experiments/compare_conjugate_gradient.py [35:232]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - BostonDataset, CaliforniaHousingDataset, YearPredictionDataset, Rcv1Dataset, ) from ridge_sketch import RidgeSketch from kernel_ridge_sketch import KernelRidgeSketch from benchmarks import compute_distribution, pad_residual_norms, update_times from experiments.plot_experiments import plot_runs_over_iterations, plot_runs_over_time class ExperimentSketches: """ Class to run the experiments for different sketches """ def __init__( self, dataset_name, X, y, regularizer, is_kernel, use_heuristic, tolerance, max_iter, solvers, sketch_size, sketch_size_formula, n_repetitions, ): self.dataset_name = dataset_name self.X = X self.y = y self.regularizer = regularizer self.is_kernel = is_kernel self.use_heuristic = use_heuristic self.tolerance = tolerance self.max_iter = max_iter # List of sketch solvers self.solvers = solvers # Only one sketch size here self.sketch_size = sketch_size self.sketch_size_formula = sketch_size_formula # For error areas 1st/3rd quartiles self.n_repetitions = n_repetitions # df of times of all runs of the experiment self.times_df = None # df of relative residual norms of all runs of the experiment self.residual_norms_df = None def run_full_exp(self, verbose=True): algo_mode = "mom" mom_beta = None step_size = None mom_eta = 0.995 # increasing momentum parameter # dict of the outputs times = defaultdict(list) residual_norms = {} counter = 1 n_settings = len(self.solvers) for solver in self.solvers: if solver == "coordinate descent": sketch_size = 1 sketch_size_formula = "1" else: sketch_size = self.sketch_size sketch_size_formula = self.sketch_size_formula run_name = self.dataset_name # solver print(f"----> Setting {counter} over {n_settings} : {solver}") ( times_distribution, residual_norms_distribution, ) = self.compute_fit_time_and_residual( solver, sketch_size, algo_mode, step_size, mom_beta, mom_eta, ) # Storing the results self.book_keeping( run_name, solver, sketch_size_formula, sketch_size, times_distribution, residual_norms_distribution, times, residual_norms, ) counter += 1 print("\n") # converting the outputs to dataframes times_df = pd.DataFrame(times) if verbose: print(f"residual_norms:\n{residual_norms}") residual_norms_df = pd.DataFrame.from_dict( residual_norms, orient="index" ).transpose() self.times_df = times_df self.residual_norms_df = residual_norms_df return times_df, residual_norms_df def compute_fit_time_and_residual( self, solver, sketch_size, algo_mode, step_size, mom_beta, mom_eta ): """ Repeats model fit for n_repetitions. Returns quartile 1, 3 and median time taken """ times = [] residual_norms = [] for repetition_idx in range(self.n_repetitions): print( f"--------> Repetition {repetition_idx+1} / " f"{self.n_repetitions}", end="\r", ) random_state = repetition_idx np.random.seed(seed=random_state) model = load_model( solver, sketch_size, algo_mode, step_size, mom_beta, mom_eta, self.regularizer, self.is_kernel, self.use_heuristic, self.tolerance, self.max_iter, random_state, ) start = default_timer() model.fit(self.X, self.y) times.append(default_timer() - start) residual_norms.append(model.residual_norms) times_distribution = compute_distribution(times) residual_norms = pad_residual_norms(residual_norms) residual_norms_distribution = compute_distribution(residual_norms) return times_distribution, residual_norms_distribution def book_keeping( self, run_name, solver, sketch_size_formula, sketch_size, times_distribution, residual_norms_distribution, run_times, run_residual_norms, ): """Store time and relative residual norms distribution for each solver""" description = { "run_name": run_name, "solver": solver, "sketch_size_formula": sketch_size_formula, "sketch_size": sketch_size, } description_string = ( f"{run_name} | {solver} | sketch_size = " f"{sketch_size_formula} = {sketch_size}" ) run_time = { **description, "time (median)": times_distribution.median, "time (1st quartile)": times_distribution.q1, "time (3rd quartile)": times_distribution.q3, } run_times = update_times(run_times, run_time) run_residual_norm = { f"{description_string} | residual_norms (median)": residual_norms_distribution.median, f"{description_string} | residual_norms (1st quartile)": residual_norms_distribution.q1, f"{description_string} | residual_norms (3rd quartile)": residual_norms_distribution.q3, } run_residual_norms.update(run_residual_norm) def save(self, exp_name, save_path): full_path = os.path.join(save_path, exp_name) # clear old results shutil.rmtree(full_path, ignore_errors=True) os.makedirs(full_path) self.save_settings(exp_name, full_path) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - experiments/different_sketches.py [39:236]: - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - BostonDataset, CaliforniaHousingDataset, YearPredictionDataset, Rcv1Dataset, ) from ridge_sketch import RidgeSketch from kernel_ridge_sketch import KernelRidgeSketch from benchmarks import compute_distribution, pad_residual_norms, update_times from experiments.plot_experiments import plot_runs_over_iterations, plot_runs_over_time class ExperimentSketches: """ Class to run the experiments for different sketches """ def __init__( self, dataset_name, X, y, regularizer, is_kernel, use_heuristic, tolerance, max_iter, solvers, sketch_size, sketch_size_formula, n_repetitions, ): self.dataset_name = dataset_name self.X = X self.y = y self.regularizer = regularizer self.is_kernel = is_kernel self.use_heuristic = use_heuristic self.tolerance = tolerance self.max_iter = max_iter # List of sketch solvers self.solvers = solvers # Only one sketch size here self.sketch_size = sketch_size self.sketch_size_formula = sketch_size_formula # For error areas 1st/3rd quartiles self.n_repetitions = n_repetitions # df of times of all runs of the experiment self.times_df = None # df of relative residual norms of all runs of the experiment self.residual_norms_df = None def run_full_exp(self, verbose=True): algo_mode = "mom" mom_beta = None step_size = None mom_eta = 0.995 # increasing momentum parameter # dict of the outputs times = defaultdict(list) residual_norms = {} counter = 1 n_settings = len(self.solvers) for solver in self.solvers: if solver == "coordinate descent": sketch_size = 1 sketch_size_formula = "1" else: sketch_size = self.sketch_size sketch_size_formula = self.sketch_size_formula run_name = self.dataset_name # solver print(f"----> Setting {counter} over {n_settings} : {solver}") ( times_distribution, residual_norms_distribution, ) = self.compute_fit_time_and_residual( solver, sketch_size, algo_mode, step_size, mom_beta, mom_eta, ) # Storing the results self.book_keeping( run_name, solver, sketch_size_formula, sketch_size, times_distribution, residual_norms_distribution, times, residual_norms, ) counter += 1 print("\n") # converting the outputs to dataframes times_df = pd.DataFrame(times) if verbose: print(f"residual_norms:\n{residual_norms}") residual_norms_df = pd.DataFrame.from_dict( residual_norms, orient="index" ).transpose() self.times_df = times_df self.residual_norms_df = residual_norms_df return times_df, residual_norms_df def compute_fit_time_and_residual( self, solver, sketch_size, algo_mode, step_size, mom_beta, mom_eta ): """ Repeats model fit for n_repetitions. Returns quartile 1, 3 and median time taken """ times = [] residual_norms = [] for repetition_idx in range(self.n_repetitions): print( f"--------> Repetition {repetition_idx+1} / " f"{self.n_repetitions}", end="\r", ) random_state = repetition_idx np.random.seed(seed=random_state) model = load_model( solver, sketch_size, algo_mode, step_size, mom_beta, mom_eta, self.regularizer, self.is_kernel, self.use_heuristic, self.tolerance, self.max_iter, random_state, ) start = default_timer() model.fit(self.X, self.y) times.append(default_timer() - start) residual_norms.append(model.residual_norms) times_distribution = compute_distribution(times) residual_norms = pad_residual_norms(residual_norms) residual_norms_distribution = compute_distribution(residual_norms) return times_distribution, residual_norms_distribution def book_keeping( self, run_name, solver, sketch_size_formula, sketch_size, times_distribution, residual_norms_distribution, run_times, run_residual_norms, ): """Store time and relative residual norms distribution for each solver""" description = { "run_name": run_name, "solver": solver, "sketch_size_formula": sketch_size_formula, "sketch_size": sketch_size, } description_string = ( f"{run_name} | {solver} | sketch_size = " f"{sketch_size_formula} = {sketch_size}" ) run_time = { **description, "time (median)": times_distribution.median, "time (1st quartile)": times_distribution.q1, "time (3rd quartile)": times_distribution.q3, } run_times = update_times(run_times, run_time) run_residual_norm = { f"{description_string} | residual_norms (median)": residual_norms_distribution.median, f"{description_string} | residual_norms (1st quartile)": residual_norms_distribution.q1, f"{description_string} | residual_norms (3rd quartile)": residual_norms_distribution.q3, } run_residual_norms.update(run_residual_norm) def save(self, exp_name, save_path): full_path = os.path.join(save_path, exp_name) # clear old results shutil.rmtree(full_path, ignore_errors=True) os.makedirs(full_path) self.save_settings(exp_name, full_path) - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - - -