mirror of
https://github.com/ltcptgeneral/cse151b-final-project.git
synced 2025-10-22 18:49:21 +00:00
Compare commits
9 Commits
ethan-test
...
e799c14ece
| Author | SHA1 | Date | |
|---|---|---|---|
|
e799c14ece | ||
|
|
bbe9a1891c | ||
|
9172326013 | ||
|
|
4836be8121 | ||
|
|
5672169073 | ||
|
|
5ec123e0f1 | ||
|
|
e9622b6f68 | ||
|
83e81722d2 | ||
|
|
320f2f81b7 |
3
.gitignore
vendored
3
.gitignore
vendored
@@ -1 +1,4 @@
|
||||
**/data/*
|
||||
**/*.zip
|
||||
**/__pycache__
|
||||
/env
|
||||
671
dqn_wordle.ipynb
Normal file
671
dqn_wordle.ipynb
Normal file
@@ -0,0 +1,671 @@
|
||||
{
|
||||
"cells": [
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 1,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"import gym\n",
|
||||
"import gym_wordle\n",
|
||||
"from stable_baselines3 import DQN, PPO, common\n",
|
||||
"import numpy as np\n",
|
||||
"import tqdm"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 2,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"<Monitor<WordleEnv instance>>\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"env = gym_wordle.wordle.WordleEnv()\n",
|
||||
"env = common.monitor.Monitor(env)\n",
|
||||
"\n",
|
||||
"print(env)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"application/vnd.jupyter.widget-view+json": {
|
||||
"model_id": "7c52630b65904d5e8e200be505d2121a",
|
||||
"version_major": 2,
|
||||
"version_minor": 0
|
||||
},
|
||||
"text/plain": [
|
||||
"Output()"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"output_type": "display_data"
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"Using cuda device\n",
|
||||
"Wrapping the env with a `Monitor` wrapper\n",
|
||||
"Wrapping the env in a DummyVecEnv.\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -175 |\n",
|
||||
"| exploration_rate | 0.525 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 10000 |\n",
|
||||
"| fps | 4606 |\n",
|
||||
"| time_elapsed | 10 |\n",
|
||||
"| total_timesteps | 49989 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -208 |\n",
|
||||
"| exploration_rate | 0.0502 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 20000 |\n",
|
||||
"| fps | 1118 |\n",
|
||||
"| time_elapsed | 89 |\n",
|
||||
"| total_timesteps | 99980 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 24.6 |\n",
|
||||
"| n_updates | 12494 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -230 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 30000 |\n",
|
||||
"| fps | 856 |\n",
|
||||
"| time_elapsed | 175 |\n",
|
||||
"| total_timesteps | 149974 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 18.7 |\n",
|
||||
"| n_updates | 24993 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -242 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 40000 |\n",
|
||||
"| fps | 766 |\n",
|
||||
"| time_elapsed | 260 |\n",
|
||||
"| total_timesteps | 199967 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 24 |\n",
|
||||
"| n_updates | 37491 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -186 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 50000 |\n",
|
||||
"| fps | 722 |\n",
|
||||
"| time_elapsed | 346 |\n",
|
||||
"| total_timesteps | 249962 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 21.5 |\n",
|
||||
"| n_updates | 49990 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -183 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 60000 |\n",
|
||||
"| fps | 694 |\n",
|
||||
"| time_elapsed | 431 |\n",
|
||||
"| total_timesteps | 299957 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 17.6 |\n",
|
||||
"| n_updates | 62489 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -181 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 70000 |\n",
|
||||
"| fps | 675 |\n",
|
||||
"| time_elapsed | 517 |\n",
|
||||
"| total_timesteps | 349953 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 26.8 |\n",
|
||||
"| n_updates | 74988 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -196 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 80000 |\n",
|
||||
"| fps | 663 |\n",
|
||||
"| time_elapsed | 603 |\n",
|
||||
"| total_timesteps | 399936 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 22.5 |\n",
|
||||
"| n_updates | 87483 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -174 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 90000 |\n",
|
||||
"| fps | 653 |\n",
|
||||
"| time_elapsed | 688 |\n",
|
||||
"| total_timesteps | 449928 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 21.1 |\n",
