ml-lab-prep 0.1.0__tar.gz

ml_lab_prep-0.1.0/PKG-INFO

@@ -0,0 +1,31 @@
+Metadata-Version: 2.4
+Name: ml-lab-prep
+Version: 0.1.0
+Summary: Practice and solved Jupyter notebooks for an ML lab exam: EDA, tuning, CV, PCA, significance tests.
+Classifier: Programming Language :: Python :: 3
+Classifier: Intended Audience :: Education
+Requires-Python: >=3.9
+Description-Content-Type: text/markdown
+Requires-Dist: numpy
+Requires-Dist: pandas
+Requires-Dist: scikit-learn>=1.2
+Requires-Dist: scipy
+Requires-Dist: matplotlib
+Requires-Dist: seaborn
+Requires-Dist: xgboost
+
+# ml-lab-prep
+
+Practice and solved notebooks for a machine learning lab exam (classification, regression, K-Means, PCA, grid vs random search, 5-fold CV, Friedman/ANOVA and post-hoc tests), plus the datasets they use.
+
+```bash
+pip install ml-lab-prep
+ml-lab-prep check            # verify scikit-learn / xgboost etc. work
+ml-lab-prep init             # copies notebooks + data into ./ML_Lab_Prep
+cd ML_Lab_Prep && jupyter notebook
+```
+
+Options: `ml-lab-prep init my_folder`, `ml-lab-prep init --practice-only`.
+Start with `practice/01_Spambase_Classification.ipynb`; `solved/` has run outputs; read `EXAM_PLAYBOOK.md` last.
+
+For study use. Follow your institution's rules on what material is allowed in an exam.

ml_lab_prep-0.1.0/README.md

@@ -0,0 +1,15 @@
+# ml-lab-prep
+
+Practice and solved notebooks for a machine learning lab exam (classification, regression, K-Means, PCA, grid vs random search, 5-fold CV, Friedman/ANOVA and post-hoc tests), plus the datasets they use.
+
+```bash
+pip install ml-lab-prep
+ml-lab-prep check            # verify scikit-learn / xgboost etc. work
+ml-lab-prep init             # copies notebooks + data into ./ML_Lab_Prep
+cd ML_Lab_Prep && jupyter notebook
+```
+
+Options: `ml-lab-prep init my_folder`, `ml-lab-prep init --practice-only`.
+Start with `practice/01_Spambase_Classification.ipynb`; `solved/` has run outputs; read `EXAM_PLAYBOOK.md` last.
+
+For study use. Follow your institution's rules on what material is allowed in an exam.

ml_lab_prep-0.1.0/pyproject.toml

@@ -0,0 +1,23 @@
+[build-system]
+requires = ["setuptools>=64"]
+build-backend = "setuptools.build_meta"
+
+[project]
+name = "ml-lab-prep"
+version = "0.1.0"
+description = "Practice and solved Jupyter notebooks for an ML lab exam: EDA, tuning, CV, PCA, significance tests."
+readme = "README.md"
+requires-python = ">=3.9"
+dependencies = [
+  "numpy", "pandas", "scikit-learn>=1.2", "scipy", "matplotlib", "seaborn", "xgboost",
+]
+classifiers = ["Programming Language :: Python :: 3", "Intended Audience :: Education"]
+
+[project.scripts]
+ml-lab-prep = "ml_lab_prep.cli:main"
+
+[tool.setuptools.packages.find]
+where = ["src"]
+
+[tool.setuptools.package-data]
+ml_lab_prep = ["content/**/*"]

ml_lab_prep-0.1.0/setup.cfg

@@ -0,0 +1,4 @@
+[egg_info]
+tag_build = 
+tag_date = 0
+

ml_lab_prep-0.1.0/src/ml_lab_prep/__init__.py

@@ -0,0 +1 @@
+__version__ = "0.1.0"

