ci: update version string in docs

BREAKING CHANGE: compiled model serialization is not backwards compatible

.github/workflows/rust.yml

@@ -781,6 +781,8 @@ jobs:
         run: python -m venv .env --clear; source .env/bin/activate; pip install -r requirements.txt; python -m ensurepip --upgrade
       - name: Build python ezkl
         run: source .env/bin/activate; unset CONDA_PREFIX; maturin develop --features python-bindings --profile=test-runs
+      - name: Cat and Dog notebook
+        run: source .env/bin/activate; cargo nextest run py_tests::tests::cat_and_dog_notebook_
       - name: All notebooks
         run: source .env/bin/activate; cargo nextest run py_tests::tests::run_notebook_ --test-threads 1
       - name: Voice tutorial

benches/poseidon.rs

@@ -23,8 +23,6 @@ use halo2curves::bn256::{Bn256, Fr};
 use rand::rngs::OsRng;
 use snark_verifier::system::halo2::transcript::evm::EvmTranscript;
 
-const L: usize = 10;
-
 #[derive(Clone, Debug)]
 struct MyCircuit {
     image: ValTensor<Fr>,
@@ -40,7 +38,7 @@ impl Circuit<Fr> for MyCircuit {
     }
 
     fn configure(cs: &mut ConstraintSystem<Fr>) -> Self::Config {
-        PoseidonChip::<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE, 10>::configure(cs, ())
+        PoseidonChip::<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE>::configure(cs, ())
     }
 
     fn synthesize(
@@ -48,7 +46,7 @@ impl Circuit<Fr> for MyCircuit {
         config: Self::Config,
         mut layouter: impl Layouter<Fr>,
     ) -> Result<(), Error> {
-        let chip: PoseidonChip<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE, L> =
+        let chip: PoseidonChip<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE> =
             PoseidonChip::new(config);
         chip.layout(&mut layouter, &[self.image.clone()], 0, &mut HashMap::new())?;
         Ok(())
@@ -59,7 +57,7 @@ fn runposeidon(c: &mut Criterion) {
     let mut group = c.benchmark_group("poseidon");
 
     for size in [64, 784, 2352, 12288].iter() {
-        let k = (PoseidonChip::<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE, L>::num_rows(*size)
+        let k = (PoseidonChip::<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE>::num_rows(*size)
             as f32)
             .log2()
             .ceil() as u32;
@@ -67,7 +65,7 @@ fn runposeidon(c: &mut Criterion) {
 
         let message = (0..*size).map(|_| Fr::random(OsRng)).collect::<Vec<_>>();
         let _output =
-            PoseidonChip::<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE, L>::run(message.to_vec())
+            PoseidonChip::<PoseidonSpec, POSEIDON_WIDTH, POSEIDON_RATE>::run(message.to_vec())
                 .unwrap();
 
         let mut image = Tensor::from(message.into_iter().map(Value::known));

docs/python/src/conf.py

@@ -1,7 +1,7 @@
 import ezkl
 
 project = 'ezkl'
-release = '0.0.0'
+release = '20.0.4'
 version = release
 
 

examples/notebooks/cat_and_dog_data.sh

@@ -0,0 +1,13 @@
+# download tess data
+# check if first argument has been set 
+if [ ! -z "$1" ]; then
+    DATA_DIR=$1
+else
+    DATA_DIR=data
+fi
+
+echo "Downloading data to $DATA_DIR"
+
+if [ ! -d "$DATA_DIR/CATDOG" ]; then
+    kaggle datasets download tongpython/cat-and-dog -p $DATA_DIR/CATDOG --unzip
+fi

src/bindings/python.rs

@@ -337,6 +337,8 @@ enum PyInputType {
     Int,
     ///
     TDim,
+    ///
+    Unknown,
 }
 
