added HyperOpt dev dependency

Author: ltiao

bore/engine.py

@@ -9,7 +9,7 @@
 # `plugins.hpbandster` and place it here.
 
 
-class Ledger:
+class Record:
 
     def __init__(self):
         self.features = []
@@ -47,13 +47,13 @@ def to_dataframe(self):
                                                         loss=self.targets)
         return frame
 
-
-def is_duplicate(x, xs, rtol=1e-5, atol=1e-8):
-    # Clever ways of doing this would involve data structs. like KD-trees
-    # or locality sensitive hashing (LSH), but these are premature
-    # optimizations at this point, especially since the `any` below does lazy
-    # evaluation, i.e. is early stopped as soon as anything returns `True`.
-    return any(np.allclose(x_prev, x, rtol=rtol, atol=atol) for x_prev in xs)
+    def is_duplicate(self, x, rtol=1e-5, atol=1e-8):
+        # Clever ways of doing this would involve data structs. like KD-trees
+        # or locality sensitive hashing (LSH), but these are premature
+        # optimizations at this point, especially since the `any` below does lazy
+        # evaluation, i.e. is early stopped as soon as anything returns `True`.
+        return any(np.allclose(x_prev, x, rtol=rtol, atol=atol)
+                   for x_prev in self.features)
 
 
 minimize_multi_start = multi_start(minimizer_fn=minimize)

bore/plotting.py

@@ -4,16 +4,6 @@
 import matplotlib.pyplot as plt
 
 
-def plot_image_grid(ax, images, shape, nrows=20, ncols=None, cmap=None):
-
-    if ncols is None:
-        ncols = nrows
-
-    grid = images[:nrows*ncols].reshape(nrows, ncols, *shape).squeeze()
-
-    return ax.imshow(np.vstack(np.dstack(grid)), cmap=cmap)
-
-
 def fill_between_stddev(X_pred, mean_pred, stddev_pred, n=1, ax=None, *args,
                         **kwargs):
 

bore/plugins/hpbandster.py

@@ -4,7 +4,7 @@
 from tensorflow.keras.losses import BinaryCrossentropy
 import scipy.stats as sps
 
-from ..engine import Ledger, minimize_multi_start, is_duplicate
+from ..engine import Record, minimize_multi_start
 from ..types import DenseConfigurationSpace, DenseConfiguration
 from ..models import DenseSequential
 from ..decorators import unbatch, value_and_gradient, numpy_io
@@ -20,19 +20,21 @@ def __init__(self, config_space, eta=3, min_budget=0.01, max_budget=1,
                  num_restarts=10, batch_size=64, num_steps_per_iter=1000,
                  optimizer="adam", num_layers=2, num_units=32,
                  activation="relu", normalize=True, method="L-BFGS-B",
-                 max_iter=100, ftol=1e-2, distortion=1e-3, seed=None, **kwargs):
+                 max_iter=100, ftol=1e-2, distortion=None, seed=None, **kwargs):
 
         if gamma is None:
             gamma = 1/eta
 
         cg = RatioEstimator(config_space=config_space, gamma=gamma,
-                            num_random_init=num_random_init, random_rate=random_rate,
-                            num_restarts=num_restarts, batch_size=batch_size,
+                            num_random_init=num_random_init,
+                            random_rate=random_rate, num_restarts=num_restarts,
+                            batch_size=batch_size,
                             num_steps_per_iter=num_steps_per_iter,
                             optimizer=optimizer, num_layers=num_layers,
                             num_units=num_units, activation=activation,
                             normalize=normalize, method=method,
-                            max_iter=max_iter, ftol=ftol, distortion=distortion, seed=seed)
+                            max_iter=max_iter, ftol=ftol,
+                            distortion=distortion, seed=seed)
         # (LT): Note this is using the *grandparent* class initializer to
         # replace the config_generator!
         super(HyperBand, self).__init__(config_generator=cg, **kwargs)
@@ -71,7 +73,7 @@ def __init__(self, config_space, gamma=1/3, num_random_init=10,
                  random_rate=0.25, num_restarts=3, batch_size=64,
                  num_steps_per_iter=1000, optimizer="adam", num_layers=2,
                  num_units=32, activation="relu", normalize=True,
-                 method="L-BFGS-B", max_iter=100, ftol=1e-2, distortion=1e-3,
+                 method="L-BFGS-B", max_iter=100, ftol=1e-2, distortion=None,
                  seed=None, **kwargs):
 
         super(RatioEstimator, self).__init__(**kwargs)
@@ -106,7 +108,7 @@ def __init__(self, config_space, gamma=1/3, num_random_init=10,
         self.batch_size = batch_size
         self.num_steps_per_iter = num_steps_per_iter
 
-        self.ledger = Ledger()
+        self.record = Record()
 
         self.seed = seed
         self.random_state = np.random.RandomState(seed)
@@ -151,9 +153,9 @@ def loss(x):
 
     def _update_model(self):
 
