[BlenderBot2] A hybrid mode (skip search when needed)

parlai/tasks/wizard_of_internet/agents.py

@@ -350,6 +350,7 @@ class WizardDialogTeacher(WizardOfInternetBaseTeacher):
     def __init__(self, opt, shared=None):
         self.prepend_gold_knowledge = opt.get('prepend_gold_knowledge')
         self.gold_knowledge_delimiter = opt.get('gold_knowledge_delimiter', '\n')
+        self.add_skip_search = opt.get('add_skip_search')
         super().__init__(opt, shared=shared)
         self.id = 'WizInternetWizardTeacher'
 
@@ -363,6 +364,12 @@ def add_cmdline_args(cls, parser: ParlaiParser, partial_opt=None) -> ParlaiParse
             default=False,
             help='If true, prepend text with checked sentences',
         )
+        arg_group.add_argument(
+            '--add-skip-search',
+            type='bool',
+            default=False,
+            help='If true, add skip search field when prepending text with checked sentences',
+        )
         return parser
 
     def custom_evaluation(
@@ -443,6 +450,8 @@ def teacher_setup_data(self, datafile) -> Message:
                         f' {self.gold_knowledge_delimiter} {text}'
                     ),
                 )
+                if self.add_skip_search:
+                    message[CONST.SKIP_SEARCH] = True
             yield message, episode_started
 
 

parlai/tasks/wizard_of_internet/constants.py

@@ -46,6 +46,7 @@ class HISTORY_TYPE:
 IS_SEARCH_QUERY = 'is_search_query'
 IS_LAST_SEARCH_QUERY = 'is_last_search_query'
 LABELS = 'labels'
+SKIP_SEARCH = 'skip_search'
 
 # Message values
 NO_SEARCH_QUERY_USED = '__no_search_used__'

projects/blenderbot2/agents/blenderbot2.py

@@ -13,13 +13,14 @@
 The Memory Decoder examines the context and generates memories to write to
 the long-term memory module.
 """
+import copy
 import torch
 import torch.nn
 import torch.nn.functional as F
 from typing import Union, Dict, List, Tuple, Optional, Any
 
 from parlai.agents.fid.fid import FidAgent, WizIntGoldDocRetrieverFiDAgent
-from parlai.agents.rag.args import DPR_ZOO_MODEL, QUERY_MODEL_TYPES
+from parlai.agents.rag.args import DPR_ZOO_MODEL, QUERY_MODEL_TYPES, RetrieverType
 from parlai.agents.rag.rag import RagAgent
 from parlai.agents.rag.model_types import (
     RagTurn,
@@ -37,8 +38,10 @@
     SELECTED_DOCS_TITLES,
     SELECTED_SENTENCES,
     NO_SELECTED_DOCS_TOKEN,
+    SKIP_SEARCH,
 )
 from parlai.utils.torch import padded_3d
+from parlai.utils.typing import TShared
 
 from .modules import (
     BlenderBot2RagModel,
@@ -99,6 +102,7 @@ def augment_batch_for_generation(
             batch.num_gold_docs,
             batch.memory_decoder_vec,
             batch.num_memory_decoder_vecs,
+            batch.skip_search,
         )
         doc_log_probs = F.log_softmax(doc_scores, dim=1)
         batch.src_text_vec = batch.text_vec
@@ -222,6 +226,12 @@ def add_cmdline_args(
             default=SELECTED_DOCS_TITLES,
             help='Field for selected docs titles.',
         )
+        bb2_group.add_argument(
+            '--skip-search-key',
+            type=str,
+            default=SKIP_SEARCH,
+            help='Field for whether to skip search or not.',
+        )
         bb2_group.add_argument(
             '--insert-gold-docs',
             type='bool',
@@ -700,6 +710,7 @@ def batchify(self, obs_batch: List[Message], sort: bool = False) -> Batch:
         batch.num_gold_docs = None
         batch.memory_decoder_vec = None
         batch.num_memory_decoder_vecs = None
+        batch.skip_search = None
         if any(ex.get('memory_vec') is not None for ex in valid_exs):
             batch = self._set_batch_memory_vec(valid_exs, batch)
         if any(ex.get('query_generator_vec') is not None for ex in valid_exs):
@@ -708,6 +719,8 @@ def batchify(self, obs_batch: List[Message], sort: bool = False) -> Batch:
             batch = self._set_batch_gold_doc_vec(valid_exs, batch)
         if any(ex.get('memory_decoder_vec') is not None for ex in valid_exs):
             batch = self._set_batch_memory_decoder_vec(valid_exs, batch)
+        if any(ex.get(self.opt['skip_search_key']) is not None for ex in valid_exs):
+            batch = self._set_batch_skip_search(valid_exs, batch)
         return batch
 
     def _set_batch_memory_vec(self, valid_exs: List[Message], batch: Batch) -> Batch:
@@ -780,6 +793,11 @@ def _set_batch_memory_decoder_vec(
         batch.num_memory_decoder_vecs = torch.LongTensor(num_memory_dec_toks)
         return batch
 
+    def _set_batch_skip_search(self, valid_exs: List[Message], batch: Batch) -> Batch:
+        skip_search = [ex.get(self.opt['skip_search_key'], False) for ex in valid_exs]
+        batch.skip_search = torch.BoolTensor(skip_search)
+        return batch
+
     def eval_step(self, batch):
         output = super().eval_step(batch)
         if output is None or not hasattr(self.model, 'retriever'):
@@ -807,6 +825,7 @@ def _model_input(
         torch.LongTensor,
         torch.LongTensor,
         torch.LongTensor,
+        torch.BoolTensor,
     ]:
         """
         Override RagAgent._model_input to include several more input vectors.
@@ -826,6 +845,7 @@ def _model_input(
             batch.num_gold_docs,
             batch.memory_decoder_vec,
             batch.num_memory_decoder_vecs,
+            batch.skip_search,
         )
 
     def compute_loss(
@@ -894,6 +914,13 @@ def build_model(self) -> Union[BlenderBot2FidModel, T5BlenderBot2FidModel]:
         return model
 
 
+class BlenderBot2SearchQueryFiDAgent(BlenderBot2FidAgent):
+    def __init__(self, opt: Opt, shared: TShared = None):
+        opt = copy.deepcopy(opt)
+        opt['rag_retriever_type'] = RetrieverType.SEARCH_ENGINE.value
+        super().__init__(opt, shared=shared)
+
+
 class BlenderBot2WizIntGoldDocRetrieverFiDAgent(
     WizIntGoldDocRetrieverFiDAgent, BlenderBot2FidAgent
 ):

projects/blenderbot2/agents/modules.py

@@ -177,6 +177,7 @@ def encoder(
         num_gold_docs: torch.LongTensor,
         memory_decoder_vec: torch.LongTensor,
         num_memory_decoder_vecs: torch.LongTensor,
+        skip_search: torch.BoolTensor,
         positions: Optional[torch.LongTensor] = None,
         segments: Optional[torch.LongTensor] = None,
     ) -> Tuple[
@@ -214,6 +215,8 @@ def encoder(
             3D [bsz, num_lines, seqlen] text to convert to memories with memory decoder
         :param num_memory_decoder_vecs:
             1D [bsz] # of memory decoder vectors for each batch item
+        :param skip_search:
+            1D [bsz] whether to skip search
         """
         # Retrieve, get expanded input
         if all([tensor is not None for tensor in [input_lengths, query_vec]]):
@@ -230,6 +233,7 @@ def encoder(
                 num_gold_docs,
                 memory_decoder_vec,
                 num_memory_decoder_vecs,
+                skip_search,
             )
         else:
             expanded_input = input
@@ -317,6 +321,7 @@ def retrieve_and_concat(
         num_gold_docs: torch.LongTensor,
         memory_decoder_vec: torch.LongTensor,
         num_memory_decoder_vecs: torch.LongTensor,
+        skip_search: torch.BoolTensor,
     ) -> Tuple[torch.LongTensor, List[List[Document]], torch.Tensor]:
         """
         Override RagModel.retrieve_and_concat to perform different retrieval, depending
@@ -360,7 +365,7 @@ def retrieve_and_concat(
         if self.should_generate_query:
             assert self.has_query_generator()
             retrieval_type, search_queries = self.query_generator.classify_retrieval(
-                query_generator_vec, num_memories, generated_memories
+                query_generator_vec, num_memories, generated_memories, skip_search
             )
             logging.debug(f'Classify Retrieval: {time.time() - start:.2f}')
         else:
@@ -844,6 +849,7 @@ def encoder(
         num_gold_docs: torch.LongTensor,
         memory_decoder_vec: torch.LongTensor,
         num_memory_decoder_vecs: torch.LongTensor,
+        skip_search: torch.BoolTensor,
         positions: Optional[torch.LongTensor] = None,
         segments: Optional[torch.LongTensor] = None,
     ) -> Tuple[
@@ -866,6 +872,7 @@ def encoder(
             num_gold_docs,
             memory_decoder_vec,
             num_memory_decoder_vecs,
+            skip_search,
             positions,
             segments,
         )  # type: ignore

projects/blenderbot2/agents/sub_modules.py

@@ -181,7 +181,7 @@ def __init__(self, opt: Opt):
             )
             assert isinstance(base_agent, TorchAgent)
             self.agents = [base_agent]
-            bsz = max(opt.get('batchsize', 1), opt.get('eval_batchsize', 1))
+            bsz = max(opt.get('batchsize') or 1, opt.get('eval_batchsize') or 1)
             rag_turn_n_turns = opt.get('rag_turn_n_turns', 1)
             if bsz > 1 or rag_turn_n_turns > 1:
                 self.agents += [
@@ -196,6 +196,7 @@ def classify_retrieval(
         input: torch.LongTensor,
         num_memories: torch.LongTensor,
         generated_memories: Optional[List[List[str]]],
+        skip_search: Optional[torch.BoolTensor],
     ) -> Tuple[torch.LongTensor, List[str]]:
         """
         Classify input and get retrieval type.
@@ -242,6 +243,8 @@ def classify_retrieval(
                 self.retrieval_type[i] = RetrievalType.MEMORY.value
             elif strip_punc(s) in NONE_STRINGS + MEMORY_STRINGS:
                 self.retrieval_type[i] = RetrievalType.NONE.value
+            elif skip_search is not None and skip_search[i]:
+                self.retrieval_type[i] = RetrievalType.NONE.value
             else:
                 self.retrieval_type[i] = RetrievalType.SEARCH.value
                 searches.append(s)