 impl From<InputType> for PyInputType {
@@ -348,6 +350,7 @@ impl From<InputType> for PyInputType {
             InputType::F64 => PyInputType::F64,
             InputType::Int => PyInputType::Int,
             InputType::TDim => PyInputType::TDim,
+            InputType::Unknown => PyInputType::Unknown,
         }
     }
 }
@@ -361,6 +364,7 @@ impl From<PyInputType> for InputType {
             PyInputType::F64 => InputType::F64,
             PyInputType::Int => InputType::Int,
             PyInputType::TDim => InputType::TDim,
+            PyInputType::Unknown => InputType::Unknown,
         }
     }
 }
@@ -375,6 +379,7 @@ impl FromStr for PyInputType {
             "f64" => Ok(PyInputType::F64),
             "int" => Ok(PyInputType::Int),
             "tdim" => Ok(PyInputType::TDim),
+            "unknown" => Ok(PyInputType::Unknown),
             _ => Err("Invalid value for InputType".to_string()),
         }
     }
@@ -592,7 +597,7 @@ fn poseidon_hash(message: Vec<PyFelt>) -> PyResult<Vec<PyFelt>> {
 /// Arguments
 /// -------
 /// message: list[str]
-///     List of field elements represnted as strings
+///     List of field elements represented as strings
 ///
 /// vk_path: str
 ///     Path to the verification key
@@ -651,7 +656,7 @@ fn kzg_commit(
 /// Arguments
 /// -------
 /// message: list[str]
-///     List of field elements represnted as strings
+///     List of field elements represented as strings
 ///
 /// vk_path: str
 ///     Path to the verification key
@@ -1945,7 +1950,7 @@ fn deploy_da_evm(
 ///     does the verifier use data attestation ?
 ///
 /// addr_vk: str
-///     The addess of the separate VK contract (if the verifier key is rendered as a separate contract)
+///     The address of the separate VK contract (if the verifier key is rendered as a separate contract)
 /// Returns
 /// -------
 /// bool

src/circuit/modules/polycommit.rs

@@ -1,7 +1,7 @@
 /*
 An easy-to-use implementation of the Poseidon Hash in the form of a Halo2 Chip. While the Poseidon Hash function
 is already implemented in halo2_gadgets, there is no wrapper chip that makes it easy to use in other circuits.
-Thanks to https://github.com/summa-dev/summa-solvency/blob/master/src/chips/poseidon/hash.rs for the inspiration (and also helping us understand how to use this).
+Thanks to https://github.com/summa-dev/summa-solvency/blob/master/zk_prover/src/chips/poseidon/hash.rs for the inspiration (and also helping us understand how to use this).
 */
 
 use std::collections::HashMap;

src/circuit/modules/poseidon.rs

@@ -1,7 +1,7 @@
 /*
 An easy-to-use implementation of the Poseidon Hash in the form of a Halo2 Chip. While the Poseidon Hash function
 is already implemented in halo2_gadgets, there is no wrapper chip that makes it easy to use in other circuits.
-Thanks to https://github.com/summa-dev/summa-solvency/blob/master/src/chips/poseidon/hash.rs for the inspiration (and also helping us understand how to use this).
+Thanks to https://github.com/summa-dev/summa-solvency/blob/master/zk_prover/src/chips/poseidon/hash.rs for the inspiration (and also helping us understand how to use this).
 */
 
 pub mod poseidon_params;

src/circuit/ops/layouts.rs

@@ -156,25 +156,6 @@ pub(crate) fn div<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     claimed_output.reshape(input_dims)?;
     // implicitly check if the prover provided output is within range
     let claimed_output = identity(config, region, &[claimed_output], true)?;
-    // check if x is too large only if the decomp would support overflow in the previous op
-    if F::from_u128(IntegerRep::MAX as u128)
-        < F::from_u128(region.base() as u128).pow([region.legs() as u64]) - F::ONE
-    {
-        // here we decompose and extract the sign of the input
-        let sign = sign(config, region, &[claimed_output.clone()])?;
-
-        let abs_value = pairwise(
-            config,
-            region,
-            &[claimed_output.clone(), sign],
-            BaseOp::Mult,
-        )?;
-        let max_val = create_constant_tensor(integer_rep_to_felt(IntegerRep::MAX), 1);
-        let less_than_max = less(config, region, &[abs_value.clone(), max_val])?;
-        // assert the result is 1
-        let comparison_unit = create_constant_tensor(F::ONE, less_than_max.len());
-        enforce_equality(config, region, &[abs_value, comparison_unit])?;
-    }
 
     let product = pairwise(
         config,
@@ -248,32 +229,6 @@ pub(crate) fn recip<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
         &[equal_zero_mask.clone(), equal_inverse_mask],
     )?;
 