-        X, z = self.ledger.load_classification_data(self.gamma)
+        X, z = self.record.load_classification_data(self.gamma)
 
-        dataset_size = self.ledger.size()
+        dataset_size = self.record.size()
         steps_per_epoch = self._get_steps_per_epoch(dataset_size)
         num_epochs = self.num_steps_per_iter // steps_per_epoch
 
@@ -189,8 +191,10 @@ def _get_maximum(self):
                               f" ({res.message})")
 
             # TODO(LT): Create Enum type for these status codes
-            if (res.status == 0 or res.status == 9) and \
-                    not is_duplicate(res.x, self.ledger.features):
+            # status == 1 signifies maximum iteration reached, which we don't
+            # want to treat as a failure condition.
+            if (res.success or res.status == 1) and \
+                    not self.record.is_duplicate(res.x):
                 # if (res_best is not None) *implies* (res.fun < res_best.fun)
                 # (i.e. material implication) is logically equivalent to below
                 if res_best is None or res.fun < res_best.fun:
@@ -200,20 +204,20 @@ def _get_maximum(self):
 
     def get_config(self, budget):
 
-        dataset_size = self.ledger.size()
+        dataset_size = self.record.size()
 
         config_random = self.config_space.sample_configuration()
         config_random_dict = config_random.get_dictionary()
 
         if dataset_size < self.num_random_init:
             self.logger.debug(f"Completed {dataset_size}/{self.num_random_init}"
-                              " initial runs. Returning random candidate...")
+                              " initial runs. Suggesting random candidate...")
             return (config_random_dict, {})
 
         if self.random_state.binomial(p=self.random_rate, n=1):
             self.logger.info("[Glob. maximum: skipped "
                              f"(prob={self.random_rate:.2f})] "
-                             "Returning random candidate ...")
+                             "Suggesting random candidate ...")
             return (config_random_dict, {})
 
         # Update model
@@ -227,11 +231,9 @@ def get_config(self, budget):
             self.logger.warn("[Glob. maximum: not found!] Either optimization "
                              f"failed in all {self.num_restarts} starts, or "
                              "all maxima found have been evaluated previously!"
-                             " Returning random candidate...")
+                             " Suggesting random candidate...")
             return (config_random_dict, {})
 
-        self.logger.info(f"[Glob. maximum: value={-opt.fun:.3f}, x={opt.x}")
-
         loc = opt.x
 
         if self.distortion is None:
@@ -246,6 +248,9 @@ def get_config(self, budget):
 
             config_opt_arr = dist.rvs(random_state=self.random_state)
 
+        self.logger.info(f"[Glob. maximum: value={-opt.fun:.3f} x={loc}] "
+                         f"Suggesting x={config_opt_arr}")
+
         config_opt_dict = self._dict_from_array(config_opt_arr)
 
         return (config_opt_dict, {})
@@ -262,4 +267,4 @@ def new_result(self, job, update_model=True):
 
         loss = job.result["loss"]
 
-        self.ledger.append(x=config_arr, y=loss, b=budget)
+        self.record.append(x=config_arr, y=loss, b=budget)

bore/utils.py

@@ -1,26 +0,0 @@
-import pandas as pd
-
-
-def dataframe_from_result(results):
-
-    rows = []
-
-    for task, config_id in enumerate(results.data):
-
-        d = results.data[config_id]
-        bracket, _, _ = config_id
-
-        for epoch in d.results:
-
-            row = dict(task=task,
-                       bracket=bracket,
-                       epoch=int(epoch),
-                       loss=d.results[epoch]["loss"],
-                       info=d.results[epoch]["info"],
-                       submitted=d.time_stamps[epoch]["submitted"],
-                       started=d.time_stamps[epoch]["started"],
-                       finished=d.time_stamps[epoch]["finished"])
-            row.update(d.config)
-            rows.append(row)
-
-    return pd.DataFrame(rows)

requirements.txt

@@ -1,12 +1,7 @@
-tensorflow-probability==0.11.0
 tqdm>=4.48.2
-# numpy>=1.17.3
 scipy>=1.4.1
 scikit-learn>=0.23.2
 pandas>=1.1.1
 seaborn>=0.11.0
 tables>=3.6.1
 statsmodels>=0.12.0
-
-GPy==1.9.9
-gpyopt==1.2.6

requirements_dev.txt

@@ -16,3 +16,7 @@ nbsphinx
 sphinx_bootstrap_theme
 sphinx-gallery>=0.7.0
 pillow
+
+GPy==1.9.9
+gpyopt==1.2.6
+hyperopt

scripts/plotting/plot_result.py

@@ -1,5 +1,6 @@
 import sys
 import click
+import yaml
 
 import pandas as pd
 
@@ -10,28 +11,27 @@
 from utils import (GOLDEN_RATIO, WIDTH, size, load_frame, extract_series,
                    merge_stack_series, get_error_mins, sanitize, get_ci)
 
