Upload fusion_t2i_CLIP_interrogator.ipynb
Browse files
Google Colab Notebooks/fusion_t2i_CLIP_interrogator.ipynb
CHANGED
|
@@ -25,20 +25,11 @@
|
|
| 25 |
"id": "cRV2YWomjMBU"
|
| 26 |
}
|
| 27 |
},
|
| 28 |
-
{
|
| 29 |
-
"cell_type": "markdown",
|
| 30 |
-
"source": [
|
| 31 |
-
"THIS IS AN OLD VERSION OF THE CLIP INTERROGATOR.\n",
|
| 32 |
-
"\n",
|
| 33 |
-
"YOU WILL FIND THE UP TO DATE VERSION HERE:https://huggingface.co/datasets/codeShare/fusion-t2i-generator-data/tree/main/Google%20Colab%20Jupyter%20Notebooks"
|
| 34 |
-
],
|
| 35 |
-
"metadata": {
|
| 36 |
-
"id": "9slWHq0JIX6D"
|
| 37 |
-
}
|
| 38 |
-
},
|
| 39 |
{
|
| 40 |
"cell_type": "code",
|
| 41 |
"source": [
|
|
|
|
|
|
|
| 42 |
"import os\n",
|
| 43 |
"home_directory = '/content/'\n",
|
| 44 |
"using_Kaggle = os.environ.get('KAGGLE_URL_BASE','')\n",
|
|
@@ -69,18 +60,17 @@
|
|
| 69 |
" %cd {home_directory}\n",
|
| 70 |
" !git clone https://huggingface.co/datasets/codeShare/fusion-t2i-generator-data\n",
|
| 71 |
" loaded = True\n",
|
| 72 |
-
"
|
| 73 |
-
"
|
| 74 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 75 |
"#------#\n",
|
| 76 |
-
"%cd {home_directory + 'fusion-t2i-generator-data/' + 'vocab'}\n",
|
| 77 |
-
"with open(f'prompts.json', 'r') as f:\n",
|
| 78 |
-
" data = json.load(f)\n",
|
| 79 |
-
" _df = pd.DataFrame({'count': data})['count']\n",
|
| 80 |
-
" prompts = {\n",
|
| 81 |
-
" key : value for key, value in _df.items()\n",
|
| 82 |
-
" }\n",
|
| 83 |
-
"#-------#\n",
|
| 84 |
"%cd {home_directory + 'fusion-t2i-generator-data/' + 'reference'}\n",
|
| 85 |
"with open(f'reference_prompts.json', 'r') as f:\n",
|
| 86 |
" data = json.load(f)\n",
|
|
@@ -95,28 +85,11 @@
|
|
| 95 |
" target_urls = {\n",
|
| 96 |
" key : value for key, value in _df.items()\n",
|
| 97 |
" }\n",
|
| 98 |
-
"from transformers import AutoTokenizer\n",
|
| 99 |
-
"tokenizer = AutoTokenizer.from_pretrained(\"openai/clip-vit-large-patch14\", clean_up_tokenization_spaces = False)\n",
|
| 100 |
-
"from transformers import CLIPProcessor, CLIPModel\n",
|
| 101 |
-
"processor = CLIPProcessor.from_pretrained(\"openai/clip-vit-large-patch14\" , clean_up_tokenization_spaces = True)\n",
|
| 102 |
-
"model = CLIPModel.from_pretrained(\"openai/clip-vit-large-patch14\")\n",
|
| 103 |
-
"logit_scale = model.logit_scale.exp() #logit_scale = 100.00000762939453\n",
|
| 104 |
-
"\n",
|
| 105 |
-
"index = 0\n",
|
| 106 |
-
"%cd {home_directory + 'fusion-t2i-generator-data/' + 'vocab'}\n",
|
| 107 |
-
"vocab_encodings = torch.load('vocab_encodings.pt', weights_only=False)\n",
|
| 108 |
-
"for key in vocab_encodings:\n",
|
| 109 |
-
" index = index + 1;\n",
|
| 110 |
-
"#------#\n",
|
| 111 |
-
"NUM_VOCAB_ITEMS = index\n",
|
| 112 |
"\n",
|
| 113 |
-
"index = 0\n",
|
| 114 |
-
"%cd {home_directory + 'fusion-t2i-generator-data/' + 'reference'}\n",
|
| 115 |
-
"for key in torch.load('reference_text_and_image_encodings.pt', weights_only=False):\n",
|
| 116 |
-
" index = index + 1;\n",
|
| 117 |
"#------#\n",
|
| 118 |
-
"
|
| 119 |
-
"\n"
|
|
|
|
| 120 |
],
|
| 121 |
"metadata": {
|
| 122 |
"id": "TC5lMJrS1HCC"
|
|
@@ -124,74 +97,151 @@
|
|
| 124 |
"execution_count": null,
|
| 125 |
"outputs": []
|
| 126 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 127 |
{
|
| 128 |
"cell_type": "code",
|
| 129 |
"source": [
|
| 130 |
-
"# @
|
| 131 |
-
"# @
|
| 132 |
-
"index = 213 # @param {type:\"slider\", min:0, max:1666, step:1}\n",
|
| 133 |
"PROMPT_INDEX = index\n",
|
| 134 |
"prompt = target_prompts[f'{PROMPT_INDEX}']\n",
|
| 135 |
"url = target_urls[f'{PROMPT_INDEX}']\n",
|
| 136 |
"if url.find('perchance')>-1:\n",
|
| 137 |
" image = Image.open(requests.get(url, stream=True).raw)\n",
|
| 138 |
"#------#\n",
|
|
|
|
|
|
|
|
|
|
| 139 |
"# @markdown ⚖️ 🖼️ encoding <-----?-----> 📝 encoding </div> <br>\n",
|
| 140 |
"C = 0.3 # @param {type:\"slider\", min:0, max:1, step:0.01}\n",
|
| 141 |
-
"
|
| 142 |
-
"prompt_strength = torch.tensor(math.pow(10 ,log_strength_1-1)).to(dtype = torch.float32)\n",
|
| 143 |
-
"reference = torch.zeros(768).to(dtype = torch.float32)\n",
|
| 144 |
-
"\n",
|
| 145 |
"%cd {home_directory + 'fusion-t2i-generator-data/' + 'reference'}\n",
|
| 146 |
"references = torch.load('reference_text_and_image_encodings.pt' , weights_only=False)\n",
|
| 147 |
-
"reference = torch.add(reference,
|
| 148 |
-
"reference = torch.add(reference,
|
| 149 |
"references = '' # Clear up memory\n",
|
| 150 |
-
"
|
| 151 |
-
"
|
| 152 |
-
"
|
| 153 |
-
"
|
| 154 |
-
|
| 155 |
-
