analysis/Hsu.et.al.git/src/unirep_evotune.py

'''
Unsupervised fine-tuning of UniRep on evolutionary data, "evo-tuning"
'''

import argparse
import os
import pathlib

import numpy as np
from sklearn.model_selection import train_test_split
import tensorflow.compat.v1 as tf
tf.disable_v2_behavior()

from unirep import babbler1900 as babbler
import utils

#os.environ["CUDA_VISIBLE_DEVICES"] = ''


def main():
    parser = argparse.ArgumentParser()
    parser.add_argument('seqs_fasta_path', type=pathlib.Path)
    parser.add_argument('save_weights_dir', type=pathlib.Path)
    parser.add_argument('--initial_weights_dir', type=pathlib.Path,
                default="weights/unirep/global")
    parser.add_argument('--batch_size', type=int, default=128)
    parser.add_argument('--max_seq_len', type=int, default=500)
    parser.add_argument('--num_steps', type=int, default=10000)
    parser.add_argument('--learning_rate', type=float, default=0.00001)
    args = parser.parse_args()

    # Set seeds
    tf.set_random_seed(0)
    np.random.seed(0)

    print("Num GPUs Available: ",
            len(tf.config.experimental.list_physical_devices('GPU')))

    # Load pre-trained models
    # Sync relevant weight files
    # !aws s3 sync --no-sign-request --quiet s3://unirep-public/1900_weights/ 1900_weights/
    # Import the mLSTM babbler model
    # Where model weights are stored.
    b = babbler(batch_size=args.batch_size, model_path=args.initial_weights_dir)
    
    # Load seqs from fasta.
    seqs_all = utils.read_fasta(args.seqs_fasta_path)
    seqs = dict()
    seqs['train'], seqs['val'] = train_test_split(seqs_all, test_size=0.2)

    bucket_ops = {
        'train': None,
        'val': None,
    }
    for mode in ['train', 'val']:
        prefix = str(args.seqs_fasta_path).replace('.a2m', '')
        formatted_seqs_path = prefix + f'_{mode}_formatted.txt'
        with open(formatted_seqs_path, "w") as destination:
            for i,seq in enumerate(seqs[mode]):
                seq = seq.upper().replace('-', 'X')
                seq = seq.replace('.', 'X')
                if b.is_valid_seq(seq, max_len=args.max_seq_len):
                    formatted = ",".join(map(str,b.format_seq(seq)))
                    destination.write(formatted)
                    destination.write('\n')
        bucket_ops[mode] = b.bucket_batch_pad(formatted_seqs_path,
                lower=100, upper=args.max_seq_len, interval=50)
    
    logits, seqloss, x_ph, y_ph, batch_size_ph, initial_state_ph = (
        b.get_babbler_ops())
    optimizer = tf.train.AdamOptimizer(args.learning_rate)
    tuning_op = optimizer.minimize(seqloss)
    
    args.save_weights_dir.mkdir(parents=True, exist_ok=True)

    train_loss = np.zeros(args.num_steps)
    val_loss = np.zeros(args.num_steps)
    with tf.Session() as sess:
        sess.run(tf.global_variables_initializer())
        sess.graph.finalize()
        for i in range(args.num_steps):
            print(f"Step {i}")
            batch_train = sess.run(bucket_ops['train'])
            train_loss[i], __, = sess.run([seqloss, tuning_op],
                    feed_dict={
                        x_ph: batch_train[:, :-1],
                        y_ph: batch_train[:, 1:],
                        batch_size_ph: args.batch_size,
                        initial_state_ph:b._zero_state
                    },
            )
            batch_val = sess.run(bucket_ops['val'])
            val_loss[i] = sess.run(seqloss,
                    feed_dict={
                        x_ph: batch_val[:, :-1],
                        y_ph: batch_val[:, 1:],
                        batch_size_ph: args.batch_size,
                        initial_state_ph:b._zero_state
                    },
            )
            print("Step {0}: {1} (train), {2} (val)".format(
                i, train_loss[i], val_loss[i]))
            # Save periodically
            if i % 1000 == 0 and i > 0:
                suffix = f'_{int(i / 1000)}k'
                savedir = os.path.join(args.save_weights_dir, suffix)
                pathlib.Path(savedir).mkdir(exist_ok=True)
                # Save weights
                b.dump_weights(sess, dir_name=savedir)
                # Save loss trajectories
                np.savetxt(
                        os.path.join(args.save_weights_dir, 'loss_trajectory_train.npy'),
                        train_loss)
                np.savetxt(
                        os.path.join(args.save_weights_dir, 'loss_trajectory_val.npy'),
                        val_loss)
        # Save final weights
        b.dump_weights(sess, dir_name=args.save_weights_dir)
        # Save loss trajectories
        np.savetxt(
                os.path.join(args.save_weights_dir, 'loss_trajectory_train.npy'),
                train_loss)
        np.savetxt(
                os.path.join(args.save_weights_dir, 'loss_trajectory_val.npy'),
                val_loss)


if __name__ == "__main__":
    main()