Added infos for config options

Author: Lucas Zimmer

.gitignore

@@ -45,3 +45,6 @@ test_predictions_for_ensemble.npy
 
 # testing
 tests.ipynb
+
+# venv
+env/

autoPyTorch/components/networks/image/mobilenet.py

@@ -194,8 +194,6 @@ def _cfg(url='', **kwargs):
 
         self.model.default_cfg = _cfg(url='', input_size=in_features, pool_size=(10, 10), crop_pct=0.904, num_classes=out_features)
 
-        self.layers = nn.Sequential(self.model.forward_features)
-
     def forward(self, x):
         # make sure channels first
         x = self.model(x)

autoPyTorch/components/training/image/trainer.py

@@ -69,7 +69,6 @@ def train(self, epoch, train_loader, metrics):
         metric_results = [0] * len(metrics)
         start_time = time.time()
         for step, (data, targets) in enumerate(train_loader):
-
             # import matplotlib.pyplot as plt
             # img = plt.imshow(data.numpy()[0,1,:])
             # plt.show()
@@ -125,6 +124,8 @@ def train(self, epoch, train_loader, metrics):
                 budget_exceeded = True
                 break
 
+        if N==0: # Fixes a bug during initialization
+            N=1
 
         if self.images_plot_count > 0:
             import tensorboard_logger as tl

autoPyTorch/core/api.py

@@ -269,17 +269,24 @@ def score(self, X_test, Y_test, return_loss_value=False):
         return metric(torch.from_numpy(Y_pred.astype(np.float32)), torch.from_numpy(Y_test.astype(np.float32)))
 
     def get_pytorch_model(self):
-        """Returns a pytorch sequential model of the current incumbent configuration
+        """Returns a pytorch sequential model of the current incumbent configuration. Not possible for all models.
         
         Arguments:
         
         Returns:
             model -- PyTorch sequential model of the current incumbent configuration
         """
-        if NetworkSelector.get_name() in self.pipeline:
-            return self.pipeline[NetworkSelector.get_name()].fit_output["network"].layers
-        else:
-            return self.pipeline[NetworkSelectorDatasetInfo.get_name()].fit_output["network"].layers
+        try:
+            if NetworkSelector.get_name() in self.pipeline:
+                return self.pipeline[NetworkSelector.get_name()].fit_output["network"].layers
+            else:
+                return self.pipeline[NetworkSelectorDatasetInfo.get_name()].fit_output["network"].layers
+        except:
+            print("Can not get PyTorch Sequential model for incumbent config. Returning Auto-PyTorch model")
+            if NetworkSelector.get_name() in self.pipeline:
+                return self.pipeline[NetworkSelector.get_name()].fit_output["network"]
+            else:
+                return self.pipeline[NetworkSelectorDatasetInfo.get_name()].fit_output["network"]
 
     def initialize_from_checkpoint(self, hyperparameter_config, checkpoint, in_features, out_features, final_activation=None):
         """

autoPyTorch/pipeline/nodes/image/create_dataset_info.py

@@ -77,8 +77,8 @@ def predict(self, pipeline_config, X):
     def get_pipeline_config_options(self):
         options = [
             ConfigOption(name="file_extensions", default=['.png', '.jpg', '.JPEG', '.pgm'], type=str, list=True),
-            ConfigOption(name="images_shape", default=[3, 32, 32], type=int, list=True),
-            ConfigOption(name="images_root_folders", default=[ConfigFileParser.get_autonet_home()], type='directory', list=True),
+            ConfigOption(name="images_shape", default=[3, 32, 32], type=int, list=True, info="Image size input to the networks, images will be rescaled to this."),
+            ConfigOption(name="images_root_folders", default=[ConfigFileParser.get_autonet_home()], type='directory', list=True, info="Directory relative to which image paths are given."),
             ConfigOption(name="max_class_size", default=None, type=int),
         ]
         return options

autoPyTorch/pipeline/nodes/image/create_image_dataloader.py

@@ -63,7 +63,7 @@ def fit(self, pipeline_config, hyperparameter_config, X, Y, train_indices, valid
 
     def get_pipeline_config_options(self):
         options = [
-            ConfigOption("default_dataset_download_dir", default=ConfigFileParser.get_autonet_home(), type='directory'),
+            ConfigOption("default_dataset_download_dir", default=ConfigFileParser.get_autonet_home(), type='directory', info="Directory default datasets will be downloaded to."),
             ConfigOption("dataloader_worker", default=1, type=int),
             ConfigOption("dataloader_cache_size_mb", default=0, type=int)
         ]

autoPyTorch/pipeline/nodes/image/image_augmentation.py

@@ -210,10 +210,12 @@ def compute_mean_std(self, pipeline_config, hyperparameter_config, X, Y, train_i
                 std = std + data.std(2).sum(0)
                 nb_samples += batch_samples
 
-        mean /= nb_samples
-        std /= nb_samples
-
-        mean, std = mean.numpy().tolist(), std.numpy().tolist()
+        if nb_samples > 0.:
+            mean /= nb_samples
+            std /= nb_samples
+            mean, std = mean.numpy().tolist(), std.numpy().tolist()
+        else:
+            mean, std = [mean], [std]
 
         log.debug('MEAN: ' + str(mean) + ' -- STD: ' + str(std))
 

autoPyTorch/pipeline/nodes/image/multiple_datasets.py

@@ -107,9 +107,9 @@ def predict(self, pipeline_config, X):
 
     def get_pipeline_config_options(self):
         options = [
-            ConfigOption('dataset_order', default=None, type=int, list=True),
+            ConfigOption('dataset_order', default=None, type=int, list=True, info="Order in which datasets are considered."),
 
             #autonet.refit sets this to false to avoid refit budget issues
-            ConfigOption('increase_number_of_trained_datasets', default=True, type=to_bool) 
+            ConfigOption('increase_number_of_trained_datasets', default=True, type=to_bool, info="Wether to increase the number of considered datasets with each successive halfing iteration.") 
         ]
         return options

autoPyTorch/pipeline/nodes/image/optimization_algorithm_no_timelimit.py

@@ -151,16 +151,16 @@ def get_pipeline_config_options(self):
             ConfigOption("task_id", default=-1, type=int, info="ID for each worker, if you run AutoNet on a cluster. Set to -1, if you run it locally. "),
             ConfigOption("algorithm", default="bohb", type=str, choices=list(self.algorithms.keys())),
             ConfigOption("budget_type", default="time", type=str, choices=['time', 'epochs']),
-            ConfigOption("min_budget", default=lambda c: 120 if c['budget_type'] == 'time' else 5, type=float, depends=True),
-            ConfigOption("max_budget", default=lambda c: 6000 if c['budget_type'] == 'time' else 150, type=float, depends=True),
+            ConfigOption("min_budget", default=lambda c: 120 if c['budget_type'] == 'time' else 5, type=float, depends=True, info="Min budget for fitting configurations."),
+            ConfigOption("max_budget", default=lambda c: 6000 if c['budget_type'] == 'time' else 150, type=float, depends=True, info="Max budget for fitting configurations."),
             ConfigOption("max_runtime", 
                 default=lambda c: ((-int(np.log(c["min_budget"] / c["max_budget"]) / np.log(c["eta"])) + 1) * c["max_budget"])
                         if c["budget_type"] == "time" else float("inf"),
-                type=float, depends=True),
+                type=float, depends=True, info="Total time for the run."),
             ConfigOption("num_iterations", 
                 default=lambda c:  (-int(np.log(c["min_budget"] / c["max_budget"]) / np.log(c["eta"])) + 1)
                         if c["budget_type"] == "epochs" else float("inf"),
-                type=float, depends=True),
+                type=float, depends=True, info="Number of successive halving iterations"),
             ConfigOption("eta", default=3, type=float, info='eta parameter of Hyperband.'),
             ConfigOption("min_workers", default=1, type=int),
             ConfigOption("working_dir", default=".", type="directory"),

autoPyTorch/pipeline/nodes/image/simple_train_node.py

@@ -295,7 +295,7 @@ def get_pipeline_config_options(self):
                 type=str, list=True, choices=list(self.batch_loss_computation_techniques.keys())),
             ConfigOption("minimize", default=self.default_minimize_value, type=to_bool, choices=[True, False]),
             ConfigOption("cuda", default=True, type=to_bool, choices=[True, False]),
-            ConfigOption("save_checkpoints", default=False, type=to_bool, choices=[True, False]),
+            ConfigOption("save_checkpoints", default=False, type=to_bool, choices=[True, False], info="Wether to save state dicts as checkpoints."),
             ConfigOption("tensorboard_min_log_interval", default=30, type=int),
             ConfigOption("tensorboard_images_count", default=0, type=int),
             ConfigOption("evaluate_on_train_data", default=True, type=to_bool),