diff --git a/README.md b/README.md
index efa1b15b..6b2254b5 100644
--- a/README.md
+++ b/README.md
@@ -5,7 +5,7 @@
 
 fastNLP is a modular Natural Language Processing system based on PyTorch, for fast development of NLP tools. It divides the NLP model based on deep learning into different modules. These modules fall into 4 categories: encoder, interaction, aggregation and decoder, while each category contains different implemented modules. Encoder modules encode the input into some abstract representation, interaction modules make the information in the representation interact with each other, aggregation modules aggregate and reduce information, and decoder modules decode the representation into the output. Most current NLP models could be built on these modules, which vastly simplifies the process of developing NLP models. The architecture of fastNLP is as the figure below:
 
-![](https://github.com/fastnlp/fastNLP/raw/master/fastnlp-architecture.pdf)
+![](https://github.com/fastnlp/fastNLP/raw/master/fastnlp-architecture.jpg)
 
 
 ## Requirements
diff --git a/fastNLP/core/metrics.py b/fastNLP/core/metrics.py
index 5f1557e1..c8d7fe52 100644
--- a/fastNLP/core/metrics.py
+++ b/fastNLP/core/metrics.py
@@ -173,6 +173,48 @@ def classification_report(y_true, y_pred, labels=None, target_names=None, digits
     raise NotImplementedError
 
 
+def accuracy_topk(y_true, y_prob, k=1):
+    """
+    Compute accuracy of y_true matching top-k probable
+    labels in y_prob.
+
+    Paras:
+        y_ture - ndarray, true label, [n_samples]
+        y_prob - ndarray, label probabilities, [n_samples, n_classes]
+        k - int, k in top-k
+    Returns:
+        accuracy of top-k
+    """
+
+    y_pred_topk = np.argsort(y_prob, axis=-1)[:, -1:-k - 1:-1]
+    y_true_tile = np.tile(np.expand_dims(y_true, axis=1), (1, k))
+    y_match = np.any(y_pred_topk == y_true_tile, axis=-1)
+    acc = np.sum(y_match) / y_match.shape[0]
+
+    return acc
+
+
+def pred_topk(y_prob, k=1):
+    """
+    Return top-k predicted labels and corresponding probabilities.
+
+    Args:
+        y_prob - ndarray, size [n_samples, n_classes], probabilities on labels
+        k - int, k of top-k
+    Returns:
+        y_pred_topk - ndarray, size [n_samples, k], predicted top-k labels
+        y_prob_topk - ndarray, size [n_samples, k], probabilities for
+            top-k labels
+    """
+
+    y_pred_topk = np.argsort(y_prob, axis=-1)[:, -1:-k - 1:-1]
+    x_axis_index = np.tile(
+        np.arange(len(y_prob))[:, np.newaxis],
+        (1, k))
+    y_prob_topk = y_prob[x_axis_index, y_pred_topk]
+    return y_pred_topk, y_prob_topk
+
+
 if __name__ == '__main__':
     y = np.array([1, 0, 1, 0, 1, 1])
     print(_label_types(y))
diff --git a/fastnlp-architecture.jpg b/fastnlp-architecture.jpg
new file mode 100644
index 00000000..5e5c504e
Binary files /dev/null and b/fastnlp-architecture.jpg differ
diff --git a/fastnlp-architecture.pdf b/fastnlp-architecture.pdf
deleted file mode 100644
index 3480f025..00000000
Binary files a/fastnlp-architecture.pdf and /dev/null differ