Merge branch 'refs/heads/random' into operators-random-exp

# Conflicts:
#	candle-onnx/tests/ops.rs

candle-onnx/src/eval.rs

@@ -971,6 +971,46 @@ pub fn simple_eval(
                 };
                 values.insert(node.output[0].clone(), output);
             }
+            "RandomUniform" => {
+                let dt: i64 = get_attr_opt(node, "dtype")?.copied().unwrap_or(1); // 1 is float
+                                                                                  // type by
+                                                                                  // default
+                let dtype = match DataType::try_from(dt as i32) {
+                    Ok(dt) => match dtype(dt) {
+                        Some(DType::U8 | DType::U32 | DType::I64) => {
+                            bail!(
+                                "unsupported 'dtype' value {dt:?}, only floats are allowed, for RandomUnifrom {}",
+                                node.name
+                            )
+                        }
+                        Some(dt) => dt,
+                        None => {
+                            bail!(
+                                "unsupported 'dtype' value {dt:?} for RandomUnifrom {}",
+                                node.name
+                            )
+                        }
+                    },
+                    Err(_) => {
+                        bail!(
+                            "unsupported 'dtype' value {dt:?} for RandomUniform {}",
+                            node.name
+                        )
+                    }
+                };
+                let low: f32 = get_attr_opt(node, "low")?.copied().unwrap_or(0.0);
+                let high: f32 = get_attr_opt(node, "high")?.copied().unwrap_or(1.0);
+                let seed: Option<f32> = get_attr_opt(node, "seed")?.copied();
+                if seed.is_some() {
+                    bail!("seed for RandomUniform is currently not supported")
+                };
+                let shape: Vec<usize> = get_attr::<[i64]>(node, "shape")?
+                    .iter()
+                    .map(|x| *x as usize)
+                    .collect();
+                let output = Tensor::rand(low, high, shape, &Device::Cpu)?.to_dtype(dtype)?;
+                values.insert(node.output[0].clone(), output);
+            }
             op_type => bail!("unsupported op_type {op_type} for op {node:?}"),
         }
     }

candle-onnx/tests/ops.rs

@@ -1878,6 +1878,152 @@ fn test_sqrt() -> Result<()> {
     Ok(())
 }
 
+// "RandomUniform"
+#[test]
+fn test_random_uniform() -> Result<()> {
+    test(vec![3, 2, 1, 4], None, None)?;
+    test(vec![2, 2, 2, 2], Some(-10.0), None)?;
+    test(vec![2, 2, 2, 2], None, Some(10.0))?;
+    test(vec![1, 2, 3, 4], Some(-10.0), Some(10.0))?;
+
+    fn test(shape: Vec<i64>, low: Option<f32>, high: Option<f32>) -> Result<()> {
+        let att_low = AttributeProto {
+            name: "low".to_string(),
+            ref_attr_name: "low".to_string(),
+            i: 0,
+            doc_string: "low".to_string(),
+            r#type: 1, // FLOAT
+            f: low.unwrap_or(0.0),
+            s: vec![],
+            t: None,
+            g: None,
+            sparse_tensor: None,
+            tp: None,
+            floats: vec![],
+            ints: vec![],
+            strings: vec![],
+            tensors: vec![],
+            graphs: vec![],
+            sparse_tensors: vec![],
+            type_protos: vec![],
+        };
+        let att_high = AttributeProto {
+            name: "high".to_string(),
+            ref_attr_name: "high".to_string(),
+            i: 0,
+            doc_string: "high".to_string(),
+            r#type: 1, // FLOAT
+            f: high.unwrap_or(1.0),
+            s: vec![],
+            t: None,
+            g: None,
+            sparse_tensor: None,
+            tp: None,
+            floats: vec![],
+            ints: vec![],
+            strings: vec![],
+            tensors: vec![],
+            graphs: vec![],
+            sparse_tensors: vec![],
+            type_protos: vec![],
+        };
+        let att_shape = AttributeProto {
+            name: "shape".to_string(),
+            ref_attr_name: "shape".to_string(),
+            i: 0,
+            doc_string: "shape".to_string(),
+            r#type: 7, // INTS
+            f: 0.0,
+            s: vec![],
+            t: None,
+            g: None,
+            sparse_tensor: None,
+            tp: None,
+            floats: vec![],
+            ints: shape,
+            strings: vec![],
+            tensors: vec![],
+            graphs: vec![],
+            sparse_tensors: vec![],
+            type_protos: vec![],
+        };
+        let att_dtype = AttributeProto {
+            name: "dtype".to_string(),
+            ref_attr_name: "dtype".to_string(),
+            i: 11, // DOUBLE
+            doc_string: "dtype".to_string(),
+            r#type: 2, // INT
+            f: 0.0,
+            s: vec![],
+            t: None,
+            g: None,
+            sparse_tensor: None,
+            tp: None,
+            floats: vec![],
+            ints: vec![],
+            strings: vec![],
+            tensors: vec![],
+            graphs: vec![],
+            sparse_tensors: vec![],
+            type_protos: vec![],
+        };
+        let attrs = {
+            let mut mut_attrs = vec![att_shape, att_dtype];
+            if low.is_some() {
+                mut_attrs.push(att_low);
+            }
+            if high.is_some() {
+                mut_attrs.push(att_high);
+            }
+            mut_attrs
+        };
+        let manual_graph = create_model_proto_with_graph(Some(GraphProto {
+            node: vec![NodeProto {
+                op_type: "RandomUniform".to_string(),
+                domain: "".to_string(),
+                attribute: attrs,
+                input: vec![],
+                output: vec![OUTPUT_Z.to_string()],
+                name: "".to_string(),
+                doc_string: "".to_string(),
+            }],
+            name: "".to_string(),
+            initializer: vec![],
+            input: vec![],
+            output: vec![ValueInfoProto {
+                name: OUTPUT_Z.to_string(),
+                doc_string: "".to_string(),
+                r#type: None,
+            }],
+            value_info: vec![],
+            doc_string: "".to_string(),
+            sparse_initializer: vec![],
+            quantization_annotation: vec![],
+        }));
+        let eval = candle_onnx::simple_eval(&manual_graph, HashMap::new())?;
+        assert_eq!(eval.len(), 1);
+        let z = eval.get(OUTPUT_Z).expect("Output 'z' not found");
+        let min = z
+            .flatten_all()?
+            .to_vec1()?
+            .into_iter()
+            .reduce(f64::min)
+            .unwrap();
+        let max = z
+            .flatten_all()?
+            .to_vec1()?
+            .into_iter()
+            .reduce(f64::max)
+            .unwrap();
+        assert!(min >= low.unwrap_or(0.0).into());
+        assert!(max <= high.unwrap_or(1.0).into());
+        assert_ne!(min, max);
+        Ok(())
+    }
+
+    Ok(())
+}
+
 // "Range"
 #[test]
 fn test_range() -> Result<()> {