_align(inputs[0], inputs[1], **new_attrs)
 
 
+def _mx_roi_pooling(inputs, attrs):
+    new_attrs = {}
+    new_attrs["pooled_size"] = attrs.get_int_tuple("pooled_size")
+    new_attrs["spatial_scale"] = attrs.get_float("spatial_scale")
+    new_attrs["layout"] = "NCHW"
+    return _op.vision.roi_pool(inputs[0], inputs[1], **new_attrs)
+
+
 def _mx_proposal(inputs, attrs):
     new_attrs = {}
     new_attrs["scales"] = attrs.get_float_tuple("scales", (4.0, 8.0, 16.0, 32.0))
@@ -641,6 +649,7 @@ def _mx_shape_array(inputs, attrs):
     "_contrib_MultiBoxPrior" : _mx_multibox_prior,
     "_contrib_MultiBoxDetection" : _mx_multibox_detection,
     "_contrib_ROIAlign" : _mx_roi_align,
+    "ROIPooling"        : _mx_roi_pooling,
     "_contrib_Proposal" : _mx_proposal,
     "_contrib_MultiProposal" : _mx_proposal,
     "_contrib_box_nms" : _mx_box_nms,

python/tvm/relay/op/vision/_rcnn.py

@@ -22,6 +22,22 @@ def schedule_roi_align(_, outs, target):
 
 reg.register_pattern("vision.roi_align", OpPattern.OUT_ELEMWISE_FUSABLE)
 
+@reg.register_compute("vision.roi_pool")
+def compute_roi_pool(attrs, inputs, _, target):
+    """Compute definition of roi_pool"""
+    assert attrs.layout == "NCHW"
+    return [topi.vision.rcnn.roi_pool_nchw(
+        inputs[0], inputs[1], pooled_size=get_const_tuple(attrs.pooled_size),
+        spatial_scale=attrs.spatial_scale)]
+
+@reg.register_schedule("vision.roi_pool")
+def schedule_roi_pool(_, outs, target):
+    """Schedule definition of roi_pool"""
+    with target:
+        return topi.generic.vision.schedule_roi_pool(outs)
+
+reg.register_pattern("vision.roi_pool", OpPattern.OUT_ELEMWISE_FUSABLE)
+
 @reg.register_compute("vision.proposal")
 def compute_proposal(attrs, inputs, _, target):
     """Compute definition of proposal"""

python/tvm/relay/op/vision/rcnn.py

@@ -32,6 +32,33 @@ def roi_align(data, rois, pooled_size, spatial_scale, sample_ratio=-1, layout='N
     return _make.roi_align(data, rois, pooled_size, spatial_scale, sample_ratio, layout)
 
 
+def roi_pool(data, rois, pooled_size, spatial_scale, layout='NCHW'):
+    """ROI pool operator.
+
+    Parameters
+    ----------
+    data : relay.Expr
+        4-D tensor with shape [batch, channel, height, width]
+
+    rois : relay.Expr
+        2-D tensor with shape [num_roi, 5]. The last dimension should be in format of
+        [batch_index, w_start, h_start, w_end, h_end]
+
+    pooled_size : list/tuple of two ints
+        output size
+
+    spatial_scale : float
+        Ratio of input feature map height (or w) to raw image height (or w). Equals the reciprocal
+        of total stride in convolutional layers, which should be in range (0.0, 1.0]
+
+    Returns
+    -------
+    output : relay.Expr
+        4-D tensor with shape [num_roi, channel, pooled_size, pooled_size]
+    """
+    return _make.roi_pool(data, rois, pooled_size, spatial_scale, layout)
+
+
 def proposal(cls_prob,
              bbox_pred,
              im_info,

src/relay/op/vision/rcnn_op.cc

@@ -63,6 +63,58 @@ RELAY_REGISTER_OP("vision.roi_align")
 .set_support_level(5)
 .add_type_rel("ROIAlign", ROIAlignRel);
 
