[Empathetic Dialogues] Switch to DialogTeacher (#4405)

* Start revising ED teacher * Convert over ED teacher * Silly * Minor * Remove episode_done * More cleanup * Fix * Test fix * Force new CI check * Note * Cleanup * Update parlai/tasks/empathetic_dialogues/agents.py Co-authored-by: Stephen Roller <roller@fb.com> * Update parlai/tasks/empathetic_dialogues/agents.py Co-authored-by: Stephen Roller <roller@fb.com> * Minor * Fixes * EDPersonaTopicifierTeacher fix * Fix ID * Hopefully fix style gen teacher PR * Add back fields * Update test_blended_skill_talk.py * Update test_blended_skill_talk.py * Convert over EDPersonaTopicifierTeacher * EDPersonaTopicifierTeacher overhaul * Minor * Minor * Remove imports * Black Co-authored-by: Stephen Roller <roller@fb.com>
facebookresearch · Mar 15, 2022 · 4f7b38e · 4f7b38e
1 parent 17afcb5
commit 4f7b38e
Show file tree

Hide file tree

Showing 11 changed files with 130 additions and 458 deletions.
diff --git a/parlai/tasks/blended_skill_talk/agents.py b/parlai/tasks/blended_skill_talk/agents.py
@@ -4,15 +4,13 @@
 # This source code is licensed under the MIT license found in the
 # LICENSE file in the root directory of this source tree.
 
-from parlai.core.params import ParlaiParser
 import copy
 import json
 import os
 import random
 import re
 from collections import defaultdict
 from typing import List, Optional, Dict, Tuple
-from tqdm import tqdm
 
 from parlai.core.opt import Opt
 from parlai.core.teachers import (
@@ -65,7 +63,12 @@ def _topic_to_persona_path(opt: Opt) -> str:
 
 
 def _cached_data_path(opt: Opt, experiencer_side_only: bool) -> str:
-    # Build the data if it doesn't exist.
+    """
+    Build the data if it doesn't exist.
+
+    See EDPersonaTopicifierTeacher in ParlAI v1.5.1 and earlier for the code to add
+    persona strings to the base EmpatheticDialogues dataset.
+    """
     build(opt)
     dt = opt['datatype'].split(':')[0]
     side_string = 'experiencer_only' if experiencer_side_only else 'both_sides'
@@ -168,101 +171,39 @@ class EDPersonaTopicifierTeacher(EmpatheticDialoguesTeacher):
     Adds persona and WoW topic to ED context strings.
     """
 
-    RECOMPILE_DEFAULT = False
-
-    @classmethod
-    def add_cmdline_args(
-        cls, parser: ParlaiParser, partial_opt: Optional[Opt] = None
-    ) -> ParlaiParser:
-        super().add_cmdline_args(parser, partial_opt=partial_opt)
-        agent = parser.add_argument_group('EDPersonaTopicifierTeacher arguments')
-        agent.add_argument(
-            '--recompile-persona-topic-data',
-            type='bool',
-            default=cls.RECOMPILE_DEFAULT,
-            help='Re-compile data with ConvAI2 personas and WoW topics added. Only useful for demonstrating how data was produced.',
-        )
-        return parser
-
     def __init__(self, opt, shared=None):
         self.persona_topicifier = PersonaTopicifier(
             opt=opt, should_have_personas=False, should_have_topics=False
         )
         super().__init__(opt, shared=shared)
+        self.id = 'parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher'
 
