init
Browse files- README.md +14 -0
- bpe_simple_vocab_16e6.txt.gz +3 -0
- config.json +12 -0
- configuration_viclip.py +5 -0
- demo.ipynb +158 -0
- model.safetensors +3 -0
- simple_tokenizer.py +135 -0
- viclip.py +281 -0
- viclip_text.py +305 -0
- viclip_vision.py +362 -0
README.md
ADDED
|
@@ -0,0 +1,14 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
datasets:
|
| 3 |
+
- OpenGVLab/InternVid
|
| 4 |
+
base_model:
|
| 5 |
+
- openai/clip-vit-base-patch16
|
| 6 |
+
tags:
|
| 7 |
+
- ViCLIP
|
| 8 |
+
---
|
| 9 |
+
|
| 10 |
+
huggingface weight of ViCLIP
|
| 11 |
+
|
| 12 |
+
remember to set your `tokenizer_path` in config.json
|
| 13 |
+
|
| 14 |
+
usage is in demo.ipynb
|
bpe_simple_vocab_16e6.txt.gz
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:924691ac288e54409236115652ad4aa250f48203de50a9e4722a6ecd48d6804a
|
| 3 |
+
size 1356917
|
config.json
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"architectures": [
|
| 3 |
+
"ViCLIP"
|
| 4 |
+
],
|
| 5 |
+
"auto_map": {
|
| 6 |
+
"AutoConfig": "configuration_viclip.Config",
|
| 7 |
+
"AutoModel": "viclip.ViCLIP"
|
| 8 |
+
},
|
| 9 |
+
"torch_dtype": "float32",
|
| 10 |
+
"size":"b",
|
| 11 |
+
"tokenizer_path":"./bpe_simple_vocab_16e6.txt.gz"
|
| 12 |
+
}
|
configuration_viclip.py
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
from transformers import PretrainedConfig
|
| 2 |
+
|
| 3 |
+
class Config(PretrainedConfig):
|
| 4 |
+
def __init__(self, **kwargs):
|
| 5 |
+
super().__init__(**kwargs)
|
demo.ipynb
ADDED
|
@@ -0,0 +1,158 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"cells": [
|
| 3 |
+
{
|
| 4 |
+
"cell_type": "code",
|
| 5 |
+
"execution_count": null,
|
| 6 |
+
"id": "a436c0a1-3410-4a7f-a186-9246075ac815",
|
| 7 |
+
"metadata": {},
|
| 8 |
+
"outputs": [],
|
| 9 |
+
"source": [
|
| 10 |
+
"from transformers import AutoModel\n",
|
| 11 |
+
"model=AutoModel.from_pretrained(\"OpenGVLab/ViCLIP-B-16-hf\",trust_remote_code=True)\n",
|
| 12 |
+
"tokenizer = model.tokenizer\n",
|
| 13 |
+
"model_tokenizer={\"viclip\":model,\"tokenizer\":tokenizer}\n",
|
| 14 |
+
"print(\"done\")"
|
| 15 |
+
]
|
| 16 |
+
},
|
| 17 |
+
{
|
| 18 |
+
"cell_type": "code",
|
| 19 |
+
"execution_count": 3,
|
| 20 |
+
"id": "a425a5da-ceaf-4b89-9845-c8ba576902d8",
|
| 21 |
+
"metadata": {},
|
| 22 |
+
"outputs": [],
|
| 23 |
+
"source": [
|
| 24 |
+
"# video data\n",
|
| 25 |
+
"import numpy as np\n",
|
| 26 |
+
"import os\n",
|
| 27 |
+
"import cv2\n",
|
| 28 |
+
"import torch\n",
|
| 29 |
+
"def _frame_from_video(video):\n",
|
| 30 |
+
" while video.isOpened():\n",
|
| 31 |
+
" success, frame = video.read()\n",
|
| 32 |
+
" if success:\n",
|
| 33 |
+
" yield frame\n",
|
| 34 |
+
" else:\n",
|
| 35 |
+
" break\n",
|
| 36 |
+
"video = cv2.VideoCapture('example1.mp4')\n",
|
| 37 |
+
"frames = [x for x in _frame_from_video(video)]"
|
| 38 |
+
]
|
| 39 |
+
},
|
| 40 |
+
{
|
| 41 |
+
"cell_type": "code",
|
| 42 |
+
"execution_count": 5,
|
| 43 |
+
"id": "aac775ce",
|
| 44 |
+
"metadata": {},
|
| 45 |
+
"outputs": [],
|
| 46 |
+
"source": [
|
| 47 |
+
"# function\n",
|
| 48 |
+
"\n",
|
| 49 |
+
"def get_text_feat_dict(texts, clip, tokenizer, text_feat_d={}):\n",
|
| 50 |
+
" for t in texts:\n",
|
| 51 |
+
" feat = clip.get_text_features(t, tokenizer, text_feat_d)\n",
|
| 52 |
+
" text_feat_d[t] = feat\n",
|
| 53 |
+
" return text_feat_d\n",
|
| 54 |
+
"\n",
|
| 55 |
+
"def get_vid_feat(frames, clip):\n",
|
| 56 |
+
" return clip.get_vid_features(frames)\n",
|
| 57 |
+
"\n",
|
| 58 |
+
"v_mean = np.array([0.485, 0.456, 0.406]).reshape(1,1,3)\n",
|
| 59 |
+
"v_std = np.array([0.229, 0.224, 0.225]).reshape(1,1,3)\n",
|
| 60 |
+
"def normalize(data):\n",
|
| 61 |
+
" return (data/255.0-v_mean)/v_std\n",
|
| 62 |
+
"\n",
|
| 63 |
+
"def frames2tensor(vid_list, fnum=8, target_size=(224, 224), device=torch.device('cuda')):\n",
|
| 64 |
+
" assert(len(vid_list) >= fnum)\n",
|
| 65 |
+
" step = len(vid_list) // fnum\n",
|
| 66 |
+
" vid_list = vid_list[::step][:fnum]\n",
|
| 67 |
+
" vid_list = [cv2.resize(x[:,:,::-1], target_size) for x in vid_list]\n",
|
| 68 |
+
" vid_tube = [np.expand_dims(normalize(x), axis=(0, 1)) for x in vid_list]\n",
|
| 69 |
+
" vid_tube = np.concatenate(vid_tube, axis=1)\n",
|
| 70 |
+
" vid_tube = np.transpose(vid_tube, (0, 1, 4, 2, 3))\n",
|
| 71 |
+
" vid_tube = torch.from_numpy(vid_tube).to(device, non_blocking=True).float()\n",
|
| 72 |
+
" return vid_tube\n",
|
| 73 |
+
"def retrieve_text(frames, \n",
|
| 74 |
+
" texts, \n",
|
| 75 |
+
" models={'viclip':None, \n",
|
| 76 |
+
" 'tokenizer':None},\n",
|
| 77 |
+
" topk=5, \n",
|
| 78 |
+
" device=torch.device('cuda')):\n",
|
| 79 |
+
" # clip, tokenizer = get_clip(name, model_cfg['size'], model_cfg['pretrained'], model_cfg['reload'])\n",
|
| 80 |
+
" assert(type(models)==dict and models['viclip'] is not None and models['tokenizer'] is not None)\n",
|
| 81 |
+
" clip, tokenizer = models['viclip'], models['tokenizer']\n",
|
| 82 |
+
" clip = clip.to(device)\n",
|
| 83 |
+
" frames_tensor = frames2tensor(frames, device=device)\n",
|
| 84 |
+
" vid_feat = get_vid_feat(frames_tensor, clip)\n",
|
| 85 |
+
"\n",
|
| 86 |
+
" text_feat_d = {}\n",
|
| 87 |
+
" text_feat_d = get_text_feat_dict(texts, clip, tokenizer, text_feat_d)\n",
|
| 88 |
+
" text_feats = [text_feat_d[t] for t in texts]\n",
|
| 89 |
+
" text_feats_tensor = torch.cat(text_feats, 0)\n",
|
| 90 |
+
" \n",
|
| 91 |
+
" probs, idxs = clip.get_predict_label(vid_feat, text_feats_tensor, top=topk)\n",
|
| 92 |
+
"\n",
|
| 93 |
+
" ret_texts = [texts[i] for i in idxs.numpy()[0].tolist()]\n",
|
| 94 |
+
" return ret_texts, probs.numpy()[0]"
|
| 95 |
+
]
|
| 96 |
+
},
|
| 97 |
+
{
|
| 98 |
+
"cell_type": "code",
|
| 99 |
+
"execution_count": null,
|
| 100 |
+
"id": "a2969ba6-19d0-4893-b071-b82fa046c312",
|
| 101 |
+
"metadata": {},
|
| 102 |
+
"outputs": [],
|
| 103 |
+
"source": [
|
| 104 |
+
"# retrieval\n",
|
| 105 |
+
"text_candidates = [\"A playful dog and its owner wrestle in the snowy yard, chasing each other with joyous abandon.\",\n",
|
| 106 |
+
" \"A man in a gray coat walks through the snowy landscape, pulling a sleigh loaded with toys.\",\n",
|
| 107 |
+
" \"A person dressed in a blue jacket shovels the snow-covered pavement outside their house.\",\n",
|
| 108 |
+
" \"A pet dog excitedly runs through the snowy yard, chasing a toy thrown by its owner.\",\n",
|
| 109 |
+
" \"A person stands on the snowy floor, pushing a sled loaded with blankets, preparing for a fun-filled ride.\",\n",
|
| 110 |
+
" \"A man in a gray hat and coat walks through the snowy yard, carefully navigating around the trees.\",\n",
|
| 111 |
+
" \"A playful dog slides down a snowy hill, wagging its tail with delight.\",\n",
|
| 112 |
+
" \"A person in a blue jacket walks their pet on a leash, enjoying a peaceful winter walk among the trees.\",\n",
|
| 113 |
+
" \"A man in a gray sweater plays fetch with his dog in the snowy yard, throwing a toy and watching it run.\",\n",
|
| 114 |
+
" \"A person bundled up in a blanket walks through the snowy landscape, enjoying the serene winter scenery.\"]\n",
|
| 115 |
+
"texts, probs = retrieve_text(frames, text_candidates, models=model_tokenizer, topk=5)\n",
|
| 116 |
+
"\n",
|
| 117 |
+
"for t, p in zip(texts, probs):\n",
|
| 118 |
+
" print(f'text: {t} ~ prob: {p:.4f}')\n",
|
| 119 |
+
" \n",
|
| 120 |
+
"\n",
|
| 121 |
+
"# text: A playful dog and its owner wrestle in the snowy yard, chasing each other with joyous abandon. ~ prob: 0.8192\n",
|
| 122 |
+
"# text: A man in a gray sweater plays fetch with his dog in the snowy yard, throwing a toy and watching it run. ~ prob: 0.1084\n",
|
| 123 |
+
"# text: A pet dog excitedly runs through the snowy yard, chasing a toy thrown by its owner. ~ prob: 0.0676\n",
|
| 124 |
+
"# text: A playful dog slides down a snowy hill, wagging its tail with delight. ~ prob: 0.0047\n",
|
| 125 |
+
"# text: A person dressed in a blue jacket shovels the snow-covered pavement outside their house. ~ prob: 0.0002"
|
| 126 |
+
]
|
| 127 |
+
},
|
| 128 |
+
{
|
| 129 |
+
"cell_type": "code",
|
| 130 |
+
"execution_count": null,
|
| 131 |
+
"id": "84922de7-b41c-41c1-87a0-b28e52da9b5d",
|
| 132 |
+
"metadata": {},
|
| 133 |
+
"outputs": [],
|
| 134 |
+
"source": []
|
| 135 |
+
}
|
| 136 |
+
],
|
| 137 |
+
"metadata": {
|
| 138 |
+
"kernelspec": {
|
| 139 |
+
"display_name": "Python 3 (ipykernel)",
|
| 140 |
+
"language": "python",
|
| 141 |
+
"name": "python3"
|
| 142 |
+
},
|
| 143 |
+
"language_info": {
|
| 144 |
+
"codemirror_mode": {
|
| 145 |
+
"name": "ipython",
|
| 146 |
+
"version": 3
|
| 147 |
+
},
|
| 148 |
+
"file_extension": ".py",
|
| 149 |
+
"mimetype": "text/x-python",
|
| 150 |
+
"name": "python",
|
| 151 |
+
"nbconvert_exporter": "python",
|
| 152 |
+
"pygments_lexer": "ipython3",
|
| 153 |
+
"version": "3.10.4"
|
| 154 |
+
}
|
| 155 |
+
},
|
| 156 |
+
"nbformat": 4,
|
| 157 |
+
"nbformat_minor": 5
|
| 158 |
+
}
|
model.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:af9cfba8b30a4d62fec6bf7a033f748514ae86d04bd53b3a47f17dd4c7af2741
|
| 3 |
+
size 598452684
|
simple_tokenizer.py
ADDED
|
@@ -0,0 +1,135 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gzip
|
| 2 |
+
import html
|
| 3 |
+
import os
|
| 4 |
+
from functools import lru_cache
|
| 5 |
+
|
| 6 |
+
import ftfy
|
| 7 |
+
import regex as re
|
| 8 |
+
|
| 9 |
+
|
| 10 |
+
@lru_cache()
|
| 11 |
+
def default_bpe():
|
| 12 |
+
return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz")
|
| 13 |
+
# @lru_cache()
|
| 14 |
+
# def default_bpe():
|
| 15 |
+
# return "bpe_simple_vocab_16e6.txt.gz"
|
| 16 |
+
|
| 17 |
+
|
| 18 |
+
@lru_cache()
|
| 19 |
+
def bytes_to_unicode():
|
| 20 |
+
"""
|
| 21 |
+
Returns list of utf-8 byte and a corresponding list of unicode strings.
|
| 22 |
+
The reversible bpe codes work on unicode strings.
|
| 23 |
+
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
|
| 24 |
+
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
|
| 25 |
+
This is a signficant percentage of your normal, say, 32K bpe vocab.
|
| 26 |
+
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
|
| 27 |
+
And avoids mapping to whitespace/control characters the bpe code barfs on.
|
| 28 |
+
"""
|
| 29 |
+
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
|
| 30 |
+
cs = bs[:]
|
| 31 |
+
n = 0
|
| 32 |
+
for b in range(2**8):
|
| 33 |
+
if b not in bs:
|
| 34 |
+
bs.append(b)
|
| 35 |
+
cs.append(2**8+n)
|
| 36 |
+
n += 1
|
| 37 |
+
cs = [chr(n) for n in cs]
|
| 38 |
+
return dict(zip(bs, cs))
|
| 39 |
+
|
| 40 |
+
|
| 41 |
+
def get_pairs(word):
|
| 42 |
+
"""Return set of symbol pairs in a word.
|
| 43 |
+
Word is represented as tuple of symbols (symbols being variable-length strings).
|
| 44 |
+
"""
|
| 45 |
+
pairs = set()
|
| 46 |
+
prev_char = word[0]
|
| 47 |
+
for char in word[1:]:
|
| 48 |
+
pairs.add((prev_char, char))
|
| 49 |
+
prev_char = char
|
| 50 |
+
return pairs
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
def basic_clean(text):
|
| 54 |
+
text = ftfy.fix_text(text)
|
| 55 |
+
text = html.unescape(html.unescape(text))
|
| 56 |
+
return text.strip()
|
| 57 |
+
|
| 58 |
+
|
| 59 |
+
def whitespace_clean(text):
|
| 60 |
+
text = re.sub(r'\s+', ' ', text)
|
| 61 |
+
text = text.strip()
|
| 62 |
+
return text
|
| 63 |
+
|
| 64 |
+
|
| 65 |
+
class SimpleTokenizer(object):
|
| 66 |
+
def __init__(self, bpe_path: str = default_bpe()):
|
| 67 |
+
self.byte_encoder = bytes_to_unicode()
|
| 68 |
+
self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
|
| 69 |
+
merges = gzip.open(bpe_path).read().decode("utf-8").split('\n')
|
| 70 |
+
merges = merges[1:49152-256-2+1]
|
| 71 |
+
merges = [tuple(merge.split()) for merge in merges]
|
| 72 |
+
vocab = list(bytes_to_unicode().values())
|
| 73 |
+
vocab = vocab + [v+'</w>' for v in vocab]
|
| 74 |
+
for merge in merges:
|
| 75 |
+
vocab.append(''.join(merge))
|
| 76 |
+
vocab.extend(['<|startoftext|>', '<|endoftext|>'])
|
| 77 |
+
self.encoder = dict(zip(vocab, range(len(vocab))))
|
| 78 |
+
self.decoder = {v: k for k, v in self.encoder.items()}
|
| 79 |
+
self.bpe_ranks = dict(zip(merges, range(len(merges))))
|
| 80 |
+
self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
|
| 81 |
+
self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""", re.IGNORECASE)
|
| 82 |
+
|
| 83 |
+
def bpe(self, token):
|
| 84 |
+
if token in self.cache:
|
| 85 |
+
return self.cache[token]
|
| 86 |
+
word = tuple(token[:-1]) + ( token[-1] + '</w>',)
|
| 87 |
+
pairs = get_pairs(word)
|
| 88 |
+
|
| 89 |
+
if not pairs:
|
| 90 |
+
return token+'</w>'
|
| 91 |
+
|
| 92 |
+
while True:
|
| 93 |
+
bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
|
| 94 |
+
if bigram not in self.bpe_ranks:
|
| 95 |
+
break
|
| 96 |
+
first, second = bigram
|
| 97 |
+
new_word = []
|
| 98 |
+
i = 0
|
| 99 |
+
while i < len(word):
|
| 100 |
+
try:
|
| 101 |
+
j = word.index(first, i)
|
| 102 |
+
new_word.extend(word[i:j])
|
| 103 |
+
i = j
|
| 104 |
+
except:
|
| 105 |
+
new_word.extend(word[i:])
|
| 106 |
+
break
|
| 107 |
+
|
| 108 |
+
if word[i] == first and i < len(word)-1 and word[i+1] == second:
|
| 109 |
+
new_word.append(first+second)
|
| 110 |
+
i += 2
|
| 111 |
+
else:
|
| 112 |
+
new_word.append(word[i])
|
| 113 |
+
i += 1
|
| 114 |
+
new_word = tuple(new_word)
|
| 115 |
+
word = new_word
|
| 116 |
+
if len(word) == 1:
|
| 117 |
+
break
|
| 118 |
+
else:
|
| 119 |
+
pairs = get_pairs(word)
|
| 120 |
+
word = ' '.join(word)
|
| 121 |
+