|
||||
"| n_updates | 99981 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -155 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 100000 |\n",
|
||||
"| fps | 645 |\n",
|
||||
"| time_elapsed | 774 |\n",
|
||||
"| total_timesteps | 499920 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 22.8 |\n",
|
||||
"| n_updates | 112479 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -153 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 110000 |\n",
|
||||
"| fps | 638 |\n",
|
||||
"| time_elapsed | 860 |\n",
|
||||
"| total_timesteps | 549916 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 16 |\n",
|
||||
"| n_updates | 124978 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -164 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 120000 |\n",
|
||||
"| fps | 633 |\n",
|
||||
"| time_elapsed | 947 |\n",
|
||||
"| total_timesteps | 599915 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 17.8 |\n",
|
||||
"| n_updates | 137478 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -145 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 130000 |\n",
|
||||
"| fps | 628 |\n",
|
||||
"| time_elapsed | 1033 |\n",
|
||||
"| total_timesteps | 649910 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 17.8 |\n",
|
||||
"| n_updates | 149977 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -154 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 140000 |\n",
|
||||
"| fps | 624 |\n",
|
||||
"| time_elapsed | 1120 |\n",
|
||||
"| total_timesteps | 699902 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 20.9 |\n",
|
||||
"| n_updates | 162475 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -192 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 150000 |\n",
|
||||
"| fps | 621 |\n",
|
||||
"| time_elapsed | 1206 |\n",
|
||||
"| total_timesteps | 749884 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 18.3 |\n",
|
||||
"| n_updates | 174970 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -170 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 160000 |\n",
|
||||
"| fps | 618 |\n",
|
||||
"| time_elapsed | 1293 |\n",
|
||||
"| total_timesteps | 799869 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 17.7 |\n",
|
||||
"| n_updates | 187467 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -233 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 170000 |\n",
|
||||
"| fps | 615 |\n",
|
||||
"| time_elapsed | 1380 |\n",
|
||||
"| total_timesteps | 849855 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 21.6 |\n",
|
||||
"| n_updates | 199963 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -146 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 180000 |\n",
|
||||
"| fps | 613 |\n",
|
||||
"| time_elapsed | 1466 |\n",
|
||||
"| total_timesteps | 899847 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 19.4 |\n",
|
||||
"| n_updates | 212461 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -142 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 190000 |\n",
|
||||
"| fps | 611 |\n",
|
||||
"| time_elapsed | 1553 |\n",
|
||||
"| total_timesteps | 949846 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 22.9 |\n",
|
||||
"| n_updates | 224961 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"----------------------------------\n",
|
||||
"| rollout/ | |\n",
|
||||
"| ep_len_mean | 5 |\n",
|
||||
"| ep_rew_mean | -171 |\n",
|
||||
"| exploration_rate | 0.05 |\n",
|
||||
"| time/ | |\n",
|
||||
"| episodes | 200000 |\n",
|
||||
"| fps | 609 |\n",
|
||||
"| time_elapsed | 1640 |\n",
|
||||
"| total_timesteps | 999839 |\n",
|
||||
"| train/ | |\n",
|
||||
"| learning_rate | 0.0001 |\n",
|
||||
"| loss | 20.3 |\n",
|
||||
"| n_updates | 237459 |\n",
|
||||
"----------------------------------\n"
|
||||
]
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/html": [
|
||||
"<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\"></pre>\n"
|
||||
],
|
||||
"text/plain": []
|
||||
},
|
||||
"metadata": {},
|
||||
"output_type": "display_data"
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/html": [
|
||||
"<pre style=\"white-space:pre;overflow-x:auto;line-height:normal;font-family:Menlo,'DejaVu Sans Mono',consolas,'Courier New',monospace\">\n",
|
||||
"</pre>\n"
|
||||
],
|
||||
"text/plain": [
|
||||
"\n"
|
||||
]
|
||||
},
|
||||
"metadata": {},
|
||||
"output_type": "display_data"
|
||||
},
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"<stable_baselines3.dqn.dqn.DQN at 0x294981ca090>"