+TVM_REGISTER_NODE_TYPE(ROIPoolAttrs);
+
+bool ROIPoolRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
+                 const TypeReporter& reporter) {
+  auto roi_pool_attrs = attrs.as<ROIPoolAttrs>();
+  CHECK_EQ(types.size(), 3);
+  const auto* data = types[0].as<TensorTypeNode>();
+  const auto* rois = types[1].as<TensorTypeNode>();
+  const auto& dshape = data->shape;
+  const auto& rshape = rois->shape;
+  CHECK(roi_pool_attrs);
+  CHECK_EQ(dshape.size(), 4) << "Input data should be 4-D.";
+  CHECK_EQ(rshape.size(), 2) << "Input rois should be 2-D.";
+  CHECK_EQ(roi_pool_attrs->layout, "NCHW") << "ROI Pool only supports NCHW layout";
+  // assign output type
+  std::vector<IndexExpr> oshape(
+      {rshape[0], dshape[1], roi_pool_attrs->pooled_size[0], roi_pool_attrs->pooled_size[1]});
+  reporter->Assign(types[2], TensorTypeNode::make(oshape, data->dtype));
+  return true;
+}
+
+Expr MakeROIPool(Expr data, Expr rois, Array<IndexExpr> pooled_size, double spatial_scale,
+                 std::string layout) {
+  auto attrs = make_node<ROIPoolAttrs>();
+  attrs->pooled_size = pooled_size;
+  attrs->spatial_scale = spatial_scale;
+  attrs->layout = layout;
+  static const Op& op = Op::Get("vision.roi_pool");
+  return CallNode::make(op, {data, rois}, Attrs(attrs), {});
+}
+
+TVM_REGISTER_API("relay.op.vision._make.roi_pool")
+.set_body([](const TVMArgs& args, TVMRetValue* rv) {
+    runtime::detail::unpack_call<Expr, 5>(MakeROIPool, args, rv);
+  });
+
+RELAY_REGISTER_OP("vision.roi_pool")
+    .describe(R"doc(ROI Pool operator.
+
+ - **data**: This depends on the `layout` parameter. Input is 4D array of shape
+             (batch_size, channels, height, width) if `layout` is `NCHW`.
+ - **rois**: 2D array of shape (num_roi, 5). The last dimension should be in format of
+             [batch_index, w_start, h_start, w_end, h_end].
+ - **out**: This depends on the `layout` parameter. Output is 4D array of shape
+            (num_roi, channels, pooled_height, pooled_width) if `layout` is `NCHW`.
+ )doc" TVM_ADD_FILELINE)
+.set_num_inputs(2)
+.add_argument("data", "Tensor", "The input tensor.")
+.add_argument("rois", "Tensor", "The input rois")
+.set_support_level(5)
+.add_type_rel("ROIPool", ROIPoolRel);
+
 TVM_REGISTER_NODE_TYPE(ProposalAttrs);
 
 bool ProposalRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,

tests/python/relay/test_op_level5.py

@@ -357,6 +357,38 @@ def verify_roi_align(data_shape, rois_shape, pooled_size, spatial_scale, sample_
     verify_roi_align((4, 4, 16, 16), (32, 5), pooled_size=7, spatial_scale=0.5, sample_ratio=2)
 
 
+def test_roi_pool():
+    def verify_roi_pool(data_shape, rois_shape, pooled_size, spatial_scale):
+        data = relay.var("data", relay.ty.TensorType(data_shape, "float32"))
+        rois = relay.var("rois", relay.ty.TensorType(rois_shape, "float32"))
+        z = relay.vision.roi_pool(data, rois, pooled_size=(pooled_size, pooled_size),
+                                   spatial_scale=spatial_scale, layout="NCHW")
+        zz = relay.ir_pass.infer_type(z)
+
+        batch, channel, in_size, _ = data_shape
+        num_roi = rois_shape[0]
+        assert zz.checked_type == relay.ty.TensorType(
+                (num_roi, channel, pooled_size, pooled_size), "float32")
+
+        func = relay.Function([data, rois], z)
+        func = relay.ir_pass.infer_type(func)
+        np_data = np.random.uniform(size=data_shape).astype("float32")
+        np_rois = np.random.uniform(size=rois_shape).astype('float32') * in_size
+        np_rois[:, 0] = np.random.randint(low = 0, high = batch, size = num_roi).astype('float32')
+        ref_res = topi.testing.roi_pool_nchw_python(np_data, np_rois, pooled_size=pooled_size,
+                                                     spatial_scale=spatial_scale)
+        for target, ctx in ctx_list():
+            intrp1 = relay.create_executor("graph", ctx=ctx, target=target)
+            op_res1 = intrp1.evaluate(func)(np_data, np_rois)
+            tvm.testing.assert_allclose(op_res1.asnumpy(), ref_res, rtol=1e-4)
+            intrp2 = relay.create_executor("debug", ctx=ctx, target=target)
+            op_res2 = intrp2.evaluate(func)(np_data, np_rois)
+            tvm.testing.assert_allclose(op_res2.asnumpy(), ref_res, rtol=1e-4)
+
+    verify_roi_pool((1, 4, 16, 16), (32, 5), pooled_size=7, spatial_scale=1.0)
+    verify_roi_pool((4, 4, 16, 16), (32, 5), pooled_size=7, spatial_scale=0.5)
+
+
 def test_proposal():
     def verify_proposal(np_cls_prob, np_bbox_pred, np_im_info, np_out, attrs):
         cls_prob = relay.var("cls_prob", relay.ty.TensorType(np_cls_prob.shape, "float32"))
@@ -464,6 +496,7 @@ def verify_yolo_reorg(shape, stride):
     test_multibox_transform_loc()
     test_get_valid_counts()
     test_roi_align()
+    test_roi_pool()
     test_proposal()
     test_yolo_reorg_infer_shape()
     test_yolo_reorg()