ml_lab_prep-0.1.0/src/ml_lab_prep/cli.py

@@ -0,0 +1,39 @@
+import argparse, runpy, shutil, sys
+from importlib.resources import files
+from pathlib import Path
+
+
+def _content() -> Path:
+    return Path(str(files("ml_lab_prep") / "content"))
+
+
+def init(dest: str, practice_only: bool) -> int:
+    out = Path(dest)
+    if out.exists() and any(out.iterdir()):
+        print(f"{out} already exists and is not empty. Pick another folder.")
+        return 1
+    ignore = shutil.ignore_patterns("solved", "__pycache__") if practice_only else shutil.ignore_patterns("__pycache__")
+    shutil.copytree(_content(), out, ignore=ignore, dirs_exist_ok=True)
+    print(f"Copied to {out.resolve()}\nNext: cd {out} && jupyter notebook   (open practice/01_Spambase_Classification.ipynb)")
+    return 0
+
+
+def main(argv=None) -> int:
+    p = argparse.ArgumentParser(prog="ml-lab-prep")
+    sub = p.add_subparsers(dest="cmd", required=True)
+    i = sub.add_parser("init", help="copy notebooks and data into a folder")
+    i.add_argument("dest", nargs="?", default="ML_Lab_Prep")
+    i.add_argument("--practice-only", action="store_true", help="skip the solved notebooks")
+    sub.add_parser("check", help="check that required libraries work")
+    sub.add_parser("path", help="print where the bundled content lives")
+    a = p.parse_args(argv)
+    if a.cmd == "init":
+        return init(a.dest, a.practice_only)
+    if a.cmd == "path":
+        print(_content()); return 0
+    runpy.run_path(str(_content() / "check_setup.py"), run_name="__main__")
+    return 0
+
+
+if __name__ == "__main__":
+    sys.exit(main())