-        if (
-            self.remove_political_convos is True
-            or self.opt.get('deepmoji') is not None
-            or self.opt.get('fasttextloc') is not None
-            or self.opt.get('prepend', -1) > 0
-        ):
-            raise NotImplementedError(
-                'Removing political conversations or using deepmoji, fasttextloc, or '
-                'prepend not supported with this teacher.'
-            )
-
-        # Running over all examples is really slow because the process of finding a WoW
-        # topic is expensive, so let's load cached data with personas and topics unless
-        # --recompile-persona-topic-data is True
-        if opt.get('recompile_persona_topic_data', self.RECOMPILE_DEFAULT):
-            self.data_path = (
-                _cached_data_path(
-                    opt=self.opt, experiencer_side_only=self.experiencer_side_only
-                )
-                + '.recompiled'
-            )
-            warn_once(f'Compiling data file for {self.data_path}.')
-            self.persona_topic_data = self._compile_data()
-            warn_once(f'Saving data to {self.data_path}.')
-            with PathManager.open(self.data_path, 'w') as f_write:
-                json.dump(self.persona_topic_data, f_write)
-        else:
-            self.data_path = _cached_data_path(
-                opt=self.opt, experiencer_side_only=self.experiencer_side_only
-            )
-            warn_once(f'Loading cached data from {self.data_path}.')
-            with PathManager.open(self.data_path, 'r') as f_read:
-                self.persona_topic_data = json.load(f_read)
-
-    def _compile_data(self) -> List[List[dict]]:
+    def _get_datafile(self, opt) -> str:
         """
-        Compile data to be saved for faster future use.
+        Specify a custom datafile path for examples with personas.
         """
-        warn_once(f'Starting to compile {self.num_episodes():d} episodes.')
-        all_data = []
-        for episode_idx in tqdm(range(self.num_episodes())):
-            episode_data = []
-            entry_idx = 0
-            while True:
-                example_data = self._get_example(
-                    episode_idx=episode_idx, entry_idx=entry_idx
-                )
-                episode_data.append(example_data)
-                if example_data['episode_done']:
-                    all_data.append(episode_data)
-                    break
-                else:
-                    entry_idx += 1
-
-        return all_data
-
-    def _get_example(self, episode_idx: int, entry_idx: Optional[int] = None):
-        """
-        Get example from the base ED teacher and add persona and WoW topic strings.
-        """
-        gotten = super().get(episode_idx, entry_idx=entry_idx)
-        if entry_idx == 0:
-            modified_text = self.persona_topicifier.get_modified_text(gotten['text'])
-            gotten['text'] = modified_text
-        return gotten
+        experiencer_side_only = self._get_experiencer_side_only(opt)
+        return _cached_data_path(opt=opt, experiencer_side_only=experiencer_side_only)
 
-    def get(self, episode_idx: int, entry_idx: Optional[int] = None) -> dict:
+    def setup_data(self, path):
         """
         Get example from the final data with personas and WoW topic strings.
         """
-        if entry_idx is None:
-            entry_idx = 0
-        return self.persona_topic_data[episode_idx][entry_idx]
+
+        warn_once(f'Loading cached data from {path}.')
+        with PathManager.open(path, 'r') as f_read:
+            persona_topic_data = json.load(f_read)
+
+        for episode in persona_topic_data:
+            for entry_idx, entry in enumerate(episode):
+
+                # For compatibility with DialogTeacher
+                del entry['episode_done']
+                if self._get_base_datatype(self.opt) == 'train':
+                    del entry['label_candidates']
+
+                new_episode = entry_idx == 0
+                yield entry, new_episode
 
 
 class PersonaTopicifier:

diff --git a/parlai/tasks/blended_skill_talk/test/blended_skill_talk_e_d_persona_topicifier_test.yml b/parlai/tasks/blended_skill_talk/test/blended_skill_talk_e_d_persona_topicifier_test.yml
@@ -5,7 +5,7 @@ acts:
     episode_done: false
     eval_labels:
     - Did you suffer any injuries?
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - I hope it goes well! If it makes you feel any better, most of them are probably
       just as nervous and are looking for any excuse to relax and let their guard
@@ -62,7 +62,7 @@ acts:
     episode_done: true
     eval_labels:
     - Why did you feel guilty? People really shouldn't drive drunk.
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - Yep! or if you're house will be hit.  Many possibilities with tornadoes!
     - 'Wow! I better! I Stopped at Associates of General Studies, because I wanted
@@ -101,7 +101,7 @@ acts:
     eval_labels:
     - Yeah i wanted to tell you about the time i was hit by a drunk driver im so happy
       to still be alive after that experience
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - Sugar will do that to kids every time.  Just don't let them on a  roller coaster
       to make it even worse.
@@ -150,7 +150,7 @@ acts:
     episode_done: true
     eval_labels:
     - Yeah he was punished hes in jail still
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - Are you really able to work from home? Finding a gig is so difficult, I'm glad
       that it is working for you.
@@ -194,7 +194,7 @@ acts:
     eval_labels:
     - sorry to hear! do you have any idea about the break up? did you think about
       it ?
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - That is insane! I've never heard of anything like that before.  You all should
       start protesting!

