Add EPO codebase

Author: 覃蔚

Alfshop/.gitignore

@@ -0,0 +1,168 @@
+# Byte-compiled / optimized / DLL files
+__pycache__/
+*.py[cod]
+*$py.class
+
+# C extensions
+*.so
+
+# Distribution / packaging
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+share/python-wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+MANIFEST
+
+# PyInstaller
+#  Usually these files are written by a python script from a template
+#  before PyInstaller builds the exe, so as to inject date/other infos into it.
+*.manifest
+*.spec
+
+# Installer logs
+pip-log.txt
+pip-delete-this-directory.txt
+
+# Unit test / coverage reports
+htmlcov/
+.tox/
+.nox/
+.coverage
+.coverage.*
+.cache
+nosetests.xml
+coverage.xml
+*.cover
+*.py,cover
+.hypothesis/
+.pytest_cache/
+cover/
+
+# Translations
+*.mo
+*.pot
+
+# Django stuff:
+*.log
+local_settings.py
+db.sqlite3
+db.sqlite3-journal
+
+# Flask stuff:
+instance/
+.webassets-cache
+
+# Scrapy stuff:
+.scrapy
+
+# Sphinx documentation
+docs/_build/
+
+# PyBuilder
+.pybuilder/
+target/
+
+# Jupyter Notebook
+.ipynb_checkpoints
+
+# IPython
+profile_default/
+ipython_config.py
+
+# pyenv
+#   For a library or package, you might want to ignore these files since the code is
+#   intended to run in multiple environments; otherwise, check them in:
+# .python-version
+
+# pipenv
+#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
+#   However, in case of collaboration, if having platform-specific dependencies or dependencies
+#   having no cross-platform support, pipenv may install dependencies that don't work, or not
+#   install all needed dependencies.
+#Pipfile.lock
+
+# poetry
+#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
+#   This is especially recommended for binary packages to ensure reproducibility, and is more
+#   commonly ignored for libraries.
+#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
+#poetry.lock
+
+# pdm
+#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
+#pdm.lock
+#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
+#   in version control.
+#   https://pdm.fming.dev/#use-with-ide
+.pdm.toml
+
+# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
+__pypackages__/
+
+# Celery stuff
+celerybeat-schedule
+celerybeat.pid
+
+# SageMath parsed files
+*.sage.py
+
+# Environments
+.env
+.venv
+env/
+venv/
+ENV/
+env.bak/
+venv.bak/
+
+# Spyder project settings
+.spyderproject
+.spyproject
+
+# Rope project settings
+.ropeproject
+
+# mkdocs documentation
+/site
+
+# mypy
+.mypy_cache/
+.dmypy.json
+dmypy.json
+
+# Pyre type checker
+.pyre/
+
+# pytype static type analyzer
+.pytype/
+
+# Cython debug symbols
+cython_debug/
+
+# PyCharm
+#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
+#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
+#  and can be added to the global gitignore or merged into this file.  For a more nuclear
+#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
+.idea/
+.vscode/
+
+*.zip
+envs/webshop/data/
+envs/webshop/search_index/
+/data/
+test.sh
+eval_agent/data/alfworld/json_2.1.1/

