def main():
params = Params()
mp.set_start_method('spawn')
count = mp.Value('i', 0)
best_acc = mp.Value('d', 0.0)
lock = mp.Lock()
shared_model = A3C_LSTM_GA()
shared_model = shared_model.share_memory()
shared_optimizer = SharedAdam(shared_model.parameters(),
lr=params.lr,
amsgrad=params.amsgrad,
weight_decay=params.weight_decay)
shared_optimizer.share_memory()
#run_sim(0, params, shared_model, None, count, lock)
#test(params, shared_model, count, lock, best_acc)
processes = []
train_process = 0
test_process = 0
for rank in range(params.n_process):
p = mp.Process(target=test,
args=(
test_process,
params,
shared_model,
count,
lock,
best_acc,
))
p.start()
processes.append(p)
test_process += 1
for i in range(2):
p = mp.Process(target=run_sim,
args=(
train_process,
params,
shared_model,
shared_optimizer,
count,
lock,
))
p.start()
processes.append(p)
train_process += 1
for p in processes:
p.join()
map_location=lambda storage, loc: storage)
shared_model.load_state_dict(saved_state)
shared_model.share_memory()
if args.shared_optimizer:
if args.optimizer == 'RMSprop':
optimizer = SharedRMSprop(shared_model.parameters(), lr=args.lr)
if args.optimizer == 'Adam':
optimizer = SharedAdam(shared_model.parameters(),
lr=args.lr,
amsgrad=args.amsgrad)
optimizer_r = SharedAdam(shared_model.r_net.parameters(),
lr=args.rl_r,
amsgrad=args.amsgrad)
optimizer.share_memory()
optimizer_r.share_memory()
else:
optimizer = None
processes = []
counter = mp.Value('i', 0)
lock = mp.Lock()
p = mp.Process(target=test,
args=(args, shared_model, env_conf, lock, counter))
p.start()
processes.append(p)
# p = mp.Process(target=train_rep, args=(args, shared_model, env_conf))
# p.start()
# processes.append(p)
time.sleep(0.1)
# Get the number of conditions as a hyperparameter.
# `args` delivers the hyperparameters to `ggm.ggm`.
args.N_conditions = len(next(iter(id_to_whole_conditions.values()))) + len(next(iter(id_to_scaffold_conditions.values())))
#lines for multiprocessing
mp.set_start_method('spawn')
torch.manual_seed(1)
#model
shared_model = ggm(args)
shared_model.share_memory() # torch.nn.Module.share_memory
#shared optimizer
shared_optimizer = SharedAdam(shared_model.parameters(), lr=args.lr, amsgrad=True)
shared_optimizer.share_memory()
print ("Model #Params: %dK" % (sum([x.nelement() for x in shared_model.parameters()]) / 1000,))
#initialize parameters of the model
if save_fpath:
initial_epoch, initial_cycle = [int(value) for value in re.findall('\d+', os.path.basename(save_fpath))]
shared_model = utils.initialize_model(shared_model, save_fpath)
else:
initial_epoch = initial_cycle = 0
shared_model = utils.initialize_model(shared_model, False)
num_cycles = int(len(id_to_smiles)/args.ncpus/args.item_per_cycle)
print(f"""\
ncpus : {args.ncpus}
OMP_NUM_THREADS : {os.environ.get('OMP_NUM_THREADS')}
Number of data : {len(id_to_smiles)}