[BB3] Option to ignore in session memories

parlai/opt_presets/gen/opt_bb3.opt

@@ -192,5 +192,6 @@
   "include_prompt": false,
   "knowledge_chunk_size": 100,
   "max_prompt_len": 1912,
-  "all_vanilla_prompt": false
+  "all_vanilla_prompt": false,
+  "ignore_in_session_memories_mkm": false
 }

parlai/opt_presets/gen/opt_pt.opt

@@ -192,5 +192,6 @@
   "include_prompt": true,
   "knowledge_chunk_size": 100,
   "max_prompt_len": 1912,
-  "all_vanilla_prompt": false
+  "all_vanilla_prompt": false,
+  "ignore_in_session_memories_mkm": false
 }

projects/bb3/agents/opt_bb3_agent.py

@@ -153,6 +153,13 @@ def add_cmdline_args(
             help='Number of times to retry on API request failures (< 0 for unlimited retry).',
         )
         parser.add_argument('--metaseq-server-timeout', default=20.0, type=float)
+        parser.add_argument(
+            '--ignore-in-session-memories-mkm',
+            type='bool',
+            default=False,
+            help='If true, we do not look at the in-session memories when '
+            'generating from the MKM',
+        )
         return parser
 
     def __init__(self, opt, shared=None):
@@ -173,6 +180,9 @@ def __init__(self, opt, shared=None):
         self.dictionary = top_agent.dictionary
         # continue
         self.max_prompt_len = opt.get('max_prompt_len', PROMPT.MAX_PROMPT_LEN)
+        self.ignore_in_session_memories_mkm = opt.get(
+            'ignore_in_session_memories_mkm', False
+        )
         self.search_agent = SearchAgent(
             {
                 'server': self.opt.get('search_server', 'default'),
@@ -415,10 +425,10 @@ def batch_act_knowledge(
             for module in Module:
                 obs = all_obs[module]
                 if module is Module.MEMORY_KNOWLEDGE and i in memory_indices:
-                    memories = MemoryUtils.get_available_memory(
-                        all_obs['raw'], self.dictionary
+                    memories = MemoryUtils.maybe_reduce_memories(
+                        all_obs['raw']['text'], available_memory[i], self.dictionary
                     )
-                    memories = '\n'.join(available_memory[i])
+                    memories = '\n'.join(memories)
                     new_prompt = self._check_and_limit_len(
                         obs['prompt'].replace(module.opt_pre_context_tok(), memories)
                     )
@@ -772,7 +782,15 @@ def batch_act(
             for _ in range(len(observations))
         ]
         # Step 1: determine whether we're searching or accessing memory
-        available_memory = [o['raw']['memories'] for o in observations]
+        all_memory = [o['raw']['memories'] for o in observations]
+        available_memory = [
+            MemoryUtils.get_available_memories(
+                o['raw']['memories'],
+                o['raw']['in_session_memories'],
+                self.ignore_in_session_memories_mkm,
+            )
+            for o in observations
+        ]
 
         batch_reply_sdm, search_indices = self.batch_act_decision(
             observations,
@@ -866,7 +884,7 @@ def batch_act(
             batch_reply_mgm_partner,
             batch_reply_knowledge,
             batch_reply_dialogue,
-            available_memory,
+            all_memory,
         )
         for i, reply in enumerate(batch_reply_final):
             reply.force_set('id', 'BlenderBot3')
@@ -900,8 +918,8 @@ def self_observe(self, self_message: Message):
                 memory_candidate,
                 MemoryUtils.get_memory_prefix(person, self.MODEL_TYPE),
             ):
-                self.memories.append(
-                    MemoryUtils.add_memory_prefix(
-                        memory_candidate, person, self.MODEL_TYPE
-                    )
+                memory_to_add = MemoryUtils.add_memory_prefix(
+                    memory_candidate, person, self.MODEL_TYPE
                 )
+                self.memories.append(memory_to_add)
+                self.in_session_memories.add(memory_to_add)

projects/bb3/agents/r2c2_bb3_agent.py

@@ -421,6 +421,7 @@ def __init__(self, opt, shared=None):
             self.agents[Module.SEARCH_KNOWLEDGE] = agent
 
         self.memories = []
+        self.in_session_memories = set()
         self.search_knowledge_responses = ['__SILENCE__']
         self.memory_knowledge_responses = ['__SILENCE__']
         self.contextual_knowledge_responses = ['__SILENCE__']
@@ -529,6 +530,7 @@ def reset(self, clones_only: bool = False):
             self.contextual_knowledge_responses = ['__SILENCE__']
             self.memory_knowledge_responses = ['__SILENCE__']
             self.memories = []
+            self.in_session_memories = set()
 
     def _construct_subagent_opts(self, opt: Opt):
         """
@@ -657,6 +659,7 @@ def observe(self, observation: Message) -> Dict[Module, Message]:
 
         raw_observation = copy.deepcopy(observation)
         raw_observation['memories'] = self.memories
+        raw_observation['in_session_memories'] = self.in_session_memories
         observations['raw'] = raw_observation
 
         if observation.get('episode_done'):
@@ -1402,15 +1405,14 @@ def self_observe(self, self_message: Message):
                 ),
                 MemoryUtils.get_memory_prefix(person, self.MODEL_TYPE),
             ):
-                self.memories.append(
-                    MemoryUtils.add_memory_prefix(
-                        self_message[
-                            f'{Module.MEMORY_GENERATOR.message_name()}_{person}'
-                        ],
-                        person,
-                        self.MODEL_TYPE,
-                    )
+                memory_to_add = MemoryUtils.add_memory_prefix(
+                    self_message[f'{Module.MEMORY_GENERATOR.message_name()}_{person}'],
+                    person,
+                    self.MODEL_TYPE,
                 )
+
+                self.memories.append(memory_to_add)
+                self.in_session_memories.add(memory_to_add)
         observation = {
             'text': clean_text(
                 self.agents[Module.SEARCH_KNOWLEDGE].history.get_history_str() or ''

projects/bb3/agents/utils.py

@@ -13,7 +13,7 @@
 import os
 import string
 import time
-from typing import List, Tuple, Optional, Dict, Any
+from typing import List, Tuple, Optional, Dict, Any, Set
 
 from parlai.agents.ir_baseline.ir_baseline import score_match, MaxPriorityQueue
 from parlai.core.dict import DictionaryAgent
@@ -332,8 +332,8 @@ def _build_query_representation(
         return rep
 
     @staticmethod
-    def get_available_memory(
-        observation: Message, dictionary: DictionaryAgent
+    def maybe_reduce_memories(
+        text: str, memories: List[str], dictionary: DictionaryAgent
     ) -> List[str]:
         """
         TFIDF-Match memories with the textual input to reduce num memories.
@@ -347,17 +347,40 @@ def get_available_memory(
             return - potentially shortened - list of memories
         """
         new_memories = []
-        mems = observation['memories']
-        if not mems or len(mems) < 32:  # 512 / 16, assuming 16 tokens max per memory
-            return mems
+        if (
+            not memories or len(memories) < 32
+        ):  # 512 / 16, assuming 16 tokens max per memory
+            return memories
         mpq = MaxPriorityQueue(1000)
-        query = MemoryUtils._build_query_representation(observation['text'], dictionary)
-        for m in mems:
+        query = MemoryUtils._build_query_representation(text, dictionary)
+        for m in memories:
             score = score_match(query, m, 0, dictionary)
             mpq.add(m, score)
         new_memories = list(reversed(mpq))[:32]
         return new_memories
 
