How to get mean embeddings from model

[timoast]

How to get mean embeddings from model

[Zhihan1996]

Hi,

Thank for posting the question.

To load the model from huggingface:

import torch
from transformers import AutoTokenizer, AutoModel

tokenizer = AutoTokenizer.from_pretrained("zhihan1996/DNABERT-2-117M", trust_remote_code=True)
model = AutoModel.from_pretrained("zhihan1996/DNABERT-2-117M", trust_remote_code=True)

To calculate the embedding of a dna sequence

dna = "ACGTAGCATCGGATCTATCTATCGACACTTGGTTATCGATCTACGAGCATCTCGTTAGC"
inputs = tokenizer(dna, return_tensors = 'pt')["input_ids"]
hidden_states = model(inputs)[0] # [1, sequence_length, 768]

# embedding with mean pooling
embedding_mean = torch.mean(hidden_states[0], dim=0)
print(embedding_mean.shape) # expect to be 768

# embedding with max pooling
embedding_max = torch.max(hidden_states[0], dim=0)[0]
print(embedding_max.shape) # expect to be 768

I will also update the model card and README for the instructions.