fixed issues 61/62

examples/cross_val_fashion.py

@@ -0,0 +1,81 @@
+"""Simple Convolution and fully connected blocks cross validation example."""
+from vulcanai import datasets
+from vulcanai.models import ConvNet, DenseNet
+from vulcanai.models.metrics import Metrics
+
+import torchvision.transforms as transforms
+from torch.utils.data import DataLoader
+
+# prepare the data
+normalize = transforms.Normalize(mean=[x/255.0 for x in [125.3, 123.0, 113.9]],
+                                 std=[x/255.0 for x in [63.0, 62.1, 66.7]])
+
+transform = transforms.Compose([transforms.ToTensor(),
+                                normalize])
+
+
+data_path = "../data"
+dataset = datasets.FashionData(root=data_path,
+                                     train=True,
+                                     transform=transform,
+                                     download=True)
+
+batch_size = 100
+
+data_loader = DataLoader(dataset=dataset,
+                          batch_size=batch_size,
+                          shuffle=True)
+
+
+
+# define neural network - 3 2D conv layers followed by a dense layer
+conv_2D_config = {
+    'conv_units': [
+                    dict(
+                        in_channels=1,
+                        out_channels=16,
+                        kernel_size=(5, 5),
+                        stride=2,
+                        dropout=0.1
+                    ),
+                    dict(
+                        in_channels=16,
+                        out_channels=32,
+                        kernel_size=(5, 5),
+                        dropout=0.1
+                    ),
+                    dict(
+                        in_channels=32,
+                        out_channels=64,
+                        kernel_size=(5, 5),
+                        pool_size=2,
+                        dropout=0.1
+                        )
+    ],
+}
+
+dense_config = {
+    'dense_units': [100, 50],
+    'dropout': 0.5,  # Single value or List
+}
+
+conv_2D = ConvNet(
+    name='conv_2D',
+    in_dim=(1, 28, 28),
+    config=conv_2D_config
+)
+
+dense_model = DenseNet(
+    name='dense_model',
+    input_networks=conv_2D,
+    config=dense_config,
+    num_classes=10,
+    early_stopping="best_validation_error",
+    early_stopping_patience=2
+)
+
+
+# cross validate on 5 folds training each fold for 2 epochs
+m = Metrics()
+
+m.cross_validate(dense_model, data_loader, 5, 2)

vulcanai/models/basenetwork.py

@@ -28,6 +28,7 @@
 
 sns.set(style='dark')
 logger = logging.getLogger(__name__)
+logger.addHandler(logging.StreamHandler())
 
 
 # Because pytorch causes a bunch of unresolved references
@@ -651,9 +652,11 @@ def fit(self, train_loader, val_loader, epochs,
                 else:
                     save_path = save_path + '/' + self.name + '_'
                 save_path = get_save_path(save_path, vis_type='train')
-            iterator = trange(epochs, desc='Epoch: ')
+            # iterator = trange(epochs, desc='Epoch: ')
 
-            for epoch in iterator:
+            for epoch in range(epochs):
+
+                logger.info('\n -------- Epoch: {}  --------\n'.format(epoch))
 
                 train_loss, train_acc = self._train_epoch(train_loader,
                                                           retain_graph)
@@ -687,7 +690,7 @@ def fit(self, train_loader, val_loader, epochs,
                         self.__dict__.update(self.load_model(
                             early_stopping.save_path).__dict__)
                         # for tqdm
-                        iterator.close()
+                        # iterator.close()
                         break
 
                 # reset from None so that a distinction can be made
@@ -697,15 +700,26 @@ def fit(self, train_loader, val_loader, epochs,
                 if not valid_acc:
                     valid_acc = np.nan
 
-                tqdm.write(
-                    "\n Epoch {}:\n"
-                    "Train Loss: {:.6f} | Val Loss: {:.6f} |"
-                    "Train Acc: {:.4f} | Val Acc: {:.4f}".format(
-                        self.epoch,
-                        train_loss,
-                        valid_loss,
-                        train_acc,
-                        valid_acc))
+
+                if epoch % valid_interv == 0:
+                    tqdm.write(
+                        "\nEpoch {} Summary:\n"
+                        "Train Loss: {:.6f} | Val Loss: {:.6f} |"
+                        "Train Acc: {:.4f} | Val Acc: {:.4f} \n".format(
+                            self.epoch,
+                            train_loss,
+                            valid_loss,
+                            train_acc,
+                            valid_acc))
+
+                else:
+                    tqdm.write(
+                        "\nEpoch {} Summary:\n"
+                        "Train Loss: {:.6f} | Train Acc: {:.4f} \n".format(
+                            self.epoch,
+                            train_loss,
+                            train_acc))
+
 
                 self.record['epoch'].append(self.epoch)
                 self.record['train_error'].append(train_loss)
@@ -723,8 +737,8 @@ def fit(self, train_loader, val_loader, epochs,
 
         except KeyboardInterrupt:
             logger.warning(
-                "\n\n**********KeyboardInterrupt: "
-                "Training stopped prematurely.**********\n\n")
+                "\n\n**********  KeyboardInterrupt: "
+                "Training stopped prematurely.  **********\n\n")
 
     def _train_epoch(self, train_loader, retain_graph):
         """

vulcanai/models/metrics.py

@@ -1069,8 +1069,9 @@ def cross_validate(network, data_loader, k, epochs,
         # all the results so far
         except KeyboardInterrupt:
             logger.info(
-                "\n\n***KeyboardInterrupt: Cross validate stopped \
-                prematurely.***\n\n")
+                "\n\n**********  KeyboardInterrupt: Cross validate stopped "
+                "prematurely.  **********\n\n")
+            cross_val_network.save_model()
 
         if average_results:
             averaged_all_results = {}