This is an automated email from the ASF dual-hosted git repository.
ekalda pushed a commit to branch main
in repository https://gitbox.apache.org/repos/asf/tvm.git
The following commit(s) were added to refs/heads/main by this push:
new 94f6b37b06 [Relay][TFLite] Fix in qnn.conv2d when parameter groups not
equal to 1 (#15472)
94f6b37b06 is described below
commit 94f6b37b063d10fceea561288b3674d9ed32cfcd
Author: padreofthegame <[email protected]>
AuthorDate: Mon Aug 14 12:20:32 2023 +0200
[Relay][TFLite] Fix in qnn.conv2d when parameter groups not equal to 1
(#15472)
Currently, qnn.conv2d only supports parameter GROUPS to be equal 1 or depth
of the input tensor,
while in any other case (which is valid for parameter GROUPS) it results in
Relay type checker
mismatch error related to types[5] argument.
This fix expands qnn.conv2d functionality for other valid values of
parameter GROUPS.
---
src/relay/qnn/op/convolution.cc | 17 +++++++++++---
tests/python/frontend/tflite/test_forward.py | 35 +++++++++++++++++++++++++++-
2 files changed, 48 insertions(+), 4 deletions(-)
diff --git a/src/relay/qnn/op/convolution.cc b/src/relay/qnn/op/convolution.cc
index f5ac6af1df..2ce5523764 100644
--- a/src/relay/qnn/op/convolution.cc
+++ b/src/relay/qnn/op/convolution.cc
@@ -75,13 +75,24 @@ bool QnnConv2DRel(const Array<Type>& types, int num_inputs,
const Attrs& attrs,
ICHECK(axis != std::string::npos) << "Kernel layout attribute is not
defined";
AssignType(types[5], DataType::Float(32), weight->shape[axis], reporter);
// weight_scale
} else {
- // Here, total number of output channels depend on depth multiplier.
size_t o_axis = param->kernel_layout.operator std::string().find('O');
size_t i_axis = param->kernel_layout.operator std::string().find('I');
+ size_t d_axis = param->data_layout.operator std::string().find('C');
ICHECK(o_axis != std::string::npos || i_axis != std::string::npos)
<< "Kernel layout attribute is not defined";
- AssignType(types[5], DataType::Float(32), weight->shape[i_axis] *
weight->shape[o_axis],
- reporter); // weight_scale
+ ICHECK(d_axis != std::string::npos) << "Data layout attribute is not
defined";
+ if (param->channels.defined() &&
+ tvm::tir::ExprDeepEqual()(param->groups, data->shape[d_axis]) &&
+ tvm::tir::ExprDeepEqual()(param->channels, weight->shape[i_axis])) {
+ // This is depthwise convolution
+ // Here, total number of output channels depend on depth multiplier.
+ AssignType(types[5], DataType::Float(32), weight->shape[i_axis] *
weight->shape[o_axis],
+ reporter); // weight_scale
+ } else {
+ // This is grouped convolution
+ AssignType(types[5], DataType::Float(32), param->channels,
+ reporter); // weight_scale
+ }
}
// Collect the input tensor and output tensor devoid of scale and zero
points to reuse Relay
diff --git a/tests/python/frontend/tflite/test_forward.py
b/tests/python/frontend/tflite/test_forward.py
index f00624dc03..567139d0c2 100644
--- a/tests/python/frontend/tflite/test_forward.py
+++ b/tests/python/frontend/tflite/test_forward.py
@@ -1001,7 +1001,13 @@ def test_forward_l2_pool2d():
def _test_tflite2_quantized_convolution(
- input_shape, kernel_shape, filters, padding="valid", data_format=None,
int_quant_dtype=tf.int8
+ input_shape,
+ kernel_shape,
+ filters,
+ padding="valid",
+ data_format=None,
+ int_quant_dtype=tf.int8,
+ groups=1,
):
"""One iteration of TFLite2 quantized convolution with given shapes and
attributes"""
data_format = "channels_last" if data_format == "NHWC" else
"channels_first"
@@ -1015,6 +1021,7 @@ def _test_tflite2_quantized_convolution(
activation=tf.nn.relu,
padding=padding,
data_format=data_format,
+ groups=groups,
)(data_in)
keras_model = tf.keras.models.Model(data_in, conv)
@@ -1076,6 +1083,32 @@ def test_forward_quantized_convolution():
int_quant_dtype=int_quant_dtype,
)
+ _test_tflite2_quantized_convolution(
+ (64, 2, 28, 28),
+ (1, 1),
+ 12,
+ data_format="NCWH",
+ int_quant_dtype=int_quant_dtype,
+ )
+
+ _test_tflite2_quantized_convolution(
+ (1, 16, 10, 10),
+ (3, 3),
+ 2,
+ data_format="NCWH",
+ int_quant_dtype=int_quant_dtype,
+ groups=2,
+ )
+
+ _test_tflite2_quantized_convolution(
+ (2, 32, 28, 28),
+ (1, 1),
+ 16,
+ data_format="NCWH",
+ int_quant_dtype=int_quant_dtype,
+ groups=8,
+ )
+
def test_forward_quantized_depthwise_convolution():
for int_quant_dtype in [tf.int8, tf.int16]:
