fix (service): fix bug in encoder service

gnes/encoder/video/incep_mixture.py

@@ -116,7 +116,10 @@ def _encode1(self, data):
                                            feed_dict={self.inputs: data})
             return end_points_[self.select_layer]
 
-        v = [_ for vi in _encode1(self, img) for _ in vi]
+        if len(img) <= self.batch_size:
+            v = [_ for _ in _encode1(self, img)]
+        else:
+            v = [_ for vi in _encode1(self, img) for _ in vi]
 
         v_input = [v[s:e] for s, e in zip(pos_start, pos_end)]
         v_input = [(vi + [[0.0] * self.input_size] * (max_len - len(vi)))[:max_len] for vi in v_input]

gnes/service/encoder.py

@@ -35,37 +35,39 @@ def embed_chunks_in_docs(self, docs: Union[List['gnes_pb2.Document'], 'gnes_pb2.
             docs = [docs]
 
         contents = []
-        chunks = []
+        ids = []
         embeds = None
 
         for d in docs:
-            if not d.chunks:
-                self.logger.warning('document (doc_id=%s) contains no chunks!' % d.doc_id)
-                continue
-
+            ids.append(len(d.chunks))
             for c in d.chunks:
                 if d.doc_type == gnes_pb2.Document.TEXT:
                     contents.append(c.text)
-                elif getattr(c, c.WhichOneof('content')) == 'blob':
-                    contents.append(blob2array(c.blob))
                 else:
-                    self.logger.warning(
-                        'chunk content is in type: %s, dont kow how to handle that, ignored' % c.WhichOneof('content'))
-                chunks.append(c)
+                    contents.append(blob2array(c.blob))
 
-        if do_encoding and contents:
+        if do_encoding:
             embeds = self._model.encode(contents)
-            if len(chunks) != embeds.shape[0]:
+            if sum(ids) != embeds.shape[0]:
                 raise ServiceError(
                     'mismatched %d chunks and a %s shape embedding, '
-                    'the first dimension must be the same' % (len(chunks), embeds.shape))
-            for idx, c in enumerate(chunks):
-                c.embedding.CopyFrom(array2blob(embeds[idx]))
+                    'the first dimension must be the same' % (sum(ids), embeds.shape))
+            idx = 0
+            for d in docs:
+                for c in d.chunks:
+                    c.embedding.CopyFrom(array2blob(embeds[idx]))
+                    idx += 1
+
         return contents, embeds
 
     @handler.register(gnes_pb2.Request.IndexRequest)
     def _handler_index(self, msg: 'gnes_pb2.Message'):
-        self.embed_chunks_in_docs(msg.request.index.docs)
+        _, embeds = self.embed_chunks_in_docs(msg.request.index.docs)
+        idx = 0
+        for d in msg.request.index.docs:
+            for c in d.chunks:
+                c.embedding.CopyFrom(array2blob(embeds[idx]))
+                idx += 1
 
     @handler.register(gnes_pb2.Request.TrainRequest)
     def _handler_train(self, msg: 'gnes_pb2.Message'):
@@ -83,4 +85,5 @@ def _handler_train(self, msg: 'gnes_pb2.Message'):
 
     @handler.register(gnes_pb2.Request.QueryRequest)
     def _handler_search(self, msg: 'gnes_pb2.Message'):
-        self.embed_chunks_in_docs(msg.request.search.query, is_input_list=False)
+        _, embeds = self.embed_chunks_in_docs(msg.request.search.query)
+        msg.request.search.query.chunk_embeddings.CopyFrom(array2blob(embeds))