self.cache[token] = word
|
| 122 |
+
return word
|
| 123 |
+
|
| 124 |
+
def encode(self, text):
|
| 125 |
+
bpe_tokens = []
|
| 126 |
+
text = whitespace_clean(basic_clean(text)).lower()
|
| 127 |
+
for token in re.findall(self.pat, text):
|
| 128 |
+
token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
|
| 129 |
+
bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
|
| 130 |
+
return bpe_tokens
|
| 131 |
+
|
| 132 |
+
def decode(self, tokens):
|
| 133 |
+
text = ''.join([self.decoder[token] for token in tokens])
|
| 134 |
+
text = bytearray([self.byte_decoder[c] for c in text]).decode('utf-8', errors="replace").replace('</w>', ' ')
|
| 135 |
+
return text
|
viclip.py
ADDED
|
@@ -0,0 +1,281 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import logging
|
| 3 |
+
|
| 4 |
+
import torch
|
| 5 |
+
from einops import rearrange
|
| 6 |
+
from torch import nn
|
| 7 |
+
import math
|
| 8 |
+
|
| 9 |
+
# from .criterions import VTC_VTM_Loss
|
| 10 |
+
from .simple_tokenizer import SimpleTokenizer as _Tokenizer
|
| 11 |
+
from .viclip_vision import clip_joint_l14, clip_joint_b16
|
| 12 |
+
from .viclip_text import clip_text_l14, clip_text_b16
|
| 13 |
+
|
| 14 |
+
# from transformers import AutoModel
|
| 15 |
+
from transformers import PreTrainedModel #new
|
| 16 |
+
from transformers import PretrainedConfig
|
| 17 |
+
|
| 18 |
+
logger = logging.getLogger(__name__)
|
| 19 |
+
|
| 20 |
+
from .configuration_viclip import Config
|
| 21 |
+
# class ViCLIP(nn.Module):
|
| 22 |
+
class ViCLIP(PreTrainedModel):
|
| 23 |
+
_auto_class="AutoModel"
|
| 24 |
+
config_class=Config
|
| 25 |
+
|
| 26 |
+
def __init__(self,
|
| 27 |
+
# tokenizer=None, # config:PretrainedConfig is the only parameter
|
| 28 |
+
# size='l',
|
| 29 |
+
# pretrain=None,
|
| 30 |
+
# freeze_text=True,
|
| 31 |
+
config=PretrainedConfig()):
|
| 32 |
+
super(ViCLIP, self).__init__(config)
|
| 33 |
+
self.config=config
|
| 34 |
+
if 'size' in config.to_dict(): ###########
|
| 35 |
+
size=config.size
|
| 36 |
+
pretrain=None
|
| 37 |
+
tokenizer_path=config.tokenizer_path
|
| 38 |
+
tokenizer=None
|
| 39 |
+
freeze_text=True
|
| 40 |
+
|
| 41 |
+
if tokenizer:
|
| 42 |
+
self.tokenizer = tokenizer
|
| 43 |
+
elif tokenizer_path:
|
| 44 |
+
self.tokenizer = _Tokenizer(tokenizer_path)
|
| 45 |
+
else:
|
| 46 |
+
self.tokenizer = _Tokenizer()
|
| 47 |
+
self.max_txt_l = 32
|
| 48 |
+
|
| 49 |
+
if size.lower() == 'l':
|
| 50 |
+
self.vision_encoder_name = 'vit_l14'
|
| 51 |
+
elif size.lower() == 'b':
|
| 52 |
+
self.vision_encoder_name = 'vit_b16'
|
| 53 |
+
else:
|
| 54 |
+
raise NotImplementedError(f"Size {size} not implemented")
|
| 55 |
+
|
| 56 |
+
self.vision_encoder_pretrained = False
|
| 57 |
+
self.inputs_image_res = 224
|
| 58 |
+
self.vision_encoder_kernel_size = 1
|
| 59 |
+
self.vision_encoder_center = True
|
| 60 |
+
self.video_input_num_frames = 8
|
| 61 |
+
self.vision_encoder_drop_path_rate = 0.1
|
| 62 |
+
self.vision_encoder_checkpoint_num = 24
|
| 63 |
+
self.is_pretrain = pretrain
|
| 64 |
+
self.vision_width = 1024
|
| 65 |
+
self.text_width = 768
|
| 66 |
+
self.embed_dim = 768
|
| 67 |
+
self.masking_prob = 0.9
|
| 68 |
+
|
| 69 |
+
if size.lower() == 'l':
|
| 70 |
+
self.text_encoder_name = 'vit_l14'
|
| 71 |
+
elif size.lower() == 'b':
|
| 72 |
+
self.text_encoder_name = 'vit_b16'
|
| 73 |
+
else:
|
| 74 |
+
raise NotImplementedError(f"Size {size} not implemented")
|
| 75 |
+
|
| 76 |
+
self.text_encoder_pretrained = False#'bert-base-uncased'
|
| 77 |
+
self.text_encoder_d_model = 768
|
| 78 |
+
|
| 79 |
+
self.text_encoder_vocab_size = 49408
|
| 80 |
+
|
| 81 |
+
# create modules.
|
| 82 |
+
self.vision_encoder = self.build_vision_encoder()
|
| 83 |
+
self.text_encoder = self.build_text_encoder()
|
| 84 |
+
|
| 85 |
+
self.temp = nn.parameter.Parameter(torch.ones([]) * 1 / 100.0)
|
| 86 |
+
self.temp_min = 1 / 100.0
|
| 87 |
+
|
| 88 |
+
if pretrain:
|
| 89 |
+
logger.info(f"Load pretrained weights from {pretrain}")
|
| 90 |
+
state_dict = torch.load(pretrain, map_location='cpu')['model']
|
| 91 |
+
self.load_state_dict(state_dict)
|
| 92 |
+
|
| 93 |
+
# Freeze weights
|
| 94 |
+
if freeze_text:
|
| 95 |
+
self.freeze_text()
|
| 96 |
+
|
| 97 |
+
|
| 98 |
+
def freeze_text(self):
|
| 99 |
+
"""freeze text encoder"""
|
| 100 |
+
for p in self.text_encoder.parameters():
|
| 101 |
+
p.requires_grad = False
|
| 102 |
+
|
| 103 |
+
def no_weight_decay(self):
|
| 104 |
+
ret = {"temp"}
|
| 105 |
+
ret.update(
|
| 106 |
+
{"vision_encoder." + k for k in self.vision_encoder.no_weight_decay()}
|
| 107 |
+
)
|
| 108 |
+
ret.update(
|
| 109 |
+
{"text_encoder." + k for k in self.text_encoder.no_weight_decay()}
|
| 110 |
+
)
|
| 111 |
+
|
| 112 |
+
return ret
|
| 113 |
+
|
| 114 |
+
def forward(self, image, text, raw_text, idx, log_generation=None, return_sims=False):
|
| 115 |
+
"""forward and calculate loss.
|
| 116 |
+
|
| 117 |
+
Args:
|
| 118 |
+
image (torch.Tensor): The input images. Shape: [B,T,C,H,W].
|
| 119 |
+
text (dict): TODO
|
| 120 |
+
idx (torch.Tensor): TODO
|
| 121 |
+
|
| 122 |
+
Returns: TODO
|
| 123 |
+
|
| 124 |
+
"""
|
| 125 |
+
self.clip_contrastive_temperature()
|
| 126 |
+
|
| 127 |
+
vision_embeds = self.encode_vision(image)
|
| 128 |
+
text_embeds = self.encode_text(raw_text)
|
| 129 |
+
if return_sims:
|
| 130 |
+
sims = torch.nn.functional.normalize(vision_embeds, dim=-1) @ \
|
| 131 |
+
torch.nn.functional.normalize(text_embeds, dim=-1).transpose(0, 1)
|
| 132 |
+
return sims
|
| 133 |
+
|
| 134 |
+
# calculate loss
|
| 135 |
+
|
| 136 |
+
## VTC loss
|
| 137 |
+
loss_vtc = self.clip_loss.vtc_loss(
|
| 138 |
+
vision_embeds, text_embeds, idx, self.temp, all_gather=True
|
| 139 |
+
)
|
| 140 |
+
|
| 141 |
+
return dict(
|
| 142 |
+
loss_vtc=loss_vtc,
|
| 143 |
+
)
|
| 144 |
+
|
| 145 |
+
def encode_vision(self, image, test=False):
|
| 146 |
+
"""encode image / videos as features.
|
| 147 |
+
|
| 148 |
+
Args:
|
| 149 |
+
image (torch.Tensor): The input images.
|
| 150 |
+
test (bool): Whether testing.
|
| 151 |
+
|
| 152 |
+
Returns: tuple.
|
| 153 |
+
- vision_embeds (torch.Tensor): The features of all patches. Shape: [B,T,L,C].
|
| 154 |
+
- pooled_vision_embeds (torch.Tensor): The pooled features. Shape: [B,T,C].
|
| 155 |
+
|
| 156 |
+
"""
|
| 157 |
+
if image.ndim == 5:
|
| 158 |
+
image = image.permute(0, 2, 1, 3, 4).contiguous()
|
| 159 |
+
else:
|
| 160 |
+
image = image.unsqueeze(2)
|
| 161 |
+
|
| 162 |
+
if not test and self.masking_prob > 0.0:
|
| 163 |
+
return self.vision_encoder(
|
| 164 |
+
image, masking_prob=self.masking_prob
|
| 165 |
+
)
|
| 166 |
+
|
| 167 |
+
return self.vision_encoder(image)
|
| 168 |
+
|
| 169 |
+
def encode_text(self, text):
|
| 170 |
+
"""encode text.
