You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2021/09/15 14:30:36 UTC
[GitHub] [tvm] quic-sanirudh opened a new pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
quic-sanirudh opened a new pull request #9017:
URL: https://github.com/apache/tvm/pull/9017
This PR implements com.microsoft.QLinearAveragePool and com.microsoft.QLinearGlobalAveragePool.
QLinearAveragePool is implemented as (dequantize -> float op -> quantize) according to its definition in [com.microsoft.QLinearAveragePool](https://github.com/microsoft/onnxruntime/blob/master/docs/ContribOperators.md#com.microsoft.QLinearGlobalAveragePool)
QLinearGlobalAveragePool is also implemented as (dequantize -> float op -> quantize) because the QNN op for global avg_pool is not yet available, but that is not how it's explicitly defined in the onnxruntime docs. Once QNN op is added, we could update this op frontend to directly generate a qnn op.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] quic-sanirudh commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
quic-sanirudh commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710043945
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -351,6 +366,13 @@ class AveragePool(Pool):
name = "avg_pool"
+class QLinearAveragePool(Pool):
Review comment:
I've refactored the code to separate out the quantized and non-quantized ops in a way that doesn't touch any of the non-quantized op classes. Let me know if this is close to what you had in mind, thanks.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] AndrewZhaoLuo commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
AndrewZhaoLuo commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710291824
##########
File path: tests/python/frontend/onnx/test_forward.py
##########
@@ -3056,6 +3056,152 @@ def verify_global_pooling(x_shape, mode):
verify_global_pooling([4, 1, 2, 6, 4], mode)
+@tvm.testing.parametrize_targets
+def test_qlinear_average_pool(target, dev):
+ def verify_qlinear_average_pool(
+ x_shape, kernel_shape, strides, pads, out_shape, auto_pad="NOTSET"
+ ):
+ input_nodes = [
+ helper.make_tensor_value_info("X", TensorProto.FLOAT, list(x_shape)),
+ ]
+
+ output_nodes = [
+ helper.make_tensor_value_info("Y", TensorProto.FLOAT, list(out_shape)),
+ ]
+
+ input_names = ["X"]
+
+ node = helper.make_node(
+ "AveragePool",
Review comment:
Ah I see, I did not notice that we run a quantization routine first
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] quic-sanirudh commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
quic-sanirudh commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710050671
##########
File path: tests/python/frontend/onnx/test_forward.py
##########
@@ -3056,6 +3056,152 @@ def verify_global_pooling(x_shape, mode):
verify_global_pooling([4, 1, 2, 6, 4], mode)
+@tvm.testing.parametrize_targets
+def test_qlinear_average_pool(target, dev):
+ def verify_qlinear_average_pool(
+ x_shape, kernel_shape, strides, pads, out_shape, auto_pad="NOTSET"
+ ):
+ input_nodes = [
+ helper.make_tensor_value_info("X", TensorProto.FLOAT, list(x_shape)),
+ ]
+
+ output_nodes = [
+ helper.make_tensor_value_info("Y", TensorProto.FLOAT, list(out_shape)),
+ ]
+
+ input_names = ["X"]
+
+ node = helper.make_node(
+ "AveragePool",
+ inputs=input_names,
+ outputs=["Y"],
+ kernel_shape=kernel_shape,
+ strides=strides,
+ )
+
+ if pads is None:
+ pad_attr = helper.make_attribute("auto_pad", auto_pad)
+ else:
+ pad_attr = helper.make_attribute("pads", pads)
+ node.attribute.append(pad_attr)
+
+ graph = helper.make_graph(
+ [node],
+ "qlinear_average_pool_test",
+ inputs=input_nodes,
+ outputs=output_nodes,
+ )
+
+ model = helper.make_model(graph, producer_name="qlinear_average_pool_Test")
+ quantize_and_verify_with_ort(model, input_names, [x_shape], target, dev)
+
+ # Pool1D
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32],
+ kernel_shape=[3],
+ strides=[1],
+ pads=[1, 1],
+ out_shape=[1, 1, 32],
+ )
+ # Pool2D
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32],
+ kernel_shape=[3, 3],
+ strides=[1, 1],
+ pads=[1, 1, 1, 1],
+ out_shape=[1, 1, 32, 32],
+ )
+
+ # Pool1D with stride
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32],
+ kernel_shape=[3],
+ strides=[2],
+ pads=[1, 1],
+ out_shape=[1, 1, 16],
+ )
+ # Pool2D with stride
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32],
+ kernel_shape=[3, 3],
