fix: use np.int64 and np.float64 instead of deprecated np.int and np.float

beir/retrieval/models/sparta.py

@@ -61,9 +61,9 @@ def encode_corpus(self, corpus: List[Dict[str, str]], batch_size: int = 16, **kw
         sentences = [(doc["title"] + self.sep + doc["text"]).strip() for doc in corpus]
         sparse_idx = 0
         num_elements = len(sentences) * self.sparse_vector_dim
-        col = np.zeros(num_elements, dtype=np.int)
-        row = np.zeros(num_elements, dtype=np.int)
-        values = np.zeros(num_elements, dtype=np.float)
+        col = np.zeros(num_elements, dtype=np.int64)
+        row = np.zeros(num_elements, dtype=np.int64)
+        values = np.zeros(num_elements, dtype=np.float64)
 
         for start_idx in trange(0, len(sentences), batch_size, desc="docs"):
             doc_embs = self._compute_sparse_embeddings(sentences[start_idx: start_idx + batch_size])
@@ -74,4 +74,4 @@ def encode_corpus(self, corpus: List[Dict[str, str]], batch_size: int = 16, **kw
                     values[sparse_idx] = score
                     sparse_idx += 1
 
-        return csr_matrix((values, (row, col)), shape=(len(self.bert_input_embeddings), len(sentences)), dtype=np.float)
+        return csr_matrix((values, (row, col)), shape=(len(self.bert_input_embeddings), len(sentences)), dtype=np.float64)

beir/retrieval/models/unicoil.py

@@ -19,7 +19,7 @@ def __init__(self, model_path: Union[str, Tuple] = None, sep: str = " ", query_m
         self.model.eval()
 
     def encode_query(self, query: str, batch_size: int = 16, **kwargs):
-        embedding = np.zeros(self.bert_input_emb, dtype=np.float)
+        embedding = np.zeros(self.bert_input_emb, dtype=np.float64)
         input_ids = self.tokenizer(query, max_length=self.query_max_length, padding='longest',
                                         truncation=True, add_special_tokens=True,
                                         return_tensors='pt').to(self.device)["input_ids"]
@@ -59,9 +59,9 @@ def encode(
                 non_zero_tokens += len(token_ids_and_embs)
                 passage_embs.append(token_ids_and_embs)
 
-        col = np.zeros(non_zero_tokens, dtype=np.int)
-        row = np.zeros(non_zero_tokens, dtype=np.int)
-        values = np.zeros(non_zero_tokens, dtype=np.float)
+        col = np.zeros(non_zero_tokens, dtype=np.int64)
+        row = np.zeros(non_zero_tokens, dtype=np.int64)
+        values = np.zeros(non_zero_tokens, dtype=np.float64)
         sparse_idx = 0    
 
         for pid, emb in enumerate(passage_embs):
@@ -71,7 +71,7 @@ def encode(
                 values[sparse_idx] = score
                 sparse_idx += 1
 
-        return csr_matrix((values, (col, row)), shape=(len(sentences), self.bert_input_emb), dtype=np.float)
+        return csr_matrix((values, (col, row)), shape=(len(sentences), self.bert_input_emb), dtype=np.float64)
 
 # class UniCOIL:
 #     def __init__(self, model_path: Union[str, Tuple] = None, sep: str = " ", **kwargs):
@@ -98,7 +98,7 @@ def encode(
 #         batch_size: int = 32,
 #         max_length: int = 512) -> np.ndarray:
 
-#         embeddings = np.zeros((len(sentences), self.sparse_vector_dim), dtype=np.float)
+#         embeddings = np.zeros((len(sentences), self.sparse_vector_dim), dtype=np.float64)
 
 #         for start_idx in trange(0, len(sentences), batch_size, desc="docs"):
 #             documents = sentences[start_idx: start_idx + batch_size]
@@ -114,7 +114,7 @@ def encode(
 #                 np.put(embeddings[start_idx + idx], batch_token_ids[idx], batch_weights[idx].flatten())
 
 #         return embeddings
-#         # return csr_matrix((values, (row, col)), shape=(len(sentences), self.sparse_vector_dim), dtype=np.float).toarray()
+#         # return csr_matrix((values, (row, col)), shape=(len(sentences), self.sparse_vector_dim), dtype=np.float64).toarray()
 
 
 # Chunks of this code has been taken from: https://github.com/castorini/pyserini/blob/master/pyserini/encode/_unicoil.py
@@ -165,4 +165,4 @@ def forward(
         sequence_output = outputs.last_hidden_state
         tok_weights = self.tok_proj(sequence_output)
         tok_weights = torch.relu(tok_weights)
-        return tok_weights
+        return tok_weights

beir/retrieval/search/dense/exact_search_multi_gpu.py

@@ -162,7 +162,7 @@ def search(self,
             query_id = query_ids[query_itr]
             for i in range(len(cos_scores_top_k_values)):
                 sub_corpus_id = cos_scores_top_k_idx[i][query_itr]
-                score = cos_scores_top_k_values[i][query_itr].item() # convert np.float to float
+                score = cos_scores_top_k_values[i][query_itr].item() # convert np.float64 to float
                 corpus_id = corpus_ids[sub_corpus_id]
                 if corpus_id != query_id:
                     self.results[query_id][corpus_id] = score

beir/retrieval/search/dense/faiss_index.py

@@ -147,7 +147,7 @@ def search(self, query_embeddings: np.ndarray, k: int, binary_k: int = 1000, rer
         if self._passage_ids is not None:
             ids_arr = self._passage_ids[ids_arr.reshape(-1)].reshape(num_queries, -1)
         else:
-            ids_arr = np.array([self.index.id_map.at(int(id_)) for id_ in ids_arr.reshape(-1)], dtype=np.int)
+            ids_arr = np.array([self.index.id_map.at(int(id_)) for id_ in ids_arr.reshape(-1)], dtype=np.int64)
             ids_arr = ids_arr.reshape(num_queries, -1)
 
         scores_arr = scores_arr[np.arange(num_queries)[:, None], sorted_indices]
@@ -171,4 +171,4 @@ def build(
         for start in trange(0, len(passage_ids), buffer_size):
             index.add(passage_embeddings[start : start + buffer_size])
 
-        return cls(index, passage_ids, passage_embeddings)
+        return cls(index, passage_ids, passage_embeddings)