|
| 156 |
-
"
|
| 157 |
-
|
| 158 |
-
|
| 159 |
-
|
| 160 |
-
|
| 161 |
-
|
| 162 |
-
|
| 163 |
-
|
| 164 |
-
"
|
| 165 |
-
"
|
| 166 |
-
"
|
| 167 |
-
"
|
| 168 |
-
"
|
| 169 |
-
"
|
| 170 |
-
"
|
| 171 |
-
"
|
| 172 |
-
"
|
| 173 |
-
"\n",
|
| 174 |
-
"
|
| 175 |
-
"
|
| 176 |
-
|
| 177 |
-
|
| 178 |
-
"
|
| 179 |
-
|
| 180 |
-
|
| 181 |
-
|
| 182 |
-
|
| 183 |
-
|
| 184 |
-
|
| 185 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 186 |
"\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 187 |
"min_wordcount = 0 # @param {type:\"slider\", min:0, max:20, step:1}\n",
|
| 188 |
-
"
|
| 189 |
-
"
|
| 190 |
-
" blacklist
|
| 191 |
" txt = _txt.lower().strip()\n",
|
| 192 |
" if len(txt)<min_wordcount: return True\n",
|
| 193 |
" if txt.isnumeric(): return True\n",
|
| 194 |
-
"
|
| 195 |
" for item in list(blacklist.split(',')):\n",
|
| 196 |
" if item.strip() == '' : continue\n",
|
| 197 |
" if txt.find(item.strip())> -1 : return True\n",
|
|
@@ -203,76 +253,112 @@
|
|
| 203 |
" if found:break\n",
|
| 204 |
" #------#\n",
|
| 205 |
" return not found\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 206 |
"\n",
|
| 207 |
-
"
|
| 208 |
-
"
|
| 209 |
-
"
|
| 210 |
-
"
|
| 211 |
-
"
|
| 212 |
-
"
|
| 213 |
-
"
|
| 214 |
-
"
|
| 215 |
-
"
|
| 216 |
-
"
|
| 217 |
-
"\n",
|
| 218 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 219 |
"\n",
|
| 220 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
| 221 |
"\n",
|
| 222 |
-
"
|
| 223 |
-
"
|
| 224 |
-
"
|
| 225 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 226 |
"\n",
|
| 227 |
-
"if (enable):\n",
|
| 228 |
-
" reference = reference/reference.norm(p=2, dim=-1, keepdim=True)\n",
|
| 229 |
-
" %cd {home_directory + 'fusion-t2i-generator-data/' + 'vocab'}\n",
|
| 230 |
-
" sims = torch.matmul(vocab_encodings.dequantize(),reference.t())\n",
|
| 231 |
-
" sorted , indices = torch.sort(sims,dim=0 , descending=True)\n",
|
| 232 |
"\n",
|
| 233 |
-
"
|
| 234 |
-
"
|
| 235 |
-
"
|
| 236 |
-
"
|
| 237 |
-
"
|
| 238 |
-
"
|
| 239 |
-
"
|
| 240 |
-
"
|
| 241 |
-
"
|
| 242 |
-
"
|
| 243 |
-
"
|
| 244 |
-
"
|
| 245 |
-
"
|
| 246 |
-
"
|
| 247 |
-
"
|
| 248 |
-
" variance = torch.add(variance, difference_to_average * difference_to_average)\n",
|
| 249 |
-
" if (key>=start_at_index + list_size) : break\n",
|
| 250 |
-
" #--------#\n",
|
| 251 |
-
" variance = variance * (1/max(1, list_size))\n",
|
| 252 |
-
" variance= variance.clone().detach();\n",
|
| 253 |
-
" print(f'The variance for the selected range is {math.sqrt(variance.item())} units from average')\n",
|
| 254 |
-
" #--------#\n",
|
| 255 |
-
"#---#\n",
|
| 256 |
-
"output = '{'\n",
|
| 257 |
-
"for _index in range(list_size):\n",
|
| 258 |
-
" tmp = prompts[f'{indices[min(_index+start_at_index,NUM_VOCAB_ITEMS-1)].item()}']\n",
|
| 259 |
-
" if isBlacklisted(tmp , SKIP): continue\n",
|
| 260 |
-
" tmp = fix_bad_symbols(tmp)\n",
|
| 261 |
-
" if output.find(tmp)>-1:continue\n",
|
| 262 |
-
" output = output + tmp + '|'\n",
|
| 263 |
"#---------#\n",
|
| 264 |
-
"
|
| 265 |
-
"print('')\n",
|
| 266 |
-
"print('')\n",
|
| 267 |
-
"for iter in range(N):\n",
|
| 268 |
-
" print(output)\n",
|
| 269 |
-
"#-------#\n",
|
| 270 |
-
"print('')\n",
|
| 271 |
-
"print('')\n",
|
| 272 |
-
"image or print('No image found')"
|
| 273 |
],
|
| 274 |
"metadata": {
|
| 275 |
-
"id": "
|
| 276 |
},
|
| 277 |
"execution_count": null,
|
| 278 |
"outputs": []
|
|
@@ -280,89 +366,50 @@
|
|
| 280 |
{
|
| 281 |
"cell_type": "code",
|
| 282 |
"source": [
|
| 283 |
-
"
|
| 284 |
-
"
|
| 285 |
-
"sorted , indices = torch.sort(
|
| 286 |
-
"
|
| 287 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 288 |
],
|
| 289 |
"metadata": {
|
| 290 |
-
"id": "
|
| 291 |
},
|
| 292 |
"execution_count": null,
|
| 293 |
"outputs": []
|
| 294 |
},
|
| 295 |
{
|
| 296 |
-
"cell_type": "
|
| 297 |
"source": [
|
| 298 |
-
"
|
| 299 |
-
"list_size = 1000 # @param {type:'number'}\n",
|
| 300 |
-
"start_at_index = 0 # @param {type:'number'}\n",
|
| 301 |
-
"print_Similarity = True # @param {type:\"boolean\"}\n",
|
| 302 |
-
"print_Prompts = True # @param {type:\"boolean\"}\n",
|
| 303 |
-
"print_Descriptions = True # @param {type:\"boolean\"}\n",
|
| 304 |
-
"compact_Output = True # @param {type:\"boolean\"}\n",
|
| 305 |
-
"newline_Separator = False # @param {type:\"boolean\"}\n",
|
| 306 |
-
"\n",
|
| 307 |
-
"import random\n",
|
| 308 |
-
"# @markdown -----------\n",
|
| 309 |
-
"# @markdown Mix with...\n",
|
| 310 |
-
"list_size2 = 1000 # @param {type:'number'}\n",
|
| 311 |
-
"start_at_index2 = 10000 # @param {type:'number'}\n",
|
| 312 |
-
"rate_percent = 0 # @param {type:\"slider\", min:0, max:100, step:1}\n",
|
| 313 |
-
"\n",
|
| 314 |
-
"# @markdown -----------\n",
|
| 315 |
-
"# @markdown Repeat output N times\n",
|
| 316 |
-
"N = 6 # @param {type:\"slider\", min:0, max:10, step:1}\n",
|
| 317 |
-
"\n",
|
| 318 |
-
"# title Show the 100 most similiar suffix and prefix text-encodings to the text encoding\n",
|
| 319 |
-
"RANGE = list_size\n",
|
| 320 |
-
"separator = '|'\n",
|
| 321 |
-
"if newline_Separator : separator = separator + '\\n'\n",
|
| 322 |
-
"\n",
|
| 323 |
-
"_prompts = ''\n",
|
| 324 |
-
"_sims = ''\n",
|
| 325 |
-
"for _index in range(start_at_index + RANGE):\n",
|
| 326 |
-
" if _index < start_at_index : continue\n",
|
| 327 |
-
" index = indices[_index].item()\n",
|
| 328 |
-
"\n",
|
| 329 |
-
" prompt = prompts[f'{index}']\n",
|
| 330 |
-
" if rate_percent >= random.randint(0,100) : prompt = prompts[f'{random.randint(start_at_index2 , start_at_index2 + list_size2)}']\n",
|
| 331 |
-
"\n",
|
| 332 |
-
" #Remove duplicates\n",
|
| 333 |
-
" if _prompts.find(prompt + separator)<=-1:\n",
|
| 334 |
-
" _sims = _sims + f'{round(100*sims[index].item(), 2)} %' + separator\n",
|
| 335 |
-
" #-------#\n",
|
| 336 |
-
" _prompts = _prompts.replace(prompt + separator,'')\n",
|
| 337 |
-
" _prompts = _prompts + prompt + separator\n",
|
| 338 |
-
" #------#\n",
|
| 339 |
-
"#------#\n",
|
| 340 |
-
"__prompts = fix_bad_symbols(__prompts)\n",
|
| 341 |
-
"__prompts = ('{' + _prompts + '}').replace(separator + '}', '}')\n",
|
| 342 |
-
"__sims = ('{' + _sims + '}').replace(separator + '}', '}')\n",
|
| 343 |
-
"#------#\n",
|
| 344 |
-
"\n",
|
| 345 |
-
"if(not print_Prompts): __prompts = ''\n",
|
| 346 |
-
"if(not print_Similarity): __sims = ''\n",
|
| 347 |
-
"\n",
|
| 348 |
-
"if(not compact_Output):\n",
|
| 349 |
-
" if(print_Descriptions):\n",
|
| 350 |
-
" print(f'The {start_at_index}-{start_at_index + RANGE} most similiar items to prompt : \\n\\n ')\n",
|
| 351 |
-
" for i in range(N) : print(__prompts)\n",
|
| 352 |
-
" print(f'The {start_at_index}-{start_at_index + RANGE} similarity % for items : \\n\\n' + __sims)\n",
|
| 353 |
-
" print('')\n",
|
| 354 |
-
" else:\n",
|
| 355 |
-
" for i in range(N) : print(__prompts)\n",
|
| 356 |
-
"else:\n",
|
| 357 |
-
" for i in range(N) : print(__prompts)\n",
|
| 358 |
-
"#-------#"
|
| 359 |
],
|
| 360 |
"metadata": {
|
| 361 |
-
"id": "
|
| 362 |
-
|
| 363 |
-
},
|
| 364 |
-
"execution_count": null,
|
| 365 |
-
"outputs": []
|
| 366 |
},
|
| 367 |
{
|
| 368 |
"cell_type": "markdown",
|
|
@@ -605,59 +652,54 @@
|
|
| 605 |
{
|
| 606 |
"cell_type": "code",
|
| 607 |
"source": [
|
| 608 |
-
"# @title \t⚄
|
| 609 |
-
"\n",
|
| 610 |
-
"def get_num_vocab_items(_url):\n",
|
| 611 |
-
" num_vocab_items = 0\n",
|
| 612 |
-
" for item in _url.split('_'):\n",
|
| 613 |
-
" if item.find('safetensors')>-1: num_vocab_items = int(item.replace('.safetensors', ''))\n",
|
| 614 |
-
" #------#\n",
|
| 615 |
-
" return num_vocab_items-1\n",
|
| 616 |
"\n",
|
|
|
|
|
|
|
|
|
|
| 617 |
"\n",
|
| 618 |
-
"
|
| 619 |
-
"
|
| 620 |
-
" _SCALE = torch.tensor(0.0043).to(dot_dtype)\n",
|
| 621 |
-
" _DIM = 768\n",
|
| 622 |
-
" _vocab = {}\n",
|
| 623 |
-
" #----#\n",
|
| 624 |
-
" inputs = tokenizer(text = _ref.strip(), truncation = True , padding=True, return_tensors=\"pt\")\n",
|
| 625 |
-
" ref = model.get_text_features(**inputs)[0]\n",
|
| 626 |
-
" ref = (ref/ref.norm(p=2, dim=-1, keepdim=True)).to(dtype = dot_dtype)\n",
|
| 627 |
-
" #-----#\n",
|
| 628 |
-
" num_vocab_items = 0\n",
|
| 629 |
-
" for url in urls:\n",
|
| 630 |
-
" num_vocab_items = num_vocab_items + get_num_vocab_items(url)\n",
|
| 631 |
-
" #------#\n",
|
| 632 |
-
" vocab = torch.zeros(num_vocab_items , _DIM).to(torch.uint8)\n",
|
| 633 |
" prompts = {}\n",
|
| 634 |
-
"
|
| 635 |
-
"
|
| 636 |
-
"
|
| 637 |
-
"
|
| 638 |
-
"
|
| 639 |
-
"
|
| 640 |
-
"
|
| 641 |
-
"
|
| 642 |
-
"
|
| 643 |
-
"
|
| 644 |
-
"
|
| 645 |
-
"
|
| 646 |
-
"
|
| 647 |
-
"
|
| 648 |
-
" for index in range(
|
| 649 |
-
"
|
| 650 |
-
"
|
| 651 |
-
"
|
| 652 |
-
"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 653 |
" #-------#\n",
|
| 654 |
-
"
|
| 655 |
-
"
|
| 656 |
-
"\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 657 |
],
|
| 658 |
"metadata": {
|
| 659 |
"cellView": "form",
|
| 660 |
-
"id": "
|
| 661 |
},
|
| 662 |
"execution_count": null,
|
| 663 |
