use super::{Node, NodeCodegen, SerializationBackend};
use crate::burn::{BurnImports, OtherType, Scope, TensorType, ToTokens, Type};
use burn::{
module::{ConstantRecord, Param, ParamId},
nn::conv::{Conv2dConfig, Conv2dRecord},
record::{PrecisionSettings, Record},
tensor::{Tensor, TensorData},
};
use proc_macro2::TokenStream;
use quote::quote;
use serde::Serialize;
#[derive(Debug, Clone)]
pub struct Conv2dNode {
pub field: OtherType,
pub input: TensorType,
pub output: TensorType,
pub data_weights: TensorData,
pub data_bias: Option<TensorData>,
pub config: Conv2dConfig,
}
impl Conv2dNode {
pub fn new<S: AsRef<str>>(
name: S,
input: TensorType,
output: TensorType,
data_weights: TensorData,
data_bias: Option<TensorData>,
config: Conv2dConfig,
) -> Self {
Self {
field: OtherType::new(
name,
quote! {
Conv2d<B>
},
),
input,
output,
data_weights,
data_bias,
config,
}
}
}
impl<PS: PrecisionSettings> NodeCodegen<PS> for Conv2dNode {
fn input_types(&self) -> Vec<Type> {
vec![Type::Tensor(self.input.clone())]
}
fn output_types(&self) -> Vec<Type> {
vec![Type::Tensor(self.output.clone())]
}
fn field_type(&self) -> Option<Type> {
Some(Type::Other(self.field.clone()))
}
fn field_init(&self) -> Option<TokenStream> {
let name = &self.field.name;
let channels = self.config.channels.to_tokens();
let kernel_size = self.config.kernel_size.to_tokens();
let stride = self.config.stride.to_tokens();
let dilation = self.config.dilation.to_tokens();
let groups = self.config.groups.to_tokens();
let padding = self.config.padding.to_tokens();
let bias = self.config.bias;
let tokens = quote! {
let #name = Conv2dConfig::new(#channels, #kernel_size)
.with_stride(#stride)
.with_padding(#padding)
.with_dilation(#dilation)
.with_groups(#groups)
.with_bias(#bias)
.init(device);
};
Some(tokens)
}
fn field_serialize<S: serde::Serializer>(&self, serializer: S) -> Result<S::Ok, S::Error> {
let device = Default::default();
let record = Conv2dRecord::<SerializationBackend> {
weight: Param::initialized(
ParamId::new(),
Tensor::from_data(
self.data_weights.clone().convert::<PS::FloatElem>(),
&device,
),
),
bias: self.data_bias.as_ref().map(|bias| {
Param::initialized(
ParamId::new(),
Tensor::from_data(bias.clone().convert::<PS::FloatElem>(), &device),
)
}),
stride: [ConstantRecord::new(); 2],
kernel_size: [ConstantRecord::new(); 2],
dilation: [ConstantRecord::new(); 2],
groups: ConstantRecord::new(),
padding: ConstantRecord::new(),
};
let item = Record::into_item::<PS>(record);
item.serialize(serializer)
}
fn forward(&self, scope: &mut Scope, node_position: usize) -> TokenStream {
let input = scope.tensor_use_owned(&self.input, node_position);
let output = &self.output.name;
let field = &self.field.name;
quote! {
let #output = self.#field.forward(#input);
}
}
fn register_imports(&self, imports: &mut BurnImports) {
imports.register("burn::nn::PaddingConfig2d");
imports.register("burn::nn::conv::Conv2d");
imports.register("burn::nn::conv::Conv2dConfig");
}
fn into_node(self) -> Node<PS> {
Node::Conv2d(self)
}
}
#[cfg(test)]
mod tests {
use super::*;
use crate::burn::{
graph::BurnGraph,
node::{conv2d::Conv2dNode, test::assert_tokens},
TensorType,
};
use burn::{nn::conv::Conv2dConfig, nn::PaddingConfig2d, record::FullPrecisionSettings};
#[test]
fn test_codegen() {
let mut graph = BurnGraph::<FullPrecisionSettings>::default();
graph.register(Conv2dNode::new(
"conv2d",
TensorType::new_float("input", 4),
TensorType::new_float("output", 4),
TensorData::from([2f32]),
None,
Conv2dConfig::new([3, 3], [3, 3]).with_padding(PaddingConfig2d::Valid),
));
graph.register_input_output(vec!["input".to_string()], vec!["output".to_string()]);
let expected = quote! {
use burn::{
module::Module,
tensor::{backend::Backend, Tensor},
};
use burn::nn::PaddingConfig2d;
use burn::nn::conv::Conv2d;
use burn::nn::conv::Conv2dConfig;
#[derive(Module, Debug)]
pub struct Model <B: Backend> {
conv2d: Conv2d<B>,
phantom: core::marker::PhantomData<B>,
device: burn::module::Ignored<B::Device>,
}
impl<B: Backend> Model <B> {
#[allow(unused_variables)]
pub fn new(device: &B::Device) -> Self {
let conv2d = Conv2dConfig::new([3, 3], [3, 3])
.with_stride([1, 1])
.with_padding(PaddingConfig2d::Valid)
.with_dilation([1, 1])
.with_groups(1)
.with_bias(true)
.init(device);
Self {
conv2d,
phantom: core::marker::PhantomData,
device: burn::module::Ignored(device.clone()),
}
}
#[allow(clippy::let_and_return, clippy::approx_constant)]
pub fn forward(&self, input: Tensor<B, 4>) -> Tensor<B, 4> {
let output = self.conv2d.forward(input);
output
}
}
};
assert_tokens(graph.codegen(), expected);
}
}