File size: 10,637 Bytes
5611f26 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 | import argparse
import json
import logging
import os
import time
from dataset import SDataset, MDataset
from network import MMPeptide, SEQPeptide, VoxPeptide, MMFPeptide, SMPeptide
from sklearn.model_selection import KFold
from train import train
import torch
import torch.nn as nn
from torch.utils.data import Subset, DataLoader
import numpy as np
from loss import MLCE, SuperLoss, LogCoshLoss
from utils import set_seed
parser = argparse.ArgumentParser(description='resnet26')
# model setting
parser.add_argument('--model', type=str, default='mm',
help='model resnet26, bi-gru')
parser.add_argument('--fusion', type=str, default='1',
help="Seed for splitting dataset (default 1)")
# task & dataset setting
parser.add_argument('--pdb-src', type=str, dest='pdb_src', default='af',
help='af or hf')
parser.add_argument('--task-type', type=str, dest='task_type', default='mlc',
help='mlc or slc')
parser.add_argument('--task', type=str, default='all',
help='task: anti toxin anti-all mechanism anti-binary anti-regression mic')
parser.add_argument('--classes', type=int, default=6,
help='model')
parser.add_argument('--max-length', dest='max_length', type=int, default=30,
help='Max length for sequence filtering')
parser.add_argument('--split', type=int, default=5,
help="Split k fold in cross validation (default: 5)")
parser.add_argument('--seed', type=int, default=1,
help="Seed for splitting dataset (default: 1)")
parser.add_argument('--threshold', type=float, default=64,
help="MIC threshold for determine labels (default: 64)")
# training setting
parser.add_argument('--gpu', type=int, default=0,
help='GPU index to use, -1 for CPU (default: 0)')
parser.add_argument('--batch-size', type=int, dest='batch_size', default=128,
help='input batch size for training (default: 128)')
parser.add_argument('--epochs', type=int, default=50,
help='number of epochs to train (default: 50)')
parser.add_argument('--lr', type=float, default=0.02,
help='learning rate (default: 0.002)')
parser.add_argument('--decay', type=float, default=0.0005,
help='weight decay (default: 0.0005)')
parser.add_argument('--warm-steps', type=int, dest='warm_steps', default=0,
help='number of warm start steps for learning rate (default: 10)')
parser.add_argument('--patience', type=int, default=10,
help='patience for early stopping (default: 10)')
parser.add_argument('--pretrain', type=str, dest='pretrain', default='',
help='path of the pretrain model') # /home/duadua/Desktop/fetal/3dpretrain/runs/e50.pth
parser.add_argument('--metric-avg', type=str, dest='metric_avg', default='micro',
help='metric average type')
# args for losses
parser.add_argument('--loss', type=str, default='mlce',
help='loss function (mlce, sl, mix)')
parser.add_argument('--bias-curri', dest='bias_curri', action='store_true', default=False,
help='directly use loss as the training data (biased) or not (unbiased)')
parser.add_argument('--anti-curri', dest='anti_curri', action='store_true', default=False,
help='easy to hard (curri), hard to easy (anti)')
parser.add_argument('--std-coff', dest='std_coff', type=float, default=1,
help='the hyper-parameter of std')
args = parser.parse_args()
def main():
set_seed(args.seed)
if args.task_type == 'slc':
if args.task == 'all':
raise ValueError('Choose one task number to run single label classification')
args.classes = 1
elif args.task_type == 'mlc':
pass
else:
raise NotImplementedError
weight_dir = "./run/" + args.task_type + '-' + args.task + "-m-" + args.model + '-' + args.loss + str(args.batch_size) + str(args.lr) + str(args.epochs) + args.pdb_src
if not os.path.exists(weight_dir):
raise ValueError
logging.basicConfig(handlers=[
# logging.FileHandler(filename=os.path.join(weight_dir, "eval.log"), encoding='utf-8', mode='w+'),
logging.StreamHandler()],
format="%(asctime)s: %(message)s", datefmt="%F %T", level=logging.INFO)
logging.info(f'saving_dir: {weight_dir}')
# with open(os.path.join(weight_dir, "config.json"), "w") as f:
# f.write(json.dumps(vars(args)))
device = torch.device("cpu" if args.gpu == -1 or not torch.cuda.is_available() else f"cuda:{args.gpu}")
logging.info('Loading Training Dataset')
if args.task_type == 'mlc':
set_all = MDataset(threshold=args.threshold, mode='train', max_length=args.max_length, pdb_src=args.pdb_src)
else:
set_all = SDataset(threshold=args.threshold, mode='train', task=args.task, max_length=args.max_length, pdb_src=args.pdb_src)
logging.info('Loading Test Dataset')
if args.task_type == 'mlc':
qlx_set = MDataset(threshold=args.threshold, mode='qlx', max_length=args.max_length, pdb_src=args.pdb_src)
