Views
No views yet
1import torch.nn.functional as F
2
3from torch import Tensor
4from transformers import AutoTokenizer, AutoModel
5
6
7def average_pool(last_hidden_states: Tensor,
8 attention_mask: Tensor) -> Tensor:
9 last_hidden = last_hidden_states.masked_fill(~attention_mask[..., None].bool(), 0.0)
10 return last_hidden.sum(dim=1) / attention_mask.sum(dim=1)[..., None]
11
12
13# Each input text should start with "query: " or "passage: ".
14# For tasks other than retrieval, you can simply use the "query: " prefix.
15input_texts = ['query: Espresso Pitcher with Handle',
16 'query: Women’s designer handbag sale',
17 "passage: Dianoo Espresso Steaming Pitcher, Espresso Milk Frothing Pitcher Stainless Steel",
18 "passage: Coach Outlet Eliza Shoulder Bag - Black - One Size"]
19
20tokenizer = AutoTokenizer.from_pretrained('Marqo/marqo-gcl-e5-large-v2-130')
21model_new = AutoModel.from_pretrained('Marqo/marqo-gcl-e5-large-v2-130')
22
23# Tokenize the input texts
24batch_dict = tokenizer(input_texts, max_length=77, padding=True, truncation=True, return_tensors='pt')
25
26outputs = model_new(**batch_dict)
27embeddings = average_pool(outputs.last_hidden_state, batch_dict['attention_mask'])
28
29# normalize embeddings
30embeddings = F.normalize(embeddings, p=2, dim=1)
31scores = (embeddings[:2] @ embeddings[2:].T) * 100
32print(scores.tolist())
33
34