|
| 171 |
+
Args:
|
| 172 |
+
text (dict): The output of huggingface's `PreTrainedTokenizer`. contains keys:
|
| 173 |
+
- input_ids (torch.Tensor): Token ids to be fed to a model. Shape: [B,L].
|
| 174 |
+
- attention_mask (torch.Tensor): The mask indicate padded tokens. Shape: [B,L]. 0 is padded token.
|
| 175 |
+
- other keys refer to "https://huggingface.co/docs/transformers/v4.21.2/en/main_classes/tokenizer#transformers.PreTrainedTokenizer.__call__".
|
| 176 |
+
Returns: tuple.
|
| 177 |
+
- text_embeds (torch.Tensor): The features of all tokens. Shape: [B,L,C].
|
| 178 |
+
- pooled_text_embeds (torch.Tensor): The pooled features. Shape: [B,C].
|
| 179 |
+
|
| 180 |
+
"""
|
| 181 |
+
device = next(self.text_encoder.parameters()).device
|
| 182 |
+
text = self.text_encoder.tokenize(
|
| 183 |
+
text, context_length=self.max_txt_l
|
| 184 |
+
).to(device)
|
| 185 |
+
text_embeds = self.text_encoder(text)
|
| 186 |
+
return text_embeds
|
| 187 |
+
|
| 188 |
+
@torch.no_grad()
|
| 189 |
+
def clip_contrastive_temperature(self, min_val=0.001, max_val=0.5):
|
| 190 |
+
"""Seems only used during pre-training"""
|
| 191 |
+
self.temp.clamp_(min=self.temp_min)
|
| 192 |
+
|
| 193 |
+
def build_vision_encoder(self):
|
| 194 |
+
"""build vision encoder
|
| 195 |
+
Returns: (vision_encoder, vision_layernorm). Each is a `nn.Module`.
|
| 196 |
+
|
| 197 |
+
"""
|
| 198 |
+
encoder_name = self.vision_encoder_name
|
| 199 |
+
if encoder_name == "vit_l14":
|
| 200 |
+
vision_encoder = clip_joint_l14(
|
| 201 |
+
pretrained=self.vision_encoder_pretrained,
|
| 202 |
+
input_resolution=self.inputs_image_res,
|
| 203 |
+
kernel_size=self.vision_encoder_kernel_size,
|
| 204 |
+
center=self.vision_encoder_center,
|
| 205 |
+
num_frames=self.video_input_num_frames,
|
| 206 |
+
drop_path=self.vision_encoder_drop_path_rate,
|
| 207 |
+
checkpoint_num=self.vision_encoder_checkpoint_num,
|
| 208 |
+
)
|
| 209 |
+
elif encoder_name == "vit_b16":
|
| 210 |
+
vision_encoder = clip_joint_b16(
|
| 211 |
+
pretrained=self.vision_encoder_pretrained,
|
| 212 |
+
input_resolution=self.inputs_image_res,
|
| 213 |
+
kernel_size=self.vision_encoder_kernel_size,
|
| 214 |
+
center=self.vision_encoder_center,
|
| 215 |
+
num_frames=self.video_input_num_frames,
|
| 216 |
+
drop_path=self.vision_encoder_drop_path_rate,
|
| 217 |
+
checkpoint_num=self.vision_encoder_checkpoint_num,
|
| 218 |
+
)
|
| 219 |
+
else:
|
| 220 |
+
raise NotImplementedError(f"Not implemented: {encoder_name}")
|
| 221 |
+
|
| 222 |
+
return vision_encoder
|
| 223 |
+
|
| 224 |
+
def build_text_encoder(self):
|
| 225 |
+
"""build text_encoder and possiblly video-to-text multimodal fusion encoder.
|
| 226 |
+
Returns: nn.Module. The text encoder
|
| 227 |
+
|
| 228 |
+
"""
|
| 229 |
+
encoder_name = self.text_encoder_name
|
| 230 |
+
|
| 231 |
+
if encoder_name == "vit_l14":
|
| 232 |
+
text_encoder = clip_text_l14(
|
| 233 |
+
pretrained=self.text_encoder_pretrained,
|
| 234 |
+
context_length=self.max_txt_l,
|
| 235 |
+
vocab_size=self.text_encoder_vocab_size,
|
| 236 |
+
checkpoint_num=0,
|
| 237 |
+
tokenizer_path=None if not 'tokenizer_path' in self.config.to_dict() else self.config.tokenizer_path
|
| 238 |
+
)
|
| 239 |
+
elif encoder_name == "vit_b16":
|
| 240 |
+
text_encoder = clip_text_b16(
|
| 241 |
+
pretrained=self.text_encoder_pretrained,
|
| 242 |
+
context_length=self.max_txt_l,
|
| 243 |
+
vocab_size=self.text_encoder_vocab_size,
|
| 244 |
+
checkpoint_num=0,
|
| 245 |
+
tokenizer_path=None if not 'tokenizer_path' in self.config.to_dict() else self.config.tokenizer_path
|
| 246 |
+
)
|
| 247 |
+
else:
|
| 248 |
+
raise NotImplementedError(f"Not implemented: {encoder_name}")
|
| 249 |
+
|
| 250 |
+
return text_encoder
|
| 251 |
+
|
| 252 |
+
def get_text_encoder(self):
|
| 253 |
+
"""get text encoder, used for text and cross-modal encoding"""
|
| 254 |
+
encoder = self.text_encoder
|
| 255 |
+
return encoder.bert if hasattr(encoder, "bert") else encoder
|
| 256 |
+
|
| 257 |
+
def get_text_features(self, input_text, tokenizer, text_feature_dict={}):
|
| 258 |
+
if input_text in text_feature_dict:
|
| 259 |
+
return text_feature_dict[input_text]
|
| 260 |
+
text_template= f"{input_text}"
|
| 261 |
+
with torch.no_grad():
|
| 262 |
+
# text_token = tokenizer.encode(text_template).cuda()
|
| 263 |
+
text_features = self.encode_text(text_template).float()
|
| 264 |
+
text_features /= text_features.norm(dim=-1, keepdim=True)
|
| 265 |
+
text_feature_dict[input_text] = text_features
|
| 266 |
+
return text_features
|
| 267 |
+
|
| 268 |
+
def get_vid_features(self, input_frames):
|
| 269 |
+
with torch.no_grad():
|
| 270 |
+
clip_feat = self.encode_vision(input_frames,test=True).float()
|
| 271 |
+
clip_feat /= clip_feat.norm(dim=-1, keepdim=True)
|
| 272 |
+
return clip_feat
|
| 273 |
+
|
| 274 |
+
def get_predict_label(self, clip_feature, text_feats_tensor, top=5):
|
| 275 |
+
label_probs = (100.0 * clip_feature @ text_feats_tensor.T).softmax(dim=-1)
|
| 276 |
+
top_probs, top_labels = label_probs.cpu().topk(top, dim=-1)
|
| 277 |
+
return top_probs, top_labels
|
| 278 |
+
|
| 279 |
+
|
| 280 |
+
if __name__ =="__main__":
|
| 281 |
+
tokenizer = _Tokenizer()
|
viclip_text.py
ADDED
|
@@ -0,0 +1,305 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
import logging
|
| 3 |
+
from collections import OrderedDict
|
| 4 |
+
from pkg_resources import packaging
|
| 5 |
+
from .simple_tokenizer import SimpleTokenizer as _Tokenizer
|
| 6 |
+
|
| 7 |
+
import numpy as np
|
| 8 |
+
import torch
|
| 9 |
+
import torch.nn.functional as F
|
| 10 |
+
from torch import nn
|
| 11 |
+
import torch.utils.checkpoint as checkpoint
|
| 12 |
+
import functools
|
| 13 |
+
|
| 14 |
+
logger = logging.getLogger(__name__)
|
| 15 |
+
|
| 16 |
+
|
| 17 |
+
# On P1, model extracted from https://huggingface.co/laion/CLIP-ViT-L-14-DataComp.XL-s13B-b90K
|
| 18 |
+
MODEL_PATH = 'https://huggingface.co/laion'
|
| 19 |
+
_MODELS = {
|
| 20 |
+
"ViT-L/14": os.path.join(MODEL_PATH, "CLIP-ViT-L-14-DataComp.XL-s13B-b90K", "vit_l14_text.pth"),
|
| 21 |
+
"ViT-B/16": os.path.join(MODEL_PATH, "CLIP-ViT-B-16-DataComp.XL-s13B-b90K", "vit_b16_text.pth"),
|
| 22 |
+
}
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class LayerNorm(nn.LayerNorm):
|
| 26 |
+
"""Subclass torch's LayerNorm to handle fp16."""