|
||||
]
|
||||
},
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"total_timesteps = 1_000_000\n",
|
||||
"model = DQN(\"MlpPolicy\", env, verbose=1, device='cuda')\n",
|
||||
"model.learn(total_timesteps=total_timesteps, log_interval=10_000, progress_bar=True)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 6,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"model.save(\"dqn_new_rewards\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 5,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stderr",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"c:\\Repository\\cse151b-final-project\\env\\Lib\\site-packages\\stable_baselines3\\common\\save_util.py:166: UserWarning: Could not deserialize object lr_schedule. Consider using `custom_objects` argument to replace this object.\n",
|
||||
"Exception: code() argument 13 must be str, not int\n",
|
||||
" warnings.warn(\n",
|
||||
"c:\\Repository\\cse151b-final-project\\env\\Lib\\site-packages\\stable_baselines3\\common\\save_util.py:166: UserWarning: Could not deserialize object exploration_schedule. Consider using `custom_objects` argument to replace this object.\n",
|
||||
"Exception: code() argument 13 must be str, not int\n",
|
||||
" warnings.warn(\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"# model = DQN.load(\"dqn_wordle\")"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 7,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"name": "stdout",
|
||||
"output_type": "stream",
|
||||
"text": [
|
||||
"0\n"
|
||||
]
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"env = gym_wordle.wordle.WordleEnv()\n",
|
||||
"\n",
|
||||
"for i in range(1000):\n",
|
||||
" \n",
|
||||
" state, info = env.reset()\n",
|
||||
"\n",
|
||||
" done = False\n",
|
||||
"\n",
|
||||
" wins = 0\n",
|
||||
"\n",
|
||||
" while not done:\n",
|
||||
"\n",
|
||||
" action, _states = model.predict(state, deterministic=True)\n",
|
||||
"\n",
|
||||
" state, reward, done, truncated, info = env.step(action)\n",
|
||||
"\n",
|
||||
" if info[\"correct\"]:\n",
|
||||
" wins += 1\n",
|
||||
"\n",
|
||||
"print(wins)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"(array([[18, 1, 20, 5, 19, 3, 3, 3, 3, 3],\n",
|
||||
" [14, 15, 9, 12, 25, 2, 3, 2, 2, 2],\n",
|
||||
" [25, 21, 3, 11, 15, 2, 3, 3, 3, 3],\n",
|
||||
" [25, 21, 3, 11, 15, 2, 3, 3, 3, 3],\n",
|
||||
" [ 1, 20, 13, 15, 19, 3, 3, 3, 3, 3],\n",
|
||||
" [25, 21, 3, 11, 15, 2, 3, 3, 3, 3]], dtype=int64),\n",
|
||||
" -130)"
|
||||
]
|
||||
},
|
||||
"execution_count": 8,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"state, reward"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 21,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": [
|
||||
"blah = (14, 1, 9, 22, 5)"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": 23,
|
||||
"metadata": {},
|
||||
"outputs": [
|
||||
{
|
||||
"data": {
|
||||
"text/plain": [
|
||||
"True"
|
||||
]
|
||||
},
|
||||
"execution_count": 23,
|
||||
"metadata": {},
|
||||
"output_type": "execute_result"
|
||||
}
|
||||
],
|
||||
"source": [
|
||||
"blah in info['guesses']"
|
||||
]
|
||||
},
|
||||
{
|
||||
"cell_type": "code",
|
||||
"execution_count": null,
|
||||
"metadata": {},
|
||||
"outputs": [],
|
||||
"source": []
|
||||
}
|
||||
],
|
||||
"metadata": {
|
||||
"kernelspec": {
|
||||
"display_name": "Python 3 (ipykernel)",
|
||||
"language": "python",
|
||||
"name": "python3"
|
||||
},
|
||||
"language_info": {
|
||||
"codemirror_mode": {
|
||||
"name": "ipython",
|
||||
"version": 3
|
||||
},
|
||||
"file_extension": ".py",
|
||||
"mimetype": "text/x-python",
|
||||
"name": "python",
|
||||
"nbconvert_exporter": "python",
|
||||
"pygments_lexer": "ipython3",
|
||||
"version": "3.11.5"
|
||||
}
|
||||
},
|
||||
"nbformat": 4,
|
||||
"nbformat_minor": 2
|
||||
}
|
||||
7
gym_wordle/__init__.py
Normal file
7
gym_wordle/__init__.py
Normal file
@@ -0,0 +1,7 @@
|
||||
from gym.envs.registration import register
|
||||
from .wordle import WordleEnv
|
||||
|
||||
register(
|
||||
id='Wordle-v0',
|
||||
entry_point='gym_wordle.wordle:WordleEnv'
|
||||
)
|
||||
12972
gym_wordle/dictionary/guess_list.csv
Normal file
12972
gym_wordle/dictionary/guess_list.csv
Normal file
File diff suppressed because it is too large
Load Diff
BIN
gym_wordle/dictionary/guess_list.npy
Normal file
BIN
gym_wordle/dictionary/guess_list.npy
Normal file
Binary file not shown.