+    @staticmethod
+    def get_available_memories(
+        memories: List[str],
+        in_session_memories: Set[str],
+        ignore_in_session_memories: bool,
+    ) -> List[str]:
+        """
+        Return available memories.
+
+        :param memories:
+            list of all memories
+        :param in_session_memories:
+            set of memories generated within the current conversation session
+        :param ignore_in_session_memories:
+            whether to ignore memories generated within the session
+        """
+        return [
+            m
+            for m in memories
+            if m not in in_session_memories or not ignore_in_session_memories
+        ]
+
 
 #################
 # OPT API UTILS #

tests/nightly/gpu/test_bb3.py

@@ -411,15 +411,81 @@ def test_memory_tfidf(self):
         agent = create_agent(self.opt)
         dictionary = agent.dictionary
         memories = self.memories * 100
-        new_memories = MemoryUtils.get_available_memory(
-            {'text': 'I wish I could see my cats again!', 'memories': memories},
+        new_memories = MemoryUtils.maybe_reduce_memories(
+            'I wish I could see my cats again!',
+            memories,
             dictionary,
         )
         assert "cats" in new_memories[0]
         assert len(new_memories) <= 32
-        new_memories = MemoryUtils.get_available_memory(
-            {'text': 'I hope the horses are faster today!', 'memories': memories},
+        new_memories = MemoryUtils.maybe_reduce_memories(
+            'I hope the horses are faster today!',
+            memories,
             dictionary,
         )
         assert "horses" in new_memories[0]
         assert len(new_memories) <= 32
+
+
+class TestIgnoreInSessionMemories(TestOptFtBase):
+    def test_in_session_memories(self):
+        opt = copy.deepcopy(self.opt)
+        opt['knowledge_conditioning'] = 'separate'
+        opt['override']['knowledge_conditioning'] = 'separate'
+        agent = create_agent(opt)
+        opt2 = copy.deepcopy(self.opt)
+        opt2['knowledge_conditioning'] = 'separate'
+        opt2['override']['knowledge_conditioning'] = 'separate'
+        opt2['ignore_in_session_memories_mkm'] = True
+        opt2['override']['ignore_in_session_memories_mkm'] = True
+        agent2 = create_agent(opt2)
+
+        # first, check with normal messages
+        agent1_acts = []
+        agent2_acts = []
+        for _ in range(5):
+            agent.observe(self.message)
+            agent2.observe(self.message)
+            agent1_acts.append(agent.act())
+            agent2_acts.append(agent2.act())
+
+        # ignore first message for agent1 since there aren't any memories
+        assert all(a[Module.MEMORY_DIALOGUE.message_name()] for a in agent1_acts[1:])
+        assert all(not a[Module.MEMORY_DIALOGUE.message_name()] for a in agent2_acts)
+
+        # Check that in session memories is strict subset of memories
+        # when using opening message
+        agent.reset()
+        original_memories = copy.deepcopy(self.memories)
+        agent.observe(self.opening_message)
+        agent.act()
+        assert all(m in agent.memories for m in agent.in_session_memories)
+        assert not any(m in agent.in_session_memories for m in agent.memories)
+
+        # set ignore in session memories to True; ensure that final returned memories
+        # still have all the memories, but that we don't use the memory module
+        agent.in_session_memories = set()
+        agent.ignore_in_session_memories_mkm = True
+        message = copy.deepcopy(self.message)
+        agent.observe(message)
+        act = agent.act()
+        assert all(
+            m in act['memories']
+            for m in original_memories + list(agent.in_session_memories)
+        )
+        assert len(agent.in_session_memories) == (
+            len(act['memories']) - len(original_memories)
+        )
+
+    def test_memory_utils(self):
+        new_memories = ['in session memory 1', 'in session memory 2']
+        memories = self.memories + new_memories
+        in_session_memories = set(new_memories)
+        available_memories = MemoryUtils.get_available_memories(
+            memories, in_session_memories, ignore_in_session_memories=False
+        )
+        assert available_memories == memories
+        available_memories = MemoryUtils.get_available_memories(
+            memories, in_session_memories, ignore_in_session_memories=True
+        )
+        assert available_memories == self.memories