ml_lab_prep-0.1.0/src/ml_lab_prep/content/EXAM_PLAYBOOK.md

@@ -0,0 +1,83 @@
+# ML lab exam playbook
+
+Your sir's rule: understand the pipeline, adapt the cheat sheet. Marks come from **implementing a substantial part yourself (aim for 75%+)** and explaining it. Below 50% implementation scores 0, so never paste a block you cannot explain.
+
+## 1 · The 10 steps (with cheat-sheet sections)
+
+| Step | What to do | Cheat sheet |
+|---|---|---|
+| 1 | Read csv, drop duplicates, shape / dtypes / missing / target counts. **Balanced or imbalanced? Classification or regression?** | §1 Pandas / IO |
+| 2 | EDA: target plot, describe, correlation heatmap, 1-2 boxplots/regplots, outlier check. Write the 3-4 line summary | §1, §5 Visualization |
+| 3 | Fix data: impute NaN, encode categoricals, log1p if skewed | §1 Fill, Scaling & Encoding |
+| 4 | `train_test_split` (`stratify=y` for classification only) | §1 Split |
+| 5 | Pipeline: impute, scale, (PCA), model | §2 Pipeline |
+| 6 | Baseline fit, predict, metrics | §2 General, §5 |
+| 7 | Hyper-parameter search with `cv=5` (Grid, or Random if the grid is big) | §4, Parameter Grids |
+| 8 | 5-fold CV: report **mean ± std**, not one split | §4 K-Fold & CV |
+| 9 | PCA with vs without (if asked): same model, same CV, compare + number of components | §6 |
+| 10 | Significance: Friedman/ANOVA, post-hoc, paired t / Wilcoxon, verdict at p < 0.05 | §7 |
+
+## 2 · What to change per dataset
+
+| Item | Classification | Regression |
+|---|---|---|
+| Split | `stratify=y` | no stratify |
+| CV splitter | `StratifiedKFold` | `KFold` |
+| Scoring string | `"accuracy"`, `"f1"`, `"f1_macro"` (3+ classes) | `"r2"`, `"neg_root_mean_squared_error"` |
+| Metrics | accuracy, precision, recall, F1, confusion matrix, ROC-AUC | MAE, MSE, RMSE, R², adjusted R² |
+| 3+ classes | `average="macro"`; `roc_auc_score(y, proba, multi_class="ovr")` | n/a |
+| Text target | `LabelEncoder` (print `classes_`: alphabetical, so B=0, M=1) | n/a |
+| Text features | `OneHotEncoder` (or `OrdinalEncoder` if ordered) via `ColumnTransformer` | same |
+| Tree grid criterion | `gini` / `entropy` | `squared_error` / `friedman_mse` |
+| Model class | `...Classifier` | `...Regressor` (SVR needs the **target** scaled too) |
+
+## 3 · Pitfalls that cost marks
+
+* **`model__` prefix.** Tuning a `Pipeline` needs `model__max_depth`; tuning a bare estimator needs `max_depth`.
+* **Leakage.** Fit scaler/PCA on train only (inside the pipeline, or `fit_transform(train)` then `transform(test)`). Never `fit_transform` on all data before splitting.
+* **Scaling by model.** Needed: KNN, SVM, Logistic, Linear with penalty, PCA, K-Means. Not needed: trees, forests, boosting, GaussianNB. MultinomialNB: non-negative only (`MinMaxScaler`).
+* **SVM grid.** `gamma` only for rbf/poly/sigmoid, `degree` only for poly: use a *list of dicts*.
+* **`solver="liblinear"` is binary-only in recent scikit-learn.** For 3+ classes use `lbfgs` (L2) or `saga` (L1/elasticnet) and a larger `max_iter`.
+* **`predict_proba` for SVC** needs `probability=True`.
+* **XGBoost grid** has 324 combinations: that is the reason to use `RandomizedSearchCV`.
+* **Class labels.** In `precision_score` the "positive" class is label 1. Know which class that is.
+* **Friedman needs at least 3 models.** For 2 models use paired t-test / Wilcoxon.
+* **Wilcoxon on 5 folds can never give p < 0.05** (minimum 0.0625). Use repeated CV (5×3 = 15 paired scores) for tests and say why.
+
+## 4 · How to say it (write these in the answer or viva)
+
+**EDA, 3-4 lines.** "The dataset has N rows and M features, K numeric and J categorical, with/without missing values. The target is balanced / imbalanced (x% vs y%), so accuracy alone is / is not enough. Features are skewed / correlated / on different scales, so I scaled them / used PCA / regularised. The strongest predictors are A, B."
+
+**What PCA does.** "PCA rotates the features onto orthogonal directions ordered by variance and keeps the first few (here N components keeping 95% of the variance). It reduces dimension, removes correlation and speeds training, but the components mix the original features, so supervised signal can be lost. With PCA accuracy went from X to Y while features dropped from M to N."
+
+**Grid vs Random.** "Grid tries every combination (G fits), Random tries n_iter sampled ones (R fits). Both reached similar best CV score (a vs b); random was k times faster. Use random when the grid is large."
+
+**Significance.** "A higher accuracy does not prove a better model. I compared models on the same 15 repeated-CV folds. Friedman gave p = … (< 0.05), so at least one model differs. Post-hoc paired tests with Bonferroni show that A is significantly better than B and C but not than D. The mean gap between A and D is only 0.4%, so in practice they are equivalent."
+
+**5-fold CV.** "Data split into 5 folds; train on 4, validate on 1, repeat 5 times; report mean ± std: 0.93 ± 0.01."
+
+**Input → intermediate → output.** Input: dataset and features. Intermediate: preprocessing, PCA (if any), model, tuning, CV. Output: predictions, metrics, best parameters, significance verdict.
+
+## 5 · Imbalance checklist
+1. Print `value_counts(normalize=True)` first.
+2. If the minority class is below about 30%: stop quoting accuracy alone. Report precision, recall, F1, confusion matrix.
+3. Fixes: `class_weight="balanced"`, oversample the minority / undersample the majority (`sklearn.utils.resample`), threshold tuning.
+4. Say which error is worse (medical: missed disease = false negative, so favour recall).
+
+## 6 · Which hyper-parameters matter (for "which values should be tested?")
+
+| Model | Key hyper-parameters | Why |
+|---|---|---|
+| Decision Tree | `max_depth`, `min_samples_leaf/split`, `criterion` | control tree size = overfitting |
+| KNN | `n_neighbors`, `weights`, `algorithm` | k is the bias/variance dial |
+| SVM | `kernel`, `C`, `gamma`, `degree` | margin hardness, reach of a point |
+| Logistic | `C`, `penalty` | regularisation strength / type |
+| Ridge / Lasso | `alpha` (ElasticNet: `l1_ratio`) | shrinkage strength |
+| Random Forest | `n_estimators`, `max_depth`, `max_features`, `min_samples_split` | tree count, tree size, decorrelation |
+| AdaBoost | `n_estimators`, `learning_rate`, base `estimator` depth | weak-learner count and step |
+| Gradient Boosting / XGBoost | `n_estimators`, `learning_rate`, `max_depth`, `subsample` | shallow trees + small steps generalise |
+| K-Means | `n_clusters` (+ `n_init`) | choose by elbow + silhouette |
+| PCA | `n_components` (0.95 or int) | variance kept |
+
+## 7 · Cheat-sheet gaps to fill from these notebooks
+`ColumnTransformer` (mixed column types), `Pipeline` step names, `RepeatedStratifiedKFold` for significance, `LabelEncoder` positive-class check, `TransformedTargetRegressor`, results-table building, Bonferroni correction.

