Hello World about protein_bert HOT 4 CLOSED

@BiochemStudent2
You can try something along these lines:

from proteinbert import load_pretrained_model

seq = 'MDDADPEERNYDNMLKMLSDLNKDLEKLLEEMEKISVQATWMAYDMVVMRTNPTLAESMRRLEDAFVNCKEEMEKNWQELLHETKQRL'
seq_len = 512

pretrained_model_generator, input_encoder = load_pretrained_model()
X = input_encoder.encode_X([seq], seq_len)
model = pretrained_model_generator.create_model(seq_len)
y_seq, y_annotations = model.predict(X, batch_size = 1)

from protein_bert.

Thank you so much! This works great.

from protein_bert.

seq_len

Hi there.

For the seq_len specified in this code, do we have to set the seq_len here, to be the same value as 'seq_len' OR the 'final_seq_len' we used during fine tuning, if we are now trying to use the fine tuned model, to now do predictions based on the fine tuning? As we have the seq_len and final_seq_len as different values. Kindly advise.

from protein_bert.

@ZakiaSalod The model is designed such that it should be quite flexible about seq_len and work reasonably well with different values even without further training of the model (that's in fact one of the motivation of pre-training it with different sequence lengths and fine-tuning it with seq_len and final_seq_len that are different). Having said that, I expect that a seq_len equal to the original seq_len (not final_seq_len) used during fine-tuning may provide slightly better performance. However, if a larger seq_len is called for (e.g. if you need to process longer sequences) don't hesitate to use larger values.

from protein_bert.