Feat/dialogrpt ru

assistant_dists/dream_russian/cpu.yml

@@ -10,3 +10,6 @@ services:
   dialogpt:
     environment:
       CUDA_VISIBLE_DEVICES: ""
+  dialogrpt:
+    environment:
+      CUDA_VISIBLE_DEVICES: ""

assistant_dists/dream_russian/dev.yml

@@ -100,6 +100,11 @@ services:
       - "./common:/src/common"
     ports:
       - 8092:8092
+  dialogrpt:
+    volumes:
+      - "./services/dialogrpt_ru:/src"
+    ports:
+      - 8122:8122
   dff-template-skill:
     volumes:
       - "./skills/dff_template_skill:/src"

assistant_dists/dream_russian/docker-compose.override.yml

@@ -290,6 +290,25 @@ services:
         reservations:
           memory: 128M
 
+  dialogrpt:
+    env_file: [ .env ]
+    build:
+      context: ./services/dialogrpt_ru/
+      args:
+        SERVICE_PORT: 8122
+        PRETRAINED_MODEL_FNAME: dialogrpt_ru_ckpt_v0.pth
+        TOKENIZER_NAME_OR_PATH: "Grossmend/rudialogpt3_medium_based_on_gpt2"
+    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:
@@ -305,4 +324,5 @@ services:
           memory: 128M
         reservations:
           memory: 128M
+
 version: '3.7'

assistant_dists/dream_russian/pipeline_conf.json

@@ -302,6 +302,19 @@
                     "skills"
                 ],
                 "state_manager_method": "add_hypothesis_annotation_batch"
+            },
+            "dialogrpt": {
+                "connector": {
+                    "protocol": "http",
+                    "timeout": 1,
+                    "url": "http://entity-detection:8122/respond"
+                },
+                "dialog_formatter": "state_formatters.dp_formatters:hypotheses_with_context_list",
+                "response_formatter": "state_formatters.dp_formatters:simple_formatter_service",
+                "previous_services": [
+                    "skills"
+                ],
+                "state_manager_method": "add_hypothesis_annotation_batch"
             }
         },
         "response_selectors": {

assistant_dists/dream_russian/test.yml

@@ -41,5 +41,8 @@ services:
   dialogpt:
     environment:
       - CUDA_VISIBLE_DEVICES=7
+  dialogrpt:
+    environment:
+      - CUDA_VISIBLE_DEVICES=7
   dff-template-skill:
 version: '3.7'

dockerfile_agent

@@ -29,7 +29,7 @@ RUN mkdir /pavlov && \
     pip install -e .
 
 RUN pip install sentry-sdk==0.16.0 requests==2.24.0 pandas GitPython==3.1.14 pyaml openpyxl==3.0.0 xlrd==1.2.0
-RUN pip install git+git://github.com/deepmipt/dp-agent.git@35960a8fb0ac8df8ecf75215c895a64c225c1490
+RUN pip install https://codeload.github.com/deepmipt/dp-agent/tar.gz/35960a8fb0ac8df8ecf75215c895a64c225c1490
 
 ENV LANG C.UTF-8
 ENV LC_ALL C.UTF-8

services/dialogrpt_ru/Dockerfile

@@ -0,0 +1,26 @@
+# syntax=docker/dockerfile:experimental
+
+FROM pytorch/pytorch:1.5-cuda10.1-cudnn7-runtime
+
+RUN apt-get update && apt-get install -y --allow-unauthenticated wget && rm -rf /var/lib/apt/lists/*
+
+WORKDIR /src
+
+ARG PRETRAINED_MODEL_FNAME
+ENV PRETRAINED_MODEL_FNAME ${PRETRAINED_MODEL_FNAME}
+ARG SERVICE_PORT
+ENV SERVICE_PORT ${SERVICE_PORT}
+ARG TOKENIZER_NAME_OR_PATH
+ENV TOKENIZER_NAME_OR_PATH ${TOKENIZER_NAME_OR_PATH}
+
+RUN mkdir /data/
+
+RUN wget -c -q http://files.deeppavlov.ai/deeppavlov_data/${PRETRAINED_MODEL_FNAME} -P /data/
+
+COPY ./requirements.txt /src/requirements.txt
+RUN pip install -r /src/requirements.txt
+
+COPY . /src
+
+CMD gunicorn --workers=1 server:app -b 0.0.0.0:${SERVICE_PORT} --timeout=300
+

services/dialogrpt_ru/README.md

@@ -0,0 +1,15 @@
+# Russian DialogRPT model
+
+Code from https://github.com/golsun/DialogRPT
+
+Trained on 827k samples (plus 95k validation samples) from Russian Pikabu web-site. 
+
+Data parsed from Pikabu by `zhirzemli` (OpenDataScience Slack nickname), code is available [on GitHub](https://github.com/alexeykarnachev/dialogs_data_parsers) 
+and the data is available [here](https://drive.google.com/file/d/1XYCprTqn_MlzDD9qgj7ANJkwFigK66mv/view?usp=sharing).
+
+Final acc=0.64 (on valid).
+
+Trained on 8 GPUs.
+```
+python src/main.py train --data=data/out/updown  --min_score_gap=20 --min_rank_gap=0.5 --max_seq_len 256 --batch 16 1>out.txt 2>&1
+```