-    let masked_output = pairwise(
-        config,
-        region,
-        &[claimed_output.clone(), not_equal_zero_mask.clone()],
-        BaseOp::Mult,
-    )?;
-
-    // check if x is too large only if the decomp would support overflow in the previous op
-    if F::from_u128(IntegerRep::MAX as u128)
-        < F::from_u128(region.base() as u128).pow([region.legs() as u64]) - F::ONE
-    {
-        // here we decompose and extract the sign of the input
-        let sign = sign(config, region, &[masked_output.clone()])?;
-        let abs_value = pairwise(
-            config,
-            region,
-            &[claimed_output.clone(), sign],
-            BaseOp::Mult,
-        )?;
-        let max_val = create_constant_tensor(integer_rep_to_felt(IntegerRep::MAX), 1);
-        let less_than_max = less(config, region, &[abs_value.clone(), max_val])?;
-        // assert the result is 1
-        let comparison_unit = create_constant_tensor(F::ONE, less_than_max.len());
-        enforce_equality(config, region, &[abs_value, comparison_unit])?;
-    }
-
     let err_func = |config: &BaseConfig<F>,
                     region: &mut RegionCtx<F>,
                     x: &ValTensor<F>|
@@ -349,7 +304,7 @@ pub fn sqrt<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
         .into()
     };
     claimed_output.reshape(input_dims)?;
-    // force the output to be positive or zero, also implicitly checks that the ouput is in range
+    // force the output to be positive or zero, also implicitly checks that the output is in range
     let claimed_output = abs(config, region, &[claimed_output.clone()])?;
     // rescaled input
     let rescaled_input = pairwise(config, region, &[input.clone(), unit_scale], BaseOp::Mult)?;
@@ -1841,7 +1796,7 @@ pub(crate) fn get_missing_set_elements<
 
         // get the difference between the two vectors
         for eval in input_evals.iter() {
-            // delete first occurence of that value
+            // delete first occurrence of that value
             if let Some(pos) = fullset_evals.iter().position(|x| x == eval) {
                 fullset_evals.remove(pos);
             }
@@ -1869,7 +1824,7 @@ pub(crate) fn get_missing_set_elements<
     region.increment(claimed_output.len());
 
     // input and claimed output should be the shuffles of fullset
-    // concatentate input and claimed output
+    // concatenate input and claimed output
     let input_and_claimed_output = input.concat(claimed_output.clone())?;
 
     // assert that this is a permutation/shuffle
@@ -3396,7 +3351,7 @@ pub fn max_pool<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
 /// Performs a deconvolution on the given input tensor.
 /// # Examples
 /// ```
-// // expected ouputs are taken from pytorch torch.nn.functional.conv_transpose2d
+// // expected outputs are taken from pytorch torch.nn.functional.conv_transpose2d
 ///
 /// use ezkl::tensor::Tensor;
 /// use ezkl::fieldutils::IntegerRep;
@@ -3624,7 +3579,7 @@ pub fn deconv<
 
 /// Applies convolution over a ND tensor of shape C x H x D1...DN (and adds a bias).
 /// ```
-/// // expected ouputs are taken from pytorch torch.nn.functional.conv2d
+/// // expected outputs are taken from pytorch torch.nn.functional.conv2d
 ///
 /// use ezkl::tensor::Tensor;
 /// use ezkl::fieldutils::IntegerRep;
@@ -3908,7 +3863,7 @@ pub(crate) fn rescale<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>
     Ok(rescaled_inputs)
 }
 
