Core functionality from Fast SeqProp, now compatible with tensorflow 2.
Download or clone the repo and use python setup.py install.
The following code loads a pretrained predictor and generates 100 sequences to maximize its output.
import tensorflow as tf
from tensorflow.keras import models
import corefsp
predictor = models.load_model('model_filename.h5')
# Loss function
def target_loss_func(y_pred):
return - tf.reduce_sum(y_pred)
# Generate sequences
# seq_vals has shape (n_seqs, seq_length, n_channels)
# and contains one hot-encoded designed sequences.
# pred_vals contains predictions for each generated sequence.
# train_history is a dictionary with keys 'loss', 'target_loss', and
# 'entropy_loss', containing loss values for each iteration.
seq_vals, pred_vals, train_history = corefsp.design_seqs(
predictor,
target_loss_func,
seq_length=500,
n_seqs=100,
target_weight=1,
learning_rate=0.1,
n_iter_max=1000,
)