ml_lab_prep-0.1.0/src/ml_lab_prep/content/HOW_TO_RUN.md

@@ -0,0 +1,84 @@
+# How to run these notebooks
+
+## What is in this folder
+
+```
+ML_Lab_Prep/
+  data/        the 5 csv files (3 from your assignments + 2 for the drills)
+  solved/      fully run notebooks WITH outputs (read these, or re-run them)
+  practice/    same notebooks with every code cell emptied and a TODO hint (type these yourself)
+  EXAM_PLAYBOOK.md   what to change per dataset, pitfalls, how to explain EDA / PCA / significance
+  check_setup.py     run once on your exam laptop
+  requirements.txt
+```
+
+| # | Notebook | Dataset | Run time (2 CPU cores) |
+|---|---|---|---|
+| 01 | Spambase classification | spambase.csv | about 3 min |
+| 02 | Breast cancer: trees, ensembles, PCA | breast_cancer.csv | about 5 to 8 min |
+| 03 | Loan amount regression | loan_data.csv | about 5 to 8 min |
+| 04 | K-Means + PCA | breast_cancer.csv | under 1 min |
+| 05 | Drill A: Wine (3 classes) | wine.csv | about 1 min |
+| 06 | Drill B: Diabetes, messy data | diabetes_messy.csv | about 1 min |
+
+The notebooks read data from `../data/...`. **Keep the folder structure as is.** If you move a notebook, change the `DATA = ...` line in the config cell.
+
+## Step 0 (do this tonight): check your laptop
+
+Your own repo `.venv` already has numpy, pandas, scikit-learn 1.9, scipy and xgboost. It has `ipykernel` but **no Jupyter Notebook interface**, so install it once.
+
+Mac Terminal:
+
+```bash
+cd ~/Dev/college/SEM-5/ML/ML_Lab_Prep
+source ../.venv/bin/activate
+pip install notebook seaborn      # only if missing
+python check_setup.py             # should end with: ALL GOOD
+```
+
+If you use Anaconda instead, open **Anaconda Prompt** (Windows) or Terminal and run `pip install xgboost seaborn`, then `python check_setup.py`.
+
+## Option 1: Anaconda Navigator (what you said you would use in the lab)
+
+1. Open **Anaconda Navigator**.
+2. Click **Launch** under **Jupyter Notebook**. A browser tab opens.
+3. In the file list go to `ML_Lab_Prep` then `practice` (or `solved`) and click a notebook.
+4. Run one cell with **Shift+Enter**. Run everything with menu **Kernel, Restart & Run All**.
+5. If you see `ModuleNotFoundError: xgboost`: Navigator, **Environments** tab, change the filter to *Not installed*, search `xgboost`, tick it, **Apply**. Or in Anaconda Prompt: `pip install xgboost`. Then restart the notebook kernel.
+
+## Option 2: Jupyter from the terminal
+
+```bash
+cd ~/Dev/college/SEM-5/ML/ML_Lab_Prep
+source ../.venv/bin/activate
+jupyter notebook
+```
+
+## Option 3: VS Code
+
+Open the `ML_Lab_Prep` folder, open a `.ipynb`, click **Select Kernel** (top right), pick the `.venv` Python, then **Run All** or run cell by cell.
+
+## Option 4: Google Colab
+
+Upload the notebook and the csv you need, then in the config cell change the path to the bare file name, for example `DATA = "spambase.csv"`. Colab already has everything except possibly `xgboost` (`!pip install xgboost` in a cell if needed).
+
+## How to practice (suggested order for tonight)
+
+1. Run `check_setup.py` (2 min).
+2. Open **practice/01**. Keep the cheat sheet next to you. For every cell read the TODO hint, type the code, run it. Peek at **solved/01** only after you are stuck for 3 minutes.
+3. Do the same for **02** (ensembles, PCA vs no PCA) and **03** (regression).
+4. If time is left: **05** and **06**, the 45-minute timed drills. These are the closest thing to the exam.
+5. Read `EXAM_PLAYBOOK.md` last, as a final revision sheet.
+
+If you only have 2 hours: 01 (classification skeleton), 03 (regression skeleton), then the playbook. Once you can write the skeleton, other models are only a changed constructor and grid.
+
+## Troubleshooting
+
+| Symptom | Cause / fix |
+|---|---|
+| `Invalid parameter 'max_depth' for estimator Pipeline` | Grid keys inside a pipeline need the step name: `model__max_depth`. For a bare estimator use `max_depth`. |
+| `FileNotFoundError: ../data/...csv` | You moved the notebook. Edit the `DATA` line. |
+| Run is very slow | Lower `n_jobs`/grid size, or use `RandomizedSearchCV` with a small `n_iter`. Your sir said random search exists for this reason. |
+| `FutureWarning: 'penalty' was deprecated` | Harmless on sklearn 1.8+; the cheat sheet syntax still works. Warnings are hidden in the notebooks. |
+| `ValueError: Input X contains NaN` | Add the `SimpleImputer` step (it is already inside `build`). |
+| MultinomialNB error about negative values | Use `MinMaxScaler`, not `StandardScaler`, before MultinomialNB. |