2315
gym_wordle/dictionary/solution_list.csv
Normal file
2315
gym_wordle/dictionary/solution_list.csv
Normal file
File diff suppressed because it is too large
Load Diff
BIN
gym_wordle/dictionary/solution_list.npy
Normal file
BIN
gym_wordle/dictionary/solution_list.npy
Normal file
Binary file not shown.
93
gym_wordle/utils.py
Normal file
93
gym_wordle/utils.py
Normal file
@@ -0,0 +1,93 @@
|
||||
import numpy as np
|
||||
import numpy.typing as npt
|
||||
|
||||
from pathlib import Path
|
||||
|
||||
|
||||
_chars = ' abcdefghijklmnopqrstuvwxyz'
|
||||
_char_d = {c: i for i, c in enumerate(_chars)}
|
||||
|
||||
|
||||
def to_english(array: npt.NDArray[np.int64]) -> str:
|
||||
"""Converts a numpy integer array into a corresponding English string.
|
||||
|
||||
Args:
|
||||
array: Word in array (int) form. It is assumed that each integer in the
|
||||
array is between 0,...,26 (inclusive).
|
||||
|
||||
Returns:
|
||||
A (lowercase) string representation of the word.
|
||||
"""
|
||||
return ''.join(_chars[i] for i in array)
|
||||
|
||||
|
||||
def to_array(word: str) -> npt.NDArray[np.int64]:
|
||||
"""Converts a string of characters into a corresponding numpy array.
|
||||
|
||||
Args:
|
||||
word: Word in string form. It is assumed that each character in the
|
||||
string is either an empty space ' ' or lowercase alphabetical
|
||||
character.
|
||||
|
||||
Returns:
|
||||
An array representation of the word.
|
||||
"""
|
||||
return np.array([_char_d[c] for c in word])
|
||||
|
||||
|
||||
def get_words(category: str, build: bool = False) -> npt.NDArray[np.int64]:
|
||||
"""Loads a list of words in array form.
|
||||
|
||||
If specified, this will recompute the list from the human-readable list of
|
||||
words, and save the results in array form.
|
||||
|
||||
Args:
|
||||
category: Either 'guess' or 'solution', which corresponds to the list
|
||||
of acceptable guess words and the list of acceptable solution words.
|
||||
build: If True, recomputes and saves the array-version of the computed
|
||||
list for future access.
|
||||
|
||||
Returns:
|
||||
An array representation of the list of words specified by the category.
|
||||
This array has two dimensions, and the number of columns is fixed at
|
||||
five.
|
||||
"""
|
||||
assert category in {'guess', 'solution'}
|
||||
|
||||
arr_path = Path(__file__).parent / f'dictionary/{category}_list.npy'
|
||||
if build:
|
||||
list_path = Path(__file__).parent / f'dictionary/{category}_list.csv'
|
||||
|
||||
with open(list_path, 'r') as f:
|
||||
words = np.array([to_array(line.strip()) for line in f])
|
||||
np.save(arr_path, words)
|
||||
|
||||
return np.load(arr_path)
|
||||
|
||||
|
||||
def play():
|
||||
"""Play Wordle yourself!"""
|
||||
import gym
|
||||
import gym_wordle
|
||||
|
||||
env = gym.make('Wordle-v0') # load the environment
|
||||
|
||||
env.reset()
|
||||
solution = to_english(env.unwrapped.solution_space[env.solution]).upper() # no peeking!