-OUTPUT_DIR = "figures/"
-
 
 @click.command()
 @click.argument("benchmark_name")
-@click.argument("input_dir", default="results",
+@click.argument("input_dir", default="results/",
+                type=click.Path(file_okay=False, dir_okay=True))
+@click.argument("output_dir", default="figures/",
                 type=click.Path(file_okay=False, dir_okay=True))
 @click.option('--num-runs', '-n', default=20)
 @click.option('--methods', '-m', multiple=True)
 @click.option('--ci')
-@click.option('--duration-key', default="info")
+@click.option('--duration-key', default=None)
 @click.option('--context', default="paper")
 @click.option('--style', default="ticks")
 @click.option('--palette', default="muted")
 @click.option('--width', '-w', type=float, default=WIDTH)
 @click.option('--aspect', '-a', type=float, default=GOLDEN_RATIO)
 @click.option('--extension', '-e', multiple=True, default=["png"])
-@click.option("--output-dir", default=OUTPUT_DIR,
-              type=click.Path(file_okay=False, dir_okay=True),
-              help="Output directory.")
-def main(benchmark_name, input_dir, num_runs, methods, ci, duration_key,
-         context, style, palette, width, aspect, extension, output_dir):
+@click.option("--config-file", type=click.File('r'))
+def main(benchmark_name, input_dir, output_dir, num_runs, methods, ci,
+         duration_key, context, style, palette, width, aspect, extension,
+         config_file):
 
     figsize = size(width, aspect)
     height = width / aspect
@@ -48,19 +48,8 @@ def main(benchmark_name, input_dir, num_runs, methods, ci, duration_key,
     output_path = Path(output_dir).joinpath(benchmark_name)
     output_path.mkdir(parents=True, exist_ok=True)
 
-    METHOD_PRETTY_NAMES = {
-        "random": "Random",
-        "tpe": "TPE",
-        "bore": "BORE",
-        "bore-steps-50": "BORE (50 steps)",
-        "bore-steps": "BORE (250 steps)",
-        "boredom": "BOREDOM",
-        "boredom-real": "BORE",
-        "bore-sigmoid-elu-ftol-1e-2-gamma-0.33333333333333333333": "BORE",
-        "boredom-real": "BORE",
-        "bore-logit-elu-ftol-1e-2-gamma-0.33333333333333333333": "BORE",
-        "bore-sigmoid-elu-ftol-1e-9-random-0.1": "BORE"
-    }
+    config = yaml.safe_load(config_file) if config_file else {}
+    method_names_mapping = config.get("names", {})
 
     loss_min = get_error_mins(benchmark_name, input_dir,
                               data_dir="datasets/fcnet_tabular_benchmarks")
@@ -78,28 +67,25 @@ def main(benchmark_name, input_dir, num_runs, methods, ci, duration_key,
             frame = load_frame(path, run, loss_min=loss_min,
                                duration_key=duration_key)
             frames.append(frame.assign(method=method))
-
             series[run] = extract_series(frame, index="elapsed", column="regret")
 
         frame_merged = merge_stack_series(series, y_key="regret")
         frames_merged.append(frame_merged.assign(method=method))
 
     data = pd.concat(frames, axis="index", ignore_index=True, sort=True)
-    data = sanitize(data, mapping=METHOD_PRETTY_NAMES)
+    data = sanitize(data, mapping=method_names_mapping)
 
     data_merged = pd.concat(frames_merged, axis="index", ignore_index=True, sort=True)
-    data_merged = sanitize(data_merged, mapping=METHOD_PRETTY_NAMES)
-
-    print(data_merged)
+    data_merged = sanitize(data_merged, mapping=method_names_mapping)
 
-    hue_order = style_order = list(map(METHOD_PRETTY_NAMES.get, methods))
+    hue_order = style_order = list(map(method_names_mapping.get, methods))
 
     fig, ax = plt.subplots()
     sns.despine(fig=fig, ax=ax, top=True)
 
     sns.lineplot(x="evaluation", y="regret",
                  hue="method",  # hue_order=hue_order,
-                 style="method",  # style_order=style_order,
+                 # style="method",  style_order=style_order,
                  # units="run", estimator=None,
                  ci=get_ci(ci), err_kws=dict(edgecolor='none'),
                  data=data, ax=ax)
@@ -120,7 +106,7 @@ def main(benchmark_name, input_dir, num_runs, methods, ci, duration_key,
 
     sns.lineplot(x="elapsed", y="regret",
                  hue="method",  # hue_order=hue_order,
-                 style="method",  # style_order=style_order,
+                 # style="method", style_order=style_order,
                  # units="run", estimator=None,
                  ci=get_ci(ci), err_kws=dict(edgecolor='none'),
                  data=data_merged, ax=ax)
@@ -155,7 +141,7 @@ def main(benchmark_name, input_dir, num_runs, methods, ci, duration_key,
     # # for ext in extension:
     # #     g.savefig(output_path.joinpath(f"error_vs_iterations_{context}_{suffix}.{ext}"))
 
-    # return 0
+    return 0
 
 
 if __name__ == "__main__":