topi/python/topi/cuda/vision.py

@@ -134,6 +134,10 @@ def schedule_multibox_detection(outs):
 def schedule_roi_align(outs):
     return schedule_pool(outs, 'NCHW')
 
+@generic.schedule_roi_pool.register(["cuda", "gpu"])
+def schedule_roi_pool(outs):
+    return schedule_pool(outs, 'NCHW')
+
 @generic.schedule_proposal.register(["cuda", "gpu"])
 def schedule_proposal(outs):
     """Schedule for proposal operator.

topi/python/topi/generic/vision.py

@@ -139,6 +139,23 @@ def schedule_roi_align(outs):
     """
     return _default_schedule(outs, False)
 
+@tvm.target.generic_func
+def schedule_roi_pool(outs):
+    """Schedule for roi_align
+
+    Parameters
+    ----------
+    outs: Array of Tensor
+      The computation graph description of roi_pool
+      in the format of an array of tensors.
+
+    Returns
+    -------
+    s: Schedule
+      The computation schedule for the op.
+    """
+    return _default_schedule(outs, False)
+
 @tvm.target.generic_func
 def schedule_proposal(outs):
     """Schedule for proposal operator.

topi/python/topi/testing/__init__.py

@@ -15,6 +15,7 @@
 from .bilinear_resize_python import bilinear_resize_python
 from .reorg_python import reorg_python
 from .roi_align_python import roi_align_nchw_python
+from .roi_pool_python import roi_pool_nchw_python
 from .lrn_python import lrn_python
 from .l2_normalize_python import l2_normalize_python
 from .gather_nd_python import gather_nd_python

topi/python/topi/testing/roi_pool_python.py

@@ -0,0 +1,47 @@
+# pylint: disable=invalid-name, too-many-nested-blocks
+"Roi pool in python"
+import math
+import numpy as np
+
+def roi_pool_nchw_python(a_np, rois_np, pooled_size, spatial_scale):
+    """Roi pool in python"""
+    _, channel, height, width = a_np.shape
+    num_roi = rois_np.shape[0]
+    b_np = np.zeros((num_roi, channel, pooled_size, pooled_size), dtype=a_np.dtype)
+
+    if isinstance(pooled_size, int):
+        pooled_size_h = pooled_size_w = pooled_size
+    else:
+        pooled_size_h, pooled_size_w = pooled_size
+
+    for i in range(num_roi):
+        roi = rois_np[i]
+        batch_index = int(roi[0])
+        roi_start_w = int(round(roi[1] * spatial_scale))
+        roi_start_h = int(round(roi[2] * spatial_scale))
+        roi_end_w = int(round(roi[3] * spatial_scale))
+        roi_end_h = int(round(roi[4] * spatial_scale))
+        roi_h = max(roi_end_h - roi_start_h + 1, 1)
+        roi_w = max(roi_end_w - roi_start_w + 1, 1)
+
+        bin_h = float(roi_h) / pooled_size_h
+        bin_w = float(roi_w) / pooled_size_w
+
+        for ph in range(pooled_size_h):
+            for pw in range(pooled_size_w):
+                hstart = int(math.floor(ph * bin_h))
+                wstart = int(math.floor(pw * bin_w))
+                hend = int(math.ceil((ph + 1) * bin_h))
+                wend = int(math.ceil((pw + 1) * bin_w))
+                hstart = min(max(hstart + roi_start_h, 0), height)
+                hend = min(max(hend + roi_start_h, 0), height)
+                wstart = min(max(wstart + roi_start_w, 0), width)
+                wend = min(max(wend + roi_start_w, 0), width)
+                is_empty = (hend <= hstart) or (wend <= wstart)
+
+                for c in range(channel):
+                    if is_empty:
+                        b_np[i, c, ph, pw] = 0.
+                    else:
+                        b_np[i, c, ph, pw] = np.max(a_np[batch_index, c, hstart:hend, wstart:wend])
+    return b_np