|
||||
|
||||
done = False
|
||||
|
||||
while not done:
|
||||
action = -1
|
||||
|
||||
# in general, the environment won't be forgiving if you input an
|
||||
# invalid word, but for this function I want to let you screw up user
|
||||
# input without consequence, so just loops until valid input is taken
|
||||
while not env.action_space.contains(action):
|
||||
guess = input('Guess: ')
|
||||
action = env.unwrapped.action_space.index_of(to_array(guess))
|
||||
|
||||
state, reward, done, info = env.step(action)
|
||||
env.render()
|
||||
|
||||
print(f"The word was {solution}")
|
||||
340
gym_wordle/wordle.py
Normal file
340
gym_wordle/wordle.py
Normal file
@@ -0,0 +1,340 @@
|
||||
import gymnasium as gym
|
||||
import numpy as np
|
||||
import numpy.typing as npt
|
||||
from sty import fg, bg, ef, rs
|
||||
|
||||
from collections import Counter
|
||||
from gym_wordle.utils import to_english, to_array, get_words
|
||||
from typing import Optional
|
||||
from collections import defaultdict
|
||||
|
||||
|
||||
class WordList(gym.spaces.Discrete):
|
||||
"""Super class for defining a space of valid words according to a specified
|
||||
list.
|
||||
|
||||
The space is a subclass of gym.spaces.Discrete, where each element
|
||||
corresponds to an index of a valid word in the word list. The obfuscation
|
||||
is necessary for more direct implementation of RL algorithms, which expect
|
||||
spaces of less sophisticated form.
|
||||
|
||||
In addition to the default methods of the Discrete space, it implements
|
||||
a __getitem__ method for easy index lookup, and an index_of method to
|
||||
convert potential words into their corresponding index (if they exist).
|
||||
"""
|
||||
|
||||
def __init__(self, words: npt.NDArray[np.int64], **kwargs):
|
||||
"""
|
||||
Args:
|
||||
words: Collection of words in array form with shape (_, 5), where
|
||||
each word is a row of the array. Each array element is an integer
|
||||
between 0,...,26 (inclusive).
|
||||
kwargs: See documentation for gym.spaces.MultiDiscrete
|
||||
"""
|
||||
super().__init__(words.shape[0], **kwargs)
|
||||
self.words = words
|
||||
|
||||
def __getitem__(self, index: int) -> npt.NDArray[np.int64]:
|
||||
"""Obtains the (int-encoded) word associated with the given index.
|
||||
|
||||
Args:
|
||||
index: Index for the list of words.
|
||||
|
||||
Returns:
|
||||
Associated word at the position specified by index.
|
||||
"""
|
||||
return self.words[index]
|
||||
|
||||
def index_of(self, word: npt.NDArray[np.int64]) -> int:
|
||||
"""Given a word, determine its index in the list (if it exists),
|
||||
otherwise returning -1 if no index exists.
|
||||
|
||||
Args:
|
||||
word: Word to find in the word list.
|
||||
|
||||
Returns:
|
||||
The index of the given word if it exists, otherwise -1.
|
||||
"""
|
||||
try:
|
||||
index, = np.nonzero((word == self.words).all(axis=1))
|
||||
return index[0]
|
||||
except:
|
||||
return -1
|
||||
|
||||
|
||||
class SolutionList(WordList):
|
||||
"""Space for *solution* words to the Wordle environment.
|
||||
|
||||
In the game Wordle, there are two different collections of words:
|
||||
|
||||
* "guesses", which the game accepts as valid words to use to guess the
|
||||
answer.
|
||||
* "solutions", which the game uses to choose solutions from.
|
||||
|
||||
Of course, the set of solutions is a strict subset of the set of guesses.
|
||||
|
||||
This class represents the set of solution words.
|
||||
"""
|
||||
|
||||
def __init__(self, **kwargs):
|
||||
"""
|
||||
Args:
|
||||
kwargs: See documentation for gym.spaces.MultiDiscrete
|
||||
"""
|
||||
words = get_words('solution')
|
||||
super().__init__(words, **kwargs)
|
||||
|
||||
|
||||
class WordleObsSpace(gym.spaces.Box):
|
||||
"""Implementation of the state (observation) space in terms of gym
|
||||
primitives, in this case, gym.spaces.Box.
|
||||
|
||||
The Wordle observation space can be thought of as a 6x5 array with two
|
||||
channels:
|
||||
|
||||
- the character channel, indicating which characters are placed on the
|
||||
board (unfilled rows are marked with the empty character, 0)
|
||||
- the flag channel, indicating the in-game information associated with
|
||||
each character's placement (green highlight, yellow highlight, etc.)
|
||||
|
||||
where there are 6 rows, one for each turn in the game, and 5 columns, since
|
||||
the solution will always be a word of length 5.