diff --git a/parlai/tasks/blended_skill_talk/test/blended_skill_talk_e_d_persona_topicifier_train.yml b/parlai/tasks/blended_skill_talk/test/blended_skill_talk_e_d_persona_topicifier_train.yml
@@ -3,8 +3,7 @@ acts:
     deepmoji_ctx: null
     emotion: sentimental
     episode_done: false
-    id: blended_skill_talk:e_d_persona_topicifier
-    label_candidates: []
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     labels:
     - Was this a friend you were in love with, or just a best friend?
     prepend_cand: null
@@ -28,8 +27,7 @@ acts:
     deepmoji_ctx: null
     emotion: sentimental
     episode_done: false
-    id: blended_skill_talk:e_d_persona_topicifier
-    label_candidates: []
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     labels:
     - Where has she gone?
     prepend_cand: null
@@ -41,8 +39,7 @@ acts:
     deepmoji_ctx: null
     emotion: sentimental
     episode_done: true
-    id: blended_skill_talk:e_d_persona_topicifier
-    label_candidates: []
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     labels:
     - Oh was this something that happened because of an argument?
     prepend_cand: null
@@ -54,8 +51,7 @@ acts:
     deepmoji_ctx: null
     emotion: sentimental
     episode_done: false
-    id: blended_skill_talk:e_d_persona_topicifier
-    label_candidates: []
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     labels:
     - This was a best friend. I miss her.
     prepend_cand: null
@@ -77,8 +73,7 @@ acts:
     deepmoji_ctx: null
     emotion: sentimental
     episode_done: true
-    id: blended_skill_talk:e_d_persona_topicifier
-    label_candidates: []
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     labels:
     - We no longer talk.
     prepend_cand: null

diff --git a/parlai/tasks/blended_skill_talk/test/blended_skill_talk_e_d_persona_topicifier_valid.yml b/parlai/tasks/blended_skill_talk/test/blended_skill_talk_e_d_persona_topicifier_valid.yml
@@ -5,7 +5,7 @@ acts:
     episode_done: false
     eval_labels:
     - Are you fine now?
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - It's amazing how many feelings and memories mementos like that can bring back.
     - Good for you! There is nothing wrong or uncool about short hair.
@@ -54,7 +54,7 @@ acts:
     episode_done: true
     eval_labels:
     - Cool :) Is your car damaged a lot?
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - I've actually never gone there. What's it like?
     - College is when the tear will start to flow. I love children.
@@ -92,7 +92,7 @@ acts:
     episode_done: false
     eval_labels:
     - That's funny, hope he didn't give you a heart attack.
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - o no what did you do?
     - I definitely recommend getting some mace and a pocket knife. Stay safe out there!
@@ -146,7 +146,7 @@ acts:
     episode_done: true
     eval_labels:
     - I would probably scream also.
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - 'That really does make it special. I''m glad you have that. '
     - It must've have been. Glad they are okay now.
@@ -184,7 +184,7 @@ acts:
     episode_done: false
     eval_labels:
     - 'Wow! That sounds amazing. Where are you going? '
-    id: blended_skill_talk:e_d_persona_topicifier
+    id: parlai.tasks.blended_skill_talk.agents:EDPersonaTopicifierTeacher
     label_candidates:
     - What kind of school was she at before?
     - Yes, I hate those kind of people. They have nothing to do expect lying about

diff --git a/parlai/tasks/empathetic_dialogues/README.md b/parlai/tasks/empathetic_dialogues/README.md
@@ -14,15 +14,5 @@ Returns examples like so:
 with additional task specific fields: 
 - [situation]: a 1-3 sentence description of the situation that the conversation is 
 - [emotion]: one of 32 emotion words 
-Other optional fields: 
-- [prepend_ctx]: fasttext prediction on context line - or None 
-- [prepend_cand]: fasttext prediction on label line (candidate) - or None 
-- [deepmoji_ctx]: vector encoding from deepmoji penultimate layer - or None 
-- [deepmoji_cand]: vector encoding from deepmoji penultimate layer for label line (candidate) - or None
-
-## EmotionClassificationSituationTeacher
-Classifier that returns the situation and emotion for each episode given by `EmpatheticDialoguesTeacher`. Examples:
-- [text]: A 1-3 sentence description of the situation that the conversation is (equivalent to [situation] for `EmpatheticDialoguesTeacher`)
-- [labels]: one of 32 emotion words (equivalent to [emotion] for `EmpatheticDialoguesTeacher`)
 
 Tags: #EmpatheticDialogues, #All, #ChitChat