-/// Dummy (no contraints) reshape layout
+/// Dummy (no constraints) reshape layout
 pub(crate) fn reshape<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     values: &[ValTensor<F>; 1],
     new_dims: &[usize],
@@ -3918,7 +3873,7 @@ pub(crate) fn reshape<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>
     Ok(t)
 }
 
-/// Dummy (no contraints) move_axis layout
+/// Dummy (no constraints) move_axis layout
 pub(crate) fn move_axis<F: PrimeField + TensorType + PartialOrd + std::hash::Hash>(
     values: &[ValTensor<F>; 1],
     source: usize,

src/circuit/ops/mod.rs

@@ -1,6 +1,8 @@
 use std::any::Any;
 
 use serde::{Deserialize, Serialize};
+#[cfg(all(feature = "ezkl", not(target_arch = "wasm32")))]
+use tract_onnx::prelude::DatumType;
 
 use crate::{
     graph::quantize_tensor,
@@ -96,6 +98,8 @@ pub enum InputType {
     Int,
     ///
     TDim,
+    ///
+    Unknown,
 }
 
 impl InputType {
@@ -132,6 +136,7 @@ impl InputType {
                 let int_input = input.clone().to_i64().unwrap();
                 *input = T::from_i64(int_input).unwrap();
             }
+            InputType::Unknown => {}
         }
     }
 }
@@ -152,6 +157,28 @@ impl std::str::FromStr for InputType {
     }
 }
 
+#[cfg(all(feature = "ezkl", not(target_arch = "wasm32")))]
+impl From<DatumType> for InputType {
+    fn from(datum_type: DatumType) -> Self {
+        match datum_type {
+            DatumType::Bool => InputType::Bool,
+            DatumType::F16 => InputType::F16,
+            DatumType::F32 => InputType::F32,
+            DatumType::F64 => InputType::F64,
+            DatumType::I8 => InputType::Int,
+            DatumType::I16 => InputType::Int,
+            DatumType::I32 => InputType::Int,
+            DatumType::I64 => InputType::Int,
+            DatumType::U8 => InputType::Int,
+            DatumType::U16 => InputType::Int,
+            DatumType::U32 => InputType::Int,
+            DatumType::U64 => InputType::Int,
+            DatumType::TDim => InputType::TDim,
+            _ => unimplemented!(),
+        }
+    }
+}
+
 ///
 #[derive(Clone, Debug, PartialEq, Eq, Hash, PartialOrd, Ord, Serialize, Deserialize)]
 pub struct Input {

src/graph/mod.rs

@@ -455,6 +455,10 @@ pub struct GraphSettings {
     pub num_blinding_factors: Option<usize>,
     /// unix time timestamp
     pub timestamp: Option<u128>,
+    /// Model inputs types (if any)
+    pub input_types: Option<Vec<InputType>>,
+    /// Model outputs types (if any)
+    pub output_types: Option<Vec<InputType>>,
 }
 
 impl GraphSettings {

src/graph/model.rs

@@ -379,9 +379,15 @@ pub struct ParsedNodes {
     pub nodes: BTreeMap<usize, NodeType>,
     inputs: Vec<usize>,
     outputs: Vec<Outlet>,
+    output_types: Vec<InputType>,
 }
 
 impl ParsedNodes {
+    /// Returns the output types of the computational graph.
+    pub fn get_output_types(&self) -> Vec<InputType> {
+        self.output_types.clone()
+    }
+
     /// Returns the number of the computational graph's inputs
     pub fn num_inputs(&self) -> usize {
         self.inputs.len()
@@ -491,6 +497,16 @@ impl Model {
         Ok(om)
     }
 