"outputs": []
|
|
@@ -665,31 +707,14 @@
|
|
| 665 |
{
|
| 666 |
"cell_type": "code",
|
| 667 |
"source": [
|
| 668 |
-
"
|
| 669 |
-
"
|
| 670 |
-
"
|
| 671 |
-
"
|
| 672 |
-
"
|
| 673 |
-
"\n",
|
| 674 |
-
" #'/content/fusion-t2i-generator-data/clip_vocab_q0043_541291.safetensors' , '/content/fusion-t2i-generator-data/lyrics_vocab_q0043_41905.safetensors' , '/content/fusion-t2i-generator-data/names_vocab_q0043_162977.safetensors' , '/content/fusion-t2i-generator-data/r34_vocab_q0043_96166.safetensors' ]\n",
|
| 675 |
-
"\n",
|
| 676 |
-
"indices , prompts , sims = get_similiar(ref , urls , LIST_SIZE)\n",
|
| 677 |
-
"\n",
|
| 678 |
-
"index = 0\n",
|
| 679 |
-
"_prompts = {}\n",
|
| 680 |
-
"for index in range(203662):\n",
|
| 681 |
-
" try:\n",
|
| 682 |
-
" key = prompts[f'{indices[index].item()}']\n",
|
| 683 |
-
" print(key)\n",
|
| 684 |
-
" except: print('Not found!')\n",
|
| 685 |
-
" #_prompts[f'{key}'] = sims[key].item()\n",
|
| 686 |
-
" index = index + 1\n",
|
| 687 |
-
" if index>LIST_SIZE:break\n",
|
| 688 |
-
"\n"
|
| 689 |
],
|
| 690 |
"metadata": {
|
| 691 |
-
"
|
| 692 |
-
"id": "Azz1kCza6LB3"
|
| 693 |
},
|
| 694 |
"execution_count": null,
|
| 695 |
"outputs": []
|
|
|
|
| 25 |
"id": "cRV2YWomjMBU"
|
| 26 |
}
|
| 27 |
},
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 28 |
{
|
| 29 |
"cell_type": "code",
|
| 30 |
"source": [
|
| 31 |
+
"# @title ⚄ Initialize\n",
|
| 32 |
+
"\n",
|
| 33 |
"import os\n",
|
| 34 |
"home_directory = '/content/'\n",
|
| 35 |
"using_Kaggle = os.environ.get('KAGGLE_URL_BASE','')\n",
|
|
|
|
| 60 |
" %cd {home_directory}\n",
|
| 61 |
" !git clone https://huggingface.co/datasets/codeShare/fusion-t2i-generator-data\n",
|
| 62 |
" loaded = True\n",
|
| 63 |
+
"\n",
|
| 64 |
+
"from transformers import AutoTokenizer\n",
|
| 65 |
+
"tokenizer = AutoTokenizer.from_pretrained(\"openai/clip-vit-large-patch14\", clean_up_tokenization_spaces = False)\n",
|
| 66 |
+
"from transformers import CLIPProcessor, CLIPModel\n",
|
| 67 |
+
"processor = CLIPProcessor.from_pretrained(\"openai/clip-vit-large-patch14\" , clean_up_tokenization_spaces = True)\n",
|
| 68 |
+
"model = CLIPModel.from_pretrained(\"openai/clip-vit-large-patch14\")\n",
|
| 69 |
+
"logit_scale = model.logit_scale.exp() #logit_scale = 100.00000762939453\n",
|
| 70 |
+
"\n",
|
| 71 |
+
"%cd {home_directory + 'fusion-t2i-generator-data/'}\n",
|
| 72 |
+
"!unzip reference.zip\n",
|
| 73 |
"#------#\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 74 |
"%cd {home_directory + 'fusion-t2i-generator-data/' + 'reference'}\n",
|
| 75 |
"with open(f'reference_prompts.json', 'r') as f:\n",
|
| 76 |
" data = json.load(f)\n",
|
|
|
|
| 85 |
" target_urls = {\n",
|
| 86 |
" key : value for key, value in _df.items()\n",
|
| 87 |
" }\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 88 |
"\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
| 89 |
"#------#\n",
|
| 90 |
+
"dot_dtype = torch.float32\n",
|
| 91 |
+
"dim = 768\n",
|
| 92 |
+
"reference = torch.zeros(dim).to(dtype = dot_dtype)"
|
| 93 |
],
|
| 94 |
"metadata": {
|
| 95 |
"id": "TC5lMJrS1HCC"
|
|
|
|
| 97 |
"execution_count": null,
|
| 98 |
"outputs": []
|
| 99 |
},
|
| 100 |
+
{
|
| 101 |
+
"cell_type": "markdown",
|
| 102 |
+
"source": [
|
| 103 |
+
"Feel free to skip these cells if you do not plan on using them\n"
|
| 104 |
+
],
|
| 105 |
+
"metadata": {
|
| 106 |
+
"id": "Xf9zoq-Za3wi"
|
| 107 |
+
}
|
| 108 |
+
},
|
| 109 |
{
|
| 110 |
"cell_type": "code",
|
| 111 |
"source": [
|
| 112 |
+
"# @markdown 🖼️+📝 Choose a pre-encoded reference (optional)\n",
|
| 113 |
+
"index = 657 # @param {type:\"slider\", min:0, max:1666, step:1}\n",
|
|
|
|
| 114 |
"PROMPT_INDEX = index\n",
|
| 115 |
"prompt = target_prompts[f'{PROMPT_INDEX}']\n",
|
| 116 |
"url = target_urls[f'{PROMPT_INDEX}']\n",
|
| 117 |
"if url.find('perchance')>-1:\n",
|
| 118 |
" image = Image.open(requests.get(url, stream=True).raw)\n",
|
| 119 |
"#------#\n",
|
| 120 |
+
"try: reference\n",
|
| 121 |
+
"except: reference = torch.zeros(dim).to(dtype = dot_dtype)\n",
|
| 122 |
+
"if reference == '': reference = torch.zeros(dim).to(dtype = dot_dtype)\n",
|
| 123 |
"# @markdown ⚖️ 🖼️ encoding <-----?-----> 📝 encoding </div> <br>\n",
|
| 124 |
"C = 0.3 # @param {type:\"slider\", min:0, max:1, step:0.01}\n",
|
| 125 |
+
"log_strength = 1 # @param {type:\"slider\", min:-5, max:5, step:0.01}\n",
|
|
|
|
|
|
|
|
|
|
| 126 |
"%cd {home_directory + 'fusion-t2i-generator-data/' + 'reference'}\n",
|
| 127 |
