Merge pull request #28 from microsoft/v0.2.4

v0.2.4

README.md

@@ -5,8 +5,8 @@
     <br>
 </p>
 
-FLAML is a Python library designed to automatically produce accurate machine
-learning models with low computational cost. It frees users from selecting
+FLAML is a lightweight Python library that finds accurate machine
+learning models automatically, efficiently and economically. It frees users from selecting
 learners and hyperparameters for each learner. It is fast and cheap.
 The simple and lightweight design makes it easy to extend, such as
 adding customized learners or metrics. FLAML is powered by a new, [cost-effective

flaml/automl.py

@@ -402,7 +402,7 @@ class AutoML:
             self._X_train_all, self._y_train_all = \
                 self._transformer.fit_transform(X, y, self._state.task)
             self._label_transformer = self._transformer.label_transformer
-
+        self._sample_weight_full = self._state.fit_kwargs.get('sample_weight')      
         if X_val is not None and y_val is not None:
             if not (isinstance(X_val, np.ndarray) or
                 issparse(X_val) or
@@ -446,7 +446,8 @@ class AutoML:
             self._X_train_all, self._y_train_all
         if issparse(X_train_all): 
             X_train_all = X_train_all.tocsr()
-        if self._state.task != 'regression':
+        if self._state.task != 'regression' and self._state.fit_kwargs.get(
+            'sample_weight') is None:
             # logger.info(f"label {pd.unique(y_train_all)}")
             label_set, counts = np.unique(y_train_all, return_counts=True)
             # augment rare classes
@@ -1093,8 +1094,9 @@ class AutoML:
                             self._state.best_loss))
             else:
                 logger.info(f"no enough budget for learner {estimator}")
-                self.estimator_list.remove(estimator)
-                self._estimator_index -= 1
+                if self._estimator_index is not None:
+                    self.estimator_list.remove(estimator)
+                    self._estimator_index -= 1
             if self._retrain_full and best_config_sig and not better and (
                 self._search_states[self._best_estimator].sample_size ==
                 self._state.data_size) and (est_retrain_time <=
@@ -1151,7 +1153,11 @@ class AutoML:
                 stacker = Stacker(estimators, best_m,
                     n_jobs=self._state.n_jobs,
                     passthrough=True)
-                stacker.fit(self._X_train_all, self._y_train_all)
+                if self._sample_weight_full is not None:
+                    self._state.fit_kwargs[
+                        'sample_weight'] = self._sample_weight_full
+                stacker.fit(self._X_train_all, self._y_train_all,
+                 **self._state.fit_kwargs)
                 logger.info(f'ensemble: {stacker}')
                 self._trained_estimator = stacker
                 self._trained_estimator.model = stacker

flaml/searcher/flow2.py

@@ -121,8 +121,8 @@ class FLOW2(Searcher):
         self._unordered_cat_hp = {}
         self._cat_hp_cost = {}
         for key, domain in self.space.items():
-            assert not isinstance(domain, dict), \
-                key+"'s domain is grid search which is not supported in FLOW2."
+            assert not (isinstance(domain, dict) and 'grid_search' in domain
+            ), key+"'s domain is grid search which is not supported in FLOW2."
             if callable(getattr(domain, 'get_sampler', None)):
                 self._tunable_keys.append(key)
                 sampler = domain.get_sampler()

flaml/tune/README.md

@@ -6,6 +6,7 @@ The API is compatible with ray tune.
 Example:
 
 ```python
+# require: pip install flaml[blendsearch]
 from flaml import tune
 import time
 
@@ -42,6 +43,7 @@ print(analysis.best_config) # the best config
 
 Or, using ray tune's API:
 ```python
+# require: pip install flaml[blendsearch] ray[tune]
 from ray import tune as raytune
 from flaml import CFO, BlendSearch
 import time
@@ -146,6 +148,7 @@ based on optimism in face of uncertainty.
 Example:
 
 ```python
+# require: pip install flaml[blendsearch]
 from flaml import BlendSearch
 tune.run(...
     search_alg = BlendSearch(points_to_evaluate=[init_config]),

flaml/version.py

@@ -1 +1 @@
-__version__ = "0.2.3"
+__version__ = "0.2.4"

notebook/finetune_transformer_demo.ipynb

@@ -6,11 +6,16 @@
    "source": [
     "This notebook uses the Huggingface transformers library to finetune a transformer model.\n",
     "\n",
-    "**Requirements.** This notebook has additional requirements:\n",
-    "\n",
-    "```bash\n",
-    "pip install -r transformers_requirements.txt\n",
-    "```"
+    "**Requirements.** This notebook has additional requirements:"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "!pip install torch transformers datasets ipywidgets"
    ]
   },
   {
@@ -699,7 +704,7 @@
    "source": [
     "### Step 3. Launch with `flaml.tune.run`\n",
     "\n",
-    "We are now ready to laungh the tuning using `flaml.tune.run`:"
+    "We are now ready to launch the tuning using `flaml.tune.run`:"
    ],
    "cell_type": "markdown",
    "metadata": {}
@@ -766,9 +771,13 @@
  ],
  "metadata": {
   "kernelspec": {
-   "display_name": "flaml",
-   "language": "python",
-   "name": "flaml"
+   "name": "python3",
+   "display_name": "Python 3.7.7 64-bit ('flaml': conda)",
+   "metadata": {
+    "interpreter": {
+     "hash": "bfcd9a6a9254a5e160761a1fd7a9e444f011592c6770d9f4180dde058a9df5dd"
+    }
+   }
   },
   "language_info": {
    "codemirror_mode": {
@@ -780,7 +789,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.7.6"
+   "version": "3.7.7-final"
   }
  },
  "nbformat": 4,

notebook/transformers_requirements.txt

@@ -1,4 +0,0 @@
-torch
-transformers
-datasets
-ipywidgets

setup.py

@@ -20,7 +20,6 @@ install_requires = [
     "scipy>=1.4.1",
     "catboost>=0.23",
     "scikit-learn>=0.23.2",
-    "optuna==2.3.0"
 ],
 
 
@@ -48,6 +47,10 @@ setuptools.setup(
             "coverage>=5.3",
             "xgboost<1.3",
             "rgf-python",
+            "optuna==2.3.0",
+        ],
+        "blendsearch": [
+            "optuna==2.3.0"
         ],
         "ray": [
             "ray[tune]==1.1.0",

test/test_pytorch_cifar10.py

@@ -1,3 +1,5 @@
+'''Require: pip install torchvision ray
+'''
 import unittest
 import os
 import time

test/test_tune.py

@@ -1,5 +1,6 @@
+'''Require: pip install flaml[test,ray]
+'''
 import unittest
-import os
 import time
 from sklearn.model_selection import train_test_split
 import sklearn.metrics
@@ -138,6 +139,7 @@ def _test_xgboost(method='BlendSearch'):
                     scheduler=scheduler, search_alg=algo)
             ray.shutdown()
             # # Load the best model checkpoint
+            # import os
             # best_bst = xgb.Booster()
             # best_bst.load_model(os.path.join(analysis.best_checkpoint,
             #  "model.xgb"))