topi/python/topi/vision/rcnn/__init__.py

@@ -1,4 +1,5 @@
 # pylint: disable=wildcard-import
 """Faster R-CNN and Mask R-CNN operators"""
 from .roi_align import *
+from .roi_pool import *
 from .proposal import *

topi/python/topi/vision/rcnn/roi_pool.py

@@ -0,0 +1,77 @@
+# pylint: disable=invalid-name
+"""ROI pool operator"""
+import tvm
+from ...util import get_const_tuple
+
+@tvm.target.generic_func
+def roi_pool_nchw(data, rois, pooled_size, spatial_scale):
+    """ROI pool operator in NCHW layout.
+
+    Parameters
+    ----------
+    data : tvm.Tensor
+        4-D with shape [batch, channel, height, width]
+
+    rois : tvm.Tensor
+        2-D with shape [num_roi, 5]. The last dimension should be in format of
+        [batch_index, w_start, h_start, w_end, h_end]
+
+    pooled_size : int or list/tuple of two ints
+        output size, or [out_height, out_width]
+
+    spatial_scale : float
+        Ratio of input feature map height (or w) to raw image height (or w). Equals the reciprocal
+        of total stride in convolutional layers, which should be in range (0.0, 1.0]
+
+    Returns
+    -------
+    output : tvm.Tensor
+        4-D with shape [num_roi, channel, pooled_size, pooled_size]
+    """
+    dtype = rois.dtype
+    _, channel, height, width = get_const_tuple(data.shape)
+    num_roi, _ = get_const_tuple(rois.shape)
+
+    if isinstance(pooled_size, int):
+        pooled_size_h = pooled_size_w = pooled_size
+    else:
+        pooled_size_h, pooled_size_w = pooled_size
+
+    def _pool(i, c, ph, pw):
+        roi = rois[i]
+        batch_index = roi[0].astype('int32')
+        roi_start_w, roi_start_h, roi_end_w, roi_end_h = roi[1], roi[2], roi[3], roi[4]
+
+        roi_start_h = tvm.round(roi_start_h * spatial_scale).astype('int32')
+        roi_start_w = tvm.round(roi_start_w * spatial_scale).astype('int32')
+        roi_end_h = tvm.round(roi_end_h * spatial_scale).astype('int32')
+        roi_end_w = tvm.round(roi_end_w * spatial_scale).astype('int32')
+
+        # force malformed ROIs to be 1x1
+        roi_h = tvm.max(roi_end_h - roi_start_h + 1, tvm.const(1, 'int32'))
+        roi_w = tvm.max(roi_end_w - roi_start_w + 1, tvm.const(1, 'int32'))
+
+        bin_h = roi_h.astype(dtype) / pooled_size_h
+        bin_w = roi_w.astype(dtype) / pooled_size_w
+
+        # use epsilon to prevent floating point precision loss in floor/ceil
+        epsilon = tvm.const(0.00001, dtype)
+        hstart = tvm.floor(ph * bin_h + epsilon).astype('int32')
+        wstart = tvm.floor(pw * bin_w + epsilon).astype('int32')
+        hend = tvm.ceil((ph + 1) * bin_h - epsilon).astype('int32')
+        wend = tvm.ceil((pw + 1) * bin_w - epsilon).astype('int32')
+        hstart = tvm.min(tvm.max(hstart + roi_start_h, 0), height)
+        wstart = tvm.min(tvm.max(wstart + roi_start_w, 0), width)
+        hend = tvm.min(tvm.max(hend + roi_start_h, 0), height)
+        wend = tvm.min(tvm.max(wend + roi_start_w, 0), width)
+
+        non_empty = tvm.all(hstart < hend, wstart < wend)
+        min_value = lambda dtype: tvm.if_then_else(non_empty, tvm.min_value(dtype),
+                                                   tvm.const(0.0, dtype))
+        # pylint: disable=unnecessary-lambda
+        _max = tvm.comm_reducer(lambda x, y: tvm.make._OpMax(x, y), min_value, name='max')
+        rh = tvm.reduce_axis((0, hend - hstart), 'rh')
+        rw = tvm.reduce_axis((0, wend - wstart), 'rw')
+        return _max(data[batch_index, c, hstart+rh, wstart+rw], axis=[rh, rw])
+
+    return tvm.compute((num_roi, channel, pooled_size_h, pooled_size_w), _pool, tag="pool,roi_pool")