Feat/infilling

.env

@@ -29,3 +29,5 @@ WIKIDATA_DIALOGUE_SERVICE_URL=http://wikidata-dial-service:8092/model
 NEWS_API_ANNOTATOR_URL=http://news-api-annotator:8112/respond
 WIKI_FACTS_URL=http://wiki-facts:8116/respond
 FACT_RANDOM_SERVICE_URL=http://fact-random:8119/respond
+INFILLING_SERVICE_URL=http://infilling:8122/respond
+

README.md

@@ -187,6 +187,12 @@ Dream Architecture is presented in the following image:
 | User Persona Extractor      | 40 MiB RAM               | determines which age category the user belongs to based on some key words                                                                                                                                                      |
 | Wiki parser                 | 100 MiB RAM              | extracts Wikidata triplets for the entities detected with Entity Linking                                                                                                                                                       |
 
+## Services
+| Name                      | Requirements            | Description                                                                                                                                                                                                        |
+|---------------------------|-------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|
+| DialoGPT                  | 1.3 GiB RAM, 1 GiB GPU  | generative service based on Transformers generative model, the model is set in docker compose argument `PRETRAINED_MODEL_NAME_OR_PATH` (for example, `microsoft/DialoGPT-small` with 0.2-0.5 sec on GPU)   |
+| Infilling                 | 1.7 GiB RAM, 1 GiB GPU  | generative service based on Infilling model, for the given utterance returns utterance where `_` from original text is replaced with generated tokens                          |
+
 ## Skills
 | Name                      | Requirements            | Description                                                                                                                                                                                                        |
 |---------------------------|-------------------------|--------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------|

assistant_dists/dream/cpu.yml

@@ -49,3 +49,6 @@ services:
   dialogpt:
     environment:
       CUDA_VISIBLE_DEVICES: ""
+  infilling:
+    environment:
+      CUDA_VISIBLE_DEVICES: ""

assistant_dists/dream/dev.yml

@@ -399,6 +399,11 @@ services:
       - "./services/dialogpt:/src"
     ports:
       - 8125:8125
+  infilling:
+    volumes:
+      - "./services/infilling:/src"
+    ports:
+      - 8122:8122
   dff-template-skill:
     volumes:
       - "./skills/dff_template_skill:/src"

assistant_dists/dream/docker-compose.override.yml

@@ -19,7 +19,7 @@ services:
           dff-funfact-skill:8104, dff-bot-persona-skill:8105, news-api-annotator:8112,
           dff-gossip-skill:8109, dff-wiki-skill:8111, dff-gaming-skill:8115, topic-recommendation:8113,
           user-persona-extractor:8114, wiki-facts:8116, dff-music-skill:8099, entity-detection:8103, dff-art-skill:8117,
-          midas-predictor:8121, dialogpt:8125, dff-template-skill:8120"
+          midas-predictor:8121, dialogpt:8125, infilling:8122, dff-template-skill:8120"
       WAIT_HOSTS_TIMEOUT: ${WAIT_TIMEOUT:-480}
   convers-evaluator-annotator:
     env_file: [.env]
@@ -1144,7 +1144,7 @@ services:
           memory: 50M
         reservations:
           memory: 50M
-  
+
   dialogpt:
     env_file: [ .env ]
     build:
@@ -1164,6 +1164,23 @@ services:
         reservations:
           memory: 2G
 
+  infilling:
+    env_file: [ .env ]
+    build:
+      context: ./services/infilling/
+      args:
+        SERVICE_PORT: 8122
+    command: flask run -h 0.0.0.0 -p 8122
+    environment:
+      - CUDA_VISIBLE_DEVICES=0
+      - FLASK_APP=server
+    deploy:
+      resources:
+        limits:
+          memory: 2.5G # ?
+        reservations:
+          memory: 2.5G # ?
+
   dff-template-skill:
     env_file: [.env]
     build:

assistant_dists/dream/gpu1.yml

@@ -189,6 +189,10 @@ services:
     restart: unless-stopped
     environment:
       - CUDA_VISIBLE_DEVICES=9
+  infilling:
+    restart: unless-stopped
+    environment:
+      - CUDA_VISIBLE_DEVICES=7
   dff-template-skill:
     restart: unless-stopped
 version: '3.7'

assistant_dists/dream/proxy.yml

@@ -593,7 +593,16 @@ services:
     environment:
       - PROXY_PASS=dream.deeppavlov.ai:8125
       - PORT=8125
-
+
+  infilling:
+    command: [ "nginx", "-g", "daemon off;" ]
+    build:
+      context: dp/proxy/
+      dockerfile: Dockerfile
+    environment:
+      - PROXY_PASS=dream.deeppavlov.ai:8122
+      - PORT=8122
+
   dff-template-skill:
     command: [ "nginx", "-g", "daemon off;" ]
     build:

assistant_dists/dream/test.yml

@@ -130,5 +130,8 @@ services:
   dialogpt:
     environment:
       - CUDA_VISIBLE_DEVICES=6
+  infilling:
+    environment:
+      - CUDA_VISIBLE_DEVICES=8
   dff-template-skill:
 version: '3.7'

common/infilling.py

@@ -0,0 +1,10 @@
+import os
+import requests
+
+
+INFILLING_SERVICE_URL = os.getenv("INFILLING_SERVICE_URL", "http://0.0.0.0:8122/respond")
+
+
+def infill_texts(texts, timeout=1):
+    result = requests.post(INFILLING_SERVICE_URL, json={"texts": texts}, timeout=timeout).json()["infilled_text"]
+    return result

services/infilling/Dockerfile

@@ -0,0 +1,34 @@
+# syntax=docker/dockerfile:experimental
+
+FROM pytorch/pytorch:1.5-cuda10.1-cudnn7-runtime
+
+RUN apt-get update && apt-get install -y --allow-unauthenticated wget
+
+WORKDIR /src
+
+ARG MODEL_DIR=/data/
+ENV MODEL_DIR ${MODEL_DIR}
+ARG SERVICE_PORT
+ENV SERVICE_PORT ${SERVICE_PORT}
+
+COPY ./requirements.txt /src/requirements.txt
+RUN pip install -r /src/requirements.txt
+
+COPY . /src
+
+RUN mkdir /data/
+RUN ls
+RUN python download_model.py model sto ilm | bash
+WORKDIR /data/
+
+RUN wget http://files.deeppavlov.ai/dream/infilling/additional_ids_to_tokens.pkl
+RUN wget http://files.deeppavlov.ai/dream/infilling/vocab.bpe
+RUN wget http://files.deeppavlov.ai/dream/infilling/encoder.json
+RUN wget http://files.deeppavlov.ai/dream/infilling/config.json
+
+WORKDIR /src
+
+HEALTHCHECK --interval=5s --timeout=90s --retries=3 CMD curl --fail 127.0.0.1:${SERVICE_PORT}/healthcheck || exit 1
+
+
+CMD gunicorn --workers=1 server:app -b 0.0.0.0:${SERVICE_PORT} --timeout=300

services/infilling/README.md

@@ -0,0 +1,5 @@
+GPU RAM = 1Gb
+cpu time = 0.5-2 sec 
+gpu time = 0.1-0.5 sec
+
+Very unstable inference time

services/infilling/constants.py

@@ -0,0 +1,3 @@
+GPT2_MODEL_NAMES = ["gpt2", "gpt2-medium", "gpt2-large", "gpt2-xl"]
+GPT2_TOKENIZER_LEN = 50257
+GPT2_EOS_TOKEN_ID = 50256