+    /// Gets the input types from the parsed nodes
+    pub fn get_input_types(&self) -> Result<Vec<InputType>, GraphError> {
+        self.graph.get_input_types()
+    }
+
+    /// Gets the output types from the parsed nodes
+    pub fn get_output_types(&self) -> Vec<InputType> {
+        self.graph.get_output_types()
+    }
+
     ///
     pub fn save(&self, path: PathBuf) -> Result<(), GraphError> {
         let f = std::fs::File::create(&path).map_err(|e| {
@@ -574,6 +590,11 @@ impl Model {
             required_range_checks: res.range_checks.into_iter().collect(),
             model_output_scales: self.graph.get_output_scales()?,
             model_input_scales: self.graph.get_input_scales(),
+            input_types: match self.get_input_types() {
+                Ok(x) => Some(x),
+                Err(_) => None,
+            },
+            output_types: Some(self.get_output_types()),
             num_dynamic_lookups: res.num_dynamic_lookups,
             total_dynamic_col_size: res.dynamic_lookup_col_coord,
             num_shuffles: res.num_shuffles,
@@ -704,6 +725,11 @@ impl Model {
             nodes,
             inputs: model.inputs.iter().map(|o| o.node).collect(),
             outputs: model.outputs.iter().map(|o| (o.node, o.slot)).collect(),
+            output_types: model
+                .outputs
+                .iter()
+                .map(|o| Ok::<InputType, GraphError>(model.outlet_fact(*o)?.datum_type.into()))
+                .collect::<Result<Vec<_>, GraphError>>()?,
         };
 
         let duration = start_time.elapsed();
@@ -862,6 +888,15 @@ impl Model {
                         nodes: subgraph_nodes,
                         inputs: model.inputs.iter().map(|o| o.node).collect(),
                         outputs: model.outputs.iter().map(|o| (o.node, o.slot)).collect(),
+                        output_types: model
+                            .outputs
+                            .iter()
+                            .map(|o| {
+                                Ok::<InputType, GraphError>(
+                                    model.outlet_fact(*o)?.datum_type.into(),
+                                )
+                            })
+                            .collect::<Result<Vec<_>, GraphError>>()?,
                     };
 
                     let om = Model {
@@ -1579,4 +1614,16 @@ impl Model {
         }
         Ok(instance_shapes)
     }
+
+    /// Input types of the computational graph's public inputs (if any)
+    pub fn instance_types(&self) -> Result<Vec<InputType>, GraphError> {
+        let mut instance_types = vec![];
+        if self.visibility.input.is_public() {
+            instance_types.extend(self.graph.get_input_types()?);
+        }
+        if self.visibility.output.is_public() {
+            instance_types.extend(self.graph.get_output_types());
+        }
+        Ok(instance_types)
+    }
 }

src/lib.rs

@@ -100,7 +100,6 @@ use std::str::FromStr;
 use circuit::{table::Range, CheckMode, Tolerance};
 #[cfg(all(feature = "ezkl", not(target_arch = "wasm32")))]
 use clap::Args;
-#[cfg(all(feature = "ezkl", not(target_arch = "wasm32")))]
 use fieldutils::IntegerRep;
 use graph::{Visibility, MAX_PUBLIC_SRS};
 use halo2_proofs::poly::{
@@ -399,6 +398,16 @@ impl RunArgs {
     pub fn validate(&self) -> Result<(), String> {
         let mut errors = Vec::new();
 
+        // check if the largest represented integer in the decomposed form overflows IntegerRep
+        //  try it with the largest possible value
+        let max_decomp = (self.decomp_base as IntegerRep).checked_pow(self.decomp_legs as u32);
+        if max_decomp.is_none() {
+            errors.push(format!(
+                "decomp_base^decomp_legs overflows IntegerRep: {}^{}",
+                self.decomp_base, self.decomp_legs
+            ));
+        }
+
         // Visibility validations
         if self.param_visibility == Visibility::Public {
             errors.push(

src/tensor/ops.rs

@@ -387,7 +387,7 @@ pub fn add<T: TensorType + Add<Output = T> + std::marker::Send + std::marker::Sy
 ) -> Result<Tensor<T>, TensorError> {
     if t.len() == 1 {
         return Ok(t[0].clone());
-    } else if t.len() == 0 {
+    } else if t.is_empty() {
         return Err(TensorError::DimMismatch("add".to_string()));
     }
 