|
||||
|
||||
For simplicity, and compatibility with stable_baselines algorithms,
|
||||
this multichannel is modeled as a 6x10 array, where the two channels are
|
||||
horizontally appended (along columns). Thus each row in the observation
|
||||
should be interpreted as c0 c1 c2 c3 c4 f0 f1 f2 f3 f4 when the word is
|
||||
c0...c4 and its associated flags are f0...f4.
|
||||
"""
|
||||
|
||||
def __init__(self, **kwargs):
|
||||
self.n_rows = 6
|
||||
self.n_cols = 5
|
||||
self.max_char = 26
|
||||
self.max_flag = 4
|
||||
|
||||
low = np.zeros((self.n_rows, 2*self.n_cols))
|
||||
high = np.c_[np.full((self.n_rows, self.n_cols), self.max_char),
|
||||
np.full((self.n_rows, self.n_cols), self.max_flag)]
|
||||
|
||||
super().__init__(low, high, dtype=np.int64, **kwargs)
|
||||
|
||||
|
||||
class GuessList(WordList):
|
||||
"""Space for *guess* words to the Wordle environment.
|
||||
|
||||
This class represents the set of guess words.
|
||||
"""
|
||||
|
||||
def __init__(self, **kwargs):
|
||||
"""
|
||||
Args:
|
||||
kwargs: See documentation for gym.spaces.MultiDiscrete
|
||||
"""
|
||||
words = get_words('guess')
|
||||
super().__init__(words, **kwargs)
|
||||
|
||||
|
||||
class WordleEnv(gym.Env):
|
||||
metadata = {'render.modes': ['human']}
|
||||
|
||||
# Character flag codes
|
||||
no_char = 0
|
||||
right_pos = 1
|
||||
wrong_pos = 2
|
||||
wrong_char = 3
|
||||
|
||||
def __init__(self):
|
||||
super().__init__()
|
||||
|
||||
self.action_space = GuessList()
|
||||
self.solution_space = SolutionList()
|
||||
|
||||
self.observation_space = WordleObsSpace()
|
||||
|
||||
self._highlights = {
|
||||
self.right_pos: (bg.green, bg.rs),
|
||||
self.wrong_pos: (bg.yellow, bg.rs),
|
||||
self.wrong_char: ('', ''),
|
||||
self.no_char: ('', ''),
|
||||
}
|
||||
|
||||
self.n_rounds = 6
|
||||
self.n_letters = 5
|
||||
self.info = {
|
||||
'correct': False,
|
||||
'guesses': set(),
|
||||
'known_positions': np.full(5, -1), # -1 for unknown, else letter index
|
||||
'known_letters': set(), # Letters known to be in the word
|
||||
'not_in_word': set(), # Letters known not to be in the word
|
||||
'tried_positions': defaultdict(set) # Positions tried for each letter
|
||||
}
|
||||
|
||||
def _highlighter(self, char: str, flag: int) -> str:
|
||||
"""Terminal renderer functionality. Properly highlights a character
|
||||
based on the flag associated with it.
|
||||
|
||||
Args:
|
||||
char: Character in question.
|
||||
flag: Associated flag, one of:
|
||||
- 0: no character (render no background)
|
||||
- 1: right position (render green background)
|
||||
- 2: wrong position (render yellow background)
|
||||
- 3: wrong character (render no background)
|
||||
|
||||
Returns:
|
||||
Correct ASCII sequence producing the desired character in the
|
||||
correct background.
|
||||
"""
|
||||
front, back = self._highlights[flag]
|
||||
return front + char + back
|
||||
|
||||
def reset(self, seed=None, options=None):
|
||||
"""Reset the environment to an initial state and returns an initial
|
||||
observation.
|
||||
|
||||
Note: The observation space instance should be a Box space.
|
||||
|
||||
Returns:
|
||||
state (object): The initial observation of the space.