services/infilling/download_model.py

@@ -0,0 +1,124 @@
+PREMASKED_DATA = {
+    "train": {
+        "sto_mixture": "https://drive.google.com/open?id=1LxlyPqz3OvAZsYRRC8yRdSoaCKGB0Ucg",
+        "abs_mixture": "https://drive.google.com/open?id=1rw45GKP4iRJLzXnRtX-rnk_NeGXOqWkU",
+        "lyr_mixture": "https://drive.google.com/open?id=1jGCjboxlFUF0jqvB0_-L0eeylhKWfZJV",
+    },
+    "valid": {
+        "sto_mixture": "https://drive.google.com/open?id=1Y4HRYrBnqwtdbziF5Q6b5WaIFxJd1v7m",
+        "abs_mixture": "https://drive.google.com/open?id=1hHdXX43qkkm-zpUCJz_iuv1vRRpyfbaP",
+        "lyr_mixture": "https://drive.google.com/open?id=1xR0LC5WHV1UQDPjWTN0HcOQ9C5jsXYef",
+    },
+    "test": {
+        # Table 1/6
+        "sto_sentence": "https://drive.google.com/open?id=1w02hGewoBk_Pq-thrtbOcU1JPGRdGL_U",
+        "abs_sentence": "https://drive.google.com/open?id=18aNMfcqC1wyC8wWJHbCfMxLY49Dbg-du",
+        "lyr_sentence": "https://drive.google.com/open?id=18Szj-HYwh3sjLmmfF8TNwAmB2oEddool",
+        # Table 3
+        "sto_document": "https://drive.google.com/open?id=1ydEjL0SMbX8p-1w6XeLWNrzeVn8TleGT",
+        "abs_document": "https://drive.google.com/open?id=1UjPh51URE8hvK-yTw3xkVwkBwEcCo6Uz",
+        "lyr_document": "https://drive.google.com/open?id=1KNvdzn1xhpw0Xdh0pMWN3CtrDKEK8d4N",
+        # Table 4
+        "sto_mixture": "https://drive.google.com/open?id=1Zsuj8Plrcs49f-5rV6dvJ5W2kIz_C30u",
+        "abs_mixture": "https://drive.google.com/open?id=1TA3ySrvcWxaNtoDPpN8Jk7uqjGKdPqda",
+        "lyr_mixture": "https://drive.google.com/open?id=1FGEL3CGzLvnWpgvUYWsHsOUgW65DVxgw",
+        # Table 5
+        "sto_paragraph": "https://drive.google.com/open?id=1MBM96hfN2cGJidG-mi_4bE0K07xgWAxT",
+        "abs_paragraph": "https://drive.google.com/open?id=1xXJfjCNzRLXYZgHgUrNimP4CtUW0Ziph",
+        "lyr_paragraph": "https://drive.google.com/open?id=10ScpFR8sG3Ur0WpWdkPYxAsT94jNNmZh",
+        # Table 7
+        "sto_ngram": "https://drive.google.com/open?id=1x8RBys_jbreSFO1zMdmwiT2ref2F8q_C",
+        "abs_ngram": "https://drive.google.com/open?id=1JJyh7clJjyPF-rm4rHFLyX7Y-l_doD0K",
+        "lyr_ngram": "https://drive.google.com/open?id=1dbCCc68TvY6segwTrrxYS1ukVbdC7zgJ",
+        # Table 8
+        "sto_word": "https://drive.google.com/open?id=178joxkympgzDwZoExnalWujRq2Jv_37P",
+        "abs_word": "https://drive.google.com/open?id=1PdVg-TnG5VQt8GCQOQA841AGw1GR44yl",
+        "lyr_word": "https://drive.google.com/open?id=1Td-yr6g5cTxW4yoz_Wv4gSi-wbu1376R",
+    },
+}
+
+PRETRAINED_MODELS = {
+    # Trained on stories
+    "sto_lm": "https://drive.google.com/open?id=1-FGKu-bodqOsCGrFCYY6Yyp2rTk2rRpc",
+    "sto_lmrev": "https://drive.google.com/open?id=1_uCgugc57tPGfFofKbU8doJN23cf4lEY",
+    "sto_lmall": "https://drive.google.com/open?id=1dPOLkggPbe-Pzn8VVkcrinuGJv2yRieR",
+    "sto_ilm": "https://drive.google.com/open?id=1oYFLxkX6mWbmpEwQH8BmgE7iKix2W0zU",
+    "sto_lmscratch": "https://drive.google.com/open?id=1vGxdfZUWtOB5ajpDgSGUXuHK5_BGY9GA",
+    "sto_lmrevscratch": "https://drive.google.com/open?id=1xbyQ5bMJpTxlsPtL1YsH2jmUUh_49gOI",
+    "sto_lmallscratch": "https://drive.google.com/open?id=1Qy13Dw60Jd5HqN89q8WvCMtwvTXJw7tj",
+    "sto_ilmscratch": "https://drive.google.com/open?id=14BFLWSaPi2JSsKsa68lcTSnCOnYV9jPm",
+    # Trained on abstracts
+    "abs_lm": "https://drive.google.com/open?id=1BSIFfuSTznmHIKa4R-AnwIxN93b1Ap-b",
+    "abs_lmrev": "https://drive.google.com/open?id=1yl36oZq9R_d3IhlFWLlMGq46n8F9Lq1q",
+    "abs_lmall": "https://drive.google.com/open?id=1qyM0OCL8pI5dL7sfAag-y9X_bnlTS_1Z",