Alfshop/README.md

@@ -0,0 +1,53 @@
+# Evaluation on WebShop and ALFWorld
+
+We build our WebShop and ALFWorld implementation on top of [EPO] (https://github.com/Yifan-Song793/ETO/tree/main).
+
+
+## Structure
+
+There are three main folders in this directory: `envs`, `eval_agent`, `fastchat`
+
+`envs`: the interaction environment of WebShop. We transform the original [WebShop](https://github.com/princeton-nlp/WebShop) repo into a package.
+
+`eval_agent`: the evaluation framework of agent tasks, which is inspired by [MINT](https://github.com/xingyaoww/mint-bench).
+
+`fastchat`: training scripts for EPO, which is a modified version of [FastChat](https://github.com/lm-sys/FastChat).
+
+
+## Setup
+
+```bash
+bash setup.sh
+```
+
+The setup script performs the following actions:
+
+- Install Python dependencies for agent training, deployment, evaluation, and the environments for WebShop, ALFWorld
+- Download data and search engine indices for WebShop
+- Download game files for ALFWorld
+
+
+## Evaluation with a strategic reasoning model
+
+First, launch the controller of FastChat
+```bash
+python -m fastchat.serve.controller
+```
+
+Then, launch the model worker of FastChat
+```bash
+python -m fastchat.serve.model_worker --model-path <YOUR_MODEL_PATH> --port 21002 --worker-address http://localhost:21002
+```
+
+Finally, evaluate the agent
+```bash
+python -m eval_agent.main --thought_agent_config fastchat --thought_model_name <REASON_MODEL_NAME> --action_agent_config openai --action_model_name <ACTION_MODEL_NAME> --exp_config <TASK_NAME> --split test --verbose
+```
+
+
+## ⚙️ How to Add a New Task
+
+1. Implement your task loader in `eval_agent/tasks`. You should implement the `load_tasks` method which returns a task generator.
+2. Implement the corresponding environment in `eval_agent/envs`. The environment should parse the action generated by the LLM agent, execute the action, and return the observation. The tool/API calling should also be implemented in the environment.
+3. Write the instruction prompt and ICL examples in `eval_agent/prompt`. The default setting is 1-shot evaluation.
+4. Write a new task config in `eval_agent/configs/task`. The config defines which task class and environment class to load, and the settings of the environment (e.g., max action steps).

Alfshop/envs/webshop/setup.py

@@ -0,0 +1,32 @@
+from setuptools import setup, find_packages
+
+setup(
+    name='webshop',
+    version='0.1',
+    packages=find_packages('src'),
+    package_dir={'': 'src'},
+    install_requires=[
+        "beautifulsoup4==4.11.1",
+        "cleantext==1.1.4",
+        "env==0.1.0",
+        "faiss-cpu==1.7.4",
+        "Flask==2.1.2",
+        "gym==0.24.0",
+        "pyserini==0.17.0",
+        "pytest",
+        "rank_bm25==0.2.2",
+        "requests_mock",
+        "scikit_learn==1.1.1",
+        "selenium==4.2.0",
+        "spacy==3.6.1",
+        "thinc==8.1.12",
+        "thefuzz==0.20.0",
+        "werkzeug==2.3.8",
+    ],
+    author='Your Name',
+    author_email='[email protected]',
+    description='webshop pip package version',
+    license='MIT',
+    keywords='sample setuptools development',
+    url='https://github.com/yourusername/mypackage'
+)

Alfshop/envs/webshop/src/webshop/__init__.py

@@ -0,0 +1,41 @@
+"""
+Test the site gym environment.
+
+TODO: move to testing dir for more rigorous tests
+"""
+import gym
+from rich import print
+from rich.markup import escape
+
+from ..web_agent_site.envs import WebAgentSiteEnv
+from ..web_agent_site.models import (
+    HumanPolicy,
+    RandomPolicy,
+)
+from ..web_agent_site.utils import DEBUG_PROD_SIZE
+
+
+if __name__ == '__main__':
+    #env = gym.make('WebAgentSite-v0')
+    #env = WebAgentSiteEnv(render=True, pause=2.0)
+    #env = WebAgentSiteEnv(observation_mode='html', render=False)
+    env = WebAgentSiteEnv(observation_mode='text', render=False, num_products=DEBUG_PROD_SIZE)
+    global_step = 0
+    
+    try:
+        #policy = HumanPolicy()
+        policy = RandomPolicy()
+    
+        observation = env.observation
+        while True:
+            print(observation)
+            available_actions = env.get_available_actions()
+            print('Available actions:', available_actions)
+            action = policy.forward(observation, available_actions)
+            observation, reward, done, info = env.step(action)
+            print(f'Taking action "{escape(action)}" -> Reward = {reward}')
+            if done:
+                break
+            global_step += 1
+    finally:
+        env.close()

