lazybug commited on
Commit
3b0389d
·
verified ·
1 Parent(s): 20f93fb

Add files using upload-large-folder tool

Browse files
This view is limited to 50 files because it contains too many changes.   See raw diff
Files changed (50) hide show
  1. .gitattributes +64 -0
  2. hf_datas/llava_v1_5_mix665k.json +3 -0
  3. hf_models/clip-vit-large-patch14-336/pytorch_model.bin +3 -0
  4. hf_models/clip-vit-large-patch14-336/tf_model.h5 +3 -0
  5. hf_models/vicuna-7b-v1.5/pytorch_model-00001-of-00002.bin +3 -0
  6. hf_models/vicuna-7b-v1.5/pytorch_model-00002-of-00002.bin +3 -0
  7. llava/__pycache__/__init__.cpython-310.pyc +0 -0
  8. llava/__pycache__/constants.cpython-310.pyc +0 -0
  9. llava/__pycache__/conversation.cpython-310.pyc +0 -0
  10. llava/__pycache__/mm_utils.cpython-310.pyc +0 -0
  11. llava/__pycache__/utils.cpython-310.pyc +0 -0
  12. llava/eval/__pycache__/m4c_evaluator.cpython-310.pyc +0 -0
  13. llava/eval/__pycache__/model_vqa_loader.cpython-310.pyc +0 -0
  14. llava/eval/__pycache__/model_vqa_mmbench.cpython-310.pyc +0 -0
  15. llava/eval/__pycache__/model_vqa_science.cpython-310.pyc +0 -0
  16. llava/eval/eval_science_qa.py +114 -0
  17. llava/eval/m4c_evaluator.py +334 -0
  18. llava/eval/model_vqa_loader.py +158 -0
  19. llava/eval/model_vqa_mmbench.py +170 -0
  20. llava/eval/model_vqa_science.py +122 -0
  21. llava/train/__pycache__/llava_trainer.cpython-310.pyc +0 -0
  22. llava/train/__pycache__/train_compress.cpython-310.pyc +0 -0
  23. llava/train/llama_flash_attn_monkey_patch.py +115 -0
  24. llava/train/llama_xformers_attn_monkey_patch.py +129 -0
  25. llava/train/train_compress.py +1014 -0
  26. playground/data.z01 +3 -0
  27. playground/data.z02 +3 -0
  28. playground/data.z04 +3 -0
  29. playground/data.z05 +3 -0
  30. playground/data.z06 +3 -0
  31. playground/data.z07 +3 -0
  32. playground/data.z08 +3 -0
  33. playground/data.z09 +3 -0
  34. playground/data.z10 +3 -0
  35. playground/data.z11 +3 -0
  36. playground/data.z12 +3 -0
  37. playground/data.z14 +3 -0
  38. playground/data.z15 +3 -0
  39. playground/data.z16 +3 -0
  40. playground/data.z18 +3 -0
  41. playground/data.z19 +3 -0
  42. playground/data.z20 +3 -0
  43. playground/data.z21 +3 -0
  44. playground/data.z23 +3 -0
  45. playground/data.z24 +3 -0
  46. playground/data.z25 +3 -0
  47. playground/data.z26 +3 -0
  48. playground/data.z27 +3 -0
  49. playground/data.z28 +3 -0
  50. playground/data.z30 +3 -0
.gitattributes CHANGED
@@ -57,3 +57,67 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
57
  # Video files - compressed
58
  *.mp4 filter=lfs diff=lfs merge=lfs -text
59
  *.webm filter=lfs diff=lfs merge=lfs -text
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
57
  # Video files - compressed
58
  *.mp4 filter=lfs diff=lfs merge=lfs -text
59
  *.webm filter=lfs diff=lfs merge=lfs -text
60
+ hf_datas/llava_v1_5_mix665k.json filter=lfs diff=lfs merge=lfs -text
61
+ playground/data.z52 filter=lfs diff=lfs merge=lfs -text
62
+ playground/data.z40 filter=lfs diff=lfs merge=lfs -text
63
+ playground/data.z05 filter=lfs diff=lfs merge=lfs -text
64
+ playground/data.z74 filter=lfs diff=lfs merge=lfs -text
65
+ playground/data.z01 filter=lfs diff=lfs merge=lfs -text
66
+ playground/data.z60 filter=lfs diff=lfs merge=lfs -text
67
+ playground/data.z65 filter=lfs diff=lfs merge=lfs -text
68
+ playground/data.z37 filter=lfs diff=lfs merge=lfs -text
69
+ playground/data.z62 filter=lfs diff=lfs merge=lfs -text
70
+ playground/data.z63 filter=lfs diff=lfs merge=lfs -text
71
+ playground/data.z21 filter=lfs diff=lfs merge=lfs -text
72
+ playground/data.z53 filter=lfs diff=lfs merge=lfs -text
73
+ playground/data.z41 filter=lfs diff=lfs merge=lfs -text
74
+ playground/data.z07 filter=lfs diff=lfs merge=lfs -text
75
+ playground/data.z35 filter=lfs diff=lfs merge=lfs -text
76
+ playground/data.z06 filter=lfs diff=lfs merge=lfs -text
77
+ playground/data.z50 filter=lfs diff=lfs merge=lfs -text
78
+ playground/data.z12 filter=lfs diff=lfs merge=lfs -text
79
+ playground/data.z15 filter=lfs diff=lfs merge=lfs -text
80
+ playground/data.z18 filter=lfs diff=lfs merge=lfs -text
81
+ playground/data.z23 filter=lfs diff=lfs merge=lfs -text
82
+ playground/data.z09 filter=lfs diff=lfs merge=lfs -text
83
+ playground/data.z11 filter=lfs diff=lfs merge=lfs -text
84
+ playground/data.z38 filter=lfs diff=lfs merge=lfs -text
85
+ playground/data.z33 filter=lfs diff=lfs merge=lfs -text
86
+ playground/data.z70 filter=lfs diff=lfs merge=lfs -text
87
+ playground/data.z02 filter=lfs diff=lfs merge=lfs -text
88
+ playground/data.z51 filter=lfs diff=lfs merge=lfs -text
89
+ playground/data.z73 filter=lfs diff=lfs merge=lfs -text
90
+ playground/data.z42 filter=lfs diff=lfs merge=lfs -text
91
+ playground/data.z30 filter=lfs diff=lfs merge=lfs -text
92
+ playground/data.z64 filter=lfs diff=lfs merge=lfs -text
93
+ playground/data.z46 filter=lfs diff=lfs merge=lfs -text
94
+ playground/data.z27 filter=lfs diff=lfs merge=lfs -text
95
+ playground/data.z19 filter=lfs diff=lfs merge=lfs -text
96
+ playground/data.z04 filter=lfs diff=lfs merge=lfs -text
97
+ playground/data.z28 filter=lfs diff=lfs merge=lfs -text
98
+ playground/data.z48 filter=lfs diff=lfs merge=lfs -text
99
+ playground/data.z16 filter=lfs diff=lfs merge=lfs -text
100
+ playground/data.z71 filter=lfs diff=lfs merge=lfs -text
101
+ playground/data.z24 filter=lfs diff=lfs merge=lfs -text
102
+ playground/data.z69 filter=lfs diff=lfs merge=lfs -text
103
+ playground/data.z61 filter=lfs diff=lfs merge=lfs -text
104
+ playground/data.z68 filter=lfs diff=lfs merge=lfs -text
105
+ playground/data.z14 filter=lfs diff=lfs merge=lfs -text
106
+ playground/data.z58 filter=lfs diff=lfs merge=lfs -text
107
+ playground/data.z75 filter=lfs diff=lfs merge=lfs -text
108
+ playground/data.z44 filter=lfs diff=lfs merge=lfs -text
109
+ playground/data.z49 filter=lfs diff=lfs merge=lfs -text
110
+ playground/data.z10 filter=lfs diff=lfs merge=lfs -text
111
+ playground/data.z43 filter=lfs diff=lfs merge=lfs -text
112
+ playground/data.z39 filter=lfs diff=lfs merge=lfs -text
113
+ playground/data.z26 filter=lfs diff=lfs merge=lfs -text
114
+ playground/data.z55 filter=lfs diff=lfs merge=lfs -text
115
+ playground/data.z36 filter=lfs diff=lfs merge=lfs -text
116
+ playground/data.z08 filter=lfs diff=lfs merge=lfs -text
117
+ playground/data.z20 filter=lfs diff=lfs merge=lfs -text
118
+ playground/data.z67 filter=lfs diff=lfs merge=lfs -text
119
+ playground/data.z25 filter=lfs diff=lfs merge=lfs -text
120
+ playground/data.z31 filter=lfs diff=lfs merge=lfs -text
121
+ playground/data.z45 filter=lfs diff=lfs merge=lfs -text
122
+ playground/data.z47 filter=lfs diff=lfs merge=lfs -text
123
+ playground/data.z57 filter=lfs diff=lfs merge=lfs -text
hf_datas/llava_v1_5_mix665k.json ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:ce959ce6e23073ee1cd1a8a2ef1c633768c10d4174327b8b2dc7113b91af6cf8
3
+ size 1029887963
hf_models/clip-vit-large-patch14-336/pytorch_model.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:c6032c2e0caae3dc2d4fba35535fa6307dbb49df59c7e182b1bc4b3329b81801
3
+ size 1711974081
hf_models/clip-vit-large-patch14-336/tf_model.h5 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d12828ca8f0f3c92194f277b7d893da7f2fb7824d0b99dedb305eb48eb46bb7f
3
+ size 1712454232
hf_models/vicuna-7b-v1.5/pytorch_model-00001-of-00002.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:4133d2fcc5f31286881ea50806d95b721d016b533036a99dedce3f8fe88520e6
3
+ size 9976634558
hf_models/vicuna-7b-v1.5/pytorch_model-00002-of-00002.bin ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d261d3c35e92d3070d1e61ed821ebfca812a847d2a880757d82728acf005c5ac
3
+ size 3500315539
llava/__pycache__/__init__.cpython-310.pyc ADDED
Binary file (199 Bytes). View file
 
llava/__pycache__/constants.cpython-310.pyc ADDED
Binary file (507 Bytes). View file
 
llava/__pycache__/conversation.cpython-310.pyc ADDED
Binary file (12.1 kB). View file
 
llava/__pycache__/mm_utils.cpython-310.pyc ADDED
Binary file (8.78 kB). View file
 
llava/__pycache__/utils.cpython-310.pyc ADDED
Binary file (4.04 kB). View file
 
llava/eval/__pycache__/m4c_evaluator.cpython-310.pyc ADDED
Binary file (10.1 kB). View file
 
llava/eval/__pycache__/model_vqa_loader.cpython-310.pyc ADDED
Binary file (5.96 kB). View file
 
llava/eval/__pycache__/model_vqa_mmbench.cpython-310.pyc ADDED
Binary file (5.45 kB). View file
 
llava/eval/__pycache__/model_vqa_science.cpython-310.pyc ADDED
Binary file (4.25 kB). View file
 
