Merge pull request #48 from robertknight/model-metadata

Add initial support for model metadata in RTen models

rten-cli/src/main.rs

@@ -3,7 +3,7 @@ use std::error::Error;
 use std::fs;
 use std::time::Instant;
 
-use rten::{Dimension, Input, Model, NodeId, Output, RunOptions};
+use rten::{Dimension, Input, Model, ModelMetadata, NodeId, Output, RunOptions};
 use rten_tensor::prelude::*;
 use rten_tensor::Tensor;
 
@@ -37,6 +37,11 @@ fn parse_args() -> Result<Args, lexopt::Error> {
 
 Usage: {bin_name} [OPTIONS] <model>
 
+Args:
+  <model>
+    Path to '.rten' model to inspect and run.
+
+Options:
   -t, --timing   Output timing info
   -v, --verbose  Enable verbose logging
   -h, --help     Print help
@@ -66,28 +71,27 @@ fn format_param_count(n: usize) -> String {
     }
 }
 
-/// Tool for inspecting converted ONNX models and running them with randomly
-/// generated inputs.
-///
-/// ```
-/// tools/convert-onnx.py model.onnx output.rten
-/// cargo run -p rten-cli --release output.rten
-/// ```
-///
-/// To get detailed timing information set the `RTEN_TIMING` env var before
-/// running. See `docs/profiling.md`.
-fn main() -> Result<(), Box<dyn Error>> {
-    let args = parse_args()?;
-    let model_bytes = fs::read(args.model)?;
-    let model = Model::load(&model_bytes)?;
+fn print_metadata(metadata: &ModelMetadata) {
+    fn print_field<T: std::fmt::Display>(name: &str, value: Option<T>) {
+        if let Some(value) = value {
+            println!("  {}: {}", name, value);
+        }
+    }
 
-    println!(
-        "Model stats: {} inputs, {} outputs, {} params",
-        model.input_ids().len(),
-        model.output_ids().len(),
-        format_param_count(model.total_params()),
-    );
+    println!("Metadata:");
+    print_field("ONNX hash", metadata.onnx_hash());
+    print_field("Description", metadata.description());
+    print_field("License", metadata.license());
+    print_field("Commit", metadata.commit());
+    print_field("Repository", metadata.code_repository());
+    print_field("Model repository", metadata.model_repository());
+    print_field("Run ID", metadata.run_id());
+    print_field("Run URL", metadata.run_url());
+}
 
+/// Generate random inputs for `model` using shape metadata and heuristics,
+/// run it, and print details of the output.
+fn run_with_random_input(model: &Model, run_opts: RunOptions) -> Result<(), Box<dyn Error>> {
     let mut rng = fastrand::Rng::new();
 
     // Generate random ints that are likely to be valid token IDs in a language
@@ -165,27 +169,21 @@ fn main() -> Result<(), Box<dyn Error>> {
             .as_ref()
             .and_then(|ni| ni.name())
             .unwrap_or("(unnamed)");
-        println!("Input \"{name}\" resolved shape {:?}", input.shape());
+        println!("  Input \"{name}\" generated shape {:?}", input.shape());
     }
 
     // Run model and summarize outputs.
     let start = Instant::now();
-    let outputs = model.run(
-        &inputs,
-        model.output_ids(),
-        Some(RunOptions {
-            timing: args.timing,
-            verbose: args.verbose,
-            ..Default::default()
-        }),
-    )?;
+    let outputs = model.run(&inputs, model.output_ids(), Some(run_opts))?;
     let elapsed = start.elapsed().as_millis();
 
+    println!();
     println!(
-        "Model returned {} outputs in {:.2}ms",
+        "  Model returned {} outputs in {:.2}ms.",
         outputs.len(),
         elapsed
     );
+    println!();
 
     let output_names: Vec<String> = model
         .output_ids()
@@ -204,11 +202,50 @@ fn main() -> Result<(), Box<dyn Error>> {
             Output::IntTensor(_) => "i32",
         };
         println!(
-            "Output {i} \"{name}\" data type {} shape: {:?}",
+            "  Output {i} \"{name}\" data type {} shape: {:?}",
             dtype,
             output.shape()
         );
     }
 
     Ok(())
 }
+
+/// Tool for inspecting converted ONNX models and running them with randomly
+/// generated inputs.
+///
+/// ```
+/// tools/convert-onnx.py model.onnx output.rten
+/// cargo run -p rten-cli --release output.rten
+/// ```
+///
+/// To get detailed timing information set the `RTEN_TIMING` env var before
+/// running. See `docs/profiling.md`.
+fn main() -> Result<(), Box<dyn Error>> {
+    let args = parse_args()?;
+    let model_bytes = fs::read(args.model)?;
+    let model = Model::load(&model_bytes)?;
+
+    println!(
+        "Model summary: {} inputs, {} outputs, {} params",
+        model.input_ids().len(),
+        model.output_ids().len(),
+        format_param_count(model.total_params()),
+    );
+    println!();
+
+    print_metadata(model.metadata());
+
+    println!();
+    println!("Running model with random inputs...");
+    run_with_random_input(
+        &model,
+        RunOptions {
+            timing: args.timing,
+            verbose: args.verbose,
+            ..Default::default()
+        },
+    )?;
+
+    Ok(())
+}