+ strides=[2, 2],
+ pads=[1, 1, 1, 1],
+ out_shape=[1, 1, 16, 16],
+ )
+
+ # Pool1D with stride and autopadding
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32],
+ kernel_shape=[3],
+ strides=[2],
+ pads=None,
+ out_shape=[1, 1, 16],
+ auto_pad="SAME_UPPER",
+ )
+ # Pool2D with stride and autopadding
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32],
+ kernel_shape=[3, 3],
+ strides=[2, 2],
+ pads=None,
+ out_shape=[1, 1, 16, 16],
+ auto_pad="SAME_UPPER",
+ )
+
+ # Pool3D with stride
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32, 32],
+ kernel_shape=[3, 3, 3],
+ strides=[2, 2, 2],
+ pads=[1, 1, 1, 1, 1, 1],
+ out_shape=[1, 1, 16, 16, 16],
+ )
+
+ # Pool3D with stride and autopadding
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32, 32],
+ kernel_shape=[3, 3, 3],
+ strides=[2, 2, 2],
+ pads=None,
+ out_shape=[1, 1, 16, 16, 16],
+ auto_pad="SAME_UPPER",
+ )
+
+
+@tvm.testing.parametrize_targets
+def test_qlinear_global_average_pool(target, dev):
+ def verify_qlinear_global_average_pool(x_shape):
+ out_shape = x_shape[:2] + [1] * (len(x_shape) - 2)
+
+ node_type = "GlobalAveragePool"
+
+ input_names = ["X"]
+
+ pool_node = helper.make_node(node_type, inputs=input_names, outputs=["Y"])
Review comment:
It should be the floating point op name as mentioned in the comment above.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] mbrookhart commented on pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
mbrookhart commented on pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#issuecomment-921318067
Thanks @quic-sanirudh @anwang2009 @AndrewZhaoLuo
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] quic-sanirudh commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
quic-sanirudh commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710043241
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -654,6 +676,40 @@ def _impl_v1(cls, inputs, attr, params):
)
+class QLinearGlobalAveragePool(OnnxOpConverter):
+ "Operator converter for QLinearGlobalAveragePool from Microsoft onnxruntime contrib opset."
+
+ @classmethod
+ def _impl_v1(cls, inputs, attr, params):
+ rank = len(infer_shape(inputs[0]))
+
+ x_scale = get_scalar(inputs[1], params)
+ x_zero_point = get_scalar(inputs[2], params, dtype="int32")
+ y_scale = fold_constant(get_scalar(inputs[3], params))
+ y_zero_point = get_scalar(inputs[4], params, dtype="int32")
+
+ input_dtype = infer_type(inputs[0]).checked_type.dtype
+
+ # Onnxruntime documentation does not mention that this global avg_pool should follow the
Review comment:
Added the TODO, Thanks.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] tmoreau89 commented on pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
tmoreau89 commented on pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#issuecomment-920161820
CC @anwang2009 @mbrookhart @AndrewZhaoLuo
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] AndrewZhaoLuo commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
AndrewZhaoLuo commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r709416154
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -654,6 +676,40 @@ def _impl_v1(cls, inputs, attr, params):
)
+class QLinearGlobalAveragePool(OnnxOpConverter):
+ "Operator converter for QLinearGlobalAveragePool from Microsoft onnxruntime contrib opset."
+
+ @classmethod
+ def _impl_v1(cls, inputs, attr, params):
+ rank = len(infer_shape(inputs[0]))
+
+ x_scale = get_scalar(inputs[1], params)
+ x_zero_point = get_scalar(inputs[2], params, dtype="int32")
+ y_scale = fold_constant(get_scalar(inputs[3], params))
+ y_zero_point = get_scalar(inputs[4], params, dtype="int32")
+
+ input_dtype = infer_type(inputs[0]).checked_type.dtype
+
+ # Onnxruntime documentation does not mention that this global avg_pool should follow the
Review comment:
I'm fine with this for now but this should be a TODO since I believe the actual implementation does not dq -> pool -> q
https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/mlas/lib/qlgavgpool.cpp
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -351,6 +366,13 @@ class AveragePool(Pool):
name = "avg_pool"
+class QLinearAveragePool(Pool):
Review comment:
I think composition rather than subclassing would be a cleaner solution. Right now all the code to handle quantization + not quantization are in the same place which makes it a bit harder to read. Please separate it.
