mirror of https://github.com/commaai/tinygrad.git
Added standalone CLIP tokenizer (#382)
* Added standalone CLIP tokenizer. * Fixed empty phrase. * Truncating long prompts. * Keeping two slots for the start and end token. * Fixed empty phrase. * Using tokenizer for empty phrase. * Typo.
This commit is contained in:
Fernand Pajot
GitHub
parent
a1810c8617
commit
ff0da4c802
Binary file not shown.
|
|
@ -2,11 +2,17 @@
|
|||
# https://github.com/ekagra-ranjan/huggingface-blog/blob/main/stable_diffusion.md
|
||||
|
||||
import os
|
||||
import gzip
|
||||
import math
|
||||
import numpy as np
|
||||
import os
|
||||
import re
|
||||
import traceback
|
||||
from tqdm import tqdm
|
||||
from functools import lru_cache
|
||||
from collections import namedtuple
|
||||
|
||||
import numpy as np
|
||||
from tqdm import tqdm
|
||||
|
||||
from extra.utils import fake_torch_load_zipped, get_child
|
||||
from tinygrad.nn import Conv2d, Linear
|
||||
from tinygrad.tensor import Tensor
|
||||
|
|
@ -481,6 +487,116 @@ class CLIPTextTransformer:
|
|||
x = self.encoder(x, Tensor(causal_attention_mask, device=x.device))
|
||||
return self.final_layer_norm(x)
|
||||
|
||||
# Clip tokenizer, taken from https://github.com/openai/CLIP/blob/main/clip/simple_tokenizer.py (MIT license)
|
||||
@lru_cache()
|
||||
def default_bpe():
|
||||
return os.path.join(os.path.dirname(os.path.abspath(__file__)), "bpe_simple_vocab_16e6.txt.gz")
|
||||
|
||||
def get_pairs(word):
|
||||
"""Return set of symbol pairs in a word.
|
||||
Word is represented as tuple of symbols (symbols being variable-length strings).
|
||||
"""
|
||||
pairs = set()
|
||||
prev_char = word[0]
|
||||
for char in word[1:]:
|
||||
pairs.add((prev_char, char))
|
||||
prev_char = char
|
||||
return pairs
|
||||
|
||||
def whitespace_clean(text):
|
||||
text = re.sub(r'\s+', ' ', text)
|
||||
text = text.strip()
|
||||
return text
|
||||
|
||||
def bytes_to_unicode():
|
||||
"""
|
||||
Returns list of utf-8 byte and a corresponding list of unicode strings.
|
||||
The reversible bpe codes work on unicode strings.
|
||||
This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
|
||||
When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
|
||||
This is a signficant percentage of your normal, say, 32K bpe vocab.
|
||||
To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
|
||||
And avoids mapping to whitespace/control characters the bpe code barfs on.