llava/eval/eval_science_qa.py ADDED
@@ -0,0 +1,114 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import json
3
+ import os
4
+ import re
5
+ import random
6
+
7
+
8
+ def get_args():
9
+ parser = argparse.ArgumentParser()
10
+ parser.add_argument('--base-dir', type=str)
11
+ parser.add_argument('--result-file', type=str)
12
+ parser.add_argument('--output-file', type=str)
13
+ parser.add_argument('--output-result', type=str)
14
+ parser.add_argument('--split', type=str, default='test')
15
+ parser.add_argument('--options', type=list, default=["A", "B", "C", "D", "E"])
16
+ return parser.parse_args()
17
+
18
+
19
+ def convert_caps(results):
20
+ fakecaps = []
21
+ for result in results:
22
+ image_id = result['question_id']
23
+ caption = result['text']
24
+ fakecaps.append({"image_id": int(image_id), "caption": caption})
25
+ return fakecaps
26
+
27
+
28
+ def get_pred_idx(prediction, choices, options):
29
+ """
30
+ Get the index (e.g. 2) from the prediction (e.g. 'C')
31
+ """
32
+ if prediction in options[:len(choices)]:
33
+ return options.index(prediction)
34
+ else:
35
+ return -1
36
+ return random.choice(range(len(choices)))
37
+
38
+
39
+ if __name__ == "__main__":
40
+ args = get_args()
41
+
42
+ base_dir = args.base_dir
43
+ split_indices = json.load(open(os.path.join(base_dir, "pid_splits.json")))[args.split]
44
+ problems = json.load(open(os.path.join(base_dir, "problems.json")))
45
+ predictions = [json.loads(line) for line in open(args.result_file)]
46
+ predictions = {pred['question_id']: pred for pred in predictions}
47
+ split_problems = {idx: problems[idx] for idx in split_indices}
48
+
49
+ results = {'correct': [], 'incorrect': []}
50
+ sqa_results = {}
51
+ sqa_results['acc'] = None
52
+ sqa_results['correct'] = None
53
+ sqa_results['count'] = None
54
+ sqa_results['results'] = {}
55
+ sqa_results['outputs'] = {}
56
+
57
+ for prob_id, prob in split_problems.items():
58
+ if prob_id not in predictions:
59
+ pred = {'text': 'FAILED', 'prompt': 'Unknown'}
60
+ pred_text = 'FAILED'
61
+ else:
62
+ pred = predictions[prob_id]
63
+ pred_text = pred['text']
64
+
65
+ if pred_text in args.options:
66
+ answer = pred_text
67
+ elif len(pred_text) >= 3 and pred_text[0] in args.options and pred_text[1:3] == ". ":
68
+ answer = pred_text[0]
69
+ else:
70
+ pattern = re.compile(r'The answer is ([A-Z]).')
71
+ res = pattern.findall(pred_text)
72
+ if len(res) == 1:
73
+ answer = res[0] # 'A', 'B', ...
74
+ else:
75
+ answer = "FAILED"
76
+
77
+ pred_idx = get_pred_idx(answer, prob['choices'], args.options)
78
+
79
+ analysis = {
80
+ 'question_id': prob_id,
81
+ 'parsed_ans': answer,
82
+ 'ground_truth': args.options[prob['answer']],
83
+ 'question': pred['prompt'],
84
+ 'pred': pred_text,
85
+ 'is_multimodal': '<image>' in pred['prompt'],
86
+ }
87
+
88
+ sqa_results['results'][prob_id] = get_pred_idx(answer, prob['choices'], args.options)
89
+ sqa_results['outputs'][prob_id] = pred_text
90
+
91
+ if pred_idx == prob['answer']:
92
+ results['correct'].append(analysis)
93
+ else:
94
+ results['incorrect'].append(analysis)
95
+
96
+ correct = len(results['correct'])
97
+ total = len(results['correct']) + len(results['incorrect'])
98
+
99
+ ###### IMG ######
100
+ multimodal_correct = len([x for x in results['correct'] if x['is_multimodal']])
101
+ multimodal_incorrect = len([x for x in results['incorrect'] if x['is_multimodal']])
102
+ multimodal_total = multimodal_correct + multimodal_incorrect
103
+ ###### IMG ######
104
+
105
+ print(f'Total: {total}, Correct: {correct}, Accuracy: {correct / total * 100:.2f}%, IMG-Accuracy: {multimodal_correct / multimodal_total * 100:.2f}%')
106
+
107
+ sqa_results['acc'] = correct / total * 100
108
+ sqa_results['correct'] = correct
109
+ sqa_results['count'] = total
110
+
111
+ with open(args.output_file, 'w') as f:
112
+ json.dump(results, f, indent=2)
113
+ with open(args.output_result, 'w') as f:
114
+ json.dump(sqa_results, f, indent=2)
llava/eval/m4c_evaluator.py ADDED
@@ -0,0 +1,334 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Copyright (c) Facebook, Inc. and its affiliates.
2
+ import re
3
+
4
+ from tqdm import tqdm
5
+
6
+
7
+ class EvalAIAnswerProcessor:
8
+ """
9
+ Processes an answer similar to Eval AI
10
+ copied from
11
+ https://github.com/facebookresearch/mmf/blob/c46b3b3391275b4181567db80943473a89ab98ab/pythia/tasks/processors.py#L897
12
+ """
13
+
14
+ CONTRACTIONS = {
15
+ "aint": "ain't",
16
+ "arent": "aren't",
17
+ "cant": "can't",
18
+ "couldve": "could've",
19
+ "couldnt": "couldn't",
20
+ "couldn'tve": "couldn't've",
21
+ "couldnt've": "couldn't've",
22
+ "didnt": "didn't",
23
+ "doesnt": "doesn't",
24
+ "dont": "don't",
25
+ "hadnt": "hadn't",
26
+ "hadnt've": "hadn't've",
27
+ "hadn'tve": "hadn't've",
28
+ "hasnt": "hasn't",
29
+ "havent": "haven't",
30
+ "hed": "he'd",
31
+ "hed've": "he'd've",
32
+ "he'dve": "he'd've",
33
+ "hes": "he's",
34
+ "howd": "how'd",
35
+ "howll": "how'll",
36
+ "hows": "how's",
37
+ "Id've": "I'd've",
38
+ "I'dve": "I'd've",
39
+ "Im": "I'm",
40
+ "Ive": "I've",
41
+ "isnt": "isn't",
42
+ "itd": "it'd",
43
+ "itd've": "it'd've",
44
+ "it'dve": "it'd've",
45
+ "itll": "it'll",
46
+ "let's": "let's",
47
+ "maam": "ma'am",
48
+ "mightnt": "mightn't",
49
+ "mightnt've": "mightn't've",
50
+ "mightn'tve": "mightn't've",
51
+ "mightve": "might've",
52
+ "mustnt": "mustn't",
53
+ "mustve": "must've",
54
+ "neednt": "needn't",
55
+ "notve": "not've",
56
+ "oclock": "o'clock",
57
+ "oughtnt": "oughtn't",
58
+ "ow's'at": "'ow's'at",
59
+ "'ows'at": "'ow's'at",
60
+ "'ow'sat": "'ow's'at",
61
+ "shant": "shan't",
62
+ "shed've": "she'd've",
63
+ "she'dve": "she'd've",
64
+ "she's": "she's",
65
+ "shouldve": "should've",
66
+ "shouldnt": "shouldn't",
67
+ "shouldnt've": "shouldn't've",
68
+ "shouldn'tve": "shouldn't've",
69
+ "somebody'd": "somebodyd",
70
+ "somebodyd've": "somebody'd've",
71
+ "somebody'dve": "somebody'd've",
72
+ "somebodyll": "somebody'll",
73
+ "somebodys": "somebody's",
74
+ "someoned": "someone'd",
75
+ "someoned've": "someone'd've",
76
+ "someone'dve": "someone'd've",
77
+ "someonell": "someone'll",
78
+ "someones": "someone's",
79
+ "somethingd": "something'd",
80
+ "somethingd've": "something'd've",
81
+ "something'dve": "something'd've",
82
+ "somethingll": "something'll",
83
+ "thats": "that's",
84
+ "thered": "there'd",
85
+ "thered've": "there'd've",
86
+ "there'dve": "there'd've",
87
+ "therere": "there're",
88
+ "theres": "there's",
89
+ "theyd": "they'd",
90
+ "theyd've": "they'd've",
91
+ "they'dve": "they'd've",
92
+ "theyll": "they'll",
93
+ "theyre": "they're",
94
+ "theyve": "they've",
95
+ "twas": "'twas",
96
+ "wasnt": "wasn't",
97
+ "wed've": "we'd've",
98
+ "we'dve": "we'd've",
99
+ "weve": "we've",
100
+ "werent": "weren't",
101
+ "whatll": "what'll",
102
+ "whatre": "what're",
103
+ "whats": "what's",
104
+ "whatve": "what've",
105
+ "whens": "when's",
106
+ "whered": "where'd",
107
+ "wheres": "where's",
108
+ "whereve": "where've",
109
+ "whod": "who'd",
110
+ "whod've": "who'd've",
111
+ "who'dve": "who'd've",
112
+ "wholl": "who'll",
113
+ "whos": "who's",
114
+ "whove": "who've",
115
+ "whyll": "why'll",
116
+ "whyre": "why're",
117
+ "whys": "why's",
118
+ "wont": "won't",
119
+ "wouldve": "would've",
120
+ "wouldnt": "wouldn't",
121
+ "wouldnt've": "wouldn't've",
122
+ "wouldn'tve": "wouldn't've",
123
+ "yall": "y'all",
124
+ "yall'll": "y'all'll",
125
+ "y'allll": "y'all'll",
126
+ "yall'd've": "y'all'd've",
127
+ "y'alld've": "y'all'd've",
128
+ "y'all'dve": "y'all'd've",
129
+ "youd": "you'd",
130
+ "youd've": "you'd've",
131
+ "you'dve": "you'd've",
132
+ "youll": "you'll",
133
+ "youre": "you're",
134
+ "youve": "you've",
135
+ }
136
+
137
+ NUMBER_MAP = {
138
+ "none": "0",
139
+ "zero": "0",
140
+ "one": "1",
141
+ "two": "2",
142
+ "three": "3",
143
+ "four": "4",
144
+ "five": "5",
145
+ "six": "6",
146
+ "seven": "7",
147
+ "eight": "8",
148
+ "nine": "9",
149
+ "ten": "10",
150
+ }
151
+ ARTICLES = ["a", "an", "the"]
152
+ PERIOD_STRIP = re.compile(r"(?!<=\d)(\.)(?!\d)")
153
+ COMMA_STRIP = re.compile(r"(?<=\d)(\,)+(?=\d)")
154
+ PUNCTUATIONS = [
155
+ ";",
156
+ r"/",
157
+ "[",
158
+ "]",
159
+ '"',
160
+ "{",
161
+ "}",
162
+ "(",
163
+ ")",
164
+ "=",
165
+ "+",
166
+ "\\",
167
+ "_",
168
+ "-",
169
+ ">",
170
+ "<",
171
+ "@",
172
+ "`",
173
+ ",",
174
+ "?",
175
+ "!",
176
+ ]
177
+
178
+ def __init__(self, *args, **kwargs):
179
+ pass
180
+
181
+ def word_tokenize(self, word):
182
+ word = word.lower()
183
+ word = word.replace(",", "").replace("?", "").replace("'s", " 's")
184
+ return word.strip()
185
+
186
+ def process_punctuation(self, in_text):
187
+ out_text = in_text
188
+ for p in self.PUNCTUATIONS:
189
+ if (p + " " in in_text or " " + p in in_text) or (
190
+ re.search(self.COMMA_STRIP, in_text) is not None
191
+ ):
192
+ out_text = out_text.replace(p, "")
193
+ else:
194
+ out_text = out_text.replace(p, " ")
195
+ out_text = self.PERIOD_STRIP.sub("", out_text, re.UNICODE)
196
+ return out_text
197
+
198
+ def process_digit_article(self, in_text):
199
+ out_text = []
200
+ temp_text = in_text.lower().split()
201
+ for word in temp_text:
202
+ word = self.NUMBER_MAP.setdefault(word, word)
203
+ if word not in self.ARTICLES:
204
+ out_text.append(word)
205
+ else:
206
+ pass
207
+ for word_id, word in enumerate(out_text):
208
+ if word in self.CONTRACTIONS:
209
+ out_text[word_id] = self.CONTRACTIONS[word]
210
+ out_text = " ".join(out_text)
211
+ return out_text
212
+
213
+ def __call__(self, item):
214
+ item = self.word_tokenize(item)
215
+ item = item.replace("\n", " ").replace("\t", " ").strip()
216
+ item = self.process_punctuation(item)
217
+ item = self.process_digit_article(item)
218
+ return item
219
+
220
+
221
+ class TextVQAAccuracyEvaluator:
222
+ def __init__(self):
223
+ self.answer_processor = EvalAIAnswerProcessor()
224
+
225
+ def _compute_answer_scores(self, raw_answers):
226
+ """
227
+ compute the accuracy (soft score) of human answers
228
+ """
229
+ answers = [self.answer_processor(a) for a in raw_answers]
230
+ assert len(answers) == 10
231
+ gt_answers = list(enumerate(answers))
232
+ unique_answers = set(answers)
233
+ unique_answer_scores = {}
234
+
235
+ for unique_answer in unique_answers:
236
+ accs = []
237
+ for gt_answer in gt_answers:
238
+ other_answers = [item for item in gt_answers if item != gt_answer]
