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 2020/11/11 11:52:20 UTC
[GitHub] [incubator-tvm] giuseros opened a new pull request #6899: Add initial support for quantized transpose convolution in Relay
giuseros opened a new pull request #6899:
URL: https://github.com/apache/incubator-tvm/pull/6899
This work is based on @jainris initial PR: https://github.com/apache/incubator-tvm/pull/6523
I added a relay.qnn.conv2d_transpose node. The strategy I followed is to
convert to int16 and invoke nn.conv2d_transpose (which already exists in
relay). Main changes:
- The node declaration lives in relay/qnn/op/convolution_transpose.cc
- Cast int8->int16 and subsequent offset removal is in tvm/relay/qnn/op/legalizations.py.
- I added and tested the operator in the tflite front-end
- I added a unit-test in Relay for qnn.conv2d_transpose
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] giuseros commented on a change in pull request #6899: Add initial support for quantized transpose convolution in Relay
Posted by GitBox <gi...@apache.org>.
giuseros commented on a change in pull request #6899:
URL: https://github.com/apache/incubator-tvm/pull/6899#discussion_r528801570
##########
File path: tests/python/frontend/tflite/test_forward.py
##########
@@ -1124,53 +1126,124 @@ def _test_transpose_conv(tensor_in_sizes, filter_in_sizes, output_shape, strides
total_size_1 *= s
for s in filter_in_sizes:
total_size_2 *= s
- # Initializes the input tensor with array containing incrementing
- # numbers from 1.
- data_array = [f * 1.0 for f in range(1, total_size_1 + 1)]
- filter_array = [f * 1.0 for f in range(1, total_size_2 + 1)]
with tf.Graph().as_default():
- in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype="float32")
- in_filter = constant_op.constant(filter_array, shape=filter_in_sizes, dtype="float32")
- strides = [1] + strides + [1]
- # in_filter layout is HWOI
- out = nn_ops.conv2d_transpose(
- in_data, in_filter, output_shape=output_shape, strides=strides, padding=padding
- )
- data_array = np.reshape(data_array, tensor_in_sizes).astype("float32")
- compare_tflite_with_tvm(data_array, "Placeholder:0", [in_data], [out])
+ if quantized:
+ # Initializes the input tensor with array containing incrementing
+ # numbers from 1.
+ data_array = [max(f, 255) for f in range(1, total_size_1 + 1)]
+ filter_array = [max(f, 255) for f in range(1, total_size_2 + 1)]
+ data_array = np.reshape(data_array, tensor_in_sizes).astype("uint8")
+ filter_array = np.reshape(filter_array, filter_in_sizes).astype("uint8")
+
+ in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype="float32", name="in_data")
+ inq_data = tf.quantization.fake_quant_with_min_max_args(
+ in_data, min=-100, max=100, name="q_data"
+ )
+ input_range = {"q_data": (-100, 100)}
+
+ in_filter = constant_op.constant(
+ filter_array, shape=filter_in_sizes, dtype="float32", name="in_filter"
+ )
+ inq_filter = tf.quantization.fake_quant_with_min_max_args(
+ in_filter, min=-100, max=100, name="q_filter"
+ )
+
+ strides = [1] + strides + [1]
+
+ out = nn_ops.conv2d_transpose(
+ inq_data, inq_filter, output_shape=output_shape, strides=strides, padding=padding
+ )
+ out = tf.quantization.fake_quant_with_min_max_args(out, min=-100, max=100, name="out")
+ compare_tflite_with_tvm(
+ [data_array], ["q_data"], [inq_data], [out], quantized=True, input_range=input_range
+ )
+ else:
+ # Initializes the input tensor with array containing incrementing
+ # numbers from 1.
