def run_gmgan_for_docking(nmodel=3, MAX_EXPNUM=1, nsamples=[1, 10, 30, 50], num_train_input=100, sup_max_epoch=10000, max_epochs=20000):
queries = np.loadtxt('data/docking_queries.csv', delimiter=',')
vmps = np.loadtxt('data/docking_weights.csv', delimiter=',')
starts = np.loadtxt('data/docking_starts.csv', delimiter=',')
goals = np.loadtxt('data/docking_goals.csv', delimiter=',')
# clean the data
wtest = np.expand_dims(vmps, axis=1)
cc, successId = evaluate_docking(wtest, queries, starts,goals)
data = np.concatenate([queries, starts, goals], axis=1)
data = data[successId, :]
vmps = vmps[successId,:]
knum = np.shape(vmps)[1]
rstates = np.random.randint(0, 100, size=MAX_EXPNUM)
nn_structure = {'d_feat': 20,
'feat_layers': [40],
'mean_layers': [60],
'scale_layers': [60],
'mixing_layers': [60],
'discriminator': [20],
'lambda': [10],
'd_response': [40,5],
'd_context': [10,5]}
gmgan = GMGAN(n_comps=nmodel, context_dim=6, response_dim=knum, nn_structure=nn_structure)
gmgan.entropy_ratio = 0.0
gmgan.lratio['entropy'] = 200
gmgan.gen_sup_lrate = 0.00005
gmgan.gen_adv_lrate = 0.00005
gmgan.dis_lrate = 0.0002
gmgan.sup_max_epoch = sup_max_epoch
csrates = np.zeros(shape=(MAX_EXPNUM, len(nsamples)))
train_input = np.random.uniform(low=np.min(queries, axis=0), high=np.max(queries, axis=0), size=(num_train_input, np.shape(queries)[1]))
for expId in range(MAX_EXPNUM):
trdata, tdata, trvmps, tvmps = train_test_split(data, vmps, test_size=0.3, random_state=rstates[expId])
trdata, _, trvmps, _ = train_test_split(trdata, trvmps, test_size=0.3, random_state=rstates[expId])
print("use {} data for training and {} data for testing".format(np.shape(trdata)[0], np.shape(tdata)[0]))
print("======== Exp: {} ========".format(expId))
trqueries = trdata[:,0:6]
gmgan.create_network()#num_real_data=np.shape(trdata)[0])
gmgan.init_train()
gmgan.train(train_context=train_input, real_context=trqueries, real_response=trvmps, max_epochs=max_epochs, is_load=False, is_save=False)
tqueries = tdata[:, 0:6]
for i in range(len(nsamples)):
wout = gmgan.generate(tqueries, nsamples[i])
starts = tdata[:, 6:8]
goals = tdata[:, 8:10]
srate, _ = evaluate_docking(wout, tqueries, starts, goals)
csrates[expId, i] = srate
srates = np.mean(csrates, axis=0)
return srates
def train_evaluate_mdgan_for_docking(mdgan,
trqueries,
trvmps,
tdata,
use_entropy=False,
max_epochs=20000):
if use_entropy:
mdgan.entropy_ratio = 0.2
else:
mdgan.entropy_ratio = 0.0
train_input = np.random.uniform(low=np.min(trqueries, axis=0),
high=np.max(trqueries, axis=0),
size=(10000, np.shape(trqueries)[1]))
mdgan.create_network(num_real_data=np.shape(trqueries)[0])
mdgan.init_train()
mdgan.train(train_context=train_input,
real_context=trqueries,
real_response=trvmps,
max_epochs=max_epochs,
is_load=False,
is_save=False)
tqueries = tdata[:, 0:6]
starts = tdata[:, 6:8]
goals = tdata[:, 8:10]
wout = mdgan.generate_multi(tqueries, 1)
srate, _ = evaluate_docking(wout, tqueries, starts, goals)
return srate
def train_evaluate_gmgan_for_docking(gmgan, trqueries, trvmps, tdata, use_entropy=False, max_epochs=20000, sup_max_epoch=0,
sample_num=1, g_lrate=0.0003, d_lrate=0.001):
if use_entropy:
gmgan.entropy_ratio = 1
else:
gmgan.entropy_ratio = 0
gmgan.lratio['entropy'] = 10
gmgan.lratio['adv_cost'] = 0
gmgan.gen_sup_lrate = g_lrate
gmgan.gen_adv_lrate = g_lrate
gmgan.dis_lrate = d_lrate
gmgan.sup_max_epoch = sup_max_epoch
train_input = np.random.uniform(low=np.min(trqueries, axis=0), high=np.max(trqueries, axis=0), size=(10000, np.shape(trqueries)[1]))
