def opt_fn(hyperparams,
data_path,
test_set,
save_dir,
args_file,
metric='percent_improved_mae',
input_file=None,
num_decode=20,
seed=1):
## Get args:
args = read_args(args_file)
run_training(data_path,save_dir,args_file,input_file,hyperparams)
vocab_file = os.path.join(save_dir,'inputs','vocab.txt')
model_file = os.path.join(save_dir,'models','model.'+str(args['epoch']-1))
if not os.path.isdir(os.path.join(save_dir,'eval')):
os.mkdir(os.path.join(save_dir,'eval'))
output_file = os.path.join(save_dir,'eval','decoded_mols.csv')
stats_file = os.path.join(save_dir,'eval','stats.pkl')
decode(test_set,vocab_file,model_file,output_file,args,
atom_vocab=common_atom_vocab,
num_decode=num_decode, ## Will not come from run input
seed=seed,
hyperparams=hyperparams)
stats,_ = evaluate_chemprop(output_file,fold_path=args['fold_path'])
with open(stats_file, 'wb') as f:
pickle.dump(stats, f, pickle.HIGHEST_PROTOCOL)
return stats[metric]
def iterate_round(args_file,
save_dir,
data_path,
chemprop_path,
constraint_file=None,
iteration_num=1,
solvent=None):
args = read_args(args_file)
save_dir1 = os.path.join(save_dir, 'iteration' + str(iteration_num))
if not os.path.isdir(save_dir1):
os.mkdir(save_dir1)
# Train model:
run_training(data_path,
save_dir1,
args_file,
chemprop_path=chemprop_path,
constraint_file=constraint_file)
# Make augment folder
if not os.path.isdir(os.path.join(save_dir1, 'augment')):
os.mkdir(os.path.join(save_dir1, 'augment'))
augment_folder = os.path.join(save_dir1, 'augment')
molfile = os.path.join(save_dir1, 'inputs', 'mols.txt')
vocab = os.path.join(save_dir1, 'inputs', 'vocab.txt')
model = os.path.join(save_dir1, 'models',
'model.' + str(args['epoch'] - 1))
args_file = os.path.join(save_dir1, 'input.dat')
# Generate new molecules based on original dataset:
decode(
molfile,
vocab,
model,
os.path.join(augment_folder, 'gen_out.csv'),
args, # Args are needed to make sure the network architecture is correct
atom_vocab=common_atom_vocab,
num_decode=args['num_decode'])
# Assign/predict molecule properties:
if solvent == None:
_, preds_tot = evaluate_chemprop(os.path.join(augment_folder,
'gen_out.csv'),
fold_path=args['fold_path'],
chemprop_path=chemprop_path)
preds_tot.to_csv(os.path.join(augment_folder, 'gen_evaluated.csv'),
index=False)
else:
_, preds_tot = evaluate_chemprop_sol(os.path.join(
augment_folder, 'gen_out.csv'),
solvent=solvent,
fold_path=args['fold_path'],
chemprop_path=chemprop_path)
preds_tot.to_csv(os.path.join(augment_folder, 'gen_evaluated.csv'),
index=False)
# Apply filters and create new datafile
update_dataset(os.path.join(augment_folder, 'gen_evaluated.csv'),
os.path.join(augment_folder, 'data.csv'),
target=args['target'],
threshold=args['cutoff_iterations'],
min_mol_wt=args['min_mol_wt'],
pairing_method=args['pairing_method'],
n_clusters=args['n_clusters'],
tan_threshold=args['tan_threshold']
) # Reusing cutoff criteria defined for pairing algorithm
# Return locations of folders
return augment_folder
from g2g_optimization.train.decode import decode
from g2g_optimization.train.args import read_args
from g2g_optimization.hgraph import common_atom_vocab
lg = rdkit.RDLogger.logger()
lg.setLevel(rdkit.RDLogger.CRITICAL)
parser = argparse.ArgumentParser()
parser.add_argument('--test',required=True)
parser.add_argument('--vocab',required=True)
parser.add_argument('--model',required=True)
parser.add_argument('--output_file',required=True)
parser.add_argument('--args_file',type=str, default=None) #Without an args file, many parameters will revert to default
parser.add_argument('--num_decode',type=int, default=20)
parser.add_argument('--seed',type=int, default=1)
args = parser.parse_args()
if args.args_file == None:
print('WARNING: You are running without an args_file')
args_file = {}
else:
args_file = read_args(args.args_file)
decode(args.test,args.vocab,args.model,args.output_file,args_file,
atom_vocab=common_atom_vocab,
num_decode=args.num_decode, ## Will not come from run input
seed=args.seed)
def run_training(data_path='data/solvation_open',save_dir='checkpoints',args_file=None,chemprop_path='/data/rsg/chemistry/cbilod/chemprop/',
constraint_file=None,
input_file=None,
hyperparams=None):
args = {}
if args_file!=None:
args = read_args(args_file)
if input_file==None:
input_file = os.path.join(data_path,'data.csv')
## Hyperparameters:-------
if hyperparams !=None:
if 'latent_size' in list(hyperparams.keys()):
args['latent_size'] = hyperparams['latent_size']
#--------------------------
if not os.path.isdir(os.path.join(save_dir,'inputs')):
os.mkdir(os.path.join(save_dir,'inputs'))
input_dir = os.path.join(save_dir,'inputs')
mol_file = os.path.join(input_dir,'mols.txt')
vocab_file = os.path.join(input_dir,'vocab.txt')
train_file = os.path.join(input_dir,'train_pairs.txt')
adjacency_file = os.path.join(input_dir,'adjacency.pkl')
if not os.path.isdir(os.path.join(save_dir,'tensors')):
os.mkdir(os.path.join(save_dir,'tensors'))
tensor_dir = os.path.join(save_dir,'tensors')
if not os.path.isdir(os.path.join(save_dir,'models')):
os.mkdir(os.path.join(save_dir,'models'))
model_dir = os.path.join(save_dir,'models')
log_file = os.path.join(save_dir,'run.log')
f = open(log_file, 'w')
f.write('Arguments:\n')
f.write(str(args)+'\n')
f.write('\n')
f.write('Starting Pair Generation \n')
start = time.time()
generate_pairs(input_file, train_file, mol_file,args,chemprop_path,constraint_file,adjacency_file)
end = time.time()
f.write('Ending Pair Generation \n')
f.write('Time Elapsed: '+str(end-start)+'\n')
f.write('\n')
f.write('Starting Vocab \n')
start = time.time()
get_vocab(mol_file,vocab_file)
end = time.time()
f.write('Ending Vocab \n')
f.write('Time Elapsed: '+str(end-start)+'\n')
f.write('\n')
f.write('Starting Preprocessing \n')
start = time.time()
generate_tensors(train_file,vocab_file,tensor_dir,args)
end = time.time()
f.write('Ending Preprocessing \n')
f.write('Time Elapsed: '+str(end-start) + '\n')
f.write('\n')
f.write('Starting Model Training \n')
start = time.time()
print(args)
gnn_train(tensor_dir,vocab_file,model_dir,args)
end = time.time()
f.write('Ending Model Training \n')
f.write('Time Elapsed: '+str(end-start)+ '\n')
f.write('\n')
if args_file!=None:
copyfile(args_file,os.path.join(save_dir,'input.dat'))