+ data_array = [f * 1.0 for f in range(1, total_size_1 + 1)]
+ filter_array = [f * 1.0 for f in range(1, total_size_2 + 1)]
+
+ in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype="float32", name="in_data")
+ in_filter = constant_op.constant(
+ filter_array, shape=filter_in_sizes, dtype="float32", name="in_filter"
+ )
+ strides = [1] + strides + [1]
+ # in_filter layout is HWOI
+ out = nn_ops.conv2d_transpose(
+ in_data, in_filter, output_shape=output_shape, strides=strides, padding=padding
+ )
+ data_array = np.reshape(data_array, tensor_in_sizes).astype("float32")
+ compare_tflite_with_tvm([data_array], ["in_data"], [in_data], [out])
def test_forward_transpose_conv():
- # kernel 3x3, padding VALID
- _test_transpose_conv([4, 32, 32, 16], [3, 3, 5, 16], [4, 34, 34, 5], [1, 1], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 65, 65, 5], [2, 2], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 65, 34, 5], [2, 1], "VALID")
-
- # kernel 3x3, padding SAME
- _test_transpose_conv([4, 32, 32, 16], [3, 3, 5, 16], [4, 32, 32, 5], [1, 1], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 64, 64, 5], [2, 2], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 64, 32, 5], [2, 1], "SAME")
-
- # kernel 2x2, padding VALID
- _test_transpose_conv([4, 32, 32, 16], [2, 2, 5, 16], [4, 33, 33, 5], [1, 1], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 64, 5], [2, 2], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 33, 5], [2, 1], "VALID")
-
- # kernel 2x2, padding SAME
- _test_transpose_conv([4, 32, 32, 16], [2, 2, 5, 16], [4, 32, 32, 5], [1, 1], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 64, 5], [2, 2], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 32, 5], [2, 1], "SAME")
-
- # kernel 1x1, padding VALID
- _test_transpose_conv([4, 32, 32, 16], [1, 1, 5, 16], [4, 32, 32, 5], [1, 1], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 63, 5], [2, 2], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 32, 5], [2, 1], "VALID")
-
- # kernel 1x1, padding SAME
- _test_transpose_conv([4, 32, 32, 16], [1, 1, 5, 16], [4, 32, 32, 5], [1, 1], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 63, 5], [2, 2], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 32, 5], [2, 1], "SAME")
+ for quantized in [True, False]:
Review comment:
There is a test for per-channel quantization in `test_op_qnn_conv2d_transpose.py`. `test_forward.py` for TFlite does not seem to contain any per-channel quantized tests for any of the operators under test (while there are some per-channel tests in the pytorch front-end).
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] mbaret commented on a change in pull request #6899: Add initial support for quantized transpose convolution in Relay
Posted by GitBox <gi...@apache.org>.
mbaret commented on a change in pull request #6899:
URL: https://github.com/apache/incubator-tvm/pull/6899#discussion_r528696443
##########
File path: python/tvm/relay/qnn/op/legalizations.py
##########
@@ -46,6 +52,25 @@ def qnn_conv2d_legalize(attrs, inputs, types):
return None
+# Generic QNN Conv2Transpose legalization function.
Review comment:
```suggestion
# Generic QNN Conv2DTranspose legalization function.
```
##########
File path: python/tvm/relay/qnn/op/legalizations.py
##########
@@ -46,6 +52,25 @@ def qnn_conv2d_legalize(attrs, inputs, types):
return None
+# Generic QNN Conv2Transpose legalization function.
+@tvm.target.generic_func
+def qnn_conv2d_transpose_legalize(attrs, inputs, types):
+ """Convert kernel and data to int16, subtract offsets upfront
+ and calls into relay.nn.conv2d_transpose."""
+
+ # Collect the input exprs.
+ data, kernel, input_zero_point, kernel_zero_point, _, _ = inputs
+
+ shift_data = relay.subtract(
+ relay.cast(data, dtype="int16"), relay.cast(input_zero_point, "int16")
+ )
+ shift_kernel = relay.subtract(
+ relay.cast(kernel, dtype="int16"), relay.cast(kernel_zero_point, "int16")
+ )
+ new_attrs = {k: attrs[k] for k in attrs.keys()}
Review comment:
Could you explain what's going on here? I'm assuming attrs is some TVM type that's not-quite-a-dict and doesn't work with the ** syntax?