239
+ matching_answers = [
240
+ item for item in other_answers if item[1] == unique_answer
241
+ ]
242
+ acc = min(1, float(len(matching_answers)) / 3)
243
+ accs.append(acc)
244
+ unique_answer_scores[unique_answer] = sum(accs) / len(accs)
245
+
246
+ return unique_answer_scores
247
+
248
+ def eval_pred_list(self, pred_list):
249
+ pred_scores = []
250
+ for entry in tqdm(pred_list):
251
+ pred_answer = self.answer_processor(entry["pred_answer"])
252
+ unique_answer_scores = self._compute_answer_scores(entry["gt_answers"])
253
+ score = unique_answer_scores.get(pred_answer, 0.0)
254
+ pred_scores.append(score)
255
+
256
+ accuracy = sum(pred_scores) / len(pred_scores)
257
+ return accuracy
258
+
259
+
260
+ class STVQAAccuracyEvaluator:
261
+ def __init__(self):
262
+ self.answer_processor = EvalAIAnswerProcessor()
263
+
264
+ def eval_pred_list(self, pred_list):
265
+ pred_scores = []
266
+ for entry in pred_list:
267
+ pred_answer = self.answer_processor(entry["pred_answer"])
268
+ gts = [self.answer_processor(a) for a in entry["gt_answers"]]
269
+ score = 1.0 if pred_answer in gts else 0.0
270
+ pred_scores.append(score)
271
+
272
+ accuracy = sum(pred_scores) / len(pred_scores)
273
+ return accuracy
274
+
275
+
276
+ class STVQAANLSEvaluator:
277
+ def __init__(self):
278
+ import editdistance # install with `pip install editdistance`
279
+
280
+ self.get_edit_distance = editdistance.eval
281
+
282
+ def get_anls(self, s1, s2):
283
+ s1 = s1.lower().strip()
284
+ s2 = s2.lower().strip()
285
+ iou = 1 - self.get_edit_distance(s1, s2) / max(len(s1), len(s2))
286
+ anls = iou if iou >= 0.5 else 0.0
287
+ return anls
288
+
289
+ def eval_pred_list(self, pred_list):
290
+ pred_scores = []
291
+ for entry in pred_list:
292
+ anls = max(
293
+ self.get_anls(entry["pred_answer"], gt) for gt in entry["gt_answers"]
294
+ )
295
+ pred_scores.append(anls)
296
+
297
+ accuracy = sum(pred_scores) / len(pred_scores)
298
+ return accuracy
299
+
300
+
301
+ class TextCapsBleu4Evaluator:
302
+ def __init__(self):
303
+ # The following script requires Java 1.8.0 and pycocotools installed.
304
+ # The pycocoevalcap can be installed with pip as
305
+ # pip install git+https://github.com/ronghanghu/coco-caption.git@python23
306
+ # Original pycocoevalcap code is at https://github.com/tylin/coco-caption
307
+ # but has no python3 support yet.
308
+ try:
309
+ from pycocoevalcap.bleu.bleu import Bleu
310
+ from pycocoevalcap.tokenizer.ptbtokenizer import PTBTokenizer
311
+ except ModuleNotFoundError:
312
+ print(
313
+ "Please install pycocoevalcap module using "
314
+ "pip install git+https://github.com/ronghanghu/coco-caption.git@python23" # noqa
315
+ )
316
+ raise
317
+
318
+ self.tokenizer = PTBTokenizer()
319
+ self.scorer = Bleu(4)
320
+
321
+ def eval_pred_list(self, pred_list):
322
+ # Create reference and hypotheses captions.
323
+ gts = {}
324
+ res = {}
325
+ for idx, entry in enumerate(pred_list):
326
+ gts[idx] = [{"caption": a} for a in entry["gt_answers"]]
327
+ res[idx] = [{"caption": entry["pred_answer"]}]
328
+
329
+ gts = self.tokenizer.tokenize(gts)
330
+ res = self.tokenizer.tokenize(res)
331
+ score, _ = self.scorer.compute_score(gts, res)
332
+
333
+ bleu4 = score[3] # score is (Bleu-1, Bleu-2, Bleu-3, Bleu-4)
334
+ return bleu4
llava/eval/model_vqa_loader.py ADDED
@@ -0,0 +1,158 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import torch
3
+ import os
4
+ import json
5
+ from tqdm import tqdm
6
+ import shortuuid
7
+
8
+ from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
9
+ from llava.conversation import conv_templates, SeparatorStyle
10
+ from llava import conversation as conversation_lib
11
+ from llava.model.builder import load_pretrained_model
12
+ from llava.utils import disable_torch_init
13
+ from llava.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
14
+ from torch.utils.data import Dataset, DataLoader
15
+
16
+ from PIL import Image
17
+ import math
18
+
19
+
20
+ def split_list(lst, n):
21
+ """Split a list into n (roughly) equal-sized chunks"""
22
+ chunk_size = math.ceil(len(lst) / n) # integer division
23
+ return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
24
+
25
+
26
+ def get_chunk(lst, n, k):
27
+ chunks = split_list(lst, n)
28
+ return chunks[k]
29
+
30
+
31
+ # Custom dataset class
32
+ class CustomDataset(Dataset):
33
+ def __init__(self, questions, image_folder, tokenizer, image_processor, model_config, voco_num):
34
+ self.questions = questions
35
+ self.image_folder = image_folder
36
+ self.tokenizer = tokenizer
37
+ self.image_processor = image_processor
38
+ self.model_config = model_config
39
+ self.voco_num = voco_num
40
+ print("voco_num is ", voco_num)
41
+
42
+ def __getitem__(self, index):
43
+ line = self.questions[index]
44
+ image_file = line["image"]
45
+ qs = line["text"]
46
+ if self.model_config.mm_use_im_start_end:
47
+ qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
48
+ else:
49
+ qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
50
+
51
+ # conv = conv_templates[args.conv_mode].copy()
52
+ conv = conversation_lib.voco_default_conversation.copy()
53
+ conv.append_message(conv.roles[0], qs)
54
+ conv.append_message(conv.roles[1], None)
55
+ prompt = conv.get_prompt()
56
+
57
+ # print(prompt)
58
+ maybe_voco_str = "".join(
59
+ ["<voco>" for _ in range(self.voco_num)]
60
+ )
61
+ prompt = f"<image>\n{maybe_voco_str}\n" + prompt.replace("\n", '').replace("<image>", '')
62
+
63
+ image = Image.open(os.path.join(self.image_folder, image_file)).convert('RGB')
64
+ image_tensor = process_images([image], self.image_processor, self.model_config)[0]
65
+
66
+ input_ids = tokenizer_image_token(prompt, self.tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt')
67
+
68
+ return input_ids, image_tensor, image.size
69
+
70
+ def __len__(self):
71
+ return len(self.questions)
72
+
73
+
74
+ def collate_fn(batch):
75
+ input_ids, image_tensors, image_sizes = zip(*batch)
76
+ input_ids = torch.stack(input_ids, dim=0)
77
+ image_tensors = torch.stack(image_tensors, dim=0)
78
+ return input_ids, image_tensors, image_sizes
79
+
80
+
81
+ # DataLoader
82
+ def create_data_loader(questions, image_folder, tokenizer, image_processor, model_config, voco_num, batch_size=1, num_workers=4):
83
+ assert batch_size == 1, "batch_size must be 1"
84
+ dataset = CustomDataset(questions, image_folder, tokenizer, image_processor, model_config, voco_num)
85
+ data_loader = DataLoader(dataset, batch_size=batch_size, num_workers=num_workers, shuffle=False, collate_fn=collate_fn)
86
+ return data_loader
87
+
88
+
89
+ def eval_model(args):
90
+ # Model
91
+ disable_torch_init()
92
+ model_path = os.path.expanduser(args.model_path)
93
+ model_name = get_model_name_from_path(model_path)
94
+ tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name, llava_model="initial")
95
+
96
+ print("*************", len(tokenizer))
97
+
98
+ questions = [json.loads(q) for q in open(os.path.expanduser(args.question_file), "r")]
99
+ questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
100
+ answers_file = os.path.expanduser(args.answers_file)
101
+ os.makedirs(os.path.dirname(answers_file), exist_ok=True)
102
+ ans_file = open(answers_file, "w")
103
+ voco_num = args.voco_num
104
+
105
+ if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:
106
+ args.conv_mode = args.conv_mode + '_mmtag'
107
+ print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')
108
+
109
+ data_loader = create_data_loader(questions, args.image_folder, tokenizer, image_processor, model.config, voco_num)
110
+
111
+ for (input_ids, image_tensor, image_sizes), line in tqdm(zip(data_loader, questions), total=len(questions)):
112
+ idx = line["question_id"]
113
+ cur_prompt = line["text"]
114
+
115
+ input_ids = input_ids.to(device='cuda', non_blocking=True)
116
+
117
+ with torch.inference_mode():
118
+ output_ids = model.generate(
119
+ input_ids,
120
+ images=image_tensor.to(dtype=torch.float16, device='cuda', non_blocking=True),
121
+ image_sizes=image_sizes,
122
+ do_sample=True if args.temperature > 0 else False,
123
+ temperature=args.temperature,
124
+ top_p=args.top_p,
125
+ num_beams=args.num_beams,
126
+ max_new_tokens=args.max_new_tokens,
127
+ use_cache=True)
128
+
129
+ outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
130
+
131
+ ans_id = shortuuid.uuid()
132
+ ans_file.write(json.dumps({"question_id": idx,
133
+ "prompt": cur_prompt,
134
+ "text": outputs,
135
+ "answer_id": ans_id,
136
+ "model_id": model_name,
137
+ "metadata": {}}) + "\n")
138
+ ans_file.flush()
139
+ ans_file.close()
140
+
141
+ if __name__ == "__main__":
142
+ parser = argparse.ArgumentParser()
143
+ parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
144
+ parser.add_argument("--model-base", type=str, default=None)
145
+ parser.add_argument("--image-folder", type=str, default="")
146
+ parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
147
+ parser.add_argument("--answers-file", type=str, default="answer.jsonl")
148
+ parser.add_argument("--conv-mode", type=str, default="llava_v1")
149
+ parser.add_argument("--num-chunks", type=int, default=1)
150
+ parser.add_argument("--chunk-idx", type=int, default=0)
151
+ parser.add_argument("--temperature", type=float, default=0.2)
152
+ parser.add_argument("--top_p", type=float, default=None)
153
+ parser.add_argument("--num_beams", type=int, default=1)
154
+ parser.add_argument("--max_new_tokens", type=int, default=128)
155
+ parser.add_argument("--voco_num", type=int, default=None)
156
+ args = parser.parse_args()
157
+
158
+ eval_model(args)
llava/eval/model_vqa_mmbench.py ADDED
@@ -0,0 +1,170 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import torch
3
+ import os
4
+ import json
5
+ import pandas as pd
6
+ from tqdm import tqdm
7
+ import shortuuid
8
+
9
+ from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
10
+ from llava.conversation import conv_templates, SeparatorStyle
11
+ from llava.model.builder import load_pretrained_model
12
+ from llava.utils import disable_torch_init
13
+ from llava.mm_utils import tokenizer_image_token, process_images, load_image_from_base64, get_model_name_from_path
14
+
15
+ from llava import conversation as conversation_lib
16
+ from PIL import Image
17
+ import math
18
+
19
+
20
+ all_options = ['A', 'B', 'C', 'D']
21
+
22
+
23
+ def split_list(lst, n):
24
+ """Split a list into n (roughly) equal-sized chunks"""
