File size: 10,852 Bytes
784a7e2 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 |
{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Augmentation by parapharsing"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Init & Load Seed Data"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import json, openai\n",
"from tqdm import tqdm "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"DOMAIN = \"drone-planning/\"\n",
"# DOMAIN = \"clean-up/\"\n",
"# DOMAIN = \"pick-and-place/\"\n",
"with open(DOMAIN + \"train_seed.jsonl\") as f:\n",
" train_seed = [json.loads(line) for line in f]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"eng_seeds = {\n",
" seed['natural']: [] for seed in train_seed\n",
"}"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Augmentation Code\n",
"prompting GPT-3 seems to work the best in this case"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"# You need to set your OPENAI API key here\n",
"# https://beta.openai.com/account/api-keys\n",
"openai.api_key = \"TO_BE_SET\""
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def normalize(sentence):\n",
" # captialize first letter and add period at the end if not present\n",
" if sentence[0].islower():\n",
" sentence = sentence[0].upper() + sentence[1:]\n",
" if sentence[-1] != '.':\n",
" sentence = sentence + '.'\n",
" return sentence\n",
"\n",
"def parse_sentences_from_response(response):\n",
" lines = response.split('\\n')\n",
" # assert len(lines) == 5\n",
" assert len(lines) == 10\n",
" lines[0] = \"1.\" + lines[0]\n",
" paraphrases = []\n",
" for idx, line in enumerate(lines):\n",
" assert line.startswith(str(idx+1) + '. ')\n",
" sentence_start_idx = len(str(idx+1) + '. ')\n",
" paraphrases.append(line[sentence_start_idx:])\n",
" for paraphrase in paraphrases:\n",
" if paraphrase[-1] == ' ':\n",
" if paraphrase[-2] == '.':\n",
" paraphrase = paraphrase[:-1]\n",
" else:\n",
" paraphrase = paraphrase[:-2] + '.'\n",
" return paraphrases\n",
"\n",
"\n",
"PROMPT = \"\"\"Rephrase the source sentence in 10 different ways. Make the outputs as diverse as possible.\n",
"\n",
"Source: \n",
"SOURCE-TO-BE-PLACED\n",
"\n",
"Outputs:\n",
"1.\"\"\"\n",
"def rephrase_a_sentence(sentence):\n",
" response = openai.Completion.create(\n",
" model=\"text-davinci-002\",\n",
" prompt=PROMPT.replace(\"SOURCE-TO-BE-PLACED\", normalize(sentence)),\n",
" temperature=0.7,\n",
" max_tokens=512,\n",
" top_p=1,\n",
" best_of=1,\n",
" frequency_penalty=0.1,\n",
" presence_penalty=0\n",
" )\n",
" output = response['choices'][0]['text']\n",
" try:\n",
" paraphrases = parse_sentences_from_response(output)\n",
" except:\n",
" print(\"Error in parsing response\")\n",
" print(output)\n",
" return output, \"ERROR\"\n",
" return parse_sentences_from_response(output)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"O = rephrase_a_sentence(\"Go to the red room or go to the green room to finally go to the blue room.\")"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"O"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Run Augmentation"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"len(eng_seeds)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"list(eng_seeds.keys())[0]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"def paraphrase_done(eng_seeds):\n",
" for eng_seed, extended in tqdm(eng_seeds.items()):\n",
" if len(extended) == 0:\n",
" return False\n",
" return True\n",
"\n",
"while not paraphrase_done(eng_seeds):\n",
" for eng_seed, extended in tqdm(eng_seeds.items()):\n",
" if len(extended) == 0:\n",
" extended += rephrase_a_sentence(eng_seed)"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"eng_seeds"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Dump as Training Data"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"train_seed[0]"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"with open(DOMAIN + \"syn-aug.train.jsonl\", 'w') as f:\n",