##########
File path: python/tvm/relay/qnn/op/qnn.py
##########
@@ -296,6 +296,101 @@ def conv2d(
)
+def conv2d_transpose(
+ data,
+ weight,
+ input_zero_point,
+ kernel_zero_point,
+ input_scale,
+ kernel_scale,
+ strides=(1, 1),
+ padding=(0, 0),
+ dilation=(1, 1),
+ groups=1,
+ channels=None,
+ kernel_size=None,
+ data_layout="NCHW",
+ kernel_layout="OIHW",
+ out_layout="",
+ output_padding=(0, 0),
+ out_dtype="",
+):
+ """This operator deconvolves quantized data with quantized kernel. The scale of
+ the output quantized tensor is the product of the kernel_scale and
+ input_scale of the input quantized tensors. The zero point of the output
+ quantized tensor is 0. By default, the dtype of output is int32. Please also
+ refer to Requantize operator to understand how to scale back the int32
+ output to (u)int8.
+
+ Parameters
+ ----------
+ data : tvm.relay.Expr
+ The input data to the operator.
+
+ weight : tvm.relay.Expr
+ The weight expressions.
+
+ strides : Tuple[int], optional
+ The strides of convolution.
+
+ padding : Tuple[int], optional
+ The padding of convolution on both sides of inputs.
+
+ dilation : Tuple[int], optional
+ Specifies the dilation rate to be used for dilated convolution.
+
+ channels : int, optional
+ Number of output channels of this convolution.
+
+ kernel_size : tuple of int, optional
+ The spatial of the convolution kernel.
Review comment:
fix/clarify
##########
File path: src/relay/qnn/op/convolution_transpose.cc
##########
@@ -0,0 +1,154 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/qnn/op/convolution.cc
+ * \brief Property def of qnn convolution operator.
Review comment:
```suggestion
* \brief Property def of qnn transpose convolution operator.
```
##########
File path: python/tvm/relay/qnn/op/qnn.py
##########
@@ -296,6 +296,101 @@ def conv2d(
)
+def conv2d_transpose(
+ data,
+ weight,
+ input_zero_point,
+ kernel_zero_point,
+ input_scale,
+ kernel_scale,
+ strides=(1, 1),
+ padding=(0, 0),
+ dilation=(1, 1),
+ groups=1,
+ channels=None,
+ kernel_size=None,
+ data_layout="NCHW",
+ kernel_layout="OIHW",
+ out_layout="",
+ output_padding=(0, 0),
+ out_dtype="",
+):
+ """This operator deconvolves quantized data with quantized kernel. The scale of
+ the output quantized tensor is the product of the kernel_scale and
+ input_scale of the input quantized tensors. The zero point of the output
+ quantized tensor is 0. By default, the dtype of output is int32. Please also
+ refer to Requantize operator to understand how to scale back the int32
+ output to (u)int8.
+
+ Parameters
+ ----------
+ data : tvm.relay.Expr
+ The input data to the operator.
+
+ weight : tvm.relay.Expr
+ The weight expressions.
+
+ strides : Tuple[int], optional
+ The strides of convolution.
+
+ padding : Tuple[int], optional
+ The padding of convolution on both sides of inputs.
Review comment:
Unclear what 'both sides of the inputs' means. I think this can also be specified 4 ways?
##########
File path: src/relay/qnn/op/convolution_transpose.cc
##########
@@ -0,0 +1,154 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/qnn/op/convolution.cc
Review comment:
```suggestion
* \file src/relay/qnn/op/convolution_transpose.cc
```
##########
File path: python/tvm/relay/qnn/op/qnn.py
##########
@@ -296,6 +296,101 @@ def conv2d(
)
+def conv2d_transpose(
+ data,
+ weight,
+ input_zero_point,
+ kernel_zero_point,
+ input_scale,
+ kernel_scale,
+ strides=(1, 1),
+ padding=(0, 0),
+ dilation=(1, 1),
+ groups=1,
+ channels=None,
+ kernel_size=None,
+ data_layout="NCHW",
+ kernel_layout="OIHW",
+ out_layout="",
+ output_padding=(0, 0),
+ out_dtype="",
+):
+ """This operator deconvolves quantized data with quantized kernel. The scale of
+ the output quantized tensor is the product of the kernel_scale and
+ input_scale of the input quantized tensors. The zero point of the output
+ quantized tensor is 0. By default, the dtype of output is int32. Please also
+ refer to Requantize operator to understand how to scale back the int32
+ output to (u)int8.