"references = torch.load('reference_text_and_image_encodings.pt' , weights_only=False)\n",
|
| 128 |
+
"reference = torch.add(reference, math.pow(10 ,log_strength-1) * C * references[index][0].dequantize().to(dtype = torch.float32))\n",
|
| 129 |
+
"reference = torch.add(reference, math.pow(10 ,log_strength-1) * (1-C) * references[index][1].dequantize().to(dtype = torch.float32))\n",
|
| 130 |
"references = '' # Clear up memory\n",
|
| 131 |
+
"ref = reference.clone().detach()\n",
|
| 132 |
+
"#------#\n",
|
| 133 |
+
"print(f'Prompt for this image : \\n\\n \"{prompt} \" \\n\\n')\n",
|
| 134 |
+
"image"
|
| 135 |
+
],
|
| 136 |
+
"metadata": {
|
| 137 |
+
"id": "BwrEs5zVB0Sb"
|
| 138 |
+
},
|
| 139 |
+
"execution_count": null,
|
| 140 |
+
"outputs": []
|
| 141 |
+
},
|
| 142 |
+
{
|
| 143 |
+
"cell_type": "code",
|
| 144 |
+
"source": [
|
| 145 |
+
"# @markdown 🖼️ Upload your own image for use as reference via URL (optional)\n",
|
| 146 |
+
"URL = '' # @param {type:'string' ,placeholder:'paste an url here'}\n",
|
| 147 |
+
"image = Image.open(requests.get(URL, stream=True).raw)\n",
|
| 148 |
+
"#---------#\n",
|
| 149 |
+
"# Get image features\n",
|
| 150 |
+
"inputs = processor(images=image, return_tensors=\"pt\")\n",
|
| 151 |
+
"image_features = model.get_image_features(**inputs)\n",
|
| 152 |
+
"image_features = image_features / image_features.norm(p=2, dim=-1, keepdim=True)\n",
|
| 153 |
+
"#-----#\n",
|
| 154 |
+
"log_strength = 1 # @param {type:\"slider\", min:-5, max:5, step:0.01}\n",
|
| 155 |
+
"ref = ref + math.pow(10,log_strength-1)*image_features\n",
|
| 156 |
+
"image"
|
| 157 |
+
],
|
| 158 |
+
"metadata": {
|
| 159 |
+
"id": "IqUsiQw2HU2C"
|
| 160 |
+
},
|
| 161 |
+
"execution_count": null,
|
| 162 |
+
"outputs": []
|
| 163 |
+
},
|
| 164 |
+
{
|
| 165 |
+
"cell_type": "code",
|
| 166 |
+
"source": [
|
| 167 |
+
"# @markdown 🖼️ Upload your own image in the /content/ folder for use as reference (optional)\n",
|
| 168 |
+
"FILENAME = '' # @param {type:'string' ,placeholder:'IMG_123.png'}\n",
|
| 169 |
+
"import cv2\n",
|
| 170 |
+
"image = cv2.imread(FILENAME)\n",
|
| 171 |
+
"image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)\n",
|
| 172 |
"\n",
|
| 173 |
+
"#---------#\n",
|
| 174 |
+
"# Get image features\n",
|
| 175 |
+
"inputs = processor(images=image, return_tensors=\"pt\")\n",
|
| 176 |
+
"image_features = model.get_image_features(**inputs)\n",
|
| 177 |
+
"image_features = image_features / image_features.norm(p=2, dim=-1, keepdim=True)\n",
|
| 178 |
+
"#-----#\n",
|
| 179 |
+
"log_strength = 1 # @param {type:\"slider\", min:-5, max:5, step:0.01}\n",
|
| 180 |
+
"ref = ref + math.pow(10,log_strength-1)*image_features\n",
|
| 181 |
+
"image"
|
| 182 |
+
],
|
| 183 |
+
"metadata": {
|
| 184 |
+
"id": "I_-GOwFPKkha"
|
| 185 |
+
},
|
| 186 |
+
"execution_count": null,
|
| 187 |
+
"outputs": []
|
| 188 |
+
},
|
| 189 |
+
{
|
| 190 |
+
"cell_type": "markdown",
|
| 191 |
+
"source": [
|
| 192 |
+
"Save the reference prior to running the Interrogator"
|
| 193 |
+
],
|
| 194 |
+
"metadata": {
|
| 195 |
+
"id": "zeu6JcM-mk9z"
|
| 196 |
+
}
|
| 197 |
+
},
|
| 198 |
+
{
|
| 199 |
+
"cell_type": "code",
|
| 200 |
+
"source": [
|
| 201 |
+
"# @title ⚄ Save the reference\n",
|
| 202 |
+
"try: ref\n",
|
| 203 |
+
"except: ref = torch.zeros(dim)\n",
|
| 204 |
+
"_ref = {}\n",
|
| 205 |
+
"_ref['weights'] = ref.to(dot_dtype)\n",
|
| 206 |
+
"%cd /content/\n",
|
| 207 |
+
"save_file(_ref , 'reference.safetensors' )"
|
| 208 |
+
],
|
| 209 |
+
"metadata": {
|
| 210 |
+
"id": "lOQuTPfBMK82"
|
| 211 |
+
},
|
| 212 |
+
"execution_count": null,
|
| 213 |
+
"outputs": []
|
| 214 |
+
},
|
| 215 |
+
{
|
| 216 |
+
"cell_type": "code",
|
| 217 |
+
"source": [
|
| 218 |
+
"# @title ⚄ Run the CLIP interrogator on the saved reference\n",
|
| 219 |
+
"LIST_SIZE = 1000 # @param {type:'number' , placeholder:'set how large the list should be'}\n",
|
| 220 |
+
"START_AT = 0 # @param {type:'number' , placeholder:'set how large the list should be'}\n",
|
| 221 |
+
"# @markdown -----\n",
|
| 222 |
+
"# @markdown Select vocab\n",
|
| 223 |
+
"general = False # @param {type:\"boolean\"}\n",
|
| 224 |
+
"civit9 = True # @param {type:\"boolean\"}\n",
|
| 225 |
+
"fanfic1 = False # @param {type:\"boolean\"}\n",
|
| 226 |
+
"fanfic2 = False # @param {type:\"boolean\"}\n",
|
| 227 |
+
"# @markdown -----\n",
|
| 228 |
+
"# @title ⚄ New interrogator code using quantized text corpus\n",
|
| 229 |
+
"%cd /content/\n",
|
| 230 |
+