25
+ chunk_size = math.ceil(len(lst) / n) # integer division
26
+ return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
27
+
28
+
29
+ def get_chunk(lst, n, k):
30
+ chunks = split_list(lst, n)
31
+ return chunks[k]
32
+
33
+
34
+ def is_none(value):
35
+ if value is None:
36
+ return True
37
+ if type(value) is float and math.isnan(value):
38
+ return True
39
+ if type(value) is str and value.lower() == 'nan':
40
+ return True
41
+ if type(value) is str and value.lower() == 'none':
42
+ return True
43
+ return False
44
+
45
+ def get_options(row, options):
46
+ parsed_options = []
47
+ for option in options:
48
+ option_value = row[option]
49
+ if is_none(option_value):
50
+ break
51
+ parsed_options.append(option_value)
52
+ return parsed_options
53
+
54
+
55
+ def eval_model(args):
56
+ # Model
57
+ disable_torch_init()
58
+ model_path = os.path.expanduser(args.model_path)
59
+ model_name = get_model_name_from_path(model_path)
60
+ tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name, llava_model="initial")
61
+
62
+ print("*************", len(tokenizer))
63
+
64
+ questions = pd.read_table(os.path.expanduser(args.question_file))
65
+ questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
66
+ answers_file = os.path.expanduser(args.answers_file)
67
+ os.makedirs(os.path.dirname(answers_file), exist_ok=True)
68
+ ans_file = open(answers_file, "w")
69
+
70
+ if 'plain' in model_name and 'finetune' not in model_name.lower() and 'mmtag' not in args.conv_mode:
71
+ args.conv_mode = args.conv_mode + '_mmtag'
72
+ print(f'It seems that this is a plain model, but it is not using a mmtag prompt, auto switching to {args.conv_mode}.')
73
+
74
+ for index, row in tqdm(questions.iterrows(), total=len(questions)):
75
+ options = get_options(row, all_options)
76
+ cur_option_char = all_options[:len(options)]
77
+
78
+ if args.all_rounds:
79
+ num_rounds = len(options)
80
+ else:
81
+ num_rounds = 1
82
+
83
+ for round_idx in range(num_rounds):
84
+ idx = row['index']
85
+ question = row['question']
86
+ hint = row['hint']
87
+ image = load_image_from_base64(row['image'])
88
+ if not is_none(hint):
89
+ question = hint + '\n' + question
90
+ for option_char, option in zip(all_options[:len(options)], options):
91
+ question = question + '\n' + option_char + '. ' + option
92
+ qs = cur_prompt = question
93
+ if model.config.mm_use_im_start_end:
94
+ qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
95
+ else:
96
+ qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
97
+
98
+ if args.single_pred_prompt:
99
+ if args.lang == 'cn':
100
+ qs = qs + '\n' + "请直接回答选项字母。"
101
+ else:
102
+ qs = qs + '\n' + "Answer with the option's letter from the given choices directly."
103
+
104
+ # conv = conv_templates[args.conv_mode].copy()
105
+ conv = conversation_lib.voco_default_conversation.copy()
106
+ conv.append_message(conv.roles[0], qs)
107
+ conv.append_message(conv.roles[1], None)
108
+ prompt = conv.get_prompt()
109
+
110
+ maybe_voco_str = "".join(
111
+ ["<voco>" for _ in range(args.voco_num)]
112
+ )
113
+ prompt = f"<image>\n{maybe_voco_str}\n" + prompt.replace("\n", '').replace("<image>", '')
114
+
115
+ input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
116
+
117
+ image_tensor = process_images([image], image_processor, model.config)[0]
118
+
119
+ with torch.inference_mode():
120
+ output_ids = model.generate(
121
+ input_ids,
122
+ images=image_tensor.unsqueeze(0).half().cuda(),
123
+ image_sizes=[image.size],
124
+ do_sample=True if args.temperature > 0 else False,
125
+ temperature=args.temperature,
126
+ top_p=args.top_p,
127
+ num_beams=args.num_beams,
128
+ # no_repeat_ngram_size=3,
129
+ max_new_tokens=1024,
130
+ use_cache=True)
131
+
132
+ outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
133
+
134
+ ans_id = shortuuid.uuid()
135
+ ans_file.write(json.dumps({"question_id": idx,
136
+ "round_id": round_idx,
137
+ "prompt": cur_prompt,
138
+ "text": outputs,
139
+ "options": options,
140
+ "option_char": cur_option_char,
141
+ "answer_id": ans_id,
142
+ "model_id": model_name,
143
+ "metadata": {}}) + "\n")
144
+ ans_file.flush()
145
+
146
+ # rotate options
147
+ options = options[1:] + options[:1]
148
+ cur_option_char = cur_option_char[1:] + cur_option_char[:1]
149
+ ans_file.close()
150
+
151
+ if __name__ == "__main__":
152
+ parser = argparse.ArgumentParser()
153
+ parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
154
+ parser.add_argument("--model-base", type=str, default=None)
155
+ parser.add_argument("--image-folder", type=str, default="")
156
+ parser.add_argument("--question-file", type=str, default="tables/question.jsonl")
157
+ parser.add_argument("--answers-file", type=str, default="answer.jsonl")
158
+ parser.add_argument("--conv-mode", type=str, default="llava_v1")
159
+ parser.add_argument("--num-chunks", type=int, default=1)
160
+ parser.add_argument("--chunk-idx", type=int, default=0)
161
+ parser.add_argument("--temperature", type=float, default=0.2)
162
+ parser.add_argument("--top_p", type=float, default=None)
163
+ parser.add_argument("--num_beams", type=int, default=1)
164
+ parser.add_argument("--all-rounds", action="store_true")
165
+ parser.add_argument("--single-pred-prompt", action="store_true")
166
+ parser.add_argument("--lang", type=str, default="en")
167
+ parser.add_argument("--voco_num", type=int, default=None)
168
+ args = parser.parse_args()
169
+
170
+ eval_model(args)
llava/eval/model_vqa_science.py ADDED
@@ -0,0 +1,122 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import argparse
2
+ import torch
3
+ import os
4
+ import json
5
+ from tqdm import tqdm
6
+ import shortuuid
7
+
8
+ from llava.constants import IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
9
+ from llava.conversation import conv_templates, SeparatorStyle
10
+ from llava.model.builder import load_pretrained_model
11
+ from llava import conversation as conversation_lib
12
+ from llava.utils import disable_torch_init
13
+ from llava.mm_utils import tokenizer_image_token, process_images, get_model_name_from_path
14
+
15
+ from PIL import Image
16
+ import math
17
+
18
+
19
+ def split_list(lst, n):
20
+ """Split a list into n (roughly) equal-sized chunks"""
21
+ chunk_size = math.ceil(len(lst) / n) # integer division
22
+ return [lst[i:i+chunk_size] for i in range(0, len(lst), chunk_size)]
23
+
24
+
25
+ def get_chunk(lst, n, k):
26
+ chunks = split_list(lst, n)
27
+ return chunks[k]
28
+
29
+
30
+ def eval_model(args):
31
+ # Model
32
+ disable_torch_init()
33
+ model_path = os.path.expanduser(args.model_path)
34
+ model_name = get_model_name_from_path(model_path)
35
+ tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, args.model_base, model_name, llava_model="initial")
36
+
37
+ print("*************", len(tokenizer))
38
+
39
+ questions = json.load(open(os.path.expanduser(args.question_file), "r"))
40
+ questions = get_chunk(questions, args.num_chunks, args.chunk_idx)
41
+ answers_file = os.path.expanduser(args.answers_file)
42
+ os.makedirs(os.path.dirname(answers_file), exist_ok=True)
43
+ ans_file = open(answers_file, "w")
44
+ for i, line in enumerate(tqdm(questions)):
45
+ idx = line["id"]
46
+ question = line['conversations'][0]
47
+ qs = question['value'].replace('<image>', '').strip()
48
+ cur_prompt = qs
49
+
50
+ if 'image' in line:
51
+ image_file = line["image"]
52
+ image = Image.open(os.path.join(args.image_folder, image_file))
53
+ image_tensor = process_images([image], image_processor, model.config)[0]
54
+ images = image_tensor.unsqueeze(0).half().cuda()
55
+ image_sizes = [image.size]
56
+ if getattr(model.config, 'mm_use_im_start_end', False):
57
+ qs = DEFAULT_IM_START_TOKEN + DEFAULT_IMAGE_TOKEN + DEFAULT_IM_END_TOKEN + '\n' + qs
58
+ else:
59
+ qs = DEFAULT_IMAGE_TOKEN + '\n' + qs
60
+ cur_prompt = '<image>' + '\n' + cur_prompt
61
+ else:
62
+ continue
63
+ images = None
64
+ image_sizes = None
65
+
66
+ if args.single_pred_prompt:
67
+ qs = qs + '\n' + "Answer with the option's letter from the given choices directly."
68
+ cur_prompt = cur_prompt + '\n' + "Answer with the option's letter from the given choices directly."
69
+
70
+ # conv = conv_templates[args.conv_mode].copy()
71
+ conv = conversation_lib.voco_default_conversation.copy()
72
+ conv.append_message(conv.roles[0], qs)
73
+ conv.append_message(conv.roles[1], None)
74
+ prompt = conv.get_prompt()
75
+
76
+ maybe_voco_str = "".join(
77
+ ["<voco>" for _ in range(args.voco_num)]
78
+ )
79
+ prompt = f"<image>\n{maybe_voco_str}\n" + prompt.replace("\n", '').replace("<image>", '')
80
+
81
+ input_ids = tokenizer_image_token(prompt, tokenizer, IMAGE_TOKEN_INDEX, return_tensors='pt').unsqueeze(0).cuda()
82
+
83
+ with torch.inference_mode():
84
+ output_ids = model.generate(
85
+ input_ids,
86
+ images=images,
87
+ image_sizes=image_sizes,
88
+ do_sample=True if args.temperature > 0 else False,
89
+ temperature=args.temperature,
90
+ max_new_tokens=1024,
91
+ use_cache=True,
92
+ )
93
+
94
+ outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
95
+
96
+ ans_id = shortuuid.uuid()
97
+ ans_file.write(json.dumps({"question_id": idx,
98
+ "prompt": cur_prompt,
99
+ "text": outputs,
100
+ "answer_id": ans_id,
101
+ "model_id": model_name,
102
+ "metadata": {}}) + "\n")
103
+ ans_file.flush()
104
+ ans_file.close()
105
+
106
+ if __name__ == "__main__":
107
+ parser = argparse.ArgumentParser()
108
+ parser.add_argument("--model-path", type=str, default="facebook/opt-350m")
109
+ parser.add_argument("--model-base", type=str, default=None)
110
+ parser.add_argument("--image-folder", type=str, default="")
111
+ parser.add_argument("--question-file", type=str, default="tables/question.json")
112
+ parser.add_argument("--answers-file", type=str, default="answer.jsonl")
113
+ parser.add_argument("--conv-mode", type=str, default="llava_v0")
114
+ parser.add_argument("--num-chunks", type=int, default=1)
115
+ parser.add_argument("--chunk-idx", type=int, default=0)
116
+ parser.add_argument("--temperature", type=float, default=0.2)
117
+ parser.add_argument("--answer-prompter", action="store_true")
118
+ parser.add_argument("--single-pred-prompt", action="store_true")
119
+ parser.add_argument("--voco_num", type=int, default=None)
120
+ args = parser.parse_args()
121
+
122
+ eval_model(args)
llava/train/__pycache__/llava_trainer.cpython-310.pyc ADDED
Binary file (11.4 kB). View file
 
