codeShare commited on
Commit
db315c0
·
verified ·
1 Parent(s): 7932cb1

Upload fusion_t2i_CLIP_interrogator.ipynb

Browse files
Google Colab Jupyter Notebooks/fusion_t2i_CLIP_interrogator.ipynb CHANGED
@@ -25,6 +25,17 @@
25
  "id": "cRV2YWomjMBU"
26
  }
27
  },
 
 
 
 
 
 
 
 
 
 
 
28
  {
29
  "cell_type": "code",
30
  "source": [
@@ -57,7 +68,18 @@
57
  " #----#\n",
58
  " %cd {home_directory}\n",
59
  " !git clone https://huggingface.co/datasets/codeShare/fusion-t2i-generator-data\n",
60
- " loaded = True\n",
 
 
 
 
 
 
 
 
 
 
 
61
  " %cd {home_directory + 'fusion-t2i-generator-data/'}\n",
62
  " !unzip vocab.zip\n",
63
  " !unzip reference.zip\n",
@@ -104,8 +126,7 @@
104
  "for key in torch.load('reference_text_and_image_encodings.pt', weights_only=False):\n",
105
  " index = index + 1;\n",
106
  "#------#\n",
107
- "NUM_REFERENCE_ITEMS = index\n",
108
- "\n"
109
  ],
110
  "metadata": {
111
  "id": "TC5lMJrS1HCC"
@@ -261,26 +282,136 @@
261
  "image or print('No image found')"
262
  ],
263
  "metadata": {
264
- "id": "NqL_I3ZSrISq"
 
265
  },
266
  "execution_count": null,
267
  "outputs": []
268
  },
 
 
 
 
 
 
 
269
  {
270
  "cell_type": "code",
271
  "source": [
272
- "# Check the average value for this set\n",
273
- "sims = torch.matmul(vocab_encodings.dequantize(),average.t())\n",
274
- "sorted , indices = torch.sort(sims,dim=0 , descending=True)\n",
275
- "for index in range(10):\n",
276
- " print(prompts[f'{indices[index].item()}'])"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
277
  ],
278
  "metadata": {
279
- "id": "XNHz0hfhHRUu"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
280
  },
281
  "execution_count": null,
282
  "outputs": []
283
  },
 
 
 
 
 
 
 
 
 
284
  {
285
  "cell_type": "code",
286
  "source": [
@@ -594,59 +725,54 @@
594
  {
595
  "cell_type": "code",
596
  "source": [
597
- "# @title \t⚄ New code (work in progress)\n",
598
- "\n",
599
- "def get_num_vocab_items(_url):\n",
600
- " num_vocab_items = 0\n",
601
- " for item in _url.split('_'):\n",
602
- " if item.find('safetensors')>-1: num_vocab_items = int(item.replace('.safetensors', ''))\n",
603
- " #------#\n",
604
- " return num_vocab_items-1\n",
605
  "\n",
 
 
 
606
  "\n",
607
- "def get_similiar(_ref , urls, _LIST_SIZE):\n",
608
- " dot_dtype = torch.float16\n",
609
- " _SCALE = torch.tensor(0.0043).to(dot_dtype)\n",
610
- " _DIM = 768\n",
611
- " _vocab = {}\n",
612
- " #----#\n",
613
- " inputs = tokenizer(text = _ref.strip(), truncation = True , padding=True, return_tensors=\"pt\")\n",
614
- " ref = model.get_text_features(**inputs)[0]\n",
615
- " ref = (ref/ref.norm(p=2, dim=-1, keepdim=True)).to(dtype = dot_dtype)\n",
616
- " #-----#\n",
617
- " num_vocab_items = 0\n",
618
- " for url in urls:\n",
619
- " num_vocab_items = num_vocab_items + get_num_vocab_items(url)\n",
620
- " #------#\n",
621
- " vocab = torch.zeros(num_vocab_items , _DIM).to(torch.uint8)\n",
622
  " prompts = {}\n",
623
- " index = 0\n",
624
- " for url in urls:\n",
625
- " __vocab = load_file(url)\n",
626
- " for key in load_file(url):\n",
627
- " vocab[index] = __vocab[key][1:_DIM+1] - __vocab[key][0]*torch.ones(_DIM).t()\n",
628
- " prompts[f'{index}'] = key\n",
629
- " index = index + 1\n",
630
- " #-------#\n",
631
- " __vocab = {}\n",
632
- " #-------#\n",
633
- " sims = torch.matmul((vocab*_SCALE).to(dot_dtype) , ref.t())\n",
634
- " sorted , indices = torch.sort(sims, dim = 0 , descending = True)\n",
635
- " return indices , prompts , sims\n",
636
- " _prompts = {}\n",
637
- " for index in range(num_vocab_items):\n",
638
- " key = prompts[f'{indices[index]}']\n",
639
- " _prompts[f'{key}'] = sims[key].item()\n",
640
- " index = index + 1\n",
641
- " if index>_LIST_SIZE:break\n",
 