You can do something like refactor the Pool impl to a new class method like _run_calculation(...) and call it from QLinearAveragePool
##########
File path: tests/python/frontend/onnx/test_forward.py
##########
@@ -3056,6 +3056,152 @@ def verify_global_pooling(x_shape, mode):
verify_global_pooling([4, 1, 2, 6, 4], mode)
+@tvm.testing.parametrize_targets
+def test_qlinear_average_pool(target, dev):
+ def verify_qlinear_average_pool(
+ x_shape, kernel_shape, strides, pads, out_shape, auto_pad="NOTSET"
+ ):
+ input_nodes = [
+ helper.make_tensor_value_info("X", TensorProto.FLOAT, list(x_shape)),
+ ]
+
+ output_nodes = [
+ helper.make_tensor_value_info("Y", TensorProto.FLOAT, list(out_shape)),
+ ]
+
+ input_names = ["X"]
+
+ node = helper.make_node(
+ "AveragePool",
Review comment:
Should these be QLinear Nodes?
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -3794,12 +3850,14 @@ def _get_convert_map(opset):
"Xor": Renamer("logical_xor"),
# defs/nn
"AveragePool": AveragePool.get_converter(opset),
+ "QLinearAveragePool": QLinearAveragePool.get_converter(opset),
Review comment:
There's a quantization section down below, you should move these there
##########
File path: tests/python/frontend/onnx/test_forward.py
##########
@@ -3056,6 +3056,152 @@ def verify_global_pooling(x_shape, mode):
verify_global_pooling([4, 1, 2, 6, 4], mode)
+@tvm.testing.parametrize_targets
+def test_qlinear_average_pool(target, dev):
+ def verify_qlinear_average_pool(
+ x_shape, kernel_shape, strides, pads, out_shape, auto_pad="NOTSET"
+ ):
+ input_nodes = [
+ helper.make_tensor_value_info("X", TensorProto.FLOAT, list(x_shape)),
+ ]
+
+ output_nodes = [
+ helper.make_tensor_value_info("Y", TensorProto.FLOAT, list(out_shape)),
+ ]
+
+ input_names = ["X"]
+
+ node = helper.make_node(
+ "AveragePool",
+ inputs=input_names,
+ outputs=["Y"],
+ kernel_shape=kernel_shape,
+ strides=strides,
+ )
+
+ if pads is None:
+ pad_attr = helper.make_attribute("auto_pad", auto_pad)
+ else:
+ pad_attr = helper.make_attribute("pads", pads)
+ node.attribute.append(pad_attr)
+
+ graph = helper.make_graph(
+ [node],
+ "qlinear_average_pool_test",
+ inputs=input_nodes,
+ outputs=output_nodes,
+ )
+
+ model = helper.make_model(graph, producer_name="qlinear_average_pool_Test")
+ quantize_and_verify_with_ort(model, input_names, [x_shape], target, dev)
+
+ # Pool1D
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32],
+ kernel_shape=[3],
+ strides=[1],
+ pads=[1, 1],
+ out_shape=[1, 1, 32],
+ )
+ # Pool2D
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32],
+ kernel_shape=[3, 3],
+ strides=[1, 1],
+ pads=[1, 1, 1, 1],
+ out_shape=[1, 1, 32, 32],
+ )
+
+ # Pool1D with stride
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32],
+ kernel_shape=[3],
+ strides=[2],
+ pads=[1, 1],
+ out_shape=[1, 1, 16],
+ )
+ # Pool2D with stride
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32],
+ kernel_shape=[3, 3],
+ strides=[2, 2],
+ pads=[1, 1, 1, 1],
+ out_shape=[1, 1, 16, 16],
+ )
+
+ # Pool1D with stride and autopadding
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32],
+ kernel_shape=[3],
+ strides=[2],
+ pads=None,
+ out_shape=[1, 1, 16],
+ auto_pad="SAME_UPPER",
+ )
+ # Pool2D with stride and autopadding
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32],
+ kernel_shape=[3, 3],
+ strides=[2, 2],
+ pads=None,
+ out_shape=[1, 1, 16, 16],
+ auto_pad="SAME_UPPER",
+ )
+
+ # Pool3D with stride
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32, 32],
+ kernel_shape=[3, 3, 3],
+ strides=[2, 2, 2],
+ pads=[1, 1, 1, 1, 1, 1],
+ out_shape=[1, 1, 16, 16, 16],
+ )
+
+ # Pool3D with stride and autopadding
+ verify_qlinear_average_pool(
+ x_shape=[1, 1, 32, 32, 32],
+ kernel_shape=[3, 3, 3],
+ strides=[2, 2, 2],
+ pads=None,
+ out_shape=[1, 1, 16, 16, 16],
+ auto_pad="SAME_UPPER",
+ )
+
+
+@tvm.testing.parametrize_targets
+def test_qlinear_global_average_pool(target, dev):
+ def verify_qlinear_global_average_pool(x_shape):
+ out_shape = x_shape[:2] + [1] * (len(x_shape) - 2)
+
+ node_type = "GlobalAveragePool"
+
+ input_names = ["X"]