" for seed in train_seed:\n",
" f.write(json.dumps(seed) + '\\n')\n",
" for aug_eng in eng_seeds[seed['natural']]:\n",
" f.write(json.dumps({\n",
" 'natural': aug_eng,\n",
" 'canonical': seed['canonical'],\n",
" 'formula': seed['formula']\n",
" }) + '\\n')"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"with open(DOMAIN + \"syn.train.jsonl\", 'w') as f:\n",
" for seed in train_seed:\n",
" f.write(json.dumps(seed) + '\\n')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Normalize the natural language form "
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"if DOMAIN == \"clean-up/\":\n",
" # in clean up, golden natural language data comes without period at the end, no capitalization in the beginning\n",
" def clean_up_normalize(sentence):\n",
" if sentence[0].isupper():\n",
" sentence = sentence[0].lower() + sentence[1:]\n",
" if sentence[-1] == '.':\n",
" sentence = sentence[:-1]\n",
" return sentence\n",
"\n",
" buffer = []\n",
" with open(DOMAIN + \"syn-aug.train.jsonl\", 'r') as f:\n",
" for l in f.readlines():\n",
" buffer.append(json.loads(l))\n",
" \n",
" with open(DOMAIN + \"syn-aug.train.jsonl\", 'w') as f:\n",
" for dp in buffer:\n",
" f.write(json.dumps({\n",
" 'natural': clean_up_normalize(dp['natural']),\n",
" 'canonical': dp['canonical'],\n",
" 'formula': dp['formula']\n",
" }) + '\\n')\n",
"\n",
"if DOMAIN == \"pick-and-place/\":\n",
" # in pick and place, golden natural language data comes without period at the end, no capitalization in the beginning\n",
" def clean_up_normalize(sentence):\n",
" if sentence[0].isupper():\n",
" sentence = sentence[0].lower() + sentence[1:]\n",
" if sentence[-1] == '.':\n",
" sentence = sentence[:-1]\n",
" return sentence\n",
"\n",
" buffer = []\n",
" with open(DOMAIN + \"syn-aug.train.jsonl\", 'r') as f:\n",
" for l in f.readlines():\n",
" buffer.append(json.loads(l))\n",
" \n",
" with open(DOMAIN + \"syn-aug.train.jsonl\", 'w') as f:\n",
" for dp in buffer:\n",
" f.write(json.dumps({\n",
" 'natural': clean_up_normalize(dp['natural']),\n",
" 'canonical': dp['canonical'],\n",
" 'formula': dp['formula']\n",
" }) + '\\n')"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"if DOMAIN == \"drone-planning/\":\n",
" # in clean up, golden natural language data comes with a \"space + period\" at the end, no capitalization in the beginning\n",
" def clean_up_normalize(sentence):\n",
" if sentence[0].isupper():\n",
" sentence = sentence[0].lower() + sentence[1:]\n",
" while sentence[-1] == ' ' or sentence[-1] == '.' or sentence[-1] == '!':\n",
" sentence = sentence[:-1]\n",
" sentence = sentence + '.'\n",
" sentence = sentence.replace('.', ' .')\n",
" sentence = sentence.replace(',', ' ,')\n",
" return sentence\n",
"\n",
" buffer = []\n",
" # with open(DOMAIN + \"syn-aug.train.jsonl\", 'r') as f:\n",
" # for l in f.readlines():\n",
" # buffer.append(json.loads(l))\n",
" \n",
" # with open(DOMAIN + \"syn-aug.train.jsonl\", 'w') as f:\n",
" # for dp in buffer:\n",
" # f.write(json.dumps({\n",
" # 'natural': clean_up_normalize(dp['natural']),\n",
" # 'canonical': dp['canonical'],\n",
" # 'formula': dp['formula']\n",
" # }) + '\\n')\n",
" with open(DOMAIN + \"syn.train.jsonl\", 'r') as f:\n",
" for l in f.readlines():\n",
" buffer.append(json.loads(l))\n",
" \n",
" with open(DOMAIN + \"syn.train.jsonl\", 'w') as f:\n",
" for dp in buffer:\n",
" f.write(json.dumps({\n",
" 'natural': clean_up_normalize(dp['natural']),\n",
" 'canonical': dp['canonical'],\n",
" 'formula': dp['formula']\n",
" }) + '\\n')"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "GPML",
"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": "3.7.13"
},
"orig_nbformat": 4,
"vscode": {
"interpreter": {
"hash": "75567ad983eac98a78c1e40a895e8d82557b42cf9969286235abec07ddbf9e7d"
}
}
},
"nbformat": 4,
"nbformat_minor": 2
}
|