@@ -441,7 +441,7 @@ pub fn sub<T: TensorType + Sub<Output = T> + std::marker::Send + std::marker::Sy
 ) -> Result<Tensor<T>, TensorError> {
     if t.len() == 1 {
         return Ok(t[0].clone());
-    } else if t.len() == 0 {
+    } else if t.is_empty() {
         return Err(TensorError::DimMismatch("sub".to_string()));
     }
     // calculate value of output
@@ -492,7 +492,7 @@ pub fn mult<T: TensorType + Mul<Output = T> + std::marker::Send + std::marker::S
 ) -> Result<Tensor<T>, TensorError> {
     if t.len() == 1 {
         return Ok(t[0].clone());
-    } else if t.len() == 0 {
+    } else if t.is_empty() {
         return Err(TensorError::DimMismatch("mult".to_string()));
     }
     // calculate value of output
@@ -1326,7 +1326,6 @@ pub fn pad<T: TensorType>(
 ///
 /// # Errors
 /// Returns a TensorError if the tensors in `inputs` have incompatible dimensions for concatenation along the specified `axis`.
-
 pub fn concat<T: TensorType + Send + Sync>(
     inputs: &[&Tensor<T>],
     axis: usize,
@@ -2102,7 +2101,6 @@ pub mod nonlinearities {
     /// let expected = Tensor::<IntegerRep>::new(Some(&[4, 25, 8, 1, 1, 0]), &[2, 3]).unwrap();
     /// assert_eq!(result, expected);
     /// ```
-
     pub fn tanh(a: &Tensor<IntegerRep>, scale_input: f64) -> Tensor<IntegerRep> {
         a.par_enum_map(|_, a_i| {
             let kix = (a_i as f64) / scale_input;

tests/py_integration_tests.rs

@@ -49,6 +49,23 @@ mod py_tests {
         std::env::set_var("VOICE_DATA_DIR", format!("{}", voice_data_dir));
     }
 
+    fn download_catdog_data() {
+        let cat_and_dog_data_dir = shellexpand::tilde("~/data/catdog_data");
+
+        DOWNLOAD_VOICE_DATA.call_once(|| {
+            let status = Command::new("bash")
+                .args([
+                    "examples/notebooks/cat_and_dog_data.sh",
+                    &cat_and_dog_data_dir,
+                ])
+                .status()
+                .expect("failed to execute process");
+            assert!(status.success());
+        });
+        // set VOICE_DATA_DIR environment variable
+        std::env::set_var("CATDOG_DATA_DIR", format!("{}", cat_and_dog_data_dir));
+    }
+
     fn setup_py_env() {
         ENV_SETUP.call_once(|| {
             // supposes that you have a virtualenv called .env and have run the following
@@ -225,6 +242,20 @@ mod py_tests {
                 anvil_child.kill().unwrap();
             }
 
+
+            #[test]
+            fn cat_and_dog_notebook_() {
+                crate::py_tests::init_binary();
+                let mut anvil_child = crate::py_tests::start_anvil(false);
+                crate::py_tests::download_catdog_data();
+                let test_dir: TempDir = TempDir::new("cat_and_dog").unwrap();
+                let path = test_dir.path().to_str().unwrap();
+                crate::py_tests::mv_test_(path, "cat_and_dog.ipynb");
+                run_notebook(path, "cat_and_dog.ipynb");
+                test_dir.close().unwrap();
+                anvil_child.kill().unwrap();
+            }
+
             #[test]
             fn reusable_verifier_notebook_() {
                 crate::py_tests::init_binary();

tests/wasm.rs

@@ -337,7 +337,7 @@ mod wasm32 {
         // Run compiled circuit validation on onnx network (should fail)
         let circuit = compiledCircuitValidation(wasm_bindgen::Clamped(NETWORK.to_vec()));
         assert!(circuit.is_err());
-        // Run compiled circuit validation on comiled network (should pass)
+        // Run compiled circuit validation on compiled network (should pass)
         let circuit = compiledCircuitValidation(wasm_bindgen::Clamped(NETWORK_COMPILED.to_vec()));
         assert!(circuit.is_ok());
         // Run input validation on witness (should fail)