|
| 27 |
+
|
| 28 |
+
def forward(self, x: torch.Tensor):
|
| 29 |
+
orig_type = x.dtype
|
| 30 |
+
ret = super().forward(x.type(torch.float32))
|
| 31 |
+
return ret.type(orig_type)
|
| 32 |
+
|
| 33 |
+
|
| 34 |
+
class QuickGELU(nn.Module):
|
| 35 |
+
def forward(self, x: torch.Tensor):
|
| 36 |
+
return x * torch.sigmoid(1.702 * x)
|
| 37 |
+
|
| 38 |
+
|
| 39 |
+
class ResidualAttentionBlock(nn.Module):
|
| 40 |
+
def __init__(self, d_model: int, n_head: int, attn_mask: torch.Tensor = None):
|
| 41 |
+
super().__init__()
|
| 42 |
+
|
| 43 |
+
self.attn = nn.MultiheadAttention(d_model, n_head)
|
| 44 |
+
self.ln_1 = LayerNorm(d_model)
|
| 45 |
+
self.mlp = nn.Sequential(OrderedDict([
|
| 46 |
+
("c_fc", nn.Linear(d_model, d_model * 4)),
|
| 47 |
+
("gelu", QuickGELU()),
|
| 48 |
+
("c_proj", nn.Linear(d_model * 4, d_model))
|
| 49 |
+
]))
|
| 50 |
+
self.ln_2 = LayerNorm(d_model)
|
| 51 |
+
self.attn_mask = attn_mask
|
| 52 |
+
|
| 53 |
+
def attention(self, x: torch.Tensor):
|
| 54 |
+
self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
|
| 55 |
+
return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
|
| 56 |
+
|
| 57 |
+
def forward(self, x: torch.Tensor):
|
| 58 |
+
x = x + self.attention(self.ln_1(x))
|
| 59 |
+
x = x + self.mlp(self.ln_2(x))
|
| 60 |
+
return x
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
class Transformer(nn.Module):
|
| 64 |
+
def __init__(self, width: int, layers: int, heads: int, attn_mask: torch.Tensor = None,
|
| 65 |
+
checkpoint_num: int = 0):
|
| 66 |
+
super().__init__()
|
| 67 |
+
self.width = width
|
| 68 |
+
self.layers = layers
|
| 69 |
+
self.resblocks = nn.Sequential(*[ResidualAttentionBlock(width, heads, attn_mask) for _ in range(layers)])
|
| 70 |
+
|
| 71 |
+
self.checkpoint_num = checkpoint_num
|
| 72 |
+
|
| 73 |
+
def forward(self, x: torch.Tensor):
|
| 74 |
+
if self.checkpoint_num > 0:
|
| 75 |
+
segments = min(self.checkpoint_num, len(self.resblocks))
|
| 76 |
+
return checkpoint.checkpoint_sequential(self.resblocks, segments, x)
|
| 77 |
+
else:
|
| 78 |
+
return self.resblocks(x)
|
| 79 |
+
|
| 80 |
+
|
| 81 |
+
class CLIP_TEXT(nn.Module):
|
| 82 |
+
def __init__(
|
| 83 |
+
self,
|
| 84 |
+
embed_dim: int,
|
| 85 |
+
context_length: int,
|
| 86 |
+
vocab_size: int,
|
| 87 |
+
transformer_width: int,
|
| 88 |
+
transformer_heads: int,
|
| 89 |
+
transformer_layers: int,
|
| 90 |
+
checkpoint_num: int,
|
| 91 |
+
tokenizer_path:str=None,
|
| 92 |
+
):
|
| 93 |
+
super().__init__()
|
| 94 |
+
|
| 95 |
+
self.context_length = context_length
|
| 96 |
+
if tokenizer_path:
|
| 97 |
+
self._tokenizer = _Tokenizer(tokenizer_path)
|
| 98 |
+
else:
|
| 99 |
+
self._tokenizer = _Tokenizer()
|
| 100 |
+
|
| 101 |
+
self.transformer = Transformer(
|
| 102 |
+
width=transformer_width,
|
| 103 |
+
layers=transformer_layers,
|
| 104 |
+
heads=transformer_heads,
|
| 105 |
+
attn_mask=self.build_attention_mask(),
|
| 106 |
+
checkpoint_num=checkpoint_num,
|
| 107 |
+
)
|
| 108 |
+
|
| 109 |
+
self.vocab_size = vocab_size
|
| 110 |
+
self.token_embedding = nn.Embedding(vocab_size, transformer_width)
|
| 111 |
+
self.positional_embedding = nn.Parameter(torch.empty(self.context_length, transformer_width))
|
| 112 |
+
self.ln_final = LayerNorm(transformer_width)
|
| 113 |
+
|
| 114 |
+
self.text_projection = nn.Parameter(torch.empty(transformer_width, embed_dim))
|
| 115 |
+
|
| 116 |
+
def no_weight_decay(self):
|
| 117 |
+
return {'token_embedding', 'positional_embedding'}
|
| 118 |
+
|
| 119 |
+
@functools.lru_cache(maxsize=None)
|
| 120 |
+
def build_attention_mask(self):
|
| 121 |
+
# lazily create causal attention mask, with full attention between the vision tokens
|
| 122 |
+
# pytorch uses additive attention mask; fill with -inf
|
| 123 |
+
mask = torch.empty(self.context_length, self.context_length)
|
| 124 |
+
mask.fill_(float("-inf"))
|
| 125 |
+
mask.triu_(1) # zero out the lower diagonal
|
| 126 |
+
return mask
|
| 127 |
+
|
| 128 |
+
def tokenize(self, texts, context_length=77, truncate=True):
|
| 129 |
+
"""
|
| 130 |
+
Returns the tokenized representation of given input string(s)
|
| 131 |
+
Parameters
|
| 132 |
+
----------
|
| 133 |
+
texts : Union[str, List[str]]
|
| 134 |
+
An input string or a list of input strings to tokenize
|
| 135 |
+
context_length : int
|
| 136 |
+
The context length to use; all CLIP models use 77 as the context length
|
| 137 |
+
truncate: bool
|
| 138 |
+
Whether to truncate the text in case its encoding is longer than the context length
|
| 139 |
+
Returns
|
| 140 |
+
-------
|
| 141 |
+
A two-dimensional tensor containing the resulting tokens, shape = [number of input strings, context_length].
|
| 142 |
+
We return LongTensor when torch version is <1.8.0, since older index_select requires indices to be long.
|
| 143 |
+
"""
|
| 144 |
+
if isinstance(texts, str):
|
| 145 |
+
texts = [texts]
|
| 146 |
+
|
| 147 |
+
sot_token = self._tokenizer.encoder["<|startoftext|>"]
|
| 148 |
+
eot_token = self._tokenizer.encoder["<|endoftext|>"]
|
| 149 |
+
all_tokens = [[sot_token] + self._tokenizer.encode(text) + [eot_token] for text in texts]
|
| 150 |
+
if packaging.version.parse(torch.__version__) < packaging.version.parse("1.8.0"):
|
| 151 |
+
result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
|
| 152 |
+
else:
|
| 153 |
+
result = torch.zeros(len(all_tokens), context_length, dtype=torch.int)
|
| 154 |
+
|
| 155 |
+
for i, tokens in enumerate(all_tokens):
|
| 156 |
+
if len(tokens) > context_length:
|
| 157 |
+
if truncate:
|
| 158 |
+
tokens = tokens[:context_length]
|
| 159 |
+
tokens[-1] = eot_token
|
| 160 |
+
else:
|
| 161 |
+
raise RuntimeError(f"Input {texts[i]} is too long for context length {context_length}")
|
| 162 |
+
result[i, :len(tokens)] = torch.tensor(tokens)
|
| 163 |
+
|
| 164 |
+
return result
|
| 165 |
+
|
| 166 |
+
def forward(self, text):
|
| 167 |
+
x = self.token_embedding(text) # [batch_size, n_ctx, d_model]
|
| 168 |
+
|
| 169 |
+
x = x + self.positional_embedding
|
| 170 |
+
x = x.permute(1, 0, 2) # NLD -> LND
|
| 171 |
+
x = self.transformer(x)
|
| 172 |
+
x = x.permute(1, 0, 2) # LND -> NLD
|
| 173 |
+
x = self.ln_final(x)
|
| 174 |
+
|
| 175 |
+
# x.shape = [batch_size, n_ctx, transformer.width]
|
| 176 |
+
# take features from the eot embedding (eot_token is the highest number in each sequence)
|
| 177 |
+
x = x[torch.arange(x.shape[0]), text.argmax(dim=-1)] @ self.text_projection
|
| 178 |
+
|
| 179 |
+
return x
|
| 180 |
+
|
| 181 |
+
|
| 182 |
+
def clip_text_b16(
|
| 183 |
+
embed_dim=512,
|
| 184 |
+
context_length=77,
|
| 185 |
+
vocab_size=49408,
|
| 186 |
+
transformer_width=512,
|
| 187 |
+
transformer_heads=8,
|
| 188 |
+
transformer_layers=12,
|
| 189 |
+
checkpoint_num=0,
|
| 190 |
+
pretrained=True,
|
| 191 |
+
tokenizer_path:str=None,
|
| 192 |
+
):
|
| 193 |
+
# raise NotImplementedError
|
| 194 |
+
model = CLIP_TEXT(
|
| 195 |
+
embed_dim,
|
| 196 |
+
context_length,
|
| 197 |
+
vocab_size,
|
| 198 |
+
transformer_width,
|
| 199 |
+
transformer_heads,
|
| 200 |
+
transformer_layers,
|
| 201 |
+
checkpoint_num,
|
| 202 |
+
tokenizer_path,
|
| 203 |
+
)
|
| 204 |
+
# pretrained = _MODELS["ViT-B/16"]
|
| 205 |
+
# logger.info(f"Load pretrained weights from {pretrained}")
|
| 206 |
+
# state_dict = torch.load(pretrained, map_location='cpu')
|
| 207 |
+
# model.load_state_dict(state_dict, strict=False)
|
| 208 |
+
# return model.eval()
|
| 209 |
+
if pretrained:
|
| 210 |
+
if isinstance(pretrained, str) and pretrained != "bert-base-uncased":
|
| 211 |
+
pretrained = _MODELS[pretrained]
|
| 212 |
+
else:
|
| 213 |
+
pretrained = _MODELS["ViT-B/16"]
|
| 214 |
+
logger.info(f"Load pretrained weights from {pretrained}")
|
| 215 |
+
state_dict = torch.load(pretrained, map_location='cpu')
|
| 216 |
+
if context_length != state_dict["positional_embedding"].size(0):
|
| 217 |
+
# assert context_length < state_dict["positional_embedding"].size(0), "Cannot increase context length."
|
| 218 |
+
print(f"Resize positional embedding from {state_dict['positional_embedding'].size(0)} to {context_length}")
|
| 219 |
+
if context_length < state_dict["positional_embedding"].size(0):
|
| 220 |
+
state_dict["positional_embedding"] = state_dict["positional_embedding"][:context_length]
|
| 221 |
+
else:
|
| 222 |
+
state_dict["positional_embedding"] = F.pad(
|
| 223 |
+
state_dict["positional_embedding"],
|
| 224 |
+
(0, 0, 0, context_length - state_dict["positional_embedding"].size(0)),
|
| 225 |
+
value=0,
|
| 226 |
+
)
|
| 227 |
+
|
| 228 |
+
message = model.load_state_dict(state_dict, strict=False)
|
| 229 |
+
print(f"Load pretrained weights from {pretrained}: {message}")
|
| 230 |
+
return model.eval()
|
| 231 |
+
|
| 232 |
+
|
| 233 |
+
def clip_text_l14(
|
| 234 |
+
embed_dim=768,
|
| 235 |
+
context_length=77,
|
| 236 |
+
vocab_size=49408,
|
| 237 |
+
transformer_width=768,
|
| 238 |
+
transformer_heads=12,
|
| 239 |
+
transformer_layers=12,
|
| 240 |
+
checkpoint_num=0,
|
| 241 |
+
pretrained=True,
|
| 242 |
+
tokenizer_path:str=None,
|
| 243 |
+
):
|
| 244 |
+
model = CLIP_TEXT(
|
| 245 |
+
embed_dim,
|
| 246 |
+
context_length,
|
| 247 |
+
vocab_size,
|
| 248 |
+
transformer_width,
|
| 249 |
+
transformer_heads,
|
| 250 |
+
transformer_layers,
|
| 251 |
+
checkpoint_num,
|
| 252 |
+
tokenizer_path,
|
| 253 |
+
)
|
| 254 |
+
if pretrained:
|
| 255 |
+
if isinstance(pretrained, str) and pretrained != "bert-base-uncased":
|
| 256 |
+
pretrained = _MODELS[pretrained]
|
| 257 |
+
else:
|
| 258 |
+
pretrained = _MODELS["ViT-L/14"]
|
| 259 |
+
logger.info(f"Load pretrained weights from {pretrained}")
|
| 260 |
+
state_dict = torch.load(pretrained, map_location='cpu')
|
| 261 |
+
if context_length != state_dict["positional_embedding"].size(0):
|
| 262 |
+
# assert context_length < state_dict["positional_embedding"].size(0), "Cannot increase context length."
|
| 263 |
+
print(f"Resize positional embedding from {state_dict['positional_embedding'].size(0)} to {context_length}")
|
| 264 |
+
if context_length < state_dict["positional_embedding"].size(0):
|
| 265 |
+
state_dict["positional_embedding"] = state_dict["positional_embedding"][:context_length]
|
| 266 |
+
else:
|
| 267 |
+
state_dict["positional_embedding"] = F.pad(
|
| 268 |
+
state_dict["positional_embedding"],
|
| 269 |
+
(0, 0, 0, context_length - state_dict["positional_embedding"].size(0)),
|
| 270 |
+
value=0,
|
| 271 |
+
)
|
| 272 |
+
|
| 273 |
+
message = model.load_state_dict(state_dict, strict=False)
|
| 274 |
+
print(f"Load pretrained weights from {pretrained}: {message}")
|
| 275 |
+
return model.eval()
|
| 276 |
+
|
| 277 |
+
|
| 278 |
+
def clip_text_l14_336(
|
| 279 |
+
embed_dim=768,
|
| 280 |
+
context_length=77,
|
| 281 |
+
vocab_size=49408,
|
| 282 |
+
transformer_width=768,
|
| 283 |
+
transformer_heads=12,
|
| 284 |
+
transformer_layers=12,
|
| 285 |
+
):
|
| 286 |
+
raise NotImplementedError
|
| 287 |
+
model = CLIP_TEXT(
|
| 288 |
+
embed_dim,
|
| 289 |
+
context_length,
|
| 290 |
+
vocab_size,
|
| 291 |
+
transformer_width,
|
| 292 |
+
transformer_heads,
|
| 293 |
+
transformer_layers
|
| 294 |
+
)
|
| 295 |
+
pretrained = _MODELS["ViT-L/14_336"]
|
| 296 |
+
logger.info(f"Load pretrained weights from {pretrained}")
|
| 297 |
+
state_dict = torch.load(pretrained, map_location='cpu')
|
| 298 |
+
model.load_state_dict(state_dict, strict=False)
|
| 299 |
+
return model.eval()
|
| 300 |
+
|
| 301 |
+
|
| 302 |
+
def build_clip(config):
|
| 303 |
+
model_cls = config.text_encoder.clip_teacher
|
| 304 |
+
model = eval(model_cls)()
|
| 305 |
+
return model
|
viclip_vision.py
ADDED
|
@@ -0,0 +1,362 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
#!/usr/bin/env python
|
| 2 |
+
import os
|
| 3 |
+
import logging
|
| 4 |
+
from collections import OrderedDict
|
| 5 |
+
|
| 6 |
+
import torch
|
| 7 |
+
from torch import nn
|
| 8 |
+
from einops import rearrange
|
| 9 |
+
from timm.models.layers import DropPath
|
| 10 |
+
from timm.models.registry import register_model
|
| 11 |
+
|
| 12 |
+
import torch.utils.checkpoint as checkpoint
|
| 13 |
+
|
| 14 |
+
# from models.utils import load_temp_embed_with_mismatch
|
| 15 |
+
|
| 16 |
+
logger = logging.getLogger(__name__)
|
| 17 |
+
|
| 18 |
+
def load_temp_embed_with_mismatch(temp_embed_old, temp_embed_new, add_zero=True):
|
| 19 |
+
"""
|
| 20 |
+
Add/Remove extra temporal_embeddings as needed.
|
| 21 |
+
https://arxiv.org/abs/2104.00650 shows adding zero paddings works.
|
| 22 |
+
|
| 23 |
+
temp_embed_old: (1, num_frames_old, 1, d)
|
| 24 |
+
temp_embed_new: (1, num_frames_new, 1, d)
|
| 25 |
+
add_zero: bool, if True, add zero, else, interpolate trained embeddings.
|
| 26 |
+
"""
|
| 27 |
+
# TODO zero pad
|
| 28 |
+
num_frms_new = temp_embed_new.shape[1]
|
| 29 |
+
num_frms_old = temp_embed_old.shape[1]
|
| 30 |
+
logger.info(f"Load temporal_embeddings, lengths: {num_frms_old}-->{num_frms_new}")
|
| 31 |
+
if num_frms_new > num_frms_old:
|
| 32 |
+
if add_zero:
|
| 33 |
+
temp_embed_new[
|
| 34 |
+
:, :num_frms_old
|
| 35 |
+
] = temp_embed_old # untrained embeddings are zeros.