+
+ Parameters
Review comment:
The qnn params are missing from this.
##########
File path: tests/python/frontend/tflite/test_forward.py
##########
@@ -1124,53 +1126,124 @@ def _test_transpose_conv(tensor_in_sizes, filter_in_sizes, output_shape, strides
total_size_1 *= s
for s in filter_in_sizes:
total_size_2 *= s
- # Initializes the input tensor with array containing incrementing
- # numbers from 1.
- data_array = [f * 1.0 for f in range(1, total_size_1 + 1)]
- filter_array = [f * 1.0 for f in range(1, total_size_2 + 1)]
with tf.Graph().as_default():
- in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype="float32")
- in_filter = constant_op.constant(filter_array, shape=filter_in_sizes, dtype="float32")
- strides = [1] + strides + [1]
- # in_filter layout is HWOI
- out = nn_ops.conv2d_transpose(
- in_data, in_filter, output_shape=output_shape, strides=strides, padding=padding
- )
- data_array = np.reshape(data_array, tensor_in_sizes).astype("float32")
- compare_tflite_with_tvm(data_array, "Placeholder:0", [in_data], [out])
+ if quantized:
+ # Initializes the input tensor with array containing incrementing
+ # numbers from 1.
+ data_array = [max(f, 255) for f in range(1, total_size_1 + 1)]
+ filter_array = [max(f, 255) for f in range(1, total_size_2 + 1)]
+ data_array = np.reshape(data_array, tensor_in_sizes).astype("uint8")
+ filter_array = np.reshape(filter_array, filter_in_sizes).astype("uint8")
+
+ in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype="float32", name="in_data")
+ inq_data = tf.quantization.fake_quant_with_min_max_args(
+ in_data, min=-100, max=100, name="q_data"
+ )
+ input_range = {"q_data": (-100, 100)}
+
+ in_filter = constant_op.constant(
+ filter_array, shape=filter_in_sizes, dtype="float32", name="in_filter"
+ )
+ inq_filter = tf.quantization.fake_quant_with_min_max_args(
+ in_filter, min=-100, max=100, name="q_filter"
+ )
+
+ strides = [1] + strides + [1]
+
+ out = nn_ops.conv2d_transpose(
+ inq_data, inq_filter, output_shape=output_shape, strides=strides, padding=padding
+ )
+ out = tf.quantization.fake_quant_with_min_max_args(out, min=-100, max=100, name="out")
+ compare_tflite_with_tvm(
+ [data_array], ["q_data"], [inq_data], [out], quantized=True, input_range=input_range
+ )
+ else:
+ # Initializes the input tensor with array containing incrementing
+ # numbers from 1.
+ data_array = [f * 1.0 for f in range(1, total_size_1 + 1)]
+ filter_array = [f * 1.0 for f in range(1, total_size_2 + 1)]
+
+ in_data = array_ops.placeholder(shape=tensor_in_sizes, dtype="float32", name="in_data")
+ in_filter = constant_op.constant(
+ filter_array, shape=filter_in_sizes, dtype="float32", name="in_filter"
+ )
+ strides = [1] + strides + [1]
+ # in_filter layout is HWOI
+ out = nn_ops.conv2d_transpose(
+ in_data, in_filter, output_shape=output_shape, strides=strides, padding=padding
+ )
+ data_array = np.reshape(data_array, tensor_in_sizes).astype("float32")
+ compare_tflite_with_tvm([data_array], ["in_data"], [in_data], [out])
def test_forward_transpose_conv():
- # kernel 3x3, padding VALID
- _test_transpose_conv([4, 32, 32, 16], [3, 3, 5, 16], [4, 34, 34, 5], [1, 1], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 65, 65, 5], [2, 2], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 65, 34, 5], [2, 1], "VALID")
-
- # kernel 3x3, padding SAME
- _test_transpose_conv([4, 32, 32, 16], [3, 3, 5, 16], [4, 32, 32, 5], [1, 1], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 64, 64, 5], [2, 2], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [3, 3, 5, 16], [1, 64, 32, 5], [2, 1], "SAME")
-
- # kernel 2x2, padding VALID
- _test_transpose_conv([4, 32, 32, 16], [2, 2, 5, 16], [4, 33, 33, 5], [1, 1], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 64, 5], [2, 2], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 33, 5], [2, 1], "VALID")