llava/train/__pycache__/train_compress.cpython-310.pyc ADDED
Binary file (28.1 kB). View file
 
llava/train/llama_flash_attn_monkey_patch.py ADDED
@@ -0,0 +1,115 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from typing import Optional, Tuple
2
+ import warnings
3
+
4
+ import torch
5
+
6
+ import transformers
7
+ from transformers.models.llama.modeling_llama import apply_rotary_pos_emb, repeat_kv
8
+
9
+ try:
10
+ from flash_attn.flash_attn_interface import flash_attn_unpadded_qkvpacked_func
11
+ except ImportError:
12
+ from flash_attn.flash_attn_interface import flash_attn_varlen_qkvpacked_func as flash_attn_unpadded_qkvpacked_func
13
+ from flash_attn.bert_padding import unpad_input, pad_input
14
+
15
+
16
+ def forward(
17
+ self,
18
+ hidden_states: torch.Tensor,
19
+ attention_mask: Optional[torch.Tensor] = None,
20
+ position_ids: Optional[torch.Tensor] = None,
21
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
22
+ output_attentions: bool = False,
23
+ use_cache: bool = False,
24
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
25
+ if output_attentions:
26
+ warnings.warn(
27
+ "Output attentions is not supported for patched `LlamaAttention`, returning `None` instead."
28
+ )
29
+
30
+ bsz, q_len, _ = hidden_states.size()
31
+
32
+ query_states = (
33
+ self.q_proj(hidden_states)
34
+ .view(bsz, q_len, self.num_heads, self.head_dim)
35
+ .transpose(1, 2)
36
+ )
37
+ key_states = (
38
+ self.k_proj(hidden_states)
39
+ .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
40
+ .transpose(1, 2)
41
+ )
42
+ value_states = (
43
+ self.v_proj(hidden_states)
44
+ .view(bsz, q_len, self.num_key_value_heads, self.head_dim)
45
+ .transpose(1, 2)
46
+ ) # shape: (b, num_heads, s, head_dim)
47
+
48
+ kv_seq_len = key_states.shape[-2]
49
+ if past_key_value is not None:
50
+ kv_seq_len += past_key_value[0].shape[-2]
51
+
52
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
53
+ query_states, key_states = apply_rotary_pos_emb(
54
+ query_states, key_states, cos, sin, position_ids
55
+ )
56
+
57
+ if past_key_value is not None:
58
+ # reuse k, v
59
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
60
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
61
+
62
+ past_key_value = (key_states, value_states) if use_cache else None
63
+
64
+ # repeat k/v heads if n_kv_heads < n_heads
65
+ key_states = repeat_kv(key_states, self.num_key_value_groups)
66
+ value_states = repeat_kv(value_states, self.num_key_value_groups)
67
+
68
+ # Transform the data into the format required by flash attention
69
+ qkv = torch.stack([query_states, key_states, value_states], dim=2)
70
+ qkv = qkv.transpose(1, 3) # shape: [b, s, 3, num_heads, head_dim]
71
+ key_padding_mask = attention_mask
72
+
73
+ if key_padding_mask is None:
74
+ qkv = qkv.reshape(-1, 3, self.num_heads, self.head_dim)
75
+ cu_q_lens = torch.arange(
76
+ 0, (bsz + 1) * q_len, step=q_len, dtype=torch.int32, device=qkv.device
77
+ )
78
+ max_s = q_len
79
+ output = flash_attn_unpadded_qkvpacked_func(
80
+ qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
81
+ )
82
+ output = output.view(bsz, q_len, -1)
83
+ else:
84
+ qkv = qkv.reshape(bsz, q_len, -1)
85
+ qkv, indices, cu_q_lens, max_s = unpad_input(qkv, key_padding_mask)
86
+ qkv = qkv.view(-1, 3, self.num_heads, self.head_dim)
87
+ output_unpad = flash_attn_unpadded_qkvpacked_func(
88
+ qkv, cu_q_lens, max_s, 0.0, softmax_scale=None, causal=True
89
+ )
90
+ output_unpad = output_unpad.reshape(-1, self.num_heads * self.head_dim)
91
+ output = pad_input(output_unpad, indices, bsz, q_len)
92
+
93
+ return self.o_proj(output), None, past_key_value
94
+
95
+
96
+ # Disable the transformation of the attention mask in LlamaModel as the flash attention
97
+ # requires the attention mask to be the same as the key_padding_mask
98
+ def _prepare_decoder_attention_mask(
99
+ self, attention_mask, input_shape, inputs_embeds, past_key_values_length
100
+ ):
101
+ # [bsz, seq_len]
102
+ return attention_mask
103
+
104
+
105
+ def replace_llama_attn_with_flash_attn():
106
+ cuda_major, cuda_minor = torch.cuda.get_device_capability()
107
+ if cuda_major < 8:
108
+ warnings.warn(
109
+ "Flash attention is only supported on A100 or H100 GPU during training due to head dim > 64 backward."
110
+ "ref: https://github.com/HazyResearch/flash-attention/issues/190#issuecomment-1523359593"
111
+ )
112
+ transformers.models.llama.modeling_llama.LlamaModel._prepare_decoder_attention_mask = (
113
+ _prepare_decoder_attention_mask
114
+ )
115
+ transformers.models.llama.modeling_llama.LlamaAttention.forward = forward
llava/train/llama_xformers_attn_monkey_patch.py ADDED
@@ -0,0 +1,129 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ """
2
+ Directly copied the code from https://raw.githubusercontent.com/oobabooga/text-generation-webui/main/modules/llama_attn_hijack.py and made some adjustments
3
+ """
4
+
5
+ import logging
6
+ import math
7
+ from typing import Optional, Tuple
8
+
9
+ import torch
10
+ import transformers.models.llama.modeling_llama
11
+ from torch import nn
12
+
13
+ try:
14
+ import xformers.ops
15
+ except ImportError:
16
+ logging.error("xformers not found! Please install it before trying to use it.")
17
+
18
+
19
+ def replace_llama_attn_with_xformers_attn():
20
+ transformers.models.llama.modeling_llama.LlamaAttention.forward = xformers_forward
21
+
22
+
23
+ def xformers_forward(
24
+ self,
25
+ hidden_states: torch.Tensor,
26
+ attention_mask: Optional[torch.Tensor] = None,
27
+ position_ids: Optional[torch.LongTensor] = None,
28
+ past_key_value: Optional[Tuple[torch.Tensor]] = None,
29
+ output_attentions: bool = False,
30
+ use_cache: bool = False,
31
+ ) -> Tuple[torch.Tensor, Optional[torch.Tensor], Optional[Tuple[torch.Tensor]]]:
32
+ # pylint: disable=duplicate-code
33
+ bsz, q_len, _ = hidden_states.size()
34
+
35
+ query_states = (
36
+ self.q_proj(hidden_states)
37
+ .view(bsz, q_len, self.num_heads, self.head_dim)
38
+ .transpose(1, 2)
39
+ )
40
+ key_states = (
41
+ self.k_proj(hidden_states)
42
+ .view(bsz, q_len, self.num_heads, self.head_dim)
43
+ .transpose(1, 2)
44
+ )
45
+ value_states = (
46
+ self.v_proj(hidden_states)
47
+ .view(bsz, q_len, self.num_heads, self.head_dim)
48
+ .transpose(1, 2)
49
+ )
50
+
51
+ kv_seq_len = key_states.shape[-2]
52
+ if past_key_value is not None:
53
+ kv_seq_len += past_key_value[0].shape[-2]
54
+ cos, sin = self.rotary_emb(value_states, seq_len=kv_seq_len)
55
+ (
56
+ query_states,
57
+ key_states,
58
+ ) = transformers.models.llama.modeling_llama.apply_rotary_pos_emb(
59
+ query_states, key_states, cos, sin, position_ids
60
+ )
61
+ # [bsz, nh, t, hd]
62
+
63
+ if past_key_value is not None:
64
+ # reuse k, v, self_attention
65
+ key_states = torch.cat([past_key_value[0], key_states], dim=2)
66
+ value_states = torch.cat([past_key_value[1], value_states], dim=2)
67
+
68
+ past_key_value = (key_states, value_states) if use_cache else None
69
+
70
+ # We only apply xformers optimizations if we don't need to output the whole attention matrix
71
+ if not output_attentions:
72
+ query_states = query_states.transpose(1, 2)
73
+ key_states = key_states.transpose(1, 2)
74
+ value_states = value_states.transpose(1, 2)
75
+
76
+ # This is a nasty hack. We know attention_mask in transformers is either LowerTriangular or all Zeros.
77
+ # We therefore check if one element in the upper triangular portion is zero. If it is, then the mask is all zeros.
78
+ if attention_mask is None or attention_mask[0, 0, 0, 1] == 0:
79
+ # input and output should be of form (bsz, q_len, num_heads, head_dim)
80
+ attn_output = xformers.ops.memory_efficient_attention(
81
+ query_states, key_states, value_states, attn_bias=None
82
+ )
83
+ else:
84
+ # input and output should be of form (bsz, q_len, num_heads, head_dim)
85
+ attn_output = xformers.ops.memory_efficient_attention(
86
+ query_states,
87
+ key_states,
88
+ value_states,
89
+ attn_bias=xformers.ops.LowerTriangularMask(),
90
+ )
91
+ attn_weights = None
92
+ else:
93
+ attn_weights = torch.matmul(
94
+ query_states, key_states.transpose(2, 3)
95
+ ) / math.sqrt(self.head_dim)
96
+
97
+ if attn_weights.size() != (bsz, self.num_heads, q_len, kv_seq_len):
98
+ raise ValueError(
99
+ f"Attention weights should be of size {(bsz * self.num_heads, q_len, kv_seq_len)}, but is"
100
+ f" {attn_weights.size()}"
101
+ )
102
+
103
+ if attention_mask is not None:
104
+ if attention_mask.size() != (bsz, 1, q_len, kv_seq_len):
105
+ raise ValueError(
106
+ f"Attention mask should be of size {(bsz, 1, q_len, kv_seq_len)}, but is {attention_mask.size()}"
107
+ )
108
+ attn_weights = attn_weights + attention_mask
109
+ attn_weights = torch.max(
110
+ attn_weights, torch.tensor(torch.finfo(attn_weights.dtype).min)
111
+ )
112
+
113
+ # upcast attention to fp32
114
+ attn_weights = nn.functional.softmax(
115
+ attn_weights, dim=-1, dtype=torch.float32
116
+ ).to(query_states.dtype)
117
+ attn_output = torch.matmul(attn_weights, value_states)
118
+
119
+ if attn_output.size() != (bsz, self.num_heads, q_len, self.head_dim):
120
+ raise ValueError(
121
+ f"`attn_output` should be of size {(bsz, self.num_heads, q_len, self.head_dim)}, but is"
122
+ f" {attn_output.size()}"
123
+ )
124
+
125
+ attn_output = attn_output.transpose(1, 2)
126
+
127
+ attn_output = attn_output.reshape(bsz, q_len, self.hidden_size)
128
+ attn_output = self.o_proj(attn_output)
129
+ return attn_output, attn_weights, past_key_value
llava/train/train_compress.py ADDED
@@ -0,0 +1,1014 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ # Adopted from https://github.com/haotian-liu/LLaVA.
2
+ # Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright:
3
+ # Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright:
4
+ # Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
5
+ #
6
+ # Licensed under the Apache License, Version 2.0 (the "License");
7
+ # you may not use this file except in compliance with the License.
8
+ # You may obtain a copy of the License at
9
+ #
10
+ # http://www.apache.org/licenses/LICENSE-2.0
11
+ #
12
+ # Unless required by applicable law or agreed to in writing, software
13
+ # distributed under the License is distributed on an "AS IS" BASIS,
14
+ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
15
+ # See the License for the specific language governing permissions and
16
+ # limitations under the License.
17
+
18
+ import os
19
+ import copy
20
+ from dataclasses import dataclass, field
21
+ import json
22
+ import logging
23
+ import pathlib
24
+ from typing import Dict, Optional, Sequence, List
25
+
26
+ import torch
27
+
28
+ import transformers
29
+ import tokenizers
30
+
31
+ from llava.constants import IGNORE_INDEX, IMAGE_TOKEN_INDEX, DEFAULT_IMAGE_TOKEN, DEFAULT_IM_START_TOKEN, DEFAULT_IM_END_TOKEN
32
+ from torch.utils.data import Dataset
33
+ from llava.train.llava_trainer import LLaVATrainer
34
+
35
+ from llava import conversation as conversation_lib
36
+ from llava.model import *
37
+ from llava.mm_utils import tokenizer_image_token
38
+
39
+ from PIL import Image
40
+
41
+
42
+ local_rank = None
43
+
44
+
45
+ def rank0_print(*args):
46
+ if local_rank == 0:
47
+ print(*args)
48
+
49
+
50
+ from packaging import version
51
+ IS_TOKENIZER_GREATER_THAN_0_14 = version.parse(tokenizers.__version__) >= version.parse('0.14')
52
+
53
+
54
+ @dataclass
55
+ class ModelArguments:
56
+ model_name_or_path: Optional[str] = field(default="facebook/opt-125m")
57
+ version: Optional[str] = field(default="v0")
58
+ freeze_backbone: bool = field(default=False)
59
+ tune_mm_mlp_adapter: bool = field(default=False)
60
+ vision_tower: Optional[str] = field(default=None)
61
+ mm_vision_select_layer: Optional[int] = field(default=-1) # default to the last layer
62
+ pretrain_mm_mlp_adapter: Optional[str] = field(default=None)
63
+ mm_projector_type: Optional[str] = field(default='linear')
64
+ mm_use_im_start_end: bool = field(default=False)
65
+ mm_use_im_patch_token: bool = field(default=True)
66
+ mm_patch_merge_type: Optional[str] = field(default='flat')
67
+ mm_vision_select_feature: Optional[str] = field(default="patch")
68
+
69
+
70
+ @dataclass
71
+ class DataArguments:
72
+ data_path: str = field(default=None,
73
+ metadata={"help": "Path to the training data."})
74
+ lazy_preprocess: bool = False
75
+ is_multimodal: bool = False
76
+ image_folder: Optional[str] = field(default=None)
77
+ image_aspect_ratio: str = 'square'
78
+
79
+
80
+ @dataclass
81
+ class TrainingArguments(transformers.TrainingArguments):
82
+ cache_dir: Optional[str] = field(default=None)
83
+ optim: str = field(default="adamw_torch")
84
