Merge pull request #1 from jiaqima/add-api

Add api.py for external API calls to OpenXAI

openxai/api.py

@@ -0,0 +1,129 @@
+"""External APIs."""
+import numpy as np
+import pandas as pd
+import torch
+
+from openxai.LoadModel import LoadModel
+from openxai.Explainer import Explainer
+from openxai.dataloader import return_loaders
+
+
+class OpenXAI(object):
+    """An OpenXAI class to serve external API calls."""
+
+    def __init__(self, data_name: str, model_name: str, explainer_name: str):
+        """Load data, model, and explainer."""
+        self.data_name = data_name
+        self.model_name = model_name
+        self.explainer_name = explainer_name
+
+        self.loader_train, self.loader_test = return_loaders(
+            data_name=data_name, download=True)
+        self.model = LoadModel(data_name=data_name, ml_model=model_name)
+
+        dataset_tensor = torch.FloatTensor(self.loader_train.dataset.data)
+        self.explainer = Explainer(method=explainer_name,
+                                   model=self.model,
+                                   dataset_tensor=dataset_tensor)
+
+        # get feature names and label name
+        self.feature_names = self.loader_train.dataset.feature_names
+        self.label_name = self.loader_train.dataset.target_name
+        self.column_names = self.feature_names
+        self.column_names += [
+            "attribution_{}".format(x) for x in self.feature_names
+        ]
+        self.column_names += ["label", "prediction", "is_test"]
+
+        # will be calculated when the first time querying the full df
+        self.df_full = None
+
+    def _get_df_full(self):
+        """Get the full dataframe."""
+        # iterate through `self.loader_train` and `self.loader_test` to get and
+        # store the data, predictions, and explanations on all samples
+        data = []
+        for X, y in self.loader_train:
+            data.append(self._get_combined_data(X, y))
+        for X, y in self.loader_test:
+            data.append(self._get_combined_data(X, y, is_test=True))
+
+        data = np.concatenate(data, axis=0)
+
+        return pd.DataFrame(data, columns=self.column_names)
+
+    def _get_combined_data(self, X, y, is_test=False):
+        """Get the combined data for a single or a batch of samples.
+
+        Let n be number of samples, d be feature dimension.
+            
+        Arguments:
+          X: feature tensor with size (n, d).
+          y: label tensor with size (n,). 
+          is_test: `True` if this batch is test data. `False` otherwise.
+
+        Returns:
+          A numpy array with n rows and 2d + 3 columns. 
+          The columns from left to right are: features (d),
+          feature attribution scores (d), label (1), predicted label (1),
+          and is_test flag (1).
+        """
+        attribution = self.explainer.get_explanation(X.to(dtype=torch.float32),
+                                                     y)
+        output = self.model(X.to(dtype=torch.float32))
+        prediction = torch.argmax(output, dim=1)
+
+        # flag indicating whether this is test data
+        if is_test:
+            t = torch.ones([X.size(0), 1])
+        else:
+            t = torch.zeros([X.size(0), 1])
+
+        data = [X, attribution, y.unsqueeze(-1), prediction.unsqueeze(-1), t]
+        data = torch.cat(data, dim=1).detach().numpy()
+        return data
+
+    def query(self, X=None, y=None):
+        """Query OpenXAI to get a pandas dataframe."""
+        if X is None:  # query the full data
+            if self.df_full is None:
+                self.df_full = self._get_df_full()
+            return self.df_full
+        else:  # query a batch or a single data point
+            if len(X.size()) == 1:  # single data sample to batch with size 1
+                X = X.unsqueeze(0)
+            if len(y.size()) == 0:  # single data sample to batch with size 1
+                y = y.unsqueeze(0)
+            data = self._get_combined_data(X, y, is_test=True)  # assume test
+            return pd.DataFrame(data, columns=self.column_names)
+
+
+if __name__ == '__main__':
+    # test full query
+    oxai = OpenXAI(data_name="german", model_name="ann", explainer_name="lime")
+    df_full = oxai.query()
+    print(df_full.head())
+
+    # test batch and single query
+    model_names = ["ann", "lr"]
+    data_names = ["compas", "adult", "german"]
+    explainer_names = ["grad", "sg", "itg", "ig", "shap", "lime"]
+    for data_name in data_names:
+        _, loader_test = return_loaders(data_name=data_name, download=True)
+        X, y = iter(loader_test).next()
+        X = X.to(dtype=torch.float32)
+        X = X[:4]  # use smaller batch
+        y = y[:4]
+        X_single = X[0:1]
+        y_single = y[0:1]
+        for model_name in model_names:
+            for explainer_name in explainer_names:
+                oxai = OpenXAI(data_name=data_name,
+                               model_name=model_name,
+                               explainer_name=explainer_name)
+                df_batch = oxai.query(X, y)
+                df_single1 = oxai.query(X_single, y_single)
+                df_single2 = oxai.query(X_single, y_single.squeeze())
+                print(data_name, model_name, explainer_name, "passed!")
+
+    print("\n---------All tests passed!---------")