"_ref = load_file('reference.safetensors' )\n",
|
| 231 |
+
"ref = _ref['weights'].to(dot_dtype)\n",
|
| 232 |
+
"# @markdown 📝 Enhance/Penalize Similarity and skip items containing word(s)\n",
|
| 233 |
+
"POS1 = '' # @param {type:'string' ,placeholder:'item1 , item2 , ...'}\n",
|
| 234 |
+
"POS2 = '' # @param {type:'string' ,placeholder:'item1 , item2 , ...'}\n",
|
| 235 |
+
"NEG = ''# @param {type:'string' ,placeholder:'item1 , item2 , ...'}\n",
|
| 236 |
+
"SKIP = '' # @param {type:'string' ,placeholder:'item1 , item2 , ...'}\n",
|
| 237 |
"min_wordcount = 0 # @param {type:\"slider\", min:0, max:20, step:1}\n",
|
| 238 |
+
"def isBlacklisted(_txt):\n",
|
| 239 |
+
" blacklist = SKIP.lower().replace('</w>' , ' ').replace('{' , '').replace('}' , '').replace('|' , ',').strip()\n",
|
| 240 |
+
" if blacklist == '': return False\n",
|
| 241 |
" txt = _txt.lower().strip()\n",
|
| 242 |
" if len(txt)<min_wordcount: return True\n",
|
| 243 |
" if txt.isnumeric(): return True\n",
|
| 244 |
+
" #-----#\n",
|
| 245 |
" for item in list(blacklist.split(',')):\n",
|
| 246 |
" if item.strip() == '' : continue\n",
|
| 247 |
" if txt.find(item.strip())> -1 : return True\n",
|
|
|
|
| 253 |
" if found:break\n",
|
| 254 |
" #------#\n",
|
| 255 |
" return not found\n",
|
| 256 |
+
"# @markdown -----\n",
|
| 257 |
+
"# @markdown logarithmic prompt strength x for value 10^(x-1)\n",
|
| 258 |
+
"_POS1 = 1 # @param {type:\"slider\", min:-5, max:5, step:0.01}\n",
|
| 259 |
+
"_POS2 = 1 # @param {type:\"slider\", min:-5, max:5, step:0.01}\n",
|
| 260 |
+
"_NEG = 1 # @param {type:\"slider\", min:-5, max:5, step:0.01}\n",
|
| 261 |
+
"# @markdown -----\n",
|
| 262 |
+
"for _item in POS1.split(','):\n",
|
| 263 |
+
" item = _item.strip()\n",
|
| 264 |
+
" if item == '':continue\n",
|
| 265 |
+
" inputs = tokenizer(text = item.strip(), truncation = True , padding=True, return_tensors=\"pt\")\n",
|
| 266 |
+
" ref = ref + math.pow(10,_POS1-1) * model.get_text_features(**inputs)[0]\n",
|
| 267 |
+
"#-------#\n",
|
| 268 |
+
"for _item in POS2.split(','):\n",
|
| 269 |
+
" item = _item.strip()\n",
|
| 270 |
+
" if item == '':continue\n",
|
| 271 |
+
" inputs = tokenizer(text = item.strip(), truncation = True , padding=True, return_tensors=\"pt\")\n",
|
| 272 |
+
" ref = ref + math.pow(10,_POS2-1) * model.get_text_features(**inputs)[0]\n",
|
| 273 |
+
"#-------#\n",
|
| 274 |
+
"for _item in NEG.split(','):\n",
|
| 275 |
+
" item = _item.strip()\n",
|
| 276 |
+
" if item == '':continue\n",
|
| 277 |
+
" inputs = tokenizer(text = item.strip(), truncation = True , padding=True, return_tensors=\"pt\")\n",
|
| 278 |
+
" ref = ref + math.pow(10,_NEG-1) * model.get_text_features(**inputs)[0]\n",
|
| 279 |
+
"#------#\n",
|
| 280 |
+
"ref = (ref/ref.norm(p=2, dim=-1, keepdim=True)).to(dtype = dot_dtype)\n",
|
| 281 |
+
"vocab_to_load = ''\n",
|
| 282 |
+
"if (general): vocab_to_load = vocab_to_load + 'general , '\n",
|
| 283 |
+
"if (civit9): vocab_to_load = vocab_to_load + 'civit9 , '\n",
|
| 284 |
+
"if (fanfic1): vocab_to_load = vocab_to_load + 'fanfic1 , '\n",
|
| 285 |
+
"if (fanfic2): vocab_to_load = vocab_to_load + 'fanfic2 , '\n",
|
| 286 |
+
"vocab_to_load = (vocab_to_load +'}').replace(' , }' , '')\n",
|
| 287 |
+
"multi = vocab_to_load.find(',')>-1\n",
|
| 288 |
"\n",
|
| 289 |
+
"#-----#\n",
|
| 290 |
+
"prompts_folder = f'{home_directory}fusion-t2i-generator-data/vocab-v2/text'\n",
|
| 291 |
+
"encodings_folder = f'{home_directory}fusion-t2i-generator-data/vocab-v2/text_encodings'\n",
|
| 292 |
+
"#----#\n",
|
| 293 |
+
"scale = 0.0043\n",
|
| 294 |
+
"size = 0\n",
|
| 295 |
+
"#------#\n",
|
| 296 |
+
"total_items = 0\n",
|
| 297 |
+
"for filename in os.listdir(prompts_folder):\n",
|
| 298 |
+
" if (not general and filename.find('general')>-1):continue\n",
|
| 299 |
+
" if (not civit9 and filename.find('civit9')>-1):continue\n",
|
| 300 |
+
" if (not fanfic1 and filename.find('fanfic1')>-1):continue\n",
|
| 301 |
+
" if (not fanfic2 and filename.find('fanfic2')>-1):continue\n",
|
| 302 |
+
" size = size + LIST_SIZE\n",
|
| 303 |
+
"#-------#\n",
|
| 304 |
+
"similiar_sims = torch.zeros(size)\n",
|
| 305 |
+
"similiar_prompts = {}\n",
|
| 306 |
+
"_index = 0\n",
|
| 307 |
+
"#-------#\n",
|
| 308 |
+
"similiar_encodings = {}\n",
|
| 309 |
+
"for filename in os.listdir(prompts_folder):\n",
|
| 310 |
+
" if (not general and filename.find('general')>-1):continue\n",
|
| 311 |
+
" if (not civit9 and filename.find('civit9')>-1):continue\n",
|
| 312 |
