File size: 6,199 Bytes
64c529c 834e995 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 | # Adapted from TinyStyler (https://github.com/zacharyhorvitz/TinyStyler).
#
# MIT License
#
# Copyright (c) the TinyStyler authors
#
# Permission is hereby granted, free of charge, to any person obtaining a copy
# of this software and associated documentation files (the "Software"), to deal
# in the Software without restriction, including without limitation the rights
# to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
# copies of the Software, and to permit persons to whom the Software is
# furnished to do so, subject to the following conditions:
#
# The above copyright notice and this permission notice shall be included in all
# copies or substantial portions of the Software.
#
# THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
# IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
# FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
# AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
# LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
# OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
# SOFTWARE.
"""
Utilities for loading style models and extracting style embeddings from texts
"""
import torch
from transformers import AutoModel, AutoTokenizer
style_model, style_tokenizer = None, None
def get_text_style_embedding(texts):
global style_model, style_tokenizer
if style_model is None:
style_model, style_tokenizer, _ = load_style_model()
style_model.eval()
style_model.to('cuda')
style_embeds = text_to_style(
model=style_model,
tokenizer=style_tokenizer,
texts=texts,
device='cuda',
model_type='style',
)
return torch.stack(style_embeds).detach().cpu()
def load_style_model(model_name='AnnaWegmann/Style-Embedding'):
"""
Load model, tokenizer, word embeddings
Adapted from: https://github.com/zacharyhorvitz/ParaGuide
Supports
- Wegmann Style Embeddings https://huggingface.co/AnnaWegmann/Style-Embedding
- StyleDistance Embeddings https://huggingface.co/StyleDistance/styledistance
"""
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModel.from_pretrained(model_name)
word_embeds = get_word_embeddings(model)
return model, tokenizer, word_embeds
def mean_pooling(model_output, attention_mask):
"""
Mean pool over non mask tokens.
From https://huggingface.co/AnnaWegmann/Style-Embedding
"""
token_embeddings = model_output[
0
] # First element of model_output contains all token embeddings
input_mask_expanded = (
attention_mask.unsqueeze(-1).expand(token_embeddings.size()).float()
)
return torch.sum(token_embeddings * input_mask_expanded, 1) / torch.clamp(
input_mask_expanded.sum(1), min=1e-9
)
def text_to_style(
*, model, tokenizer, texts, device, model_type='style', max_length=512
):
"""
Applies get_style_embedding() to a list of texts
Args:
model (torch.nn.Module): PyTorch model
tokenizer (transformers.Tokenizer): tokenizer
texts (List[str]): list of texts
device (str): device
model_type (str): model type (currently only 'style' is supported)
max_length (int): max length of input text
Returns:
List[torch.Tensor]: style embeddings
"""
inputs = tokenizer(
texts, return_tensors='pt', padding=True, truncation=True, max_length=max_length
)
inputs = {k: v.to(device) for k, v in inputs.items()}
embeds = get_style_embedding(
model=model,
input_tokens=inputs['input_ids'],
attention_mask=inputs['attention_mask'],
model_type=model_type,
)
embeds = [x for x in embeds]
return embeds
def get_style_embedding(
*,
model,
inputs_embeds=None,
input_tokens=None,
attention_mask=None,
model_type='style',
):
"""
Get style embeddings from a model
Adapted from: https://github.com/zacharyhorvitz/ParaGuide
Args:
model (torch.nn.Module): PyTorch model
inputs_embeds (torch.Tensor): input token embeddings **optional**
input_tokens (torch.Tensor): input tokens **optional**
attention_mask (torch.Tensor): attention mask **optional**
model_type (str): model type (currently only 'style' is supported)
Returns:
torch.Tensor: style embeddings (batch_size, style embed_dim)
"""
if inputs_embeds is not None and input_tokens is not None:
raise ValueError('inputs_embeds and input_tokens cannot be both not None')
if model_type != 'style':
raise ValueError(
f'Unknown model type {model_type}, only model type currently supported'
)
if inputs_embeds is not None:
if attention_mask is None:
attention_mask = torch.ones(*inputs_embeds.shape[:-1]).to(
inputs_embeds.device
) # this may be why I have issues when i insert padding tokens
attention_mask = attention_mask.to(inputs_embeds.device)
return mean_pooling(
model(inputs_embeds=inputs_embeds, attention_mask=attention_mask),
attention_mask=attention_mask,
)
else:
if attention_mask is None:
attention_mask = torch.ones(*input_tokens.shape).to(input_tokens.device)
attention_mask = attention_mask.to(input_tokens.device)
return mean_pooling(
model(input_tokens, attention_mask=attention_mask),
attention_mask=attention_mask,
)
def get_word_embeddings(model):
"""
returns the word embeddings from a PyTorch model
model.get_input_embeddings() also works for most Hugging Face models
Args:
model (torch.nn.Module): PyTorch model
Returns:
torch.Tensor: word embeddings
"""
state_dict = model.state_dict()
params = []
for key in state_dict:
if 'word_embeddings' in key:
params.append((key, state_dict[key]))
assert len(params) == 1, f'Found {params}'
return params[0][1]
|