+ remove_unused_columns: bool = field(default=False)
85
+ freeze_mm_mlp_adapter: bool = field(default=False)
86
+ mpt_attn_impl: Optional[str] = field(default="triton")
87
+ model_max_length: int = field(
88
+ default=512,
89
+ metadata={
90
+ "help":
91
+ "Maximum sequence length. Sequences will be right padded (and possibly truncated)."
92
+ },
93
+ )
94
+ double_quant: bool = field(
95
+ default=True,
96
+ metadata={"help": "Compress the quantization statistics through double quantization."}
97
+ )
98
+ quant_type: str = field(
99
+ default="nf4",
100
+ metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
101
+ )
102
+ bits: int = field(
103
+ default=16,
104
+ metadata={"help": "How many bits to use."}
105
+ )
106
+ lora_enable: bool = False
107
+ lora_r: int = 64
108
+ lora_alpha: int = 16
109
+ lora_dropout: float = 0.05
110
+ lora_weight_path: str = ""
111
+ lora_bias: str = "none"
112
+ mm_projector_lr: Optional[float] = None
113
+ group_by_modality_length: bool = field(default=False)
114
+
115
+
116
+ def maybe_zero_3(param, ignore_status=False, name=None):
117
+ from deepspeed import zero
118
+ from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
119
+ if hasattr(param, "ds_id"):
120
+ if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
121
+ if not ignore_status:
122
+ logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}")
123
+ with zero.GatheredParameters([param]):
124
+ param = param.data.detach().cpu().clone()
125
+ else:
126
+ param = param.detach().cpu().clone()
127
+ return param
128
+
129
+
130
+ # Borrowed from peft.utils.get_peft_model_state_dict
131
+ def get_peft_state_maybe_zero_3(named_params, bias):
132
+ if bias == "none":
133
+ to_return = {k: t for k, t in named_params if "lora_" in k}
134
+ elif bias == "all":
135
+ to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
136
+ elif bias == "lora_only":
137
+ to_return = {}
138
+ maybe_lora_bias = {}
139
+ lora_bias_names = set()
140
+ for k, t in named_params:
141
+ if "lora_" in k:
142
+ to_return[k] = t
143
+ bias_name = k.split("lora_")[0] + "bias"
144
+ lora_bias_names.add(bias_name)
145
+ elif "bias" in k:
146
+ maybe_lora_bias[k] = t
147
+ for k, t in maybe_lora_bias:
148
+ if bias_name in lora_bias_names:
149
+ to_return[bias_name] = t
150
+ else:
151
+ raise NotImplementedError
152
+ to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()}
153
+ return to_return
154
+
155
+
156
+ def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
157
+ to_return = {k: t for k, t in named_params if "lora_" not in k}
158
+ if require_grad_only:
159
+ to_return = {k: t for k, t in to_return.items() if t.requires_grad}
160
+ to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
161
+ return to_return
162
+
163
+
164
+ def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
165
+ to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
166
+ to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
167
+ return to_return
168
+
169
+
170
+ def find_all_linear_names(model):
171
+ cls = torch.nn.Linear
172
+ lora_module_names = set()
173
+ multimodal_keywords = ['mm_projector', 'vision_tower', 'vision_resampler']
174
+ for name, module in model.named_modules():
175
+ if any(mm_keyword in name for mm_keyword in multimodal_keywords):
176
+ continue
177
+ if isinstance(module, cls):
178
+ names = name.split('.')
179
+ lora_module_names.add(names[0] if len(names) == 1 else names[-1])
180
+
181
+ if 'lm_head' in lora_module_names: # needed for 16-bit
182
+ lora_module_names.remove('lm_head')
183
+ return list(lora_module_names)
184
+
185
+
186
+ def safe_save_model_for_hf_trainer(trainer: transformers.Trainer,
187
+ output_dir: str):
188
+ """Collects the state dict and dump to disk."""
189
+
190
+ if getattr(trainer.args, "tune_mm_mlp_adapter", False):
191
+ # Only save Adapter
192
+ keys_to_match = ['mm_projector']
193
+ if getattr(trainer.args, "use_im_start_end", False):
194
+ keys_to_match.extend(['embed_tokens', 'embed_in'])
195
+
196
+ weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match)
197
+ trainer.model.config.save_pretrained(output_dir)
198
+
199
+ current_folder = output_dir.split('/')[-1]
200
+ parent_folder = os.path.dirname(output_dir)
201
+ if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
202
+ if current_folder.startswith('checkpoint-'):
203
+ mm_projector_folder = os.path.join(parent_folder, "mm_projector")
204
+ os.makedirs(mm_projector_folder, exist_ok=True)
205
+ torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin'))
206
+ else:
207
+ torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
208
+ return
209
+
210
+ if trainer.deepspeed:
211
+ torch.cuda.synchronize()
212
+ trainer.save_model(output_dir)
213
+ return
214
+
215
+ state_dict = trainer.model.state_dict()
216
+ if trainer.args.should_save:
217
+ cpu_state_dict = {
218
+ key: value.cpu()
219
+ for key, value in state_dict.items()
220
+ }
221
+ del state_dict
222
+ trainer._save(output_dir, state_dict=cpu_state_dict) # noqa
223
+
224
+
225
+ def smart_tokenizer_and_embedding_resize(
226
+ special_tokens_dict: Dict,
227
+ tokenizer: transformers.PreTrainedTokenizer,
228
+ model: transformers.PreTrainedModel,
229
+ ):
230
+ """Resize tokenizer and embedding.
231
+
232
+ Note: This is the unoptimized version that may make your embedding size not be divisible by 64.
233
+ """
234
+ num_new_tokens = tokenizer.add_special_tokens(special_tokens_dict)
235
+ model.resize_token_embeddings(len(tokenizer))
236
+
237
+ if num_new_tokens > 0:
238
+ input_embeddings = model.get_input_embeddings().weight.data
239
+ output_embeddings = model.get_output_embeddings().weight.data
240
+
241
+ input_embeddings_avg = input_embeddings[:-num_new_tokens].mean(
242
+ dim=0, keepdim=True)
243
+ output_embeddings_avg = output_embeddings[:-num_new_tokens].mean(
244
+ dim=0, keepdim=True)
245
+
246
+ input_embeddings[-num_new_tokens:] = input_embeddings_avg
247
+ output_embeddings[-num_new_tokens:] = output_embeddings_avg
248
+
249
+
250
+ def _tokenize_fn(strings: Sequence[str],
251
+ tokenizer: transformers.PreTrainedTokenizer) -> Dict:
252
+ """Tokenize a list of strings."""
253
+ tokenized_list = [
254
+ tokenizer(
255
+ text,
256
+ return_tensors="pt",
257
+ padding="longest",
258
+ max_length=tokenizer.model_max_length,
259
+ truncation=True,
260
+ ) for text in strings
261
+ ]
262
+ input_ids = labels = [
263
+ tokenized.input_ids[0] for tokenized in tokenized_list
264
+ ]
265
+ input_ids_lens = labels_lens = [
266
+ tokenized.input_ids.ne(tokenizer.pad_token_id).sum().item()
267
+ for tokenized in tokenized_list
268
+ ]
269
+ return dict(
270
+ input_ids=input_ids,
271
+ labels=labels,
272
+ input_ids_lens=input_ids_lens,
273
+ labels_lens=labels_lens,
274
+ )
275
+
276
+
277
+ def _mask_targets(target, tokenized_lens, speakers):
278
+ # cur_idx = 0
279
+ cur_idx = tokenized_lens[0]
280
+ tokenized_lens = tokenized_lens[1:]
281
+ target[:cur_idx] = IGNORE_INDEX
282
+ for tokenized_len, speaker in zip(tokenized_lens, speakers):
283
+ if speaker == "human":
284
+ target[cur_idx+2:cur_idx + tokenized_len] = IGNORE_INDEX
285
+ cur_idx += tokenized_len
286
+
287
+
288
+ def _add_speaker_and_signal(header, source, get_conversation=True):
289
+ """Add speaker and start/end signal on each round."""
290
+ BEGIN_SIGNAL = "### "
291
+ END_SIGNAL = "\n"
292
+ conversation = header
293
+ for sentence in source:
294
+ from_str = sentence["from"]
295
+ if from_str.lower() == "human":
296
+ from_str = conversation_lib.default_conversation.roles[0]
297
+ elif from_str.lower() == "gpt":
298
+ from_str = conversation_lib.default_conversation.roles[1]
299
+ else:
300
+ from_str = 'unknown'
301
+ sentence["value"] = (BEGIN_SIGNAL + from_str + ": " +
302
+ sentence["value"] + END_SIGNAL)
303
+ if get_conversation:
304
+ conversation += sentence["value"]
305
+ conversation += BEGIN_SIGNAL
306
+ return conversation
307
+
308
+
309
+ def preprocess_multimodal(
310
+ sources: Sequence[str],
311
+ data_args: DataArguments
312
+ ) -> Dict:
313
+ is_multimodal = data_args.is_multimodal
314
+ if not is_multimodal:
315
+ return sources
316
+
317
+ for source in sources:
318
+ for sentence in source:
319
+ if DEFAULT_IMAGE_TOKEN in sentence['value']:
320
+ sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '').strip()
321
+ sentence['value'] = DEFAULT_IMAGE_TOKEN + '\n' + sentence['value']
322
+ sentence['value'] = sentence['value'].strip()
323
+ if "mmtag" in conversation_lib.default_conversation.version:
324
+ sentence['value'] = sentence['value'].replace(DEFAULT_IMAGE_TOKEN, '<Image>' + DEFAULT_IMAGE_TOKEN + '</Image>')
325
+ replace_token = DEFAULT_IMAGE_TOKEN
326
+ if data_args.mm_use_im_start_end:
327
+ replace_token = DEFAULT_IM_START_TOKEN + replace_token + DEFAULT_IM_END_TOKEN
328
+ sentence["value"] = sentence["value"].replace(DEFAULT_IMAGE_TOKEN, replace_token)
329
+
330
+ return sources
331
+
332
+
333
+ def preprocess_llama_2(
334
+ sources,
335
+ tokenizer: transformers.PreTrainedTokenizer,
336
+ has_image: bool = False
337
+ ) -> Dict:
338
+ conv = conversation_lib.default_conversation.copy()
339
+ roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
340
+
341
+ # Apply prompt templates
342
+ conversations = []
343
+ for i, source in enumerate(sources):
344
+ if roles[source[0]["from"]] != conv.roles[0]:
345
+ # Skip the first one if it is not from human
346
+ source = source[1:]
347
+
348
+ conv.messages = []
349
+ for j, sentence in enumerate(source):
350
+ role = roles[sentence["from"]]
351
+ assert role == conv.roles[j % 2], f"{i}"
352
+ conv.append_message(role, sentence["value"])
353
+ conversations.append(conv.get_prompt())
354
+
355
+ # Tokenize conversations
356
+
357
+ if has_image:
358
+ input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
359
+ else:
360
+ input_ids = tokenizer(
361
+ conversations,
362
+ return_tensors="pt",
363
+ padding="longest",
364
+ max_length=tokenizer.model_max_length,
365
+ truncation=True,
366
+ ).input_ids
367
+
368
+ targets = input_ids.clone()
369
+
370
+ assert conv.sep_style == conversation_lib.SeparatorStyle.LLAMA_2
371
+
372
+ # Mask targets
373
+ sep = "[/INST] "
374
+ for conversation, target in zip(conversations, targets):
375
+ total_len = int(target.ne(tokenizer.pad_token_id).sum())
376
+
377
+ rounds = conversation.split(conv.sep2)
378
+ cur_len = 1
379
+ target[:cur_len] = IGNORE_INDEX
380
+ for i, rou in enumerate(rounds):
381
+ if rou == "":
382
+ break
383
+
384
+ parts = rou.split(sep)
385
+ if len(parts) != 2:
386
+ break
387
+ parts[0] += sep
388
+
389
+ if has_image:
390
+ round_len = len(tokenizer_image_token(rou, tokenizer))
391
+ instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2
392
+ else:
393
+ round_len = len(tokenizer(rou).input_ids)
394
+ instruction_len = len(tokenizer(parts[0]).input_ids) - 2
395
+
396
+ target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
397
+
398
+ cur_len += round_len
399
+ target[cur_len:] = IGNORE_INDEX
400
+
401
+ if cur_len < tokenizer.model_max_length:
402
+ if cur_len != total_len:
403
+ target[:] = IGNORE_INDEX
404
+ print(
405
+ f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
406
+ f" (ignored)"
407
+ )
408
+
409
+ return dict(
410
+ input_ids=input_ids,
411
+ labels=targets,
412
+ )
413
+
414
+
415
+ def preprocess_v1(
416
+ sources,
417
+ tokenizer: transformers.PreTrainedTokenizer,
418
+ has_image: bool = False
419
+ ) -> Dict:
420
+ conv = conversation_lib.voco_default_conversation.copy()
421
+ roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
422
+
423
+ # Apply prompt templates
424
+ conversations = []
425
+ for i, source in enumerate(sources):
426
+ if roles[source[0]["from"]] != conv.roles[0]:
427
+ # Skip the first one if it is not from human
428
+ source = source[1:]
429
+
430
+ conv.messages = []
431
+ for j, sentence in enumerate(source):
432
+ role = roles[sentence["from"]]
433
+ assert role == conv.roles[j % 2], f"{i}"
434
+ conv.append_message(role, sentence["value"])
435
+ conversations.append(conv.get_prompt())
436
+ # The assistant gives helpful, detailed, and polite answers to the user's questions.
437
+ # + <image> + USER: Q + ASSITENT: A + ...
438
+
439
+ # Tokenize conversations
440
+ # token num
441
+ if has_image:
442
+ maybe_voco_str = "".join(
443
+ ["<voco>" for _ in range(2)]
444
+ )
445
+ # conversations = [f"<image>\n{maybe_voco_str}\n" + conversations[0].replace("<image>\n", '')]
446
+ conversations = [f"<image>\n{maybe_voco_str}\n" + conversations[0].replace("<image>", '').replace("\n", '')]
447
+ input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
448
+ else:
449
+ input_ids = tokenizer(
450
+ conversations,
451