Alfshop/envs/webshop/src/webshop/run_envs/run_web_agent_site_env.py

@@ -0,0 +1,32 @@
+"""
+Test the text gym environment.
+
+TODO: move to testing dir for more rigorous tests
+"""
+import gym
+from rich import print
+from rich.markup import escape
+
+from web_agent_site.envs import WebAgentTextEnv
+from web_agent_site.models import RandomPolicy
+from web_agent_site.utils import DEBUG_PROD_SIZE
+
+if __name__ == '__main__':
+    env = gym.make('WebAgentTextEnv-v0', observation_mode='text', num_products=DEBUG_PROD_SIZE)
+    env.reset()
+    
+    try:
+        policy = RandomPolicy()
+    
+        observation = env.observation
+        while True:
+            print(observation)
+            available_actions = env.get_available_actions()
+            print('Available actions:', available_actions)
+            action = policy.forward(observation, available_actions)
+            observation, reward, done, info = env.step(action)
+            print(f'Taking action "{escape(action)}" -> Reward = {reward}')
+            if done:
+                break
+    finally:
+        env.close()

Alfshop/envs/webshop/src/webshop/run_envs/run_web_agent_text_env.py

@@ -0,0 +1,15 @@
+import json
+from pyserini.search.lucene import LuceneSearcher
+from rich import print
+
+
+searcher = LuceneSearcher('indexes')
+hits = searcher.search('rubber sole shoes', k=20)
+
+for hit in hits:
+    doc = searcher.doc(hit.docid)
+    print(doc)
+    obj = json.loads(doc.raw())['product']['Title']
+    print(obj)
+
+print(len(hits))

Alfshop/envs/webshop/src/webshop/search_engine/lucene_searcher.py

@@ -0,0 +1,23 @@
+# Sim-to-real Transfer
+This folder contains code for transferring agents trained on WebShop to perform on third party websites, specifically [Amazon](http://amazon.com) and [eBay](http://ebay.com). The imitation learning and reinforcement learning agents exercised by the transfer code can be found on WebShop's Hugging Face [page](https://huggingface.co/webshop).
+
+Interact with a demo of the transfer code, deployed as a 🤗 Hugging Face space [here](https://huggingface.co/spaces/webshop/amazon_shop)!
+
+## 🛠️ Usage
+The Gradio app deployed as the aforementioned Hugging Face space can be started locally by running `python app.py` in this folder. The initial `setup.sh` script should have installed all the required dependencies.
+
+## ➡️ Transfer Logic
+The Sim-to-real transfer code follows this general logical flow:
+
+<img src="../assets/transfer-logic.png" width="100%">
+
+The contents of this directory each serve the following purposes:
+* `app.py`: Run to launch interactive [Gradio](https://gradio.app/) demo of app
+* `predict_help.py`: Amazon, eBay web scraping code
+* `webshop_lite.py`: A condensed version of WebShop's templating engine
+
+If you are interested in *transferring an agent's functionality to an new website or platform*, you will need to...
+1. implement two new functions:  `parse_results_<platform>.py` and `parse_item_page_<platform>.py`. The corresponding interfaces and working examples for Amazon can be found [here](https://github.com/princeton-nlp/webshop/tree/master/transfer/predict_help.py#L262) and [here](https://github.com/princeton-nlp/webshop/tree/master/transfer/predict_help.py#L296).
+2. Invoke these functions in the [`run_episode`](https://github.com/princeton-nlp/webshop/tree/master/transfer/app.py#L105) function in the `app.py` file. Specifically, you should add a single call to...
+     * `parse_results...` in the [conditional]((https://github.com/princeton-nlp/webshop/tree/master/transfer/predict_help.py#L220)) handling `Page.RESULTS` page types
+     * `parse_item_page...` in the [conditional]((https://github.com/princeton-nlp/webshop/tree/master/transfer/predict_help.py#L240)) handling `Page.ITEMS` page types