:q

Author: jannisborn

README.md

@@ -82,7 +82,7 @@ where `-checkpoint` specifies which `.pt` file to pick for the evaluation (based
 
 ## Attention visualization
 The model uses a self-attention mechanism that can highlight chemical motifs used for the predictions.
-In [notebooks/toxicity_attention.ipynb](notebooks/toxicity_attention.ipynb) we share a tutorial on how to create such plots:
+In [notebooks/toxicity_attention_plot.ipynb](notebooks/toxicity_attention_plot.ipynb) we share a tutorial on how to create such plots:
 ![Attention](assets/attention.gif "toxicophore attention")
 
 

toxsmi/models/mca.py

@@ -11,7 +11,6 @@
     smiles_projection,
 )
 from paccmann_predictor.utils.utils import get_device
-
 from toxsmi.utils.hyperparams import ACTIVATION_FN_FACTORY, LOSS_FN_FACTORY
 from toxsmi.utils.layers import EnsembleLayer
 
@@ -231,6 +230,7 @@ def __init__(self, params: dict, *args, **kwargs):
         self.loss_name = params.get(
             "loss_fn", "binary_cross_entropy_ignore_nan_and_sum"
         )
+
         final_activation = (
             ACTIVATION_FN_FACTORY["sigmoid"]
             if "cross" in self.loss_name

toxsmi/utils/performance.py

@@ -106,7 +106,7 @@ def process_data(self, labels: np.array, preds: np.array):
         preds = preds[~np.isnan(labels)]
         labels = labels[~np.isnan(labels)]
 
-        return labels, preds
+        return labels.astype(float), preds.astype(float)
 
     def performance_report_binary_classification(
         self, labels: np.array, preds: np.array, loss: float, model: Callable
@@ -125,10 +125,10 @@ def performance_report_binary_classification(
 
         bin_preds, youden = binarize_predictions(preds, labels, return_youden=True)
         report = classification_report(labels, bin_preds, output_dict=True)
-        negative_precision = report["0.0"]["precision"]
-        negative_recall = report["0.0"]["recall"]
-        positive_precision = report["1.0"]["precision"]
-        positive_recall = report["1.0"]["recall"]
+        negative_precision = report.get("0.0", {}).get("precision", -1)
+        negative_recall = report.get("0.0", {}).get("recall", -1)
+        positive_precision = report.get("1.0", {}).get("precision", -1)
+        positive_recall = report.get("1.0", {}).get("recall", -1)
         f1 = fbeta_score(
             labels, bin_preds, beta=self.beta, pos_label=1, average="binary"
         )
@@ -237,10 +237,10 @@ def inference_report_binary_classification(
         precision, recall, _ = precision_recall_curve(labels, preds)
         precision_recall = average_precision_score(labels, preds)
         report = classification_report(labels, bin_preds, output_dict=True)
-        negative_precision = report["0.0"]["precision"]
-        negative_recall = report["0.0"]["recall"]
-        positive_precision = report["1.0"]["precision"]
-        positive_recall = report["1.0"]["recall"]
+        negative_precision = report.get("0.0", {}).get("precision", -1)
+        negative_recall = report.get("0.0", {}).get("recall", -1)
+        positive_precision = report.get("1.0", {}).get("precision", -1)
+        positive_recall = report.get("1.0", {}).get("recall", -1)
         accuracy = accuracy_score(labels, bin_preds)
         bal_accuracy = balanced_accuracy_score(labels, bin_preds)
         f1 = fbeta_score(labels, bin_preds, beta=0.5, pos_label=1, average="binary")