+ return_tensors="pt",
452
+ padding="longest",
453
+ max_length=tokenizer.model_max_length,
454
+ truncation=True,
455
+ ).input_ids
456
+
457
+ targets = input_ids.clone() # [1, len]
458
+
459
+ assert conv.sep_style == conversation_lib.SeparatorStyle.TWO
460
+
461
+ # Mask targets
462
+ sep = conv.sep + conv.roles[1] + ": "
463
+ for conversation, target in zip(conversations, targets):
464
+ total_len = int(target.ne(tokenizer.pad_token_id).sum())
465
+
466
+ rounds = conversation.split(conv.sep2)
467
+ cur_len = 1
468
+ target[:cur_len] = IGNORE_INDEX
469
+ for i, rou in enumerate(rounds):
470
+ if rou == "":
471
+ break
472
+
473
+ parts = rou.split(sep)
474
+ if len(parts) != 2:
475
+ break
476
+ parts[0] += sep
477
+ if has_image:
478
+ round_len = len(tokenizer_image_token(rou, tokenizer)) # 1 + tokenize(rou), 其中<image>这个token被特殊编码为200
479
+ instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 2 # 问题token
480
+ else:
481
+ round_len = len(tokenizer(rou).input_ids)
482
+ instruction_len = len(tokenizer(parts[0]).input_ids) - 2
483
+
484
+ if i != 0 and not tokenizer.legacy and IS_TOKENIZER_GREATER_THAN_0_14:
485
+ round_len -= 1
486
+ instruction_len -= 1
487
+
488
+ target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
489
+
490
+ cur_len += round_len
491
+ target[cur_len:] = IGNORE_INDEX
492
+ if cur_len < tokenizer.model_max_length:
493
+ if cur_len != total_len:
494
+ target[:] = IGNORE_INDEX
495
+ print(
496
+ f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
497
+ f" (ignored)"
498
+ )
499
+
500
+ return dict(
501
+ input_ids=input_ids,
502
+ labels=targets,
503
+ )
504
+
505
+
506
+ def preprocess_mpt(
507
+ sources,
508
+ tokenizer: transformers.PreTrainedTokenizer,
509
+ has_image: bool = False
510
+ ) -> Dict:
511
+ conv = conversation_lib.default_conversation.copy()
512
+ roles = {"human": conv.roles[0], "gpt": conv.roles[1]}
513
+
514
+ # Apply prompt templates
515
+ conversations = []
516
+ for i, source in enumerate(sources):
517
+ if roles[source[0]["from"]] != conv.roles[0]:
518
+ # Skip the first one if it is not from human
519
+ source = source[1:]
520
+
521
+ conv.messages = []
522
+ for j, sentence in enumerate(source):
523
+ role = roles[sentence["from"]]
524
+ assert role == conv.roles[j % 2], f"{i}"
525
+ conv.append_message(role, sentence["value"])
526
+ conversations.append(conv.get_prompt())
527
+
528
+ # Tokenize conversations
529
+
530
+ if has_image:
531
+ input_ids = torch.stack([tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations], dim=0)
532
+ else:
533
+ input_ids = tokenizer(
534
+ conversations,
535
+ return_tensors="pt",
536
+ padding="longest",
537
+ max_length=tokenizer.model_max_length,
538
+ truncation=True,
539
+ ).input_ids
540
+
541
+ targets = input_ids.clone()
542
+ assert conv.sep_style == conversation_lib.SeparatorStyle.MPT
543
+
544
+ # Mask targets
545
+ sep = conv.sep + conv.roles[1]
546
+ for conversation, target in zip(conversations, targets):
547
+ total_len = int(target.ne(tokenizer.pad_token_id).sum())
548
+
549
+ rounds = conversation.split(conv.sep)
550
+ re_rounds = [conv.sep.join(rounds[:3])] # system + user + gpt
551
+ for conv_idx in range(3, len(rounds), 2):
552
+ re_rounds.append(conv.sep.join(rounds[conv_idx:conv_idx+2])) # user + gpt
553
+ cur_len = 0
554
+ target[:cur_len] = IGNORE_INDEX
555
+ for i, rou in enumerate(re_rounds):
556
+ if rou == "":
557
+ break
558
+
559
+ parts = rou.split(sep)
560
+ if len(parts) != 2:
561
+ break
562
+ parts[0] += sep
563
+
564
+ if has_image:
565
+ round_len = len(tokenizer_image_token(rou, tokenizer))
566
+ instruction_len = len(tokenizer_image_token(parts[0], tokenizer)) - 1
567
+ else:
568
+ round_len = len(tokenizer(rou).input_ids)
569
+ instruction_len = len(tokenizer(parts[0]).input_ids) - 1
570
+
571
+ if i != 0 and getattr(tokenizer, 'legacy', False) and IS_TOKENIZER_GREATER_THAN_0_14:
572
+ round_len += 1
573
+ instruction_len += 1
574
+
575
+ target[cur_len : cur_len + instruction_len] = IGNORE_INDEX
576
+
577
+ cur_len += round_len
578
+ target[cur_len:] = IGNORE_INDEX
579
+
580
+ if cur_len < tokenizer.model_max_length:
581
+ if cur_len != total_len:
582
+ target[:] = IGNORE_INDEX
583
+ print(
584
+ f"WARNING: tokenization mismatch: {cur_len} vs. {total_len}."
585
+ f" (ignored)"
586
+ )
587
+
588
+ return dict(
589
+ input_ids=input_ids,
590
+ labels=targets,
591
+ )
592
+
593
+
594
+ def preprocess_plain(
595
+ sources: Sequence[str],
596
+ tokenizer: transformers.PreTrainedTokenizer,
597
+ ) -> Dict:
598
+ # add end signal and concatenate together
599
+ conversations = []
600
+ for source in sources:
601
+ assert len(source) == 2
602
+ assert DEFAULT_IMAGE_TOKEN in source[0]['value']
603
+ source[0]['value'] = DEFAULT_IMAGE_TOKEN
604
+ conversation = source[0]['value'] + source[1]['value'] + conversation_lib.default_conversation.sep
605
+ conversations.append(conversation)
606
+ # tokenize conversations
607
+ input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
608
+ targets = copy.deepcopy(input_ids)
609
+ for target, source in zip(targets, sources):
610
+ tokenized_len = len(tokenizer_image_token(source[0]['value'], tokenizer))
611
+ target[:tokenized_len] = IGNORE_INDEX
612
+
613
+ return dict(input_ids=input_ids, labels=targets)
614
+
615
+
616
+ def preprocess(
617
+ sources: Sequence[str],
618
+ tokenizer: transformers.PreTrainedTokenizer,
619
+ has_image: bool = False
620
+ ) -> Dict:
621
+ """
622
+ Given a list of sources, each is a conversation list. This transform:
623
+ 1. Add signal '### ' at the beginning each sentence, with end signal '\n';
624
+ 2. Concatenate conversations together;
625
+ 3. Tokenize the concatenated conversation;
626
+ 4. Make a deepcopy as the target. Mask human words with IGNORE_INDEX.
627
+ """
628
+ if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.PLAIN:
629
+ return preprocess_plain(sources, tokenizer)
630
+ if conversation_lib.default_conversation.sep_style == conversation_lib.SeparatorStyle.LLAMA_2:
631
+ return preprocess_llama_2(sources, tokenizer, has_image=has_image)
632
+ if conversation_lib.default_conversation.version.startswith("v1"):
633
+ return preprocess_v1(sources, tokenizer, has_image=has_image)
634
+ if conversation_lib.default_conversation.version == "mpt":
635
+ return preprocess_mpt(sources, tokenizer, has_image=has_image)
636
+ # add end signal and concatenate together
637
+ conversations = []
638
+ for source in sources:
639
+ header = f"{conversation_lib.default_conversation.system}\n\n"
640
+ conversation = _add_speaker_and_signal(header, source)
641
+ conversations.append(conversation)
642
+ # tokenize conversations
643
+ def get_tokenize_len(prompts):
644
+ return [len(tokenizer_image_token(prompt, tokenizer)) for prompt in prompts]
645
+
646
+ if has_image:
647
+ input_ids = [tokenizer_image_token(prompt, tokenizer, return_tensors='pt') for prompt in conversations]
648
+ else:
649
+ conversations_tokenized = _tokenize_fn(conversations, tokenizer)
650
+ input_ids = conversations_tokenized["input_ids"]
651
+
652
+ targets = copy.deepcopy(input_ids)
653
+ for target, source in zip(targets, sources):
654
+ if has_image:
655
+ tokenized_lens = get_tokenize_len([header] + [s["value"] for s in source])
656
+ else:
657
+ tokenized_lens = _tokenize_fn([header] + [s["value"] for s in source], tokenizer)["input_ids_lens"]
658
+ speakers = [sentence["from"] for sentence in source]
659
+ _mask_targets(target, tokenized_lens, speakers)
660
+
661
+ return dict(input_ids=input_ids, labels=targets)
662
+
663
+
664
+ class LazySupervisedDataset(Dataset):
665
+ """Dataset for supervised fine-tuning."""
666
+
667
+ def __init__(self, data_path: str,
668
+ tokenizer: transformers.PreTrainedTokenizer,
669
+ data_args: DataArguments,
670
+ voco_token):
671
+ super(LazySupervisedDataset, self).__init__()
672
+ list_data_dict = json.load(open(data_path, "r"))
673
+
674
+ rank0_print("Formatting inputs...Skip in lazy mode")
675
+ self.tokenizer = tokenizer
676
+ self.list_data_dict = list_data_dict
677
+ self.data_args = data_args
678
+ self.voco_token = voco_token
679
+
680
+ def __len__(self):
681
+ return len(self.list_data_dict)
682
+
683
+ @property
684
+ def lengths(self):
685
+ length_list = []
686
+ for sample in self.list_data_dict:
687
+ img_tokens = 128 if 'image' in sample else 0
688
+ length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens)
689
+ return length_list
690
+
691
+ @property
692
+ def modality_lengths(self):
693
+ length_list = []
694
+ for sample in self.list_data_dict:
695
+ cur_len = sum(len(conv['value'].split()) for conv in sample['conversations'])
696
+ cur_len = cur_len if 'image' in sample else -cur_len
697
+ length_list.append(cur_len)
698
+ return length_list
699
+
700
+ def __getitem__(self, i) -> Dict[str, torch.Tensor]:
701
+ sources = self.list_data_dict[i]
702
+ if isinstance(i, int):
703
+ sources = [sources]
704
+ assert len(sources) == 1, "Don't know why it is wrapped to a list" # FIXME
705
+ if 'image' in sources[0]:
706
+ image_file = self.list_data_dict[i]['image']
707
+ image_folder = self.data_args.image_folder
708
+ processor = self.data_args.image_processor
709
+ # print(image_folder, image_file)
710
+ image = Image.open(os.path.join(image_folder, image_file)).convert('RGB')
711
+ if self.data_args.image_aspect_ratio == 'pad':
712
+ def expand2square(pil_img, background_color):
713
+ width, height = pil_img.size
714
+ if width == height:
715
+ return pil_img
716
+ elif width > height:
717
+ result = Image.new(pil_img.mode, (width, width), background_color)
718
+ result.paste(pil_img, (0, (width - height) // 2))
719
+ return result
720
+ else:
721
+ result = Image.new(pil_img.mode, (height, height), background_color)
722
+ result.paste(pil_img, ((height - width) // 2, 0))
723
+ return result
724
+ image = expand2square(image, tuple(int(x*255) for x in processor.image_mean))
725
+ image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0] # [3, 336, 336]
726
+ else:
727
+ image = processor.preprocess(image, return_tensors='pt')['pixel_values'][0]
728
+ sources = preprocess_multimodal(
729
+ copy.deepcopy([e["conversations"] for e in sources]),
730
+ self.data_args)
731
+ else:
732
+ sources = copy.deepcopy([e["conversations"] for e in sources])
733
+ data_dict = preprocess(
734
+ sources,
735
+ self.tokenizer,
736
+ has_image=('image' in self.list_data_dict[i]))
737
+ if isinstance(i, int):
738
+ data_dict = dict(input_ids=data_dict["input_ids"][0],
739
+ labels=data_dict["labels"][0])
740
+
741
+ # image exist in the data
742
+ if 'image' in self.list_data_dict[i]:
743
+ data_dict['image'] = image
744
+ elif self.data_args.is_multimodal:
745
+ # image does not exist in the data, but the model is multimodal
746
+ crop_size = self.data_args.image_processor.crop_size
747
+ data_dict['image'] = torch.zeros(3, crop_size['height'], crop_size['width'])
748
+ return data_dict
749
+
750
+ @dataclass
751
+ class DataCollatorForSupervisedDataset(object):
752
+ """Collate examples for supervised fine-tuning."""
753
+
754
+ tokenizer: transformers.PreTrainedTokenizer
755
+
756
+ def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
757
+ input_ids, labels = tuple([instance[key] for instance in instances]
758
+ for key in ("input_ids", "labels"))
759
+ input_ids = torch.nn.utils.rnn.pad_sequence(
760
+ input_ids,
761
+ batch_first=True,
762
+ padding_value=self.tokenizer.pad_token_id)
763
+ labels = torch.nn.utils.rnn.pad_sequence(labels,
764
+ batch_first=True,
765
+ padding_value=IGNORE_INDEX)
766
+ input_ids = input_ids[:, :self.tokenizer.model_max_length]
767
+ labels = labels[:, :self.tokenizer.model_max_length]
768
+ batch = dict(
769
+ input_ids=input_ids,
770
+ labels=labels,
771
+ attention_mask=input_ids.ne(self.tokenizer.pad_token_id),
772
+ )
773
+
774
+ if 'image' in instances[0]:
775
+ images = [instance['image'] for instance in instances]
776
+ if all(x is not None and x.shape == images[0].shape for x in images):
777
+ batch['images'] = torch.stack(images)
778
+ else:
779
+ batch['images'] = images
780
+
781
+ return batch
782
+
783
+
784
+ def make_supervised_data_module(tokenizer: transformers.PreTrainedTokenizer,
785
+ data_args, voco_token) -> Dict:
786
+ """Make dataset and collator for supervised fine-tuning."""
787
+ train_dataset = LazySupervisedDataset(tokenizer=tokenizer,
788
+ data_path=data_args.data_path,
789
+ data_args=data_args,
790
+ voco_token=voco_token)
791
+ data_collator = DataCollatorForSupervisedDataset(tokenizer=tokenizer)
792
+ return dict(train_dataset=train_dataset,
793
+ eval_dataset=None,
794
+ data_collator=data_collator)
795
+
796
+
797