|
| 36 |
+
else:
|
| 37 |
+
temp_embed_new = interpolate_temporal_pos_embed(temp_embed_old, num_frms_new)
|
| 38 |
+
elif num_frms_new < num_frms_old:
|
| 39 |
+
temp_embed_new = temp_embed_old[:, :num_frms_new]
|
| 40 |
+
else: # =
|
| 41 |
+
temp_embed_new = temp_embed_old
|
| 42 |
+
return temp_embed_new
|
| 43 |
+
|
| 44 |
+
|
| 45 |
+
# On P1, model extracted from https://huggingface.co/laion/CLIP-ViT-L-14-DataComp.XL-s13B-b90K
|
| 46 |
+
MODEL_PATH = ''
|
| 47 |
+
_MODELS = {
|
| 48 |
+
"ViT-L/14": os.path.join(MODEL_PATH, "ViCLIP-L_InternVid-FLT-10M.pth"),
|
| 49 |
+
"ViT-B/16": os.path.join(MODEL_PATH, "ViCLIP-B-InternVid-FLT-10M.pth"),
|
| 50 |
+
}
|
| 51 |
+
|
| 52 |
+
|
| 53 |
+
class QuickGELU(nn.Module):
|
| 54 |
+
def forward(self, x):
|
| 55 |
+
return x * torch.sigmoid(1.702 * x)
|
| 56 |
+
|
| 57 |
+
|
| 58 |
+
class ResidualAttentionBlock(nn.Module):
|
| 59 |
+
def __init__(self, d_model, n_head, drop_path=0., attn_mask=None, dropout=0.):
|
| 60 |
+
super().__init__()
|
| 61 |
+
|
| 62 |
+
self.drop_path1 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
| 63 |
+
self.drop_path2 = DropPath(drop_path) if drop_path > 0. else nn.Identity()
|
| 64 |
+
# logger.info(f'Droppath: {drop_path}')
|
| 65 |
+
self.attn = nn.MultiheadAttention(d_model, n_head, dropout=dropout)
|
| 66 |
+
self.ln_1 = nn.LayerNorm(d_model)
|
| 67 |
+
self.mlp = nn.Sequential(OrderedDict([
|
| 68 |
+
("c_fc", nn.Linear(d_model, d_model * 4)),
|
| 69 |
+
("gelu", QuickGELU()),
|
| 70 |
+
("drop1", nn.Dropout(dropout)),
|
| 71 |
+
("c_proj", nn.Linear(d_model * 4, d_model)),
|
| 72 |
+
("drop2", nn.Dropout(dropout)),
|
| 73 |
+
]))
|
| 74 |
+
self.ln_2 = nn.LayerNorm(d_model)
|
| 75 |
+
self.attn_mask = attn_mask
|
| 76 |
+
|
| 77 |
+
def attention(self, x):
|
| 78 |
+
self.attn_mask = self.attn_mask.to(dtype=x.dtype, device=x.device) if self.attn_mask is not None else None
|
| 79 |
+
return self.attn(x, x, x, need_weights=False, attn_mask=self.attn_mask)[0]
|
| 80 |
+
|
| 81 |
+
def forward(self, x):
|
| 82 |
+
x = x + self.drop_path1(self.attention(self.ln_1(x)))
|
| 83 |
+
x = x + self.drop_path2(self.mlp(self.ln_2(x)))
|
| 84 |
+
return x
|
| 85 |
+
|
| 86 |
+
|
| 87 |
+
class Transformer(nn.Module):
|
| 88 |
+
def __init__(self, width, layers, heads, drop_path=0., checkpoint_num=0, dropout=0.):
|
| 89 |
+
super().__init__()
|
| 90 |
+
dpr = [x.item() for x in torch.linspace(0, drop_path, layers)]
|
| 91 |
+
self.resblocks = nn.ModuleList()
|
| 92 |
+
for idx in range(layers):
|
| 93 |
+
self.resblocks.append(ResidualAttentionBlock(width, heads, drop_path=dpr[idx], dropout=dropout))
|
| 94 |
+
self.checkpoint_num = checkpoint_num
|
| 95 |
+
|
| 96 |
+
def forward(self, x):
|
| 97 |
+
for idx, blk in enumerate(self.resblocks):
|
| 98 |
+
if idx < self.checkpoint_num:
|
| 99 |
+
x = checkpoint.checkpoint(blk, x)
|
| 100 |
+
else:
|
| 101 |
+
x = blk(x)
|
| 102 |
+
return x
|
| 103 |
+
|
| 104 |
+
|
| 105 |
+
class VisionTransformer(nn.Module):
|
| 106 |
+
def __init__(
|
| 107 |
+
self, input_resolution, patch_size, width, layers, heads, output_dim=None,
|
| 108 |
+
kernel_size=1, num_frames=8, drop_path=0, checkpoint_num=0, dropout=0.,
|
| 109 |
+
temp_embed=True,
|
| 110 |
+
):
|
| 111 |
+
super().__init__()
|
| 112 |
+
self.output_dim = output_dim
|
| 113 |
+
self.conv1 = nn.Conv3d(
|
| 114 |
+
3, width,
|
| 115 |
+
(kernel_size, patch_size, patch_size),
|
| 116 |
+
(kernel_size, patch_size, patch_size),
|
| 117 |
+
(0, 0, 0), bias=False
|
| 118 |
+
)
|
| 119 |
+
|
| 120 |
+
scale = width ** -0.5
|
| 121 |
+
self.class_embedding = nn.Parameter(scale * torch.randn(width))
|
| 122 |
+
self.positional_embedding = nn.Parameter(scale * torch.randn((input_resolution // patch_size) ** 2 + 1, width))
|
| 123 |
+
self.ln_pre = nn.LayerNorm(width)
|
| 124 |
+
if temp_embed:
|
| 125 |
+
self.temporal_positional_embedding = nn.Parameter(torch.zeros(1, num_frames, width))
|
| 126 |
+
|
| 127 |
+
self.transformer = Transformer(
|
| 128 |
+
width, layers, heads, drop_path=drop_path, checkpoint_num=checkpoint_num,
|
| 129 |
+
dropout=dropout)
|
| 130 |
+
|
| 131 |
+
self.ln_post = nn.LayerNorm(width)
|
| 132 |
+
if output_dim is not None:
|
| 133 |
+
self.proj = nn.Parameter(torch.empty(width, output_dim))
|
| 134 |
+
else:
|
| 135 |
+
self.proj = None
|
| 136 |
+
|
| 137 |
+
self.dropout = nn.Dropout(dropout)
|
| 138 |
+
|
| 139 |
+
def get_num_layers(self):
|
| 140 |
+
return len(self.transformer.resblocks)
|
| 141 |
+
|
| 142 |
+
@torch.jit.ignore
|
| 143 |
+
def no_weight_decay(self):
|
| 144 |
+
return {'positional_embedding', 'class_embedding', 'temporal_positional_embedding'}
|
| 145 |
+
|
| 146 |
+
def mask_tokens(self, inputs, masking_prob=0.0):
|
| 147 |
+
B, L, _ = inputs.shape
|
| 148 |
+
|
| 149 |
+
# This is different from text as we are masking a fix number of tokens
|
| 150 |
+
Lm = int(masking_prob * L)
|
| 151 |
+
masked_indices = torch.zeros(B, L)
|
| 152 |
+
indices = torch.argsort(torch.rand_like(masked_indices), dim=-1)[:, :Lm]
|
| 153 |
+
batch_indices = (
|
| 154 |
+
torch.arange(masked_indices.shape[0]).unsqueeze(-1).expand_as(indices)
|
| 155 |
+
)
|
| 156 |
+
masked_indices[batch_indices, indices] = 1
|
| 157 |
+
|
| 158 |
+
masked_indices = masked_indices.bool()
|
| 159 |
+
|
| 160 |
+
return inputs[~masked_indices].reshape(B, -1, inputs.shape[-1])
|
| 161 |
+
|
| 162 |
+
def forward(self, x, masking_prob=0.0):
|
| 163 |
+
x = self.conv1(x) # shape = [*, width, grid, grid]
|
| 164 |
+
B, C, T, H, W = x.shape
|
| 165 |
+
x = x.permute(0, 2, 3, 4, 1).reshape(B * T, H * W, C)
|
| 166 |
+
|
| 167 |
+
x = torch.cat([self.class_embedding.to(x.dtype) + torch.zeros(x.shape[0], 1, x.shape[-1], dtype=x.dtype, device=x.device), x], dim=1) # shape = [*, grid ** 2 + 1, width]
|
| 168 |
+
x = x + self.positional_embedding.to(x.dtype)
|
| 169 |
+
|
| 170 |
+
# temporal pos
|
| 171 |
+
cls_tokens = x[:B, :1, :]
|
| 172 |
+
x = x[:, 1:]
|
| 173 |
+
x = rearrange(x, '(b t) n m -> (b n) t m', b=B, t=T)
|
| 174 |
+
if hasattr(self, 'temporal_positional_embedding'):
|
| 175 |
+
if x.size(1) == 1:
|
| 176 |
+
# This is a workaround for unused parameter issue
|
| 177 |
+
x = x + self.temporal_positional_embedding.mean(1)
|
| 178 |
+
else:
|
| 179 |
+
x = x + self.temporal_positional_embedding
|
| 180 |
+
x = rearrange(x, '(b n) t m -> b (n t) m', b=B, t=T)
|
| 181 |
+
|
| 182 |
+
if masking_prob > 0.0:
|
| 183 |
+
x = self.mask_tokens(x, masking_prob)
|
| 184 |
+
|
| 185 |
+
x = torch.cat((cls_tokens, x), dim=1)
|
| 186 |
+
|
| 187 |
+
x = self.ln_pre(x)
|
| 188 |
+
|
| 189 |
+
x = x.permute(1, 0, 2) #BND -> NBD
|
| 190 |
+
x = self.transformer(x)
|
| 191 |
+
|
| 192 |
+
x = self.ln_post(x)
|
| 193 |
+
|
| 194 |
+
if self.proj is not None:
|
| 195 |
+
x = self.dropout(x[0]) @ self.proj
|
| 196 |
+
else:
|
| 197 |
+
x = x.permute(1, 0, 2) #NBD -> BND
|
| 198 |
+
|
| 199 |
+
return x
|
| 200 |
+
|
| 201 |
+
|
| 202 |
+
def inflate_weight(weight_2d, time_dim, center=True):
|
| 203 |
+
logger.info(f'Init center: {center}')
|
| 204 |
+
if center:
|
| 205 |
+
weight_3d = torch.zeros(*weight_2d.shape)
|
| 206 |
+
weight_3d = weight_3d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
