Added example

examples/cross_validation/perovskites/fit.py

@@ -0,0 +1,81 @@
+from sklearn.model_selection import RepeatedKFold
+from sklearn.preprocessing import StandardScaler
+from multilearn import models, utils
+from torch import optim, nn
+
+
+def main():
+
+    save_dir = 'outputs'
+    lr = 1e-4
+    batch_size = 32
+    n_epochs = 200
+    train_size = 0.8  # Traning fraction
+    val_size = 1.0-train_size  # Validation fraction
+    print_n = n_epochs//10
+
+    # Combine data to load
+    tasks = ['asr', 'opband', 'stability']
+    locations = [f'../../sample_data/{i}.csv' for i in tasks]
+
+    # Load data in dictionary (make sure to keep order for loading items)
+    data = utils.load(
+                      locations,
+                      names=tasks,  # User defined name
+                      targets=['y']*len(tasks),  # Target names
+                      )
+
+    # Scalers and loss corresponding to loaded Xs and ys
+    for key, value in data.items():
+        value['scaler'] = StandardScaler()
+        value['loss'] = nn.L1Loss()
+
+    # A single model that combines nodes during training
+    model = models.MultiNet(
+                            tasks=tasks,
+                            input_arch={100: 1},
+                            mid_arch={100: 1, 50: 1},
+                            out_arch={50: 1, 10: 1}
+                            )
+
+    # The optimizer for the NN model
+    optimizer = optim.Adam
+
+    # Do CV to assess
+    utils.cv(
+             data,
+             model,
+             optimizer,
+             RepeatedKFold(n_repeats=1),
+             train_size=train_size,
+             val_size=val_size,
+             save_dir=save_dir,
+             lr=lr,
+             batch_size=batch_size,
+             n_epochs=n_epochs,
+             print_n=print_n,
+             )
+
+    # Save one model to all data
+    model = utils.full_fit(
+                           data,
+                           model,
+                           optimizer,
+                           train_size=train_size,
+                           val_size=val_size,
+                           save_dir=save_dir,
+                           lr=lr,
+                           batch_size=batch_size,
+                           n_epochs=n_epochs,
+                           print_n=print_n,
+                           )
+
+    name = tasks[1]
+    X_inference = data[name]['X']  # Data with dropped columns
+
+    print(f'Model used for predicting {name}')
+    print(model.predict(X_inference, name))
+
+
+if __name__ == '__main__':
+    main()

examples/cross_validation/perovskites/run.sh

@@ -0,0 +1,3 @@
+export PYTHONPATH=$(pwd)/../../../src/:$PYTHONPATH
+rm -rf outputs
+torchrun fit.py

examples/cross_validation/perovskites/submit.sh

@@ -0,0 +1,9 @@
+#PBS -S /bin/bash
+#PBS -q bardeen
+#PBS -l select=1:ncpus=16:mpiprocs=16
+#PBS -l walltime=720:00:00
+#PBS -N job
+
+cd $PBS_O_WORKDIR 
+
+bash run.sh

examples/cross_validation/synthetic/fit.py

@@ -1,6 +1,6 @@
 from sklearn.model_selection import RepeatedKFold
 from sklearn.preprocessing import StandardScaler
-from multilearn import datasets, models, utils
+from multilearn import models, utils
 from torch import optim, nn
 
 import pandas as pd
@@ -13,8 +13,7 @@ def main():
     save_dir = 'outputs'
     lr = 1e-4
     batch_size = 32
-    n_epochs = 100
-    patience = 10  # Learning loop patience
+    n_epochs = 1500
     train_size = 0.8  # Traning fraction
     val_size = 1.0-train_size  # Validation fraction
     print_n = n_epochs//10
@@ -53,12 +52,12 @@ def main():
 
     # Load data in dictionary (make sure to keep order for loading items)
     tasks = ['name1', 'name2']
-    data = datasets.load(
-                         locations,
-                         names=tasks,  # User defined name
-                         targets=['target_1', 'target_2'],  # Target names
-                         drops=[None, ['5', '6']],  # Columns to drop
-                         )
+    data = utils.load(
+                      locations,
+                      names=tasks,  # User defined name
+                      targets=['target_1', 'target_2'],  # Target names
+                      drops=[None, ['5', '6']],  # Columns to drop
+                      )
 
     # Clean generated csv file
     [os.remove(i) for i in locations]
@@ -72,8 +71,8 @@ def main():
     model = models.MultiNet(
                             tasks=tasks,
                             input_arch={100: 1},
-                            mid_arch={100: 1, 50: 1},
-                            out_arch={50: 1, 10: 1}
+                            mid_arch={50: 1},
+                            out_arch={25: 1}
                             )
 
     # The optimizer for the NN model
@@ -91,7 +90,6 @@ def main():
              lr=lr,
              batch_size=batch_size,
              n_epochs=n_epochs,
-             patience=patience,
              print_n=print_n,
              )
 