+
" if (not fanfic1 and filename.find('fanfic1')>-1):continue\n",
|
| 313 |
+
" if (not fanfic2 and filename.find('fanfic2')>-1):continue\n",
|
| 314 |
+
" #------#\n",
|
| 315 |
+
" root_filename = filename.replace('.json', '')\n",
|
| 316 |
+
" %cd {prompts_folder}\n",
|
| 317 |
+
" prompts = {}\n",
|
| 318 |
+
" with open(f'{root_filename}.json', 'r') as f:\n",
|
| 319 |
+
" data = json.load(f).items()\n",
|
| 320 |
+
" for key,value in data:\n",
|
| 321 |
+
" prompts[key] = value\n",
|
| 322 |
+
" num_items = int(prompts['num_items'])\n",
|
| 323 |
+
" total_items = total_items + num_items\n",
|
| 324 |
"\n",
|
| 325 |
+
" #------#\n",
|
| 326 |
+
" try:vocab_loaded\n",
|
| 327 |
+
" except:\n",
|
| 328 |
+
" vocab_loaded = 'first'\n",
|
| 329 |
+
" #-----#\n",
|
| 330 |
"\n",
|
| 331 |
+
" if vocab_loaded == 'first' or (vocab_loaded != vocab_to_load and not multi):\n",
|
| 332 |
+
" %cd {encodings_folder}\n",
|
| 333 |
+
" _text_encodings = load_file(f'{root_filename}.safetensors')['weights'].to(torch.uint8)\n",
|
| 334 |
+
" text_encodings = torch.zeros(num_items , dim)\n",
|
| 335 |
+
" tmp = torch.ones(dim).to(dot_dtype)\n",
|
| 336 |
+
" for index in range(num_items):\n",
|
| 337 |
+
" text_encodings[index] = torch.sub(_text_encodings[index][1:dim+1].to(dot_dtype) , tmp , alpha= _text_encodings[index][0].to(dot_dtype))\n",
|
| 338 |
+
" vocab_loaded = vocab_to_load\n",
|
| 339 |
+
" #------#\n",
|
| 340 |
"\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 341 |
"\n",
|
| 342 |
+
" sims = torch.matmul(text_encodings*scale, ref.t())\n",
|
| 343 |
+
" sorted , indices = torch.sort(sims , dim=0 , descending = True)\n",
|
| 344 |
+
" #-----#\n",
|
| 345 |
+
" for index in range(LIST_SIZE + START_AT):\n",
|
| 346 |
+
" if index<START_AT: continue\n",
|
| 347 |
+
" key = indices[index].item()\n",
|
| 348 |
+
" try:prompt = prompts[f'{key}']\n",
|
| 349 |
+
" except:continue\n",
|
| 350 |
+
" if(isBlacklisted(prompt)):continue\n",
|
| 351 |
+
" #-------#\n",
|
| 352 |
+
" similiar_sims[_index] = torch.tensor(round(sims[key].item(), 5))\n",
|
| 353 |
+
" similiar_prompts[f'{_index}'] = prompt\n",
|
| 354 |
+
" _index = _index + 1\n",
|
| 355 |
+
" #-------#\n",
|
| 356 |
+
" continue\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 357 |
"#---------#\n",
|
| 358 |
+
"print(f'\\n\\nProcessed entire list of {total_items} items to find closest match. Saved closest matching indices {START_AT} to {START_AT + LIST_SIZE} as the dict \"similiar_prompts\" with {LIST SIZE} items. \\n\\n')\n"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 359 |
],
|
| 360 |
"metadata": {
|
| 361 |
+
"id": "kOYZ8Ajn-DD8"
|
| 362 |
},
|
| 363 |
"execution_count": null,
|
| 364 |
"outputs": []
|
|
|
|
| 366 |
{
|
| 367 |
"cell_type": "code",
|
| 368 |
"source": [
|
| 369 |
+
"\n",
|
| 370 |
+
"# @title ⚄ Printing results from text corpus\n",
|
| 371 |
+
"sorted , indices = torch.sort(similiar_sims , dim=0 , descending = True)\n",
|
| 372 |
+
"include_similiarity = False # @param {type:\"boolean\"}\n",
|
| 373 |
+
"print_as_list = False # @param {type:\"boolean\"}\n",
|
| 374 |
+
"N = 7 # @param {type:\"slider\", min:0, max:10, step:1}\n",
|
| 375 |
+
"\n",
|
| 376 |
+
"if(print_as_list):\n",
|
| 377 |
+
" for index in range(LIST_SIZE):\n",
|
| 378 |
+
" key = indices[index].item()\n",
|
| 379 |
+
" sim = similiar_sims[key].item()\n",
|
| 380 |
+
" prompt = similiar_prompts[f'{key}']\n",
|
| 381 |
+
" if include_similiarity :print(f'{prompt} - {round(sim*100,1)} %')\n",
|
| 382 |
+
" else: print(f'{prompt}')\n",
|
| 383 |
+
"#-------#\n",
|
| 384 |
+
"else:\n",
|
| 385 |
+
" prompt = ''\n",
|
| 386 |
+
" for iter in range(N):\n",
|
| 387 |
+
" prompt = prompt + '{'\n",
|
| 388 |
+
" for index in range(LIST_SIZE):\n",
|
| 389 |
+
" key = indices[index].item()\n",
|
| 390 |
+
" sim = similiar_sims[key].item()\n",
|
| 391 |
+
" prompt = prompt + fix_bad_symbols(similiar_prompts[f'{key}']) + '|'\n",
|
| 392 |
+
" #-----#\n",
|
| 393 |
+
" prompt = (prompt + '}').replace('|}', '} ')\n",
|
| 394 |
+
" #------#\n",
|
| 395 |
+
" print(f'\\ Similiar prompts: \\n\\n {prompt} \\n\\n')\n",
|
| 396 |
+
" image\n",
|
| 397 |
+
"#-----#\n"
|
| 398 |
],
|
| 399 |
"metadata": {
|
| 400 |
+
"id": "XOMkIKc9-wZz"
|
| 401 |
},
|
| 402 |
"execution_count": null,
|
| 403 |
"outputs": []
|
| 404 |
},
|
| 405 |
{
|
| 406 |
+
"cell_type": "markdown",
|
| 407 |
"source": [
|
| 408 |
+
"OTHER STUFF BELOW - Code for the modules below are work-in-progress."