|
| 207 |
+
middle_idx = time_dim // 2
|
| 208 |
+
weight_3d[:, :, middle_idx, :, :] = weight_2d
|
| 209 |
+
else:
|
| 210 |
+
weight_3d = weight_2d.unsqueeze(2).repeat(1, 1, time_dim, 1, 1)
|
| 211 |
+
weight_3d = weight_3d / time_dim
|
| 212 |
+
return weight_3d
|
| 213 |
+
|
| 214 |
+
|
| 215 |
+
def load_state_dict(model, state_dict, input_resolution=224, patch_size=16, center=True):
|
| 216 |
+
state_dict_3d = model.state_dict()
|
| 217 |
+
for k in state_dict.keys():
|
| 218 |
+
if k in state_dict_3d.keys() and state_dict[k].shape != state_dict_3d[k].shape:
|
| 219 |
+
if len(state_dict_3d[k].shape) <= 2:
|
| 220 |
+
logger.info(f'Ignore: {k}')
|
| 221 |
+
continue
|
| 222 |
+
logger.info(f'Inflate: {k}, {state_dict[k].shape} => {state_dict_3d[k].shape}')
|
| 223 |
+
time_dim = state_dict_3d[k].shape[2]
|
| 224 |
+
state_dict[k] = inflate_weight(state_dict[k], time_dim, center=center)
|
| 225 |
+
|
| 226 |
+
pos_embed_checkpoint = state_dict['positional_embedding']
|
| 227 |
+
embedding_size = pos_embed_checkpoint.shape[-1]
|
| 228 |
+
num_patches = (input_resolution // patch_size) ** 2
|
| 229 |
+
orig_size = int((pos_embed_checkpoint.shape[-2] - 1) ** 0.5)
|
| 230 |
+
new_size = int(num_patches ** 0.5)
|
| 231 |
+
if orig_size != new_size:
|
| 232 |
+
logger.info(f'Pos_emb from {orig_size} to {new_size}')
|
| 233 |
+
extra_tokens = pos_embed_checkpoint[:1]
|
| 234 |
+
pos_tokens = pos_embed_checkpoint[1:]
|
| 235 |
+
pos_tokens = pos_tokens.reshape(-1, orig_size, orig_size, embedding_size).permute(0, 3, 1, 2)
|
| 236 |
+
pos_tokens = torch.nn.functional.interpolate(
|
| 237 |
+
pos_tokens, size=(new_size, new_size), mode='bicubic', align_corners=False)
|
| 238 |
+
pos_tokens = pos_tokens.permute(0, 2, 3, 1).flatten(0, 2)
|
| 239 |
+
new_pos_embed = torch.cat((extra_tokens, pos_tokens), dim=0)
|
| 240 |
+
state_dict['positional_embedding'] = new_pos_embed
|
| 241 |
+
|
| 242 |
+
message = model.load_state_dict(state_dict, strict=False)
|
| 243 |
+
logger.info(f"Load pretrained weights: {message}")
|
| 244 |
+
|
| 245 |
+
|
| 246 |
+
@register_model
|
| 247 |
+
def clip_joint_b16(
|
| 248 |
+
pretrained=False, input_resolution=224, kernel_size=1,
|
| 249 |
+
center=True, num_frames=8, drop_path=0., checkpoint_num=0,
|
| 250 |
+
dropout=0.,
|
| 251 |
+
):
|
| 252 |
+
model = VisionTransformer(
|
| 253 |
+
input_resolution=input_resolution, patch_size=16,
|
| 254 |
+
width=768, layers=12, heads=12, output_dim=512,
|
| 255 |
+
kernel_size=kernel_size, num_frames=num_frames,
|
| 256 |
+
drop_path=drop_path, checkpoint_num=checkpoint_num,
|
| 257 |
+
dropout=dropout,
|
| 258 |
+
)
|
| 259 |
+
# raise NotImplementedError
|
| 260 |
+
if pretrained:
|
| 261 |
+
if isinstance(pretrained, str):
|
| 262 |
+
model_name = pretrained
|
| 263 |
+
else:
|
| 264 |
+
model_name = "ViT-B/16"
|
| 265 |
+
|
| 266 |
+
logger.info('load pretrained weights')
|
| 267 |
+
state_dict = torch.load(_MODELS[model_name], map_location='cpu')
|
| 268 |
+
load_state_dict(model, state_dict, input_resolution=input_resolution, patch_size=16, center=center)
|
| 269 |
+
return model.eval()
|
| 270 |
+
|
| 271 |
+
|
| 272 |
+
@register_model
|
| 273 |
+
def clip_joint_l14(
|
| 274 |
+
pretrained=False, input_resolution=224, kernel_size=1,
|
| 275 |
+
center=True, num_frames=8, drop_path=0., checkpoint_num=0,
|
| 276 |
+
dropout=0.,
|
| 277 |
+
):
|
| 278 |
+
model = VisionTransformer(
|
| 279 |
+
input_resolution=input_resolution, patch_size=14,
|
| 280 |
+
width=1024, layers=24, heads=16, output_dim=768,
|
| 281 |
+
kernel_size=kernel_size, num_frames=num_frames,
|
| 282 |
+
drop_path=drop_path, checkpoint_num=checkpoint_num,
|
| 283 |
+
dropout=dropout,
|
| 284 |
+
)
|
| 285 |
+
|
| 286 |
+
if pretrained:
|
| 287 |
+
if isinstance(pretrained, str):
|
| 288 |
+
model_name = pretrained
|
| 289 |
+
else:
|
| 290 |
+
model_name = "ViT-L/14"
|
| 291 |
+
logger.info('load pretrained weights')
|
| 292 |
+
state_dict = torch.load(_MODELS[model_name], map_location='cpu')
|
| 293 |
+
load_state_dict(model, state_dict, input_resolution=input_resolution, patch_size=14, center=center)
|
| 294 |
+
return model.eval()
|
| 295 |
+
|
| 296 |
+
|
| 297 |
+
@register_model
|
| 298 |
+
def clip_joint_l14_336(
|
| 299 |
+
pretrained=True, input_resolution=336, kernel_size=1,
|
| 300 |
+
center=True, num_frames=8, drop_path=0.
|
| 301 |
+
):
|
| 302 |
+
raise NotImplementedError
|
| 303 |
+
model = VisionTransformer(
|
| 304 |
+
input_resolution=input_resolution, patch_size=14,
|
| 305 |
+
width=1024, layers=24, heads=16, output_dim=768,
|
| 306 |
+
kernel_size=kernel_size, num_frames=num_frames,
|
| 307 |
+
drop_path=drop_path,
|
| 308 |
+
)
|
| 309 |
+
if pretrained:
|
| 310 |
+
logger.info('load pretrained weights')
|
| 311 |
+
state_dict = torch.load(_MODELS["ViT-L/14_336"], map_location='cpu')
|
| 312 |
+
load_state_dict(model, state_dict, input_resolution=input_resolution, patch_size=14, center=center)
|
| 313 |
+
return model.eval()
|
| 314 |
+
|
| 315 |
+
|
| 316 |
+
def interpolate_pos_embed_vit(state_dict, new_model):
|
| 317 |
+
key = "vision_encoder.temporal_positional_embedding"
|
| 318 |
+
if key in state_dict:
|
| 319 |
+
vision_temp_embed_new = new_model.state_dict()[key]
|
| 320 |
+
vision_temp_embed_new = vision_temp_embed_new.unsqueeze(2) # [1, n, d] -> [1, n, 1, d]
|
| 321 |
+
vision_temp_embed_old = state_dict[key]
|
| 322 |
+
vision_temp_embed_old = vision_temp_embed_old.unsqueeze(2)
|
| 323 |
+
|
| 324 |
+
state_dict[key] = load_temp_embed_with_mismatch(
|
| 325 |
+
vision_temp_embed_old, vision_temp_embed_new, add_zero=False
|
| 326 |
+
).squeeze(2)
|
| 327 |
+
|
| 328 |
+
key = "text_encoder.positional_embedding"
|
| 329 |
+
if key in state_dict:
|
| 330 |
+
text_temp_embed_new = new_model.state_dict()[key]
|
| 331 |
+
text_temp_embed_new = text_temp_embed_new.unsqueeze(0).unsqueeze(2) # [n, d] -> [1, n, 1, d]
|
| 332 |
+
text_temp_embed_old = state_dict[key]
|
| 333 |
+
text_temp_embed_old = text_temp_embed_old.unsqueeze(0).unsqueeze(2)
|
| 334 |
+
|
| 335 |
+
state_dict[key] = load_temp_embed_with_mismatch(
|
| 336 |
+
text_temp_embed_old, text_temp_embed_new, add_zero=False
|
| 337 |
+
).squeeze(2).squeeze(0)
|
| 338 |
+
return state_dict
|
| 339 |
+
|
| 340 |
+
|
| 341 |
+
if __name__ == '__main__':
|
| 342 |
+
import time
|
| 343 |
+
from fvcore.nn import FlopCountAnalysis
|
| 344 |
+
from fvcore.nn import flop_count_table
|
| 345 |
+
import numpy as np
|
| 346 |
+
|
| 347 |
+
seed = 4217
|
| 348 |
+
np.random.seed(seed)
|
| 349 |
+
torch.manual_seed(seed)
|
| 350 |
+
torch.cuda.manual_seed(seed)
|
| 351 |
+
torch.cuda.manual_seed_all(seed)
|
| 352 |
+
num_frames = 8
|
| 353 |
+
|
| 354 |
+
# model = clip_joint_b16(pretrained=True, kernel_size=1, num_frames=8, num_classes=400, drop_path=0.1)
|
| 355 |
+
# logger.info(model)
|
| 356 |
+
model = clip_joint_l14(pretrained=False)
|
| 357 |
+
|
| 358 |
+
flops = FlopCountAnalysis(model, torch.rand(1, 3, num_frames, 224, 224))
|
| 359 |
+
s = time.time()
|
| 360 |
+
logger.info(flop_count_table(flops, max_depth=1))
|
| 361 |
+
logger.info(time.time()-s)
|
| 362 |
+
# logger.info(model(torch.rand(1, 3, num_frames, 224, 224)).shape)
|