Spaces:
Sleeping
Sleeping
File size: 2,042 Bytes
08545c6 |
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 |
import pickle
import torch
import transformers
import gradio as gr
# XLM model functions
import transformers
# Our model definition
class MultilingualClipEdited(torch.nn.Module):
def __init__(self, model_name, tokenizer_name, head_name, weights_dir='head_weights/', cache_dir=None,in_features=None,out_features=None):
super().__init__()
self.model_name = model_name
self.tokenizer_name = tokenizer_name
self.head_path = weights_dir + head_name
self.tokenizer = transformers.AutoTokenizer.from_pretrained(tokenizer_name, cache_dir=cache_dir)
self.transformer = transformers.AutoModel.from_pretrained(model_name, cache_dir=cache_dir)
self.clip_head = torch.nn.Linear(in_features=in_features, out_features=out_features)
self._load_head()
def forward(self, txt):
txt_tok = self.tokenizer(txt, padding=True, return_tensors='pt')
embs = self.transformer(**txt_tok)[0]
att = txt_tok['attention_mask']
embs = (embs * att.unsqueeze(2)).sum(dim=1) / att.sum(dim=1)[:, None]
return self.clip_head(embs)
def _load_head(self):
with open(self.head_path, 'rb') as f:
lin_weights = pickle.loads(f.read())
self.clip_head.weight = torch.nn.Parameter(torch.tensor(lin_weights[0]).float().t())
self.clip_head.bias = torch.nn.Parameter(torch.tensor(lin_weights[1]).float())
AVAILABLE_MODELS = {
'bert-base-arabertv2-ViT-B-16-SigLIP-512-epoch-155-trained-2M':{
'model_name': 'Arabic-Clip/bert-base-arabertv2-ViT-B-16-SigLIP-512-epoch-155-trained-2M',
'tokenizer_name': 'Arabic-Clip/bert-base-arabertv2-ViT-B-16-SigLIP-512-epoch-155-trained-2M',
'head_name': 'arabertv2-vit-B-16-siglibheads_of_the_model_arabertv2-ViT-B-16-SigLIP-512-155_.pickle'
},
}
def load_model(name, cache_dir=None,in_features=None,out_features=None):
config = AVAILABLE_MODELS[name]
return MultilingualClipEdited(**config, cache_dir=cache_dir, in_features= in_features, out_features=out_features) |