|
||||
"""
|
||||
self.round = 0
|
||||
self.solution = self.solution_space.sample()
|
||||
self.soln_hash = set(self.solution_space[self.solution])
|
||||
|
||||
self.state = np.zeros((self.n_rounds, 2 * self.n_letters), dtype=np.int64)
|
||||
|
||||
self.info = {
|
||||
'correct': False,
|
||||
'guesses': set(),
|
||||
'known_positions': np.full(5, -1),
|
||||
'known_letters': set(),
|
||||
'not_in_word': set(),
|
||||
'tried_positions': defaultdict(set)
|
||||
}
|
||||
|
||||
self.simulate_first_guess()
|
||||
|
||||
return self.state, self.info
|
||||
|
||||
def simulate_first_guess(self):
|
||||
fixed_first_guess = "rates"
|
||||
fixed_first_guess_array = to_array(fixed_first_guess)
|
||||
|
||||
# Simulate the feedback for each letter in the fixed first guess
|
||||
feedback = np.zeros(self.n_letters, dtype=int) # Initialize feedback array
|
||||
for i, letter in enumerate(fixed_first_guess_array):
|
||||
if letter in self.solution_space[self.solution]:
|
||||
if letter == self.solution_space[self.solution][i]:
|
||||
feedback[i] = 1 # Correct position
|
||||
else:
|
||||
feedback[i] = 2 # Correct letter, wrong position
|
||||
else:
|
||||
feedback[i] = 3 # Letter not in word
|
||||
|
||||
# Update the state to reflect the fixed first guess and its feedback
|
||||
self.state[0, :self.n_letters] = fixed_first_guess_array
|
||||
self.state[0, self.n_letters:] = feedback
|
||||
|
||||
# Update self.info based on the feedback
|
||||
for i, flag in enumerate(feedback):
|
||||
if flag == self.right_pos:
|
||||
# Mark letter as correctly placed
|
||||
self.info['known_positions'][i] = fixed_first_guess_array[i]
|
||||
elif flag == self.wrong_pos:
|
||||
# Note the letter is in the word but in a different position
|
||||
self.info['known_letters'].add(fixed_first_guess_array[i])
|
||||
elif flag == self.wrong_char:
|
||||
# Note the letter is not in the word
|
||||
self.info['not_in_word'].add(fixed_first_guess_array[i])
|
||||
|
||||
# Since we're simulating the first guess, increment the round counter
|
||||
self.round = 1
|
||||
|
||||
def render(self, mode: str = 'human'):
|
||||
"""Renders the Wordle environment.
|
||||
|
||||
Currently supported render modes:
|
||||
- human: renders the Wordle game to the terminal.
|
||||
|
||||
Args:
|
||||
mode: the mode to render with.
|
||||
"""
|
||||
if mode == 'human':
|
||||
for row in self.state:
|
||||
text = ''.join(map(
|
||||
self._highlighter,
|
||||
to_english(row[:self.n_letters]).upper(),
|
||||
row[self.n_letters:]
|
||||
))
|
||||
print(text)
|
||||
else:
|
||||
super().render(mode=mode)
|
||||
|
||||
def step(self, action):
|
||||
assert self.action_space.contains(action), 'Invalid word!'
|
||||
|
||||
guessed_word = self.action_space[action]
|
||||
solution_word = self.solution_space[self.solution]
|
||||
|
||||
reward = 0
|
||||
correct_guess = np.array_equal(guessed_word, solution_word)
|
||||
|
||||
# Initialize flags for current guess
|
||||
current_flags = np.full(self.n_letters, self.wrong_char)
|
||||
|
||||
# Track newly discovered information
|
||||
new_info = False
|
||||
|
||||
for i in range(self.n_letters):
|
||||
guessed_letter = guessed_word[i]
|
||||
if guessed_letter in solution_word:
|
||||
# Penalize for reusing a letter found to not be in the word
|
||||
if guessed_letter in self.info['not_in_word']:
|
||||
reward -= 2
|
||||
|
||||
# Handle correct letter in the correct position
|
||||
if guessed_letter == solution_word[i]:
|
||||
current_flags[i] = self.right_pos
|
||||
if self.info['known_positions'][i] != guessed_letter:
|
||||
reward += 10 # Large reward for new correct placement
|
||||
new_info = True
|
||||
self.info['known_positions'][i] = guessed_letter
|
||||
else:
|
||||
reward += 20 # Large reward for repeating correct placement
|
||||
else:
|
||||
current_flags[i] = self.wrong_pos
|
||||