src/lib.rs

@@ -28,6 +28,7 @@ mod gemm;
 mod graph;
 mod iter_util;
 mod model;
+mod model_metadata;
 mod number;
 mod slice_reductions;
 mod timer;
@@ -44,6 +45,7 @@ pub mod ops;
 
 pub use graph::{Dimension, NodeId, RunOptions};
 pub use model::{DefaultOperatorFactory, Model, ModelLoadError, NodeInfo, OpRegistry};
+pub use model_metadata::ModelMetadata;
 pub use ops::{FloatOperators, Input, Operators, Output};
 pub use timer::Timer;
 pub use timing::TimingSort;

src/model.rs

@@ -9,6 +9,7 @@ use rten_tensor::Tensor;
 use smallvec::smallvec;
 
 use crate::graph::{Dimension, Graph, Node, NodeId, RunError, RunOptions};
+use crate::model_metadata::ModelMetadata;
 use crate::ops;
 use crate::ops::{
     BoxOrder, CoordTransformMode, DataType, Direction, Input, NearestMode, Operator, Output,
@@ -52,6 +53,7 @@ pub struct Model {
     input_ids: Vec<NodeId>,
     output_ids: Vec<NodeId>,
     graph: Graph,
+    metadata: ModelMetadata,
 }
 
 /// Provides access to metadata about a graph node.
@@ -144,6 +146,11 @@ impl Model {
         self.graph.get_node(id).map(|node| NodeInfo { node })
     }
 
+    /// Return metadata about the model.
+    pub fn metadata(&self) -> &ModelMetadata {
+        &self.metadata
+    }
+
     /// Return the IDs of input nodes.
     pub fn input_ids(&self) -> &[NodeId] {
         &self.input_ids
@@ -1108,11 +1115,17 @@ fn load_model(data: &[u8], registry: &OpRegistry) -> Result<Model, ModelLoadErro
         }
     }
 
+    let metadata = model
+        .metadata()
+        .map(ModelMetadata::deserialize)
+        .unwrap_or_default();
+
     let model = Model {
         node_ids: node_id_from_name,
         input_ids,
         output_ids,
         graph,
+        metadata,
     };
     Ok(model)
 }
@@ -1126,7 +1139,7 @@ mod tests {
 
     use crate::graph::{Dimension, RunError};
     use crate::model::Model;
-    use crate::model_builder::{ModelBuilder, OpType};
+    use crate::model_builder::{MetadataArgs, ModelBuilder, OpType};
     use crate::ops;
     use crate::ops::{BoxOrder, CoordTransformMode, NearestMode, OpError, ResizeMode, Scalar};
 
@@ -1157,6 +1170,10 @@ mod tests {
         );
         builder.add_operator("relu", OpType::Relu, &[Some(concat_out)], &[output_node]);
 
+        builder.add_metadata(MetadataArgs {
+            onnx_hash: Some("abc".to_string()),
+        });
+
         builder.finish()
     }
 
@@ -1198,6 +1215,14 @@ mod tests {
         assert_eq!(shape, &[1, 2, 2].map(Dimension::Fixed));
     }
 
+    #[test]
+    fn test_metadata() {
+        let buffer = generate_model_buffer();
+        let model = Model::load(&buffer).unwrap();
+        assert_eq!(model.metadata().onnx_hash(), Some("abc"));
+        assert_eq!(model.metadata().description(), None);
+    }
+
     #[test]
     fn test_input_shape() {
         let buffer = generate_model_buffer();

src/model_builder.rs

@@ -118,6 +118,7 @@ pub struct ModelBuilder<'a> {
     nodes: Vec<WIPOffset<sg::Node<'a>>>,
     input_ids: Vec<u32>,
     output_ids: Vec<u32>,
+    metadata: Option<WIPOffset<sg::Metadata<'a>>>,
 }
 
 enum NodeData<'a> {
@@ -126,6 +127,11 @@ enum NodeData<'a> {
     Operator(WIPOffset<sg::OperatorNode<'a>>),
 }
 
+/// Arguments for [ModelBuilder::add_metadata].
+pub struct MetadataArgs {
+    pub onnx_hash: Option<String>,
+}
+
 struct PadArgs {
     pad_mode: sg::PadMode,
     pads: Option<Vec<usize>>,
@@ -152,6 +158,7 @@ impl<'a> ModelBuilder<'a> {
             nodes: Vec::new(),
             input_ids: Vec::new(),
             output_ids: Vec::new(),
+            metadata: None,
         }
     }
 
@@ -683,6 +690,19 @@ impl<'a> ModelBuilder<'a> {
         self.output_ids.push(node_id);
     }
 
+    /// Add model metadata
+    pub fn add_metadata(&mut self, metadata: MetadataArgs) {
+        let hash = metadata
+            .onnx_hash
+            .as_ref()
+            .map(|hash| self.builder.create_string(hash));
+        let mut meta_builder = sg::MetadataBuilder::new(&mut self.builder);
+        if let Some(hash) = hash {
+            meta_builder.add_onnx_hash(hash);
+        }
+        self.metadata = Some(meta_builder.finish());
+    }
+
     /// Finish writing the model data to the buffer and return the buffer's contents.
     pub fn finish(mut self) -> Vec<u8> {
         let inputs_vec = self.builder.create_vector(&self.input_ids[..]);
@@ -703,6 +723,7 @@ impl<'a> ModelBuilder<'a> {
             &sg::ModelArgs {
                 schema_version: 1,
                 graph: Some(graph),
+                metadata: self.metadata,
             },
         );