fix and test preprocessing examples (#1520)

montanalow · web-flow · commit a8d8218f698f · 2024-06-11T13:44:34.000-07:00
diff --git a/.github/workflows/ubuntu-packages-and-docker-image.yml b/.github/workflows/ubuntu-packages-and-docker-image.yml
@@ -4,7 +4,7 @@ on:
   workflow_dispatch:
     inputs:
       packageVersion:
-        default: "2.8.2"
+        default: "2.9.1"
 jobs:
   #
   # PostgresML extension.
diff --git a/pgml-cms/docs/resources/developer-docs/contributing.md b/pgml-cms/docs/resources/developer-docs/contributing.md
@@ -127,7 +127,7 @@ SELECT pgml.version();
 postgres=# select pgml.version();
       version
 -------------------
- 2.7.4
+ 2.9.1
 (1 row)
 ```
 {% endtab %}
diff --git a/pgml-cms/docs/resources/developer-docs/installation.md b/pgml-cms/docs/resources/developer-docs/installation.md
@@ -132,7 +132,7 @@ CREATE EXTENSION
 pgml_test=# SELECT pgml.version();
  version 
 ---------
- 2.7.4
+ 2.9.1
 (1 row)
 ```
 
diff --git a/pgml-cms/docs/resources/developer-docs/quick-start-with-docker.md b/pgml-cms/docs/resources/developer-docs/quick-start-with-docker.md
@@ -80,7 +80,7 @@ Time: 41.520 ms
 postgresml=# SELECT pgml.version();
  version 
 ---------
- 2.7.13
+ 2.9.1
 (1 row)
 ```
 
diff --git a/pgml-cms/docs/resources/developer-docs/self-hosting/pooler.md b/pgml-cms/docs/resources/developer-docs/self-hosting/pooler.md
@@ -115,6 +115,6 @@ Type "help" for help.
 postgresml=> SELECT pgml.version();
  version
 ---------
- 2.7.9
+ 2.9.1
 (1 row)
 ```
diff --git a/pgml-extension/Cargo.lock b/pgml-extension/Cargo.lock
diff --git a/pgml-extension/Cargo.toml b/pgml-extension/Cargo.toml
@@ -1,6 +1,6 @@
 [package]
 name = "pgml"
-version = "2.9.0"
+version = "2.9.1"
 edition = "2021"
 
 [lib]
diff --git a/pgml-extension/examples/preprocessing.sql b/pgml-extension/examples/preprocessing.sql
@@ -0,0 +1,33 @@
+-- load the diamonds dataset, that contains text categorical variables
+SELECT pgml.load_dataset('jdxcosta/diamonds');
+
+-- view the data
+SELECT * FROM pgml."jdxcosta/diamonds" LIMIT 10;
+
+-- drop the Unamed column, since it's not useful for training (you could create a view instead)
+ALTER TABLE pgml."jdxcosta/diamonds" DROP COLUMN "Unnamed: 0";
+
+-- train a model using preprocessors to scale the numeric variables, and target encode the categoricals
+SELECT pgml.train(
+       project_name => 'Diamond prices',
+       task => 'regression',
+       relation_name => 'pgml.jdxcosta/diamonds',
+       y_column_name => 'price',
+       algorithm => 'lightgbm',
+       preprocess => '{
+                      "carat": {"scale": "standard"},
+                      "depth": {"scale": "standard"},
+                      "table": {"scale": "standard"},
+                      "cut": {"encode": "target", "scale": "standard"},
+                      "color": {"encode": "target", "scale": "standard"},
+                      "clarity": {"encode": "target", "scale": "standard"}
+                  }'
+);
+
+-- run some predictions, notice we're passing a heterogeneous row (tuple) as input, rather than a homogenous ARRAY[].
+SELECT price, pgml.predict('Diamond prices', (carat, cut, color, clarity, depth, "table", x, y, z)) AS prediction
+FROM pgml."jdxcosta/diamonds"
+LIMIT 10;
+
+-- This is a difficult dataset for more algorithms, which makes it a good challenge for preprocessing, and additional
+-- feature engineering. What's next?
diff --git a/pgml-extension/sql/pgml--2.9.0--2.9.1.sql b/pgml-extension/sql/pgml--2.9.0--2.9.1.sql
diff --git a/pgml-extension/src/bindings/transformers/mod.rs b/pgml-extension/src/bindings/transformers/mod.rs
@@ -380,7 +380,7 @@ pub fn load_dataset(
         .ok_or(anyhow!("dataset `data` key is not an object"))?;
     let column_names = types
         .iter()
-        .map(|(name, _type)| name.clone())
+        .map(|(name, _type)| format!("\"{}\"", name))
         .collect::<Vec<String>>()
         .join(", ");
     let column_types = types
@@ -393,13 +393,14 @@ pub fn load_dataset(
                 "int64" => "INT8",
                 "int32" => "INT4",
                 "int16" => "INT2",
+                "int8" => "INT2",
                 "float64" => "FLOAT8",
                 "float32" => "FLOAT4",
                 "float16" => "FLOAT4",
                 "bool" => "BOOLEAN",
                 _ => bail!("unhandled dataset feature while reading dataset: {type_}"),
             };
-            Ok(format!("{name} {type_}"))
+            Ok(format!("\"{name}\" {type_}"))
         })
         .collect::<Result<Vec<String>>>()?
         .join(", ");
@@ -455,7 +456,7 @@ pub fn load_dataset(
                         .into_datum(),
                 )),
                 "dict" | "list" => row.push((PgBuiltInOids::JSONBOID.oid(), JsonB(value.clone()).into_datum())),
-                "int64" | "int32" | "int16" => row.push((
+                "int64" | "int32" | "int16" | "int8" => row.push((
                     PgBuiltInOids::INT8OID.oid(),
                     value
                         .as_i64()
diff --git a/pgml-extension/src/orm/model.rs b/pgml-extension/src/orm/model.rs
@@ -344,13 +344,12 @@ impl Model {
             ).unwrap().first();
 