@@ -106,7 +104,6 @@ def main():
                            lr=lr,
                            batch_size=batch_size,
                            n_epochs=n_epochs,
-                           patience=patience,
                            print_n=print_n,
                            )
 

examples/cross_validation/synthetic/run.sh

@@ -1,2 +1,3 @@
 export PYTHONPATH=$(pwd)/../../../src/:$PYTHONPATH
+rm -rf outputs
 torchrun fit.py

setup.py

@@ -2,7 +2,7 @@
 
 # Package information
 name = 'multilearn'
-version = '0.0.7'  # Need to increment every time to push to PyPI
+version = '0.0.8'  # Need to increment every time to push to PyPI
 description = 'Multi-task learning with Pytorch.'
 url = 'https://github.com/leschultz/multilearn'
 author = 'Lane E. Schultz'

src/multilearn/models.py

@@ -26,11 +26,15 @@ def __init__(self, model, device=None):
 
     def fit(self, data, optimizer, **kwargs):
 
-        print_n = kwargs['print_n']
         n_epochs = kwargs['n_epochs']
         batch_size = kwargs['batch_size']
         lr = kwargs['lr']
-        patience = kwargs['patience']
+
+        if 'print_n' in kwargs.keys():
+            print_n = kwargs['print_n']
+            print_n_cond = True
+        else:
+            print_n_cond = False
 
         data = copy.deepcopy(data)  # Avoids editing original data
         optimizer = optimizer(self.model.parameters(), lr=lr)
@@ -115,28 +119,12 @@ def fit(self, data, optimizer, **kwargs):
                         d = (epoch, loss, indx, split)
                         df_loss.append(d)
 
-                        all_loss += loss
-
-                    else:
-                        all_loss += loss
-
-            # Early stopping
-            if all_loss < best_loss:
-                best_model = copy.deepcopy(self.model)
-                best_loss = all_loss
-                no_improv = 0
-
-            else:
-                no_improv += 1
-
-            if no_improv >= patience:
-                break
-
-            if epoch % print_n == 0:
-                print(f'Epoch {epoch}/{n_epochs}: {split} loss {loss:.2f}')
+                    # Loss from validation set if defined
+                    all_loss += loss
 
-        if patience is not None:
-            self.model = best_model
+            if print_n_cond:
+                if epoch % print_n == 0:
+                    print(f'Epoch {epoch}/{n_epochs}: {split} loss {loss:.2f}')
 
         # Loss curve
         columns = ['epoch', 'loss', 'data', 'split']

src/multilearn/plots.py

@@ -8,7 +8,7 @@
 import os
 
 # Font styles
-font = {'font.size': 16, 'lines.markersize': 10}
+font = {'font.size': 16}
 matplotlib.rcParams.update(font)
 
 

src/multilearn/utils.py

@@ -11,6 +11,56 @@
 import os
 
 
+def load(locations, names=None, targets=None, drops=None):
+    '''
+    Load data included with the package.
+
+    Args:
+        locations (list): The loctions of data to load.
+        names (list): The names for each dataset.
+        targets (list): The name of the target variable for each of names.
+        drops (list): A list of list for columns to drop.
+
+    Returns:
+        Dict: A dictionary with features, targets, and names.
+    '''
+
+    namescond = names is None
+    dropscond = drops is not None
+
+    data = {}
+    for count in range(len(locations)):
+
+        if namescond:
+            name = count
+        else:
+            name = names[count]
+
+        if dropscond:
+            drop = drops[count]
+
+        else:
+            drop = None
+
+        target = targets[count]
+
+        df = locations[count]
+        df = pd.read_csv(df)
+
+        if drop is None:
+            X = df.drop(target, axis=1).values
+        else:
+            X = df.drop([target]+drop, axis=1).values
+
+        y = df[target].values
+
+        data[name] = {}
+        data[name]['X'] = X
+        data[name]['y'] = y
+
+    return data
+
+
 def find(where, match):
     paths = list(map(str, Path(where).rglob(match)))
     return paths

tests/test_fit.py

@@ -1,6 +1,6 @@
 from sklearn.model_selection import RepeatedKFold
 from sklearn.preprocessing import StandardScaler
-from multilearn import datasets, models, utils
+from multilearn import models, utils
 from torch import optim, nn
 
 import pandas as pd
@@ -58,12 +58,12 @@ def test_ml(self):
 
         # Load data in dictionary (make sure to keep order for loading items)
         tasks = ['name1', 'name2']
-        data = datasets.load(
-                             locations,
-                             names=tasks,  # User defined name
-                             targets=['target_1', 'target_2'],  # Target names
-                             drops=[None, ['5', '6']],  # Columns to drop
-                             )
+        data = utils.load(
+                          locations,
+                          names=tasks,  # User defined name
+                          targets=['target_1', 'target_2'],  # Target names
+                          drops=[None, ['5', '6']],  # Columns to drop
+                          )
 
         # Clean generated csv file
         [os.remove(i) for i in locations]