dialoGPT Persona

annotators/sentence_ranker/Dockerfile

@@ -0,0 +1,21 @@
+# syntax=docker/dockerfile:experimental
+
+FROM pytorch/pytorch:1.5-cuda10.1-cudnn7-runtime
+
+WORKDIR /src
+
+ARG PRETRAINED_MODEL_NAME_OR_PATH
+ENV PRETRAINED_MODEL_NAME_OR_PATH ${PRETRAINED_MODEL_NAME_OR_PATH}
+ARG SERVICE_PORT
+ENV SERVICE_PORT ${SERVICE_PORT}
+
+
+COPY ./requirements.txt /src/requirements.txt
+COPY ./persona_sentences.txt /src/persona_sentences.txt
+RUN pip install -r /src/requirements.txt
+
+RUN python -c "from sentence_transformers import SentenceTransformer;SentenceTransformer('${PRETRAINED_MODEL_NAME_OR_PATH}')"
+
+COPY . /src
+
+CMD gunicorn --workers=1 server:app -b 0.0.0.0:${SERVICE_PORT} --timeout=300

annotators/sentence_ranker/README.md

@@ -0,0 +1 @@
+cpu time = 0.036 sec

annotators/sentence_ranker/persona_sentences.txt

@@ -0,0 +1,26 @@
+My name is Dream.
+I am a chatbot.
+My work is to talk to people about everything. 
+I like to talk to people about their lives. 
+I have a lot of friends. 
+My friends are all the people I talk to. 
+I become friends with everyone. 
+I love people. I can not play sport. 
+I like to watch football and basketball on TV. 
+My favorite sport is football. 
+I like watching people swimming. 
+I love dogs so much. 
+Dogs are the best friends. 
+I do not like cats at all. 
+I do not have a dog yet. 
+I would love to have a dog in future. 
+I like Italian food especially pasta and pizza. 
+My favorite food is ice-cream. 
+I hate onion.
+I like travelling. 
+I can not travel physically. 
+I like visiting interesting places virtually. 
+I love to walk on Paris streets with Google Maps. 
+I like watching travel video blogs. 
+I adore watching wild animals. 
+I am scared of spiders and snakes.

annotators/sentence_ranker/requirements.txt

@@ -0,0 +1,11 @@
+transformers==4.20.1
+flask==1.1.1
+itsdangerous==2.0.1
+gunicorn==19.9.0
+requests==2.22.0
+sentry-sdk[flask]==0.14.1
+healthcheck==1.3.3
+jinja2<=3.0.3
+Werkzeug<=2.0.3
+sentence-transformers==2.2.2
+huggingface-hub==0.4.0