 
 
 
 
 
 
642
  " #-------#\n",
643
- " return _prompts\n",
644
- "#-------#\n",
645
- "\n"
 
 
 
 
 
646
  ],
647
  "metadata": {
648
  "cellView": "form",
649
- "id": "uDzsk02CbMFc"
650
  },
651
  "execution_count": null,
652
  "outputs": []
@@ -654,31 +780,14 @@
654
  {
655
  "cell_type": "code",
656
  "source": [
657
- "vocab = {}\n",
658
- "# @title \t⚄ New code (work in progress)\n",
659
- "ref = 'impressionist painting by luis royo' # @param {type:'string' , placeholder:'type a single prompt to match'}\n",
660
- "LIST_SIZE = 1000 # @param {type:'number' , placeholder:'set how large the list should be'}\n",
661
- "urls = [ '/content/fusion-t2i-generator-data/civitai_vocab_q0043_203663.safetensors' ,]\n",
662
- "\n",
663
- " #'/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",
664
- "\n",
665
- "indices , prompts , sims = get_similiar(ref , urls , LIST_SIZE)\n",
666
- "\n",
667
- "index = 0\n",
668
- "_prompts = {}\n",
669
- "for index in range(203662):\n",
670
- " try:\n",
671
- " key = prompts[f'{indices[index].item()}']\n",
672
- " print(key)\n",
673
- " except: print('Not found!')\n",
674
- " #_prompts[f'{key}'] = sims[key].item()\n",
675
- " index = index + 1\n",
676
- " if index>LIST_SIZE:break\n",
677
- "\n"
678
  ],
679
  "metadata": {
680
- "cellView": "form",
681
- "id": "Azz1kCza6LB3"
682
  },
683
  "execution_count": null,
684
  "outputs": []
 
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": [
 
68
  " #----#\n",
69
  " %cd {home_directory}\n",
70
  " !git clone https://huggingface.co/datasets/codeShare/fusion-t2i-generator-data\n",
71
+ " loaded = True"
72
+ ],
73
+ "metadata": {
74
+ "id": "A30Xl4BswyEr"
75
+ },
76
+ "execution_count": null,
77
+ "outputs": []
78
+ },
79
+ {
80
+ "cell_type": "code",
81
+ "source": [
82
+ "\n",
83
  " %cd {home_directory + 'fusion-t2i-generator-data/'}\n",
84
  " !unzip vocab.zip\n",
85
  " !unzip reference.zip\n",
 
126
  "for key in torch.load('reference_text_and_image_encodings.pt', weights_only=False):\n",
127
  " index = index + 1;\n",
128
  "#------#\n",
129
+ "NUM_REFERENCE_ITEMS = index"
 
130
  ],
131
  "metadata": {
132
  "id": "TC5lMJrS1HCC"
 