             if !result.is_empty() {
-                let project_id = result.get(2).unwrap().unwrap();
-                let project = Project::find(project_id).unwrap();
-                let snapshot_id = result.get(3).unwrap().unwrap();
-                let snapshot = Snapshot::find(snapshot_id).unwrap();
-                let algorithm = Algorithm::from_str(result.get(4).unwrap().unwrap()).unwrap();
-                let runtime = Runtime::from_str(result.get(5).unwrap().unwrap()).unwrap();
-
+                let project_id = result.get(2).unwrap().expect("project_id is i64");
+                let project = Project::find(project_id).expect("project doesn't exist");
+                let snapshot_id = result.get(3).unwrap().expect("snapshot_id is i64");
+                let snapshot = Snapshot::find(snapshot_id).expect("snapshot doesn't exist");
+                let algorithm = Algorithm::from_str(result.get(4).unwrap().unwrap()).expect("algorithm is malformed");
+                let runtime = Runtime::from_str(result.get(5).unwrap().unwrap()).expect("runtime is malformed");
                 let data = Spi::get_one_with_args::<Vec<u8>>(
                     "
                         SELECT data
diff --git a/pgml-extension/src/orm/sampling.rs b/pgml-extension/src/orm/sampling.rs
@@ -55,7 +55,7 @@ impl Sampling {
             Sampling::stratified => {
                 format!(
                     "
-                    SELECT *
+                    SELECT {col_string}
                     FROM (
                         SELECT 
                             *,
@@ -125,7 +125,7 @@ mod tests {
         let columns = get_column_fixtures();
         let sql = sampling.get_sql("my_table", columns);
         let expected_sql = "
-                    SELECT *
+                    SELECT \"col1\", \"col2\"
                     FROM (
                         SELECT 
                             *,
diff --git a/pgml-extension/src/orm/snapshot.rs b/pgml-extension/src/orm/snapshot.rs
@@ -230,16 +230,24 @@ impl Column {
         if self.preprocessor.encode == Encode::target {
             let categories = self.statistics.categories.as_mut().unwrap();
             let mut sums = vec![0_f32; categories.len() + 1];
+            let mut total = 0.;
             Zip::from(array).and(target).for_each(|&value, &target| {
+                total += target;
                 sums[value as usize] += target;
             });
+            let avg_target = total / categories.len() as f32;
             for category in categories.values_mut() {
-                let sum = sums[category.value as usize];
-                category.value = sum / category.members as f32;
+                if category.members > 0 {
+                    let sum = sums[category.value as usize];
+                    category.value = sum / category.members as f32;
+                } else {
+                    // use avg target for categories w/ no members, e.g. __NULL__ category in a complete dataset
+                    category.value = avg_target;
+                }
             }
         }
 