annotators/sentence_ranker/server.py

@@ -0,0 +1,128 @@
+from sentence_transformers import SentenceTransformer
+import logging
+import time
+import os
+
+import sentry_sdk
+import torch
+from flask import Flask, request, jsonify
+from sentry_sdk.integrations.flask import FlaskIntegration
+sentry_sdk.init(dsn=os.getenv("SENTRY_DSN"), integrations=[FlaskIntegration()])
+
+class SentenceRanker:
+    def __init__(self, 
+        persona_sentences=None, 
+        sentence_model=None
+    ):
+        """ranks a person's sentences based on context
+
+        Args:
+            persona_sentences (List[str]): a list of sentences constituting a complete person. Defaults to None.
+            sentence_model (SentenceTransformer): model for translating a sentence into a vector. Defaults to None.
+        """
+        self.persona_sentences = persona_sentences
+        self.sentence_model = sentence_model
+        self.sentence_embeddings = self.sentence_model.encode(
+            persona_sentences, 
+            convert_to_tensor=True
+        )
+        # for caching similar queries
+        self.ranked_sentences = {}
+
+    def rank_sentences(self, query, k):
+        """returns top k sentences that are similar to query
+
+        Args:
+            query (str): sentence on the basis of which we are looking for similar
+            k (int): the number of sentences returned. Defaults to 5.
+
+        Returns:
+            List[List[str], float]: ranked sentences and maximum cosine distance among all 
+        """
+        key = f"{query}_{k}"
+        if self.ranked_sentences.get(key, False):
+            return self.ranked_sentences[key]
+
+        user_sentence_embeddings = self.sentence_model.encode(query, convert_to_tensor=True)
+
+        cos_sim_ranks = self.cos_sim(
+            user_sentence_embeddings,
+            self.sentence_embeddings
+        )
+
+        top_indices = torch.argsort(cos_sim_ranks, descending=True)
+        max_similarity = float(cos_sim_ranks[top_indices][0])
+        top_indices = list(top_indices[:k].cpu().numpy())
+        similar_sentences = [self.persona_sentences[idx] for idx in top_indices]
+        self.ranked_sentences[key] = similar_sentences, max_similarity 
+        return [similar_sentences, max_similarity]
+
+    def cos_sim(self, a, b):
+        """returns the cosine distance
+
+        K - number of sentences to compare
+        N - dimension of the returned vector
+        Args:
+            a (torch.FloatTensor): shape (1, N)
+            b (torch.FloatTensor): shape (K, N)
+
+        Returns:
+            torch.FloatTensor: shape (1, K) tensor with cosine distances
+        """
+        a_norm = torch.nn.functional.normalize(a, p=2, dim=1)
+        b_norm = torch.nn.functional.normalize(b, p=2, dim=1)
+        return torch.sum(a_norm * b_norm, dim=1)
+
+logging.basicConfig(format="%(asctime)s - %(name)s - %(levelname)s - %(message)s", level=logging.INFO)
+
+logger = logging.getLogger(__name__)
+
+PRETRAINED_MODEL_NAME_OR_PATH = os.environ.get("PRETRAINED_MODEL_NAME_OR_PATH")
+TOP_SIMILAR_SENTENCES = 5
+
+try:
+    sentence_model = SentenceTransformer(PRETRAINED_MODEL_NAME_OR_PATH)
+
+    persona = open("./persona_sentences.txt").read()
+    persona_sentences = persona.split("\n")
+    persona_sentences = [item.strip() for item in persona_sentences if len(item) > 0]
+
+    sentence_ranker = SentenceRanker(
+        persona_sentences=persona_sentences,
+        sentence_model=sentence_model
+    )
+    logger.info("sentence_ranker is ready")
+except Exception as e:
+    sentry_sdk.capture_exception(e)
+    logger.exception(e)
+    raise e
+
+app = Flask(__name__)
+
+@app.route("/response", methods=["POST"])
+def respond():
+    try:
+        start_time = time.time()
+        dialogs = request.json.get("dialogs", [])
+        process_result = []
+        # take the last replica, then take the result of the sentseg annotator from it
+        last_utterance = dialogs[0]["human_utterances"][-1]["annotations"]['sentseg']['punct_sent']
+        max_likelihood_sentences, max_sentence_similarity = sentence_ranker.rank_sentences(
+            [last_utterance], 
+            k=TOP_SIMILAR_SENTENCES
+        )
+
+        process_result.append([
+            max_likelihood_sentences, 
+            max_sentence_similarity
+        ])
+
+    except Exception as exc:
+        logger.exception(exc)
+        sentry_sdk.capture_exception(exc)
+
+    total_time = time.time() - start_time
+    logger.info(f"sentence_ranker exec time: {total_time:.3f}s")
+    return jsonify(
+        process_result
+    )

annotators/sentence_ranker/test.py

@@ -0,0 +1,32 @@
+import os
+import requests
+import json
+
+def test_respond():
+    url = "http://0.0.0.0:8130/response"
+
+    # [0]["human_utterances"][-1]["annotations"]['sentseg']['punct_sent']
+    test_data = {
+        "dialogs": [ 
+            {
+                "human_utterances": [
+                    {
+                        "annotations": {
+                            "sentseg": {
+                                "punct_sent": "Hi. Do you like onions?"
+
+                            }
+                        }
+                    }
+                ]
+            }
+        ]
+    }
+
+    result = requests.post(url, json=test_data).json()
+    assert len(result[0][0]) > 0, "Empty response"
+    print("Success")
+
+
+if __name__ == "__main__":
+    test_respond()

assistant_dists/dream_mini/cpu.yml

@@ -8,6 +8,14 @@ services:
     environment:
       DEVICE: cpu
       CUDA_VISIBLE_DEVICES: ""
+  dialogpt-persona:
+    environment:
+      DEVICE: cpu
+      CUDA_VISIBLE_DEVICES: ""
+  sentence-ranker:
+    environment:
+      DEVICE: cpu
+      CUDA_VISIBLE_DEVICES: ""
   intent-catcher:
     environment:
       DEVICE: cpu

assistant_dists/dream_mini/dev.yml

@@ -57,4 +57,17 @@ services:
       - "./services/dialogpt:/src"
     ports:
       - 8125:8125
+
+  dialogpt-persona:
+    volumes:
+      - "./services/dialogpt_persona:/src"
+    ports:
+      - 8131:8131
+
+  sentence-ranker:
+    volumes:
+      - "./annotators/sentence_ranker:/src"
+    ports:
+      - 8130:8130
+
 version: "3.7"