282
  "image or print('No image found')"
283
  ],
284
  "metadata": {
285
+ "id": "NqL_I3ZSrISq",
286
+ "cellView": "form"
287
  },
288
  "execution_count": null,
289
  "outputs": []
290
  },
291
+ {
292
+ "cell_type": "markdown",
293
+ "source": [],
294
+ "metadata": {
295
+ "id": "ouE3KYiJefac"
296
+ }
297
+ },
298
  {
299
  "cell_type": "code",
300
  "source": [
301
+ "# @title New interrogator code using quantized text corpus\n",
302
+ "ref = '' # @param {type:'string' , placeholder:'type a single prompt to match'}\n",
303
+ "LIST_SIZE = 1000 # @param {type:'number' , placeholder:'set how large the list should be'}\n",
304
+ "\n",
305
+ "# @markdown Select vocab\n",
306
+ "fanfic = False # @param {type:\"boolean\"}\n",
307
+ "civitai = True # @param {type:\"boolean\"}\n",
308
+ "names = True # @param {type:\"boolean\"}\n",
309
+ "r34 = True # @param {type:\"boolean\"}\n",
310
+ "\n",
311
+ "from transformers import AutoTokenizer\n",
312
+ "tokenizer = AutoTokenizer.from_pretrained(\"openai/clip-vit-large-patch14\", clean_up_tokenization_spaces = False)\n",
313
+ "from transformers import CLIPProcessor, CLIPModel\n",
314
+ "processor = CLIPProcessor.from_pretrained(\"openai/clip-vit-large-patch14\" , clean_up_tokenization_spaces = True)\n",
315
+ "model = CLIPModel.from_pretrained(\"openai/clip-vit-large-patch14\")\n",
316
+ "logit_scale = model.logit_scale.exp() #logit_scale = 100.00000762939453\n",
317
+ "dot_dtype = torch.float32\n",
318
+ "inputs = tokenizer(text = ref.strip(), truncation = True , padding=True, return_tensors=\"pt\")\n",
319
+ "ref = model.get_text_features(**inputs)[0]\n",
320
+ "ref = (ref/ref.norm(p=2, dim=-1, keepdim=True)).to(dtype = dot_dtype)\n",
321
+ "#-----#\n",
322
+ "prompts_folder = f'{home_directory}fusion-t2i-generator-data/vocab/text'\n",
323
+ "encodings_folder = f'{home_directory}fusion-t2i-generator-data/vocab/text_encodings'\n",
324
+ "#----#\n",
325
+ "dim = 768\n",
326
+ "scale = 0.0043\n",
327
+ "size = 0\n",
328
+ "#------#\n",
329
+ "for filename in os.listdir(prompts_folder):\n",
330
+ " if (not civitai and filename.find('civitai')>-1):continue\n",
331
+ " if (not fanfic and filename.find('fanfic')>-1):continue\n",
332
+ " if (not r34 and filename.find('r34')>-1):continue\n",
333
+ " if (not names and filename.find('names')>-1):continue\n",
334
+ " size = size + LIST_SIZE\n",
335
+ "#-------#\n",
336
+ "similiar_sims = torch.zeros(size)\n",
337
+ "similiar_prompts = {}\n",
338
+ "_index = 0\n",
339
+ "#-------#\n",
340
+ "similiar_encodings = {}\n",
341
+ "for filename in os.listdir(prompts_folder):\n",
342
+ " if (not civitai and filename.find('civitai')>-1):continue\n",
343
+ " if (not fanfic and filename.find('fanfic')>-1):continue\n",
344
+ " if (not r34 and filename.find('r34')>-1):continue\n",
345
+ " if (not names and filename.find('names')>-1):continue\n",
346
+ " #------#\n",
347
+ " root_filename = filename.replace('.json', '')\n",
348
+ " %cd {prompts_folder}\n",
349
+ " prompts = {}\n",
350
+ " with open(f'{root_filename}.json', 'r') as f:\n",
351
+ " data = json.load(f).items()\n",
352
+ " for key,value in data:\n",
353
+ " prompts[key] = value\n",
354
+ " num_items = int(prompts['num_items'])\n",
355
+ " #------#\n",
356
+ " %cd {encodings_folder}\n",
357
+ " _text_encodings = load_file(f'{root_filename}.safetensors')['weights'].to(torch.uint8)\n",
358
+ "\n",
359
+ " text_encodings = torch.zeros(num_items , dim)\n",
360
+ " tmp = torch.ones(dim).to(dot_dtype)\n",
361
+ " for index in range(num_items):\n",
362
+ " text_encodings[index] = torch.sub(_text_encodings[index][1:dim+1].to(dot_dtype) , tmp , alpha= _text_encodings[index][0].to(dot_dtype))\n",
363
+ " #------#\n",
364
+ " sims = torch.matmul(text_encodings*scale, ref.t())\n",
365
+ " sorted , indices = torch.sort(sims , dim=0 , descending = True)\n",
366
+ " for index in range(LIST_SIZE):\n",
367
+ " key = indices[index].item()\n",
368
+ " prompt = prompts[f'{key}']\n",
369
+ " #-------#\n",
370
+ " similiar_sims[_index] = torch.tensor(round(sims[key].item(), 5))\n",
371
+ " similiar_prompts[f'{_index}'] = prompt\n",
372
+ " _index = _index + 1\n",
373
+ " #-------#\n",
374
+ " continue\n",
375
+ "#---------#\n"
376
  ],
377
  "metadata": {
378
+ "cellView": "form",
379
+ "id": "w2dfozFY5IwM"
380
+ },
381
+ "execution_count": null,
382
+ "outputs": []
383
+ },
384
+ {
385
+ "cell_type": "code",
386
+ "source": [
387
+ "# @title ⚄ Printing results from text corpus\n",
388
+ "sorted , indices = torch.sort(similiar_sims , dim=0 , descending = True)\n",
389
+ "\n",
390
+ "include_similiarity = False # @param {type:\"boolean\"}\n",
391
+ "for index in range(LIST_SIZE):\n",
392
+ " key = indices[index].item()\n",
393
+ " sim = similiar_sims[key].item()\n",
394
+ " prompt = similiar_prompts[f'{key}']\n",
395
+ " #-------#\n",
396
+ " if include_similiarity :print(f'{prompt} - {round(sim*100,1)} %')\n",
397
+ " else: print(f'{prompt}')"
398
+ ],
399
+ "metadata": {
400
+ "cellView": "form",
401
+ "id": "E3kfOKXITDI9"
402
  },
403
  "execution_count": null,
404
  "outputs": []
405
  },
406
+ {
407
+ "cell_type": "markdown",
408
+ "source": [
409
+ "OTHER STUFF BELOW"
410
+ ],
411
+ "metadata": {
412
+ "id": "FRIqYJDEebpf"
413
+ }
414
+ },
415
  {
416
  "cell_type": "code",
417
  "source": [
 