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 409 |
],
|
| 410 |
"metadata": {
|
| 411 |
+
"id": "FRIqYJDEebpf"
|
| 412 |
+
}
|
|
|
|
|
|
|
|
|
|
| 413 |
},
|
| 414 |
{
|
| 415 |
"cell_type": "markdown",
|
|
|
|
| 652 |
{
|
| 653 |
"cell_type": "code",
|
| 654 |
"source": [
|
| 655 |
+
"# @title \t⚄ Quick fix for normalizing encoded text corpus tensors\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 656 |
"\n",
|
| 657 |
+
"import os\n",
|
| 658 |
+
"my_mkdirs('/content/output')\n",
|
| 659 |
+
"my_mkdirs('/content/output/text_encodings')\n",
|
| 660 |
"\n",
|
| 661 |
+
"for filename in os.listdir(f'{prompts_folder}'):\n",
|
| 662 |
+
" %cd {prompts_folder}\n",
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 663 |
" prompts = {}\n",
|
| 664 |
+
" with open(f'{filename}', 'r') as f:\n",
|
| 665 |
+
" data = json.load(f).items()\n",
|
| 666 |
+
" for key,value in data:\n",
|
| 667 |
+
" prompts[key] = value\n",
|
| 668 |
+
" #------#\n",
|
| 669 |
+
" num_items = int(prompts['num_items'])\n",
|
| 670 |
+
"\n",
|
| 671 |
+
" %cd {encodings_folder}\n",
|
| 672 |
+
" enc_filename = filename.replace('json', 'safetensors')\n",
|
| 673 |
+
" _text_encodings = load_file(f'{enc_filename}')['weights'].to(torch.uint8)\n",
|
| 674 |
+
" text_encodings = torch.zeros(num_items , dim)\n",
|
| 675 |
+
" tmp = torch.ones(dim)\n",
|
| 676 |
+
" tmp2 = torch.tensor(1/0.0043)\n",
|
| 677 |
+
" zero_point = 0\n",
|
| 678 |
+
" for index in range(num_items):\n",
|
| 679 |
+
" text_encodings[index] = torch.tensor(0.0043) * torch.sub(_text_encodings[index][1:dim+1] , tmp , alpha= _text_encodings[index][0]).to(torch.float32)\n",
|
| 680 |
+
" text_encodings[index] = tmp2*text_encodings[index]/text_encodings[index].norm(p=2, dim=-1, keepdim = True)\n",
|
| 681 |
+
" test = torch.round( torch.add(text_encodings[index],tmp*zero_point))\n",
|
| 682 |
+
" less_than_zero = test<0\n",
|
| 683 |
+
" while(torch.any(less_than_zero).item()):\n",
|
| 684 |
+
" zero_point = zero_point + 1\n",
|
| 685 |
+
" test = torch.round( torch.add(text_encodings[index],tmp*zero_point))\n",
|
| 686 |
+
" less_than_zero = test<0\n",
|
| 687 |
+
" #------#\n",
|
| 688 |
+
" _text_encodings[index][0] = zero_point\n",
|
| 689 |
+
" _text_encodings[index][1:dim+1] = test\n",
|
| 690 |
" #-------#\n",
|
| 691 |
+
" %cd /content/output/text_encodings\n",
|
| 692 |
+
"\n",
|
| 693 |
+
" tmp = {}\n",
|
| 694 |
+
" tmp['weights'] = _text_encodings.to(torch.uint8)\n",
|
| 695 |
+
" tmp['num_items'] = torch.tensor(num_items).to(torch.uint8)\n",
|
| 696 |
+
" tmp['scale'] = torch.tensor(0.0043)\n",
|
| 697 |
+
" save_file(tmp , f'{enc_filename}')\n",
|
| 698 |
+
"#------#"
|
| 699 |
],
|
| 700 |
"metadata": {
|
| 701 |
"cellView": "form",
|
| 702 |
+
"id": "9qgHW1Wr7kZn"
|
| 703 |
},
|
| 704 |
"execution_count": null,
|
| 705 |
"outputs": []
|
|
|
|
| 707 |
{
|
| 708 |
"cell_type": "code",
|
| 709 |
"source": [
|
| 710 |
+
"# Check the average value for this set\n",
|
| 711 |
+
"sims = torch.matmul(vocab_encodings.dequantize(),average.t())\n",
|
| 712 |
+
"sorted , indices = torch.sort(sims,dim=0 , descending=True)\n",
|
| 713 |
+
"for index in range(10):\n",
|
| 714 |
+
" print(prompts[f'{indices[index].item()}'])"
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 715 |
],
|
| 716 |
"metadata": {
|
| 717 |
+
"id": "XNHz0hfhHRUu"
|
|
|
|
| 718 |
},
|
| 719 |
"execution_count": null,
|
| 720 |
"outputs": []
|