-
- # kernel 2x2, padding SAME
- _test_transpose_conv([4, 32, 32, 16], [2, 2, 5, 16], [4, 32, 32, 5], [1, 1], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 64, 5], [2, 2], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [2, 2, 5, 16], [1, 64, 32, 5], [2, 1], "SAME")
-
- # kernel 1x1, padding VALID
- _test_transpose_conv([4, 32, 32, 16], [1, 1, 5, 16], [4, 32, 32, 5], [1, 1], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 63, 5], [2, 2], "VALID")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 32, 5], [2, 1], "VALID")
-
- # kernel 1x1, padding SAME
- _test_transpose_conv([4, 32, 32, 16], [1, 1, 5, 16], [4, 32, 32, 5], [1, 1], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 63, 5], [2, 2], "SAME")
- _test_transpose_conv([1, 32, 32, 16], [1, 1, 5, 16], [1, 63, 32, 5], [2, 1], "SAME")
+ for quantized in [True, False]:
Review comment:
Does any of this test the per-channel quantization?
##########
File path: python/tvm/relay/qnn/op/qnn.py
##########
@@ -296,6 +296,101 @@ def conv2d(
)
+def conv2d_transpose(
+ data,
+ weight,
+ input_zero_point,
+ kernel_zero_point,
+ input_scale,
+ kernel_scale,
+ strides=(1, 1),
+ padding=(0, 0),
+ dilation=(1, 1),
+ groups=1,
+ channels=None,
+ kernel_size=None,
+ data_layout="NCHW",
+ kernel_layout="OIHW",
+ out_layout="",
+ output_padding=(0, 0),
+ out_dtype="",
+):
+ """This operator deconvolves quantized data with quantized kernel. The scale of
+ the output quantized tensor is the product of the kernel_scale and
+ input_scale of the input quantized tensors. The zero point of the output
+ quantized tensor is 0. By default, the dtype of output is int32. Please also
+ refer to Requantize operator to understand how to scale back the int32
+ output to (u)int8.
+
+ Parameters
+ ----------
+ data : tvm.relay.Expr
+ The input data to the operator.
+
+ weight : tvm.relay.Expr
+ The weight expressions.
+
+ strides : Tuple[int], optional
+ The strides of convolution.
+
+ padding : Tuple[int], optional
+ The padding of convolution on both sides of inputs.
+
+ dilation : Tuple[int], optional
+ Specifies the dilation rate to be used for dilated convolution.
+
+ channels : int, optional
+ Number of output channels of this convolution.
+
+ kernel_size : tuple of int, optional
+ The spatial of the convolution kernel.
+
+ groups : int, optional
+ Number of groups for grouped convolution.
+
+ data_layout : str, optional
+ Layout of the input.
+
+ kernel_layout : str, optional
+ Layout of the weight.
+
+ out_layout : Optional[str]
+ Layout of the output, by default, out_layout is the same as data_layout
+
+ output_padding : Tuple[int], optional
+ Used to disambiguate the output shape.
Review comment:
I don't understand this one
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] mbaret commented on pull request #6899: Add initial support for quantized transpose convolution in Relay
Posted by GitBox <gi...@apache.org>.
mbaret commented on pull request #6899:
URL: https://github.com/apache/incubator-tvm/pull/6899#issuecomment-732404232
cc @siju-samuel @FrozenGene if you're interested
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [incubator-tvm] giuseros commented on pull request #6899: Add initial support for quantized transpose convolution in Relay
Posted by GitBox <gi...@apache.org>.
giuseros commented on pull request #6899:
URL: https://github.com/apache/incubator-tvm/pull/6899#issuecomment-725380519
cc @anijain2305 @u99127
----------------------------------------------------------------
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.
For queries about this service, please contact Infrastructure at:
users@infra.apache.org