725
  {
726
  "cell_type": "code",
727
  "source": [
728
+ "# @title \t⚄ Quick fix for normalizing encoded text corpus tensors\n",
 
 
 
 
 
 
 
729
  "\n",
730
+ "import os\n",
731
+ "my_mkdirs('/content/output')\n",
732
+ "my_mkdirs('/content/output/text_encodings')\n",
733
  "\n",
734
+ "for filename in os.listdir(f'{prompts_folder}'):\n",
735
+ " %cd {prompts_folder}\n",
 
 
 
 
 
 
 
 
 
 
 
 
 
736
  " prompts = {}\n",
737
+ " with open(f'{filename}', 'r') as f:\n",
738
+ " data = json.load(f).items()\n",
739
+ " for key,value in data:\n",
740
+ " prompts[key] = value\n",
741
+ " #------#\n",
742
+ " num_items = int(prompts['num_items'])\n",
743
+ "\n",
744
+ " %cd {encodings_folder}\n",
745
+ " enc_filename = filename.replace('json', 'safetensors')\n",
746
+ " _text_encodings = load_file(f'{enc_filename}')['weights'].to(torch.uint8)\n",
747
+ " text_encodings = torch.zeros(num_items , dim)\n",
748
+ " tmp = torch.ones(dim)\n",
749
+ " tmp2 = torch.tensor(1/0.0043)\n",
750
+ " zero_point = 0\n",
751
+ " for index in range(num_items):\n",
752
+ " 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",
753
+ " text_encodings[index] = tmp2*text_encodings[index]/text_encodings[index].norm(p=2, dim=-1, keepdim = True)\n",
754
+ " test = torch.round( torch.add(text_encodings[index],tmp*zero_point))\n",
755
+ " less_than_zero = test<0\n",
756
+ " while(torch.any(less_than_zero).item()):\n",
757
+ " zero_point = zero_point + 1\n",
758
+ " test = torch.round( torch.add(text_encodings[index],tmp*zero_point))\n",
759
+ " less_than_zero = test<0\n",
760
+ " #------#\n",
761
+ " _text_encodings[index][0] = zero_point\n",
762
+ " _text_encodings[index][1:dim+1] = test\n",
763
  " #-------#\n",
764
+ " %cd /content/output/text_encodings\n",
765
+ "\n",
766
+ " tmp = {}\n",
767
+ " tmp['weights'] = _text_encodings.to(torch.uint8)\n",
768
+ " tmp['num_items'] = torch.tensor(num_items).to(torch.uint8)\n",
769
+ " tmp['scale'] = torch.tensor(0.0043)\n",
770
+ " save_file(tmp , f'{enc_filename}')\n",
771
+ "#------#"
772
  ],
773
  "metadata": {
774
  "cellView": "form",
775
+ "id": "9qgHW1Wr7kZn"
776
  },
777
  "execution_count": null,
778
  "outputs": []
 
780
  {
781
  "cell_type": "code",
782
  "source": [
783
+ "# Check the average value for this set\n",
784
+ "sims = torch.matmul(vocab_encodings.dequantize(),average.t())\n",
785
+ "sorted , indices = torch.sort(sims,dim=0 , descending=True)\n",
786
+ "for index in range(10):\n",
787
+ " print(prompts[f'{indices[index].item()}'])"
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
788
  ],
789
  "metadata": {
790
+ "id": "XNHz0hfhHRUu"
 
791
  },
792
  "execution_count": null,
793
  "outputs": []