+    "abs_ilm": "https://drive.google.com/open?id=1FBY9DR60WWX05orILaFHuyZYlB4ChTpS",
+    "abs_lmscratch": "https://drive.google.com/open?id=103Cw2ZSb5g5PlTKslmbmhqCaxn3N65OO",
+    "abs_lmrevscratch": "https://drive.google.com/open?id=1HeuxA2A6iEs6SW26jlCom3x_tFQHnIGu",
+    "abs_lmallscratch": "https://drive.google.com/open?id=1XU61GMduqJeCzYqDk8BQ7S4M8tbzqF9g",
+    "abs_ilmscratch": "https://drive.google.com/open?id=1ZTZOO5fVTlnPBw7EC_4OOEzHmcs6tAFO",
+    # Trained on lyrics
+    "lyr_lm": "https://drive.google.com/open?id=1FJBgz26lZPcanZTEf0iWxZCXIEM6esu6",
+    "lyr_lmrev": "https://drive.google.com/open?id=1XAug1jhm7sa5lksDV6GMyF8sFQLwk1Y6",
+    "lyr_lmall": "https://drive.google.com/open?id=1nrNkd4cBsdZS0eajA3wD1i5b6t6R6bow",
+    "lyr_ilm": "https://drive.google.com/open?id=1nYuYCS5fDP2_vB7A92guk0PWh5CC2I5x",
+    "lyr_lmscratch": "https://drive.google.com/open?id=1JzDRUSWVeyGnNaWKVYM8t1BPAs58t6uB",
+    "lyr_lmrevscratch": "https://drive.google.com/open?id=1Kkli5Brmc3D6qE0b5ww5daZdZroaN1YB",
+    "lyr_lmallscratch": "https://drive.google.com/open?id=18JYIBOtDfnksZPl4TW9cOzjOh_qDBCJP",
+    "lyr_ilmscratch": "https://drive.google.com/open?id=1RObPpSttNtMw4UQ1bGiVzEM-94QqkwHT",
+}
+
+PRETRAINED_MODEL_CONFIG_JSON = "https://drive.google.com/open?id=15JnXi7L6LeEB2fq4dFK2WRvDKyX46hVi"
+PRETRAINED_SPECIAL_VOCAB_PKL = "https://drive.google.com/open?id=1nTQVe2tfkWV8dumbrLIHzMgPwpLIbYUd"
+
+PAPER_TASK_TO_INTERNAL = {
+    "lm": "lm",
+    "lmrev": "reverse_lm",
+    "lmall": "naive",
+    "ilm": "ilm",
+}
+
+_DOWNLOAD_TEMPLATE = """
+wget -nc --load-cookies /tmp/cookies.txt "https://docs.google.com/uc?export=download&confirm=$(wget \
+    --quiet --save-cookies /tmp/cookies.txt \
+    --keep-session-cookies \
+    --no-check-certificate \
+    'https://docs.google.com/uc?export=download&id={gdrive_id}' \
+    -O- | sed -rn 's/.*confirm=([0-9A-Za-z_]+).*/\\1\\n/p')&id={gdrive_id}" \
+    -O {local_path} && rm -rf /tmp/cookies.txt
+""".strip()
+
+
+if __name__ == "__main__":
+    import os
+    import sys
+    from pathlib import Path
+
+    MODEL_DIR = Path(os.environ.get("MODEL_DIR", "/data/"))
+
+    if sys.argv[1] == "model":
+        data_tag, model_type = sys.argv[2:]
+        model_tag = "{}_{}".format(data_tag[:3], model_type)
+        gdrive_urls = [PRETRAINED_MODELS[model_tag], PRETRAINED_MODEL_CONFIG_JSON, PRETRAINED_SPECIAL_VOCAB_PKL]
+        local_fns = ["pytorch_model.bin", "config.json", "additional_ids_to_tokens.pkl"]
+    elif sys.argv[1] == "data_train":
+        data_tag = sys.argv[2][:3]
+        out_dir = MODEL_DIR.joinpath("data")
+        gdrive_urls = [PREMASKED_DATA[s]["{}_mixture".format(data_tag)] for s in ["train", "valid"]]
+        local_fns = ["{}_mixture_{}.pkl".format(data_tag, s) for s in ["train", "valid"]]
+    elif sys.argv[1] == "data_eval":
+        data_tag = sys.argv[2][:3]
+        out_dir = MODEL_DIR.joinpath("data")
+        gdrive_urls = [
+            PREMASKED_DATA["test"]["{}_{}".format(data_tag, g)]
+            for g in ["mixture", "document", "paragraph", "sentence", "ngram", "word"]
+        ]
+        local_fns = [
+            "{}_{}_test.pkl".format(data_tag, g)
+            for g in ["mixture", "document", "paragraph", "sentence", "ngram", "word"]
+        ]
+
+    print("mkdir -p {}".format(str(MODEL_DIR)))
+    for gdrive_url, local_fn in zip(gdrive_urls, local_fns):
+        print(
+            _DOWNLOAD_TEMPLATE.format(gdrive_id=gdrive_url.split("=")[1], local_path=str(MODEL_DIR.joinpath(local_fn)))
+        )