saap_set = MDataset(threshold=args.threshold, mode='saap', max_length=args.max_length, pdb_src=args.pdb_src)
else:
qlx_set = SDataset(threshold=args.threshold, mode='qlx', task=args.task, max_length=args.max_length, pdb_src=args.pdb_src)
saap_set = SDataset(threshold=args.threshold, mode='saap', task=args.task, max_length=args.max_length, pdb_src=args.pdb_src)
best_perform_list = [[] for i in range(5)]
qlx_perform_list = [[] for i in range(5)]
saap_perform_list = [[] for i in range(5)]
kf = KFold(n_splits=5, shuffle=True, random_state=42)
for fold, (train_idx, val_idx) in enumerate(kf.split(set_all)):
# train_set= Subset(set_all, train_idx)
valid_set = Subset(set_all, val_idx)
# train_loader = DataLoader(train_set, batch_size=args.batch_size, shuffle=True, drop_last=True)
valid_loader = DataLoader(valid_set, batch_size=args.batch_size, shuffle=False)
qlx_loader = DataLoader(qlx_set, batch_size=1, shuffle=False)
saap_loader = DataLoader(saap_set, batch_size=1, shuffle=False)
if args.model == 'seq':
model = SEQPeptide(classes=set_all.num_classes, q_encoder='mlp')
elif args.model == 'voxel':
model = VoxPeptide(classes=set_all.num_classes)
elif args.model == 'mm':
# model = MMPeptide(classes=train_set.num_classes, q_encoder='tf', ) # attention='hamburger'
# model = SMPeptide(classes=train_set.num_classes, q_encoder='mlp', max_length=30) # attention='hamburger'
model = MMPeptide(classes=set_all.num_classes, q_encoder='mlp', max_length=args.max_length)# , attention='hamburger'
elif args.model == 'mmf':
model = MMFPeptide(classes=set_all.num_classes, q_encoder='mlp', max_length=args.max_length) # attention='hamburger'
if len(args.pretrain) != 0:
logging.info('loading pretrain model')
# model = load_pretrain_model(model, torch.load(args.pretrain))
model_state = model.state_dict()
pretrained_state = torch.load(args.pretrain)
pretrained_state = {k: v for k, v in pretrained_state.items() if
k in model_state and v.size() == model_state[k].size()}
model_state.update(pretrained_state)
model.load_state_dict(model_state)
# model.load_state_dict(torch.load(args.pretrain), strict=False)
model.to(device)
weights_path = f"{weight_dir}/model_{fold + 1}.pth"
model.load_state_dict(torch.load(weights_path))
# early_stopping = EarlyStopping(patience=args.patience, path=weights_path)
logging.info(f'Running Cross Validation {fold + 1}')
logging.info(f'Fold {fold + 1} Valid set:{len(valid_set)}, Test set:qlx {len(qlx_set)} saap {len(saap_set)}')
if args.task_type in ('mlc', 'slc') :
_, ap, f1, acc, auc = train(args, None, model, None, valid_loader, device, None, None)
logging.info(f'ap: {ap:.3f}, f1: {f1:.3f}, acc: {acc:.3f}, auc: {auc:.3f}')
best_perform_list[fold] = np.asarray([ap, f1, acc, auc])
_, qlx_ap, qlx_f1, qlx_acc, qlx_auc = train(args, None, model, None, qlx_loader, device, None, None)
logging.info(f'QLX results, ap: {qlx_ap:.3f}, f1: {qlx_f1:.3f}, acc: {qlx_acc:.3f}, auc: {qlx_auc:.3f}')
qlx_perform_list[fold] = np.asarray([qlx_ap, qlx_f1, qlx_acc, qlx_auc])
_, saap_ap, saap_f1, saap_acc, saap_auc = train(args, None, model, None, saap_loader, device, None, None)
logging.info(f'SAAP results, ap: {saap_ap:.3f}, f1: {saap_f1:.3f}, acc: {saap_acc:.3f}, auc: {saap_auc:.3f}')
saap_perform_list[fold] = np.asarray([saap_ap, saap_f1, saap_acc, saap_auc])
else:
raise NotImplementedError
logging.info(f'Cross Validation Finished!')
best_perform_list = np.asarray(best_perform_list)
qlx_perform_list = np.asarray(qlx_perform_list)
saap_perform_list = np.asarray(saap_perform_list)
logging.info('Best validation perform list\n%s', best_perform_list)
logging.info('mean: %s', np.round(np.mean(best_perform_list, 0), 3))
logging.info('std: %s', np.round(np.std(best_perform_list, 0), 3))
logging.info('Best qlx perform list\n%s', qlx_perform_list)
logging.info('mean: %s', np.round(np.mean(qlx_perform_list, 0), 3))
logging.info('std: %s', np.round(np.std(qlx_perform_list, 0), 3))
logging.info('Best saap perform list\n%s', saap_perform_list)
logging.info('mean: %s', np.round(np.mean(saap_perform_list, 0), 3))
logging.info('std: %s', np.round(np.std(saap_perform_list, 0), 3))
perform = open(weight_dir+'/eval.txt', 'w')
perform.write('Valid\n')
perform.write(','.join([str(i) for i in np.mean(best_perform_list, 0)])+'\n')
perform.write(','.join([str(i) for i in np.std(best_perform_list, 0)])+'\n')
perform.write('qlx\n')
perform.write(','.join([str(i) for i in np.mean(qlx_perform_list, 0)])+'\n')
perform.write(','.join([str(i) for i in np.std(qlx_perform_list, 0)])+'\n')
perform.write('saap\n')
perform.write(','.join([str(i) for i in np.mean(saap_perform_list, 0)])+'\n')
perform.write(','.join([str(i) for i in np.std(saap_perform_list, 0)])+'\n')
if __name__ == "__main__":
main()
|