Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Make leela2onnx Conv-nodes compatible with onnx2pytorch #1924

Merged
merged 9 commits into from
Nov 14, 2023
Merged

Make leela2onnx Conv-nodes compatible with onnx2pytorch #1924

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
9 commits
Select commit Hold shift + click to select a range
78ca4c9
Accomodating onnx2pytorch for conv-blocks
patrik-ha Oct 10, 2023
94c5d7c
Get kernel dims from weights, not as argument
patrik-ha Oct 11, 2023
8762927
alternate onnx layernorm implementation with explicit fp32 casting
borg323 Sep 17, 2023
0858bcd
Make alternative LayerNorm default for ONNX
patrik-ha Oct 11, 2023
4e01eac
Revert "Make alternative LayerNorm default for ONNX"
patrik-ha Oct 12, 2023
5a0c48f
Switch for supporting onnx2pytorch
patrik-ha Oct 12, 2023
c73e7fb
Make option for alt-ln more readable
patrik-ha Oct 13, 2023
b7e250d
Make option for alt-ln more readable
patrik-ha Oct 13, 2023
f7fe6ea
Merge branch 'onnx-pytorch' of https://github.com/patrik-ha/lc0 into …
patrik-ha Oct 13, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
5 changes: 4 additions & 1 deletion src/neural/onnx/builder.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -135,12 +135,15 @@ std::string PopulateStdNodeFields(pblczero::NodeProto* node,
std::string OnnxBuilder::Conv(const std::string& name,
const std::string& input_name,
const OnnxConst& kernel_weights,
const OnnxConst& bias_weights, int pads) {
const OnnxConst& bias_weights,
int filters,
Copy link
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why not just do a 'kernel_weights.GetDimensions().back()` instead of passing the number of filters?

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Agreed, that's better

int pads) {
auto* node = model_.mutable_graph()->add_node();
auto out = PopulateStdNodeFields(node, name, input_name, "Conv");
node->add_input(AddInitializer(name + "/w/kernel", kernel_weights));
node->add_input(AddInitializer(name + "/w/bias", bias_weights));
AddIntsAttribute(node, "pads", {pads, pads, pads, pads});
AddIntsAttribute(node, "kernel_shape", {filters, filters});
return out;
}

Expand Down
2 changes: 1 addition & 1 deletion src/neural/onnx/builder.h
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -61,7 +61,7 @@ class OnnxBuilder {
// node name.
std::string Conv(const std::string& name, const std::string& input_name,
const OnnxConst& kernel_weights,
const OnnxConst& bias_weights, int pads = 1);
const OnnxConst& bias_weights, int filters, int pads = 1);
std::string Add(const std::string& name, const std::string& input1,
const std::string& input2);
std::string Add(const std::string& name, const std::string& input1,
Expand Down
1 change: 1 addition & 0 deletions src/neural/onnx/converter.cc
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -268,6 +268,7 @@ std::string Converter::MakeConvBlock(
*GetWeghtsConverter(weights.weights,
{output_channels, input_channels, filters, filters}),
*GetWeghtsConverter(weights.biases, {output_channels}),
filters,
(filters - 1) / 2);

if (seunit) flow = MakeSqueezeAndExcite(builder, *seunit, flow, name + "/se");
Expand Down