ml_lab_prep-0.1.0/src/ml_lab_prep/content/check_setup.py

@@ -0,0 +1,35 @@
+"""Run this on your lab laptop BEFORE the exam:  python check_setup.py"""
+import importlib, sys
+
+print("Python", sys.version.split()[0])
+need = ["numpy", "pandas", "sklearn", "scipy", "matplotlib", "seaborn", "xgboost"]
+bad = []
+for m in need:
+    try:
+        mod = importlib.import_module(m)
+        print(f"  OK   {m:12s}{getattr(mod, '__version__', '')}")
+    except Exception as e:
+        print(f"  FAIL {m:12s}{e}")
+        bad.append(m)
+
+if not bad:
+    # exercises every cheat-sheet call that has changed between versions
+    import warnings; warnings.filterwarnings("ignore")
+    import numpy as np
+    from sklearn.ensemble import AdaBoostClassifier, BaggingClassifier, StackingClassifier
+    from sklearn.tree import DecisionTreeClassifier
+    from sklearn.linear_model import LogisticRegression
+    from sklearn.model_selection import RepeatedStratifiedKFold, cross_val_score
+    from xgboost import XGBClassifier
+    from sklearn.datasets import load_wine
+    X, y = load_wine(return_X_y=True); y = (y == 0).astype(int)   # binary: liblinear is binary-only in new sklearn
+    for name, m in [("AdaBoost(estimator=)", AdaBoostClassifier(estimator=DecisionTreeClassifier(max_depth=1))),
+                    ("Bagging(estimator=)", BaggingClassifier(estimator=DecisionTreeClassifier())),
+                    ("XGBClassifier", XGBClassifier(eval_metric="logloss")),
+                    ("LogisticRegression(penalty=l1)", LogisticRegression(penalty="l1", solver="liblinear", max_iter=2000))]:
+        try:
+            cross_val_score(m, X, y, cv=3); print(f"  OK   {name}")
+        except Exception as e:
+            print(f"  FAIL {name}: {str(e)[:80]}"); bad.append(name)
+
+print("\nALL GOOD" if not bad else f"\nFIX THESE: {bad}\n  pip install -r requirements.txt   (or: pip install xgboost)")