gmgan.create_network()
gmgan.init_train()
gmgan.train(train_context=train_input, real_context=trqueries, real_response=trvmps, max_epochs=max_epochs, is_load=False, is_save=False)
tqueries = tdata[:, 0:6]
starts = tdata[:, 6:8]
goals = tdata[:, 8:10]
wout = gmgan.generate(tqueries, 10000, n_output=sample_num)
srate, _ = evaluate_docking(wout, tqueries, starts, goals)
return srate
def train_evaluate_mdnmp_for_docking(mdnmp,
trqueries,
trvmps,
tdata,
max_epochs=20000,
sample_num=1,
learning_rate=0.00003):
weights = np.ones(shape=(np.shape(trvmps)[0], 1))
train_weights = np.copy(weights)
mdnmp.build_mdn(learning_rate=learning_rate)
mdnmp.init_train()
isSuccess, nlls, ents = mdnmp.train(trqueries,
trvmps,
train_weights,
max_epochs=max_epochs,
is_load=False,
is_save=False)
srate = 0
if isSuccess:
tqueries = tdata[:, 0:6]
wout, _ = mdnmp.predict(tqueries, sample_num)
starts = tdata[:, 6:8]
goals = tdata[:, 8:10]
srate, _ = evaluate_docking(wout, tqueries, starts, goals)
return srate, nlls, ents
def train_evaluate_baseline_for_docking(model_name,
trqueries,
trvmps,
tdata,
sample_num=1):
model = possible_models[model_name]
if model is not None:
model.fit(trqueries, trvmps)
tqueries = tdata[:, 0:6]
starts = tdata[:, 6:8]
goals = tdata[:, 8:10]
wmeans = model.predict(tqueries)
if model is not None:
wouts = sample_baseline(wmeans, sample_num)
else:
wouts = np.random.uniform(low=trvmps.min(),
high=trvmps.max(),
size=(100, sample_num, np.shape(trvmps)[1]))
srate, _ = evaluate_docking(wouts, tqueries, starts, goals)
return srate
parser = OptionParser()
parser.add_option("-m", "--nmodel", dest="nmodel", type="int", default=3)
parser.add_option("-n", "--num_exp", dest="expnum", type="int", default=1)
parser.add_option("-d", "--result_dir", dest="result_dir", type="string", default="results_compare_docking")
(options, args) = parser.parse_args(sys.argv)
queries = np.loadtxt('../data/docking_queries.csv', delimiter=',')
vmps = np.loadtxt('../data/docking_weights.csv', delimiter=',')
starts = np.loadtxt('../data/docking_starts.csv', delimiter=',')
goals = np.loadtxt('../data/docking_goals.csv', delimiter=',')
wtest = np.expand_dims(vmps, axis=1)
cc, successId = evaluate_docking(wtest, queries, starts,goals)
data = np.concatenate([queries, starts, goals], axis=1)
data = data[successId, :]
vmps = vmps[successId,:]
knum = np.shape(vmps)[1]
rstates = np.random.randint(0, 100, size=options.expnum)
# create mdnmp
mdnmp_struct = {'d_feat': 20,
'feat_layers': [40], #[60]
'mean_layers': [60], #[60]
'scale_layers': [60],
'mixing_layers': [20]}
mdnmp = MDNMP(n_comps=options.nmodel, d_input=6, d_output=knum, nn_structure=mdnmp_struct, scaling=1, var_init=VAR_INIT)
# create gmgan
def run_mdnmp_for_docking(nmodel=3,
MAX_EXPNUM=20,
use_entropy_cost=[False, True],
model_names=["Original MDN", "Entropy MDN"],
nsamples=[1, 10, 30, 50, 70]):
queries = np.loadtxt('data/docking_queries.csv', delimiter=',')
vmps = np.loadtxt('data/docking_weights.csv', delimiter=',')
starts = np.loadtxt('data/docking_starts.csv', delimiter=',')
goals = np.loadtxt('data/docking_goals.csv', delimiter=',')
# clean the data
wtest = np.expand_dims(vmps, axis=1)
cc, successId = evaluate_docking(wtest, queries, starts, goals)
data = np.concatenate([queries, starts, goals], axis=1)
data = data[successId, :]
vmps = vmps[successId, :]
knum = np.shape(vmps)[1]
rstates = np.random.randint(0, 100, size=MAX_EXPNUM)
srates = {}
nn_structure = {
'd_feat': 20,
'feat_layers': [40],
'mean_layers': [60],
'scale_layers': [