+
+ pool_node = helper.make_node(node_type, inputs=input_names, outputs=["Y"])
Review comment:
Should these be QLinear Nodes?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] quic-sanirudh commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
quic-sanirudh commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710042850
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -3794,12 +3850,14 @@ def _get_convert_map(opset):
"Xor": Renamer("logical_xor"),
# defs/nn
"AveragePool": AveragePool.get_converter(opset),
+ "QLinearAveragePool": QLinearAveragePool.get_converter(opset),
Review comment:
Thanks for the review, I've made the change.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] quic-sanirudh commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
quic-sanirudh commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710050405
##########
File path: tests/python/frontend/onnx/test_forward.py
##########
@@ -3056,6 +3056,152 @@ def verify_global_pooling(x_shape, mode):
verify_global_pooling([4, 1, 2, 6, 4], mode)
+@tvm.testing.parametrize_targets
+def test_qlinear_average_pool(target, dev):
+ def verify_qlinear_average_pool(
+ x_shape, kernel_shape, strides, pads, out_shape, auto_pad="NOTSET"
+ ):
+ input_nodes = [
+ helper.make_tensor_value_info("X", TensorProto.FLOAT, list(x_shape)),
+ ]
+
+ output_nodes = [
+ helper.make_tensor_value_info("Y", TensorProto.FLOAT, list(out_shape)),
+ ]
+
+ input_names = ["X"]
+
+ node = helper.make_node(
+ "AveragePool",
Review comment:
Actually it should be the non-quantized names, since we're only creating the floating point ops. When we quantize the graph, it gets converted to QLinearAveragePool. I printed the node that gets generated after quantization to verify, and I got the below output:
```
[node {
input: "X_quantized"
input: "X_scale"
input: "X_zero_point"
input: "Y_scale"
input: "Y_zero_point"
output: "Y_quantized"
op_type: "QLinearAveragePool"
attribute {
name: "auto_pad"
s: "SAME_UPPER"
type: STRING
}
attribute {
name: "kernel_shape"
ints: 3
ints: 3
ints: 3
type: INTS
}
attribute {
name: "strides"
ints: 2
ints: 2
ints: 2
type: INTS
}
domain: "com.microsoft"
}](url)
```
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] mbrookhart merged pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
mbrookhart merged pull request #9017:
URL: https://github.com/apache/tvm/pull/9017
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] AndrewZhaoLuo commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
AndrewZhaoLuo commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710288421
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -351,6 +366,13 @@ class AveragePool(Pool):
name = "avg_pool"
+class QLinearAveragePool(Pool):
Review comment:
Yeah this is fine
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] mbrookhart commented on a change in pull request #9017: [ONNX] QLinearAveragePool and QLinearGlobalAveragePool contrib op
Posted by GitBox <gi...@apache.org>.
mbrookhart commented on a change in pull request #9017:
URL: https://github.com/apache/tvm/pull/9017#discussion_r710553130
##########
File path: python/tvm/relay/frontend/onnx.py
##########
@@ -654,6 +676,40 @@ def _impl_v1(cls, inputs, attr, params):
)
+class QLinearGlobalAveragePool(OnnxOpConverter):
+ "Operator converter for QLinearGlobalAveragePool from Microsoft onnxruntime contrib opset."
+
+ @classmethod
+ def _impl_v1(cls, inputs, attr, params):
+ rank = len(infer_shape(inputs[0]))
+
+ x_scale = get_scalar(inputs[1], params)
+ x_zero_point = get_scalar(inputs[2], params, dtype="int32")
+ y_scale = fold_constant(get_scalar(inputs[3], params))
+ y_zero_point = get_scalar(inputs[4], params, dtype="int32")
+
+ input_dtype = infer_type(inputs[0]).checked_type.dtype
+
+ # Onnxruntime documentation does not mention that this global avg_pool should follow the
Review comment:
I have AvgPool in FakeQuantizationToInteger already, we can use that to convert to integer if needed. I haven't done GlobalAvgPool yet.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org