-        // Data is filtered for NaN because it is not well defined statistically, and they are counted as separate stat
+        // Data is filtered for NaN because it is not well-defined statistically, and they are counted as separate stat
         let mut data = array
             .iter()
             .filter_map(|n| if n.is_nan() { None } else { Some(*n) })
@@ -404,7 +412,8 @@ impl Snapshot {
                 .first();
             if !result.is_empty() {
                 let jsonb: JsonB = result.get(7).unwrap().unwrap();
-                let columns: Vec<Column> = serde_json::from_value(jsonb.0).unwrap();
+                let columns: Vec<Column> =
+                    serde_json::from_value(jsonb.0).expect("invalid json description of columns");
                 // let jsonb: JsonB = result.get(8).unwrap();
                 // let analysis: Option<IndexMap<String, f32>> = Some(serde_json::from_value(jsonb.0).unwrap());
                 let mut s = Snapshot {
@@ -500,9 +509,10 @@ impl Snapshot {
 
         let preprocessors: HashMap<String, Preprocessor> = serde_json::from_value(preprocess.0).expect("is valid");
 
+        let mut position = 0; // Postgres column positions are not updated when other columns are dropped, but we expect consecutive positions when we read the table.
         Spi::connect(|mut client| {
             let mut columns: Vec<Column> = Vec::new();
-            client.select("SELECT column_name::TEXT, udt_name::TEXT, is_nullable::BOOLEAN, ordinal_position::INTEGER FROM information_schema.columns WHERE table_schema = $1 AND table_name = $2 ORDER BY ordinal_position ASC",
+            client.select("SELECT column_name::TEXT, udt_name::TEXT, is_nullable::BOOLEAN FROM information_schema.columns WHERE table_schema = $1 AND table_name = $2 ORDER BY ordinal_position ASC",
                 None,
                 Some(vec![
                     (PgBuiltInOids::TEXTOID.oid(), schema_name.into_datum()),
@@ -520,7 +530,7 @@ impl Snapshot {
                     pg_type = pg_type[1..].to_string() + "[]";
                 }
                 let nullable = row[3].value::<bool>().unwrap().unwrap();
-                let position = row[4].value::<i32>().unwrap().unwrap() as usize;
+                position += 1;
                 let label = match y_column_name {
                     Some(ref y_column_name) => y_column_name.contains(&name),
                     None => false,
@@ -1158,7 +1168,7 @@ impl Snapshot {
     pub fn numeric_encoded_dataset(&mut self) -> Dataset {
         let mut data = None;
         Spi::connect(|client| {
-            // Postgres Arrays arrays are 1 indexed and so are SPI tuples...
+            // Postgres arrays are 1 indexed and so are SPI tuples...
             let result = client.select(&self.select_sql(), None, None).unwrap();
             let num_rows = result.len();
             let (num_train_rows, num_test_rows) = self.train_test_split(num_rows);
diff --git a/pgml-extension/tests/test.sql b/pgml-extension/tests/test.sql
@@ -30,5 +30,6 @@ SELECT pgml.load_dataset('wine');
 \i examples/regression.sql
 \i examples/vectors.sql
 \i examples/chunking.sql
+\i examples/preprocessing.sql
 -- transformers are generally too slow to run in the test suite
 --\i examples/transformers.sql

Original file line number	Diff line number	Diff line change
`@@ -4,7 +4,7 @@ on:`
`4`	`4`	`workflow_dispatch:`
`5`	`5`	`inputs:`
`6`	`6`	`packageVersion:`
`7`		`- default: "2.8.2"`
	`7`	`+ default: "2.9.1"`
`8`	`8`	`jobs:`
`9`	`9`	`#`
`10`	`10`	`# PostgresML extension.`
-Original file line number
+Diff line change
 pgml_test=# SELECT pgml.version();
  version
 ---------
 - 2.7.4
 + 2.9.1
 (1 row)
 ```