|
||||
"""
|
||||
bs = list(range(ord("!"), ord("~")+1))+list(range(ord("¡"), ord("¬")+1))+list(range(ord("®"), ord("ÿ")+1))
|
||||
cs = bs[:]
|
||||
n = 0
|
||||
for b in range(2**8):
|
||||
if b not in bs:
|
||||
bs.append(b)
|
||||
cs.append(2**8+n)
|
||||
n += 1
|
||||
cs = [chr(n) for n in cs]
|
||||
return dict(zip(bs, cs))
|
||||
|
||||
class ClipTokenizer(object):
|
||||
def __init__(self, bpe_path: str = default_bpe()):
|
||||
self.byte_encoder = bytes_to_unicode()
|
||||
merges = gzip.open(bpe_path).read().decode("utf-8").split('\n')
|
||||
merges = merges[1:49152-256-2+1]
|
||||
merges = [tuple(merge.split()) for merge in merges]
|
||||
vocab = list(bytes_to_unicode().values())
|
||||
vocab = vocab + [v+'</w>' for v in vocab]
|
||||
for merge in merges:
|
||||
vocab.append(''.join(merge))
|
||||
vocab.extend(['<|startoftext|>', '<|endoftext|>'])
|
||||
self.encoder = dict(zip(vocab, range(len(vocab))))
|
||||
self.bpe_ranks = dict(zip(merges, range(len(merges))))
|
||||
self.cache = {'<|startoftext|>': '<|startoftext|>', '<|endoftext|>': '<|endoftext|>'}
|
||||
self.pat = self.pat = re.compile(r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[^\s]+""", re.IGNORECASE)
|
||||
|
||||
def bpe(self, token):
|
||||
if token in self.cache:
|
||||
return self.cache[token]
|
||||
word = tuple(token[:-1]) + ( token[-1] + '</w>',)
|
||||
pairs = get_pairs(word)
|
||||
|
||||
if not pairs:
|
||||
return token+'</w>'
|
||||
|
||||
while True:
|
||||
bigram = min(pairs, key = lambda pair: self.bpe_ranks.get(pair, float('inf')))
|
||||
if bigram not in self.bpe_ranks:
|
||||
break
|
||||
first, second = bigram
|
||||
new_word = []
|
||||
i = 0
|
||||
while i < len(word):
|
||||
try:
|
||||
j = word.index(first, i)
|
||||
new_word.extend(word[i:j])
|
||||
i = j
|
||||
except:
|
||||
new_word.extend(word[i:])
|
||||
break
|
||||
|
||||
if word[i] == first and i < len(word)-1 and word[i+1] == second:
|
||||
new_word.append(first+second)
|
||||
i += 2
|
||||
else:
|
||||
new_word.append(word[i])
|
||||
i += 1
|
||||
new_word = tuple(new_word)
|
||||
word = new_word
|
||||
if len(word) == 1:
|
||||
break
|
||||
else:
|
||||
pairs = get_pairs(word)
|
||||
word = ' '.join(word)
|
||||
self.cache[token] = word
|
||||
return word
|
||||
|
||||
def encode(self, text):
|
||||
bpe_tokens = []
|
||||
text = whitespace_clean(text.strip()).lower()
|
||||
for token in re.findall(self.pat, text):
|
||||
token = ''.join(self.byte_encoder[b] for b in token.encode('utf-8'))
|
||||
bpe_tokens.extend(self.encoder[bpe_token] for bpe_token in self.bpe(token).split(' '))
|
||||
# Truncation, keeping two slots for start and end tokens.
|
||||
if len(bpe_tokens) > 75:
|
||||
bpe_tokens = bpe_tokens[:75]
|
||||
return [49406] + bpe_tokens + [49407] * (77 - len(bpe_tokens) - 1)
|
||||
|
||||
class StableDiffusion:
|
||||
def __init__(self):
|
||||
self.alphas_cumprod = Tensor.empty(1000)
|
||||
|
|
@ -532,16 +648,14 @@ if __name__ == "__main__":
|
|||
|
||||
# run through CLIP to get context
|
||||
|
||||
# "a horse sized cat eating a bagel"
|
||||
phrase = [49406, 320, 4558, 9832, 2368, 4371, 320, 28777, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407]
|
||||
|
||||
# "penguin with fire extinguisher"
|
||||
#phrase = [49406, 14952, 593, 1769, 38567, 4510, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407]
|
||||
tokenizer = ClipTokenizer()
|
||||
phrase = tokenizer.encode("a horse sized cat eating a bagel")
|
||||
# phrase = tokenizer.encode("penguin with fire extinguisher")
|
||||
|
||||
context = model.cond_stage_model.transformer.text_model(phrase).realize()
|
||||
print("got CLIP context", context.shape)
|
||||
|
||||
phrase = [49406, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407, 49407]
|
||||
phrase = tokenizer.encode("")
|
||||
unconditional_context = model.cond_stage_model.transformer.text_model(phrase).realize()
|
||||
print("got unconditional CLIP context", unconditional_context.shape)
|
||||
|
||||
|
|
|
|||
Loading…
Reference in New Issue