+ def train(attn_implementation=None):
798
+ global local_rank
799
+
800
+ parser = transformers.HfArgumentParser(
801
+ (ModelArguments, DataArguments, TrainingArguments))
802
+ model_args, data_args, training_args = parser.parse_args_into_dataclasses()
803
+ local_rank = training_args.local_rank
804
+ compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
805
+
806
+ bnb_model_from_pretrained_args = {}
807
+ if training_args.bits in [4, 8]:
808
+ from transformers import BitsAndBytesConfig
809
+ bnb_model_from_pretrained_args.update(dict(
810
+ device_map={"": training_args.device},
811
+ load_in_4bit=training_args.bits == 4,
812
+ load_in_8bit=training_args.bits == 8,
813
+ quantization_config=BitsAndBytesConfig(
814
+ load_in_4bit=training_args.bits == 4,
815
+ load_in_8bit=training_args.bits == 8,
816
+ llm_int8_skip_modules=["mm_projector"],
817
+ llm_int8_threshold=6.0,
818
+ llm_int8_has_fp16_weight=False,
819
+ bnb_4bit_compute_dtype=compute_dtype,
820
+ bnb_4bit_use_double_quant=training_args.double_quant,
821
+ bnb_4bit_quant_type=training_args.quant_type # {'fp4', 'nf4'}
822
+ )
823
+ ))
824
+
825
+ if model_args.vision_tower is not None:
826
+ if 'mpt' in model_args.model_name_or_path:
827
+ config = transformers.AutoConfig.from_pretrained(model_args.model_name_or_path, trust_remote_code=True)
828
+ config.attn_config['attn_impl'] = training_args.mpt_attn_impl
829
+ model = LlavaMptForCausalLM.from_pretrained(
830
+ model_args.model_name_or_path,
831
+ config=config,
832
+ cache_dir=training_args.cache_dir,
833
+ **bnb_model_from_pretrained_args
834
+ )
835
+ else:
836
+ print("use LlavaLlamaForCausalLM!!!")
837
+ model = LlavaLlamaForCausalLM.from_pretrained(
838
+ model_args.model_name_or_path,
839
+ cache_dir=training_args.cache_dir,
840
+ attn_implementation=attn_implementation,
841
+ torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
842
+ **bnb_model_from_pretrained_args
843
+ )
844
+ else:
845
+ model = transformers.LlamaForCausalLM.from_pretrained(
846
+ model_args.model_name_or_path,
847
+ cache_dir=training_args.cache_dir,
848
+ attn_implementation=attn_implementation,
849
+ torch_dtype=(torch.bfloat16 if training_args.bf16 else None),
850
+ **bnb_model_from_pretrained_args
851
+ )
852
+ model.config.use_cache = False
853
+
854
+ if model_args.freeze_backbone:
855
+ model.model.requires_grad_(False)
856
+
857
+ if training_args.bits in [4, 8]:
858
+ from peft import prepare_model_for_kbit_training
859
+ model.config.torch_dtype=(torch.float32 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
860
+ model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=training_args.gradient_checkpointing)
861
+
862
+ if training_args.gradient_checkpointing:
863
+ if hasattr(model, "enable_input_require_grads"):
864
+ model.enable_input_require_grads()
865
+ else:
866
+ def make_inputs_require_grad(module, input, output):
867
+ output.requires_grad_(True)
868
+ model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
869
+
870
+ if training_args.lora_enable:
871
+ from peft import LoraConfig, get_peft_model
872
+ lora_config = LoraConfig(
873
+ r=training_args.lora_r,
874
+ lora_alpha=training_args.lora_alpha,
875
+ target_modules=find_all_linear_names(model),
876
+ lora_dropout=training_args.lora_dropout,
877
+ bias=training_args.lora_bias,
878
+ task_type="CAUSAL_LM",
879
+ )
880
+ if training_args.bits == 16:
881
+ if training_args.bf16:
882
+ model.to(torch.bfloat16)
883
+ if training_args.fp16:
884
+ model.to(torch.float16)
885
+ rank0_print("Adding LoRA adapters...")
886
+ model = get_peft_model(model, lora_config)
887
+
888
+ if 'mpt' in model_args.model_name_or_path:
889
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
890
+ model_args.model_name_or_path,
891
+ cache_dir=training_args.cache_dir,
892
+ model_max_length=training_args.model_max_length,
893
+ padding_side="right"
894
+ )
895
+ else:
896
+ tokenizer = transformers.AutoTokenizer.from_pretrained(
897
+ model_args.model_name_or_path,
898
+ cache_dir=training_args.cache_dir,
899
+ model_max_length=training_args.model_max_length,
900
+ padding_side="left",
901
+ use_fast=False,
902
+ )
903
+
904
+ if model_args.version == "v0":
905
+ if tokenizer.pad_token is None:
906
+ smart_tokenizer_and_embedding_resize(
907
+ special_tokens_dict=dict(pad_token="[PAD]"),
908
+ tokenizer=tokenizer,
909
+ model=model,
910
+ )
911
+ elif model_args.version == "v0.5":
912
+ tokenizer.pad_token = tokenizer.unk_token
913
+ else:
914
+ tokenizer.pad_token = tokenizer.unk_token
915
+ if model_args.version in conversation_lib.conv_templates:
916
+ conversation_lib.default_conversation = conversation_lib.conv_templates[model_args.version]
917
+ else:
918
+ conversation_lib.default_conversation = conversation_lib.conv_templates["vicuna_v1"]
919
+
920
+ if len(tokenizer) == 32000 + 1:
921
+ assert (
922
+ model.model.embed_tokens.weight.shape[0]
923
+ == 32000 + 1
924
+ )
925
+ assert model.lm_head.weight.shape[0] == 32000 + 1
926
+ else:
927
+ print('add_voco_token 32001')
928
+ # Initialize voco token
929
+ tokenizer.add_special_tokens({"additional_special_tokens": ["<voco>"]})
930
+ model.resize_token_embeddings(len(tokenizer))
931
+ voco_token = tokenizer.additional_special_tokens_ids[-1]
932
+
933
+ if model_args.vision_tower is not None:
934
+ model.get_model().initialize_vision_modules(
935
+ model_args=model_args,
936
+ fsdp=training_args.fsdp
937
+ )
938
+
939
+ vision_tower = model.get_vision_tower()
940
+ vision_tower.to(dtype=torch.bfloat16 if training_args.bf16 else torch.float16, device=training_args.device)
941
+
942
+ data_args.image_processor = vision_tower.image_processor
943
+ data_args.is_multimodal = True
944
+
945
+ model.config.image_aspect_ratio = data_args.image_aspect_ratio
946
+ model.config.tokenizer_padding_side = tokenizer.padding_side
947
+ model.config.tokenizer_model_max_length = tokenizer.model_max_length
948
+
949
+ model.config.tune_mm_mlp_adapter = training_args.tune_mm_mlp_adapter = model_args.tune_mm_mlp_adapter
950
+ if model_args.tune_mm_mlp_adapter:
951
+ model.requires_grad_(False)
952
+ for p in model.get_model().mm_projector.parameters():
953
+ p.requires_grad = True
954
+
955
+ model.config.freeze_mm_mlp_adapter = training_args.freeze_mm_mlp_adapter
956
+ if training_args.freeze_mm_mlp_adapter:
957
+ for p in model.get_model().mm_projector.parameters():
958
+ p.requires_grad = False
959
+
960
+ if training_args.bits in [4, 8]:
961
+ model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device)
962
+
963
+ model.config.mm_use_im_start_end = data_args.mm_use_im_start_end = model_args.mm_use_im_start_end
964
+ model.config.mm_projector_lr = training_args.mm_projector_lr
965
+ training_args.use_im_start_end = model_args.mm_use_im_start_end
966
+ model.config.mm_use_im_patch_token = model_args.mm_use_im_patch_token
967
+ model.initialize_vision_tokenizer(model_args, tokenizer=tokenizer)
968
+
969
+ if training_args.bits in [4, 8]:
970
+ from peft.tuners.lora import LoraLayer
971
+ for name, module in model.named_modules():
972
+ if isinstance(module, LoraLayer):
973
+ if training_args.bf16:
974
+ module = module.to(torch.bfloat16)
975
+ if 'norm' in name:
976
+ module = module.to(torch.float32)
977
+ if 'lm_head' in name or 'embed_tokens' in name:
978
+ if hasattr(module, 'weight'):
979
+ if training_args.bf16 and module.weight.dtype == torch.float32:
980
+ module = module.to(torch.bfloat16)
981
+
982
+ data_module = make_supervised_data_module(tokenizer=tokenizer,
983
+ data_args=data_args,
984
+ voco_token=voco_token)
985
+ trainer = LLaVATrainer(model=model,
986
+ tokenizer=tokenizer,
987
+ args=training_args,
988
+ **data_module)
989
+ if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
990
+ trainer.train(resume_from_checkpoint=True)
991
+ else:
992
+ trainer.train()
993
+ trainer.save_state()
994
+
995
+ model.config.use_cache = True
996
+
997
+ if training_args.lora_enable:
998
+ state_dict = get_peft_state_maybe_zero_3(
999
+ model.named_parameters(), training_args.lora_bias
1000
+ )
1001
+ non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3(
1002
+ model.named_parameters()
1003
+ )
1004
+ if training_args.local_rank == 0 or training_args.local_rank == -1:
1005
+ model.config.save_pretrained(training_args.output_dir)
1006
+ model.save_pretrained(training_args.output_dir, state_dict=state_dict)
1007
+ torch.save(non_lora_state_dict, os.path.join(training_args.output_dir, 'non_lora_trainables.bin'))
1008
+ else:
1009
+ safe_save_model_for_hf_trainer(trainer=trainer,
1010
+ output_dir=training_args.output_dir)
1011
+
1012
+
1013
+ if __name__ == "__main__":
1014
+ train()
playground/data.z01 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e1ceb2163b22aa3902a15ce2437f706aa40fc1db138a7c4b716ed142198467cb
3
+ size 2147483648
playground/data.z02 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8e138f45d610b3d95a4fc0d80d58405dbc71ccde5aa40bd3f1cefd66732426cd
3
+ size 2147483648
playground/data.z04 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8ed07a3fd033e030566485904b51dad83a7c7ba33f84b7f695234268e4bc88f3
3
+ size 2147483648
playground/data.z05 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e05dd44b56489f31294ee72cd53584f8696989b28d004967e0f1a21346fcf0d8
3
+ size 2147483648
playground/data.z06 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:9d46ed522f29fd1f5c9005802169aadf02f79f33a47a1fc4d92f7ae164729a72
3
+ size 2147483648
playground/data.z07 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:5f497e6b0591cd77b9c98f347ddaa30033fc68d7c11d7285e438e4e15bac4030
3
+ size 2147483648
playground/data.z08 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:8b86e4ec1984deb85e4328a651d577864d52a2d26b671bf0e14c24e6b0c7355a
3
+ size 2147483648
playground/data.z09 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:73ebf3eaf895f6b2887fd31715476d5c002aa35fd7a639d827d3c62f849c5fb4
3
+ size 2147483648
playground/data.z10 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b3021d11b512ecef45411d98c050f4624c08e9b8ae9c1f92080193fba6ae2893
3
+ size 2147483648
playground/data.z11 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9a561fe234189d270e848c9a20bb333b571d8d9be301478506098a696146748
3
+ size 2147483648
playground/data.z12 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6b7fc0a9d324138c93d23e9625c05e93f848ebb10ab97f69adcd3a31df1edf9d
3
+ size 2147483648
playground/data.z14 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:a43c6a02297e630589bfdae80253abd36844e0a4bc10878f1605ac49111c14ce
3
+ size 2147483648
playground/data.z15 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:e2e6ba5d89dc588bbc9961b487e9dc9f7eb0179b7d8b349b4135f5a279741ec1
3
+ size 2147483648
playground/data.z16 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:688f74622bf5d37b662328ffe3ef7633832239e85e59dad53dd05755e6aae80f
3
+ size 2147483648
playground/data.z18 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:734c4ddaf5495e9d46a46f94f5e5c66ca5ce99726dbfb07f910343066d7a2a5e
3
+ size 2147483648
playground/data.z19 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:d4dab9b7cd28541bce4280501891d0f416dc5fae6c752f6cef953b319df3d2b5
3
+ size 2147483648
playground/data.z20 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:85e2c888eaae0ea32b2d7165df9396741f5de2e60538147b096ed8de1272761b
3
+ size 2147483648
playground/data.z21 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:b9d0a980e330cb5c3ba0b0b11c18e6f423e2e6e9a6dd471a3cc9a2964f07f49d
3
+ size 2147483648
playground/data.z23 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:2b73e57e422c0b13a245bffae386710156395896275be0f1a79f2d82f960cd9a
3
+ size 2147483648
playground/data.z24 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:78c020681da40773468f8acb64367b4a34434d387dd98e19d2bc92fb459547c7
3
+ size 2147483648
playground/data.z25 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:cf9201ad221bbea10b9ca674aa09749e9ac85c1aa38faef43ccb68635f6c2d00
3
+ size 2147483648
playground/data.z26 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:74f50f8cbad2263539c7abd33c238a7a5d3fee2925cacc14c939c5eab0abd345
3
+ size 2147483648
playground/data.z27 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:6e466a0f229e226b1f7375854e8828cf5f633c5c787e63affc75d56b965aaeda
3
+ size 2147483648
playground/data.z28 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:107c1a049083baad33211500368dbca8a309cceb28fe9c39d34203684b81cc53
3
+ size 2147483648
playground/data.z30 ADDED
@@ -0,0 +1,3 @@
 
 
 
 
1
+ version https://git-lfs.github.com/spec/v1
2
+ oid sha256:f114630e80b9f0f943340b98042f797f81b186d53e4ebb6e87f0a661a8bfaf15
3
+ size 2147483648