if guessed_letter not in self.info['known_letters'] or i not in self.info['tried_positions'][guessed_letter]:
|
||||
reward += 10 # Reward for guessing a letter in a new position
|
||||
new_info = True
|
||||
else:
|
||||
reward -= 20 # Penalize for not leveraging known information
|
||||
self.info['known_letters'].add(guessed_letter)
|
||||
self.info['tried_positions'][guessed_letter].add(i)
|
||||
else:
|
||||
# New incorrect letter
|
||||
if guessed_letter not in self.info['not_in_word']:
|
||||
reward -= 2 # Penalize for guessing a letter not in the word
|
||||
self.info['not_in_word'].add(guessed_letter)
|
||||
new_info = True
|
||||
else:
|
||||
reward -= 15 # Larger penalty for repeating an incorrect letter
|
||||
|
||||
# Update observation state with the current guess and flags
|
||||
self.state[self.round, :self.n_letters] = guessed_word
|
||||
self.state[self.round, self.n_letters:] = current_flags
|
||||
|
||||
# Check if the game is over
|
||||
done = self.round == self.n_rounds - 1 or correct_guess
|
||||
self.info['correct'] = correct_guess
|
||||
|
||||
if correct_guess:
|
||||
reward += 100 # Major reward for winning
|
||||
elif done:
|
||||
reward -= 50 # Penalty for losing without using new information effectively
|
||||
elif not new_info:
|
||||
reward -= 10 # Penalty if no new information was used in this guess
|
||||
|
||||
self.round += 1
|
||||
|
||||
return self.state, reward, done, False, self.info
|
||||
208
test.ipynb
208
test.ipynb
File diff suppressed because one or more lines are too long
61
test.py
61
test.py
@@ -1,61 +0,0 @@
|
||||
|
||||
from torch.utils.data import Dataset
|
||||
from transformers import BertGenerationEncoder, BertGenerationDecoder, EncoderDecoderModel, BertTokenizer
|
||||
from tqdm import tqdm as progress_bar
|
||||
import torch
|
||||
import matplotlib
|
||||
|
||||
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
||||
print(device)
|
||||
|
||||
encoder = BertGenerationEncoder.from_pretrained("google-bert/bert-base-uncased", bos_token_id=101, eos_token_id=102)
|
||||
# add cross attention layers and use BERT's cls token as BOS token and sep token as EOS token
|
||||
decoder = BertGenerationDecoder.from_pretrained("google-bert/bert-base-uncased", add_cross_attention=True, is_decoder=True, bos_token_id=101, eos_token_id=102)
|
||||
model = EncoderDecoderModel(encoder=encoder, decoder=decoder)
|
||||
|
||||
# create tokenizer...
|
||||
tokenizer = BertTokenizer.from_pretrained("google-bert/bert-large-uncased")
|
||||
|
||||
import json
|
||||
|
||||
class CodeDataset(Dataset):
|
||||
def __init__(self):
|
||||
with open("data/conala-train.json") as f:
|
||||
self.data = json.load(f)
|
||||
|
||||
def __len__(self):
|
||||
return len(self.data)
|
||||
|
||||
def __getitem__(self, idx):
|
||||
intent = self.data[idx]["rewritten_intent"] if self.data[idx]["rewritten_intent"] else self.data[idx]["intent"]
|
||||
return intent, self.data[idx]["snippet"]
|
||||
|
||||
|
||||
optimizer = torch.optim.AdamW(params=model.parameters(), lr=1e-3)
|
||||
dataloader = CodeDataset()
|
||||
model = model.to(device)
|
||||
|
||||
losses = []
|
||||
epochs = 10
|
||||
for i in range(epochs):
|
||||
|
||||
epoch_loss = 0
|
||||
|
||||
for idx, (question, answer) in progress_bar(enumerate(dataloader), total=len(dataloader)):
|
||||
|
||||
input_ids = tokenizer(question, add_special_tokens=False, return_tensors="pt").input_ids.to(device)
|
||||
label_ids = tokenizer(answer, return_tensors="pt").input_ids.to(device)
|
||||
|
||||
loss = model(input_ids=input_ids, decoder_input_ids=label_ids, labels=label_ids).loss
|
||||
|
||||
optimizer.zero_grad()
|
||||
loss.backward()
|
||||
optimizer.step()
|
||||
|
||||
epoch_loss += loss.item()
|
||||
|
||||
losses.append(epoch_loss)
|
||||
|
||||
plt.plot(losses, color="green", label="Training Loss")
|
||||
plt.legend(loc = 'upper left')
|
||||
plt.savefig("plot.png")
|
||||
Reference in New Issue
Block a user