Improve NN runtime by reducing num_categories_to_keep (#246)

* changed default num_categories_to_keep hyperparameter of NN to improve runtimes

* allow NN to be trained for 0 epochs

autogluon/utils/tabular/ml/models/tabular_nn/hyperparameters/parameters.py

@@ -13,7 +13,7 @@ def get_fixed_params():
         # Options: [3,4,10, 100, 1000]
         'proc.impute_strategy': 'median',  # # strategy argument of sklearn.SimpleImputer() used to impute missing numeric values
         # Options: ['median', 'mean', 'most_frequent']
-        'proc.max_category_levels': 500,  # maximum number of allowed levels per categorical feature
+        'proc.max_category_levels': 100,  # maximum number of allowed levels per categorical feature
         # Options: [10, 100, 200, 300, 400, 500, 1000, 10000]
         'proc.skew_threshold': 0.99,  # numerical features whose absolute skewness is greater than this receive special power-transform preprocessing. Choose big value to avoid using power-transforms
         # Options: [0.2, 0.3, 0.5, 0.8, 1.0, 10.0, 100.0]

autogluon/utils/tabular/ml/models/tabular_nn/tabular_nn_model.py

@@ -293,6 +293,23 @@ def train_net(self, params, train_dataset, test_dataset=None,
         else:
             verbose_eval = 1
 
+        if num_epochs == 0: # use dummy training loop that stops immediately (useful for using NN just for data preprocessing / debugging)
+            logger.log(20, "Not training Neural Net since num_epochs == 0.  Neural network architecture is:")
+            for batch_idx, data_batch in enumerate(train_dataset.dataloader):
+                data_batch = train_dataset.format_batch_data(data_batch, self.ctx)
+                with autograd.record():
+                    output = self.model(data_batch)
+                    labels = data_batch['label']
+                    loss = self.loss_func(output, labels) / loss_scaling_factor
+                    # print(str(nd.mean(loss).asscalar()), end="\r") # prints per-batch losses
+                loss.backward()
+                self.optimizer.step(labels.shape[0])
+                if batch_idx > 0:
+                    break
+            self.model.save_parameters(self.net_filename)
+            logger.log(15, "untrained Neural Net saved to file")
+            return
+
         # Training Loop:
         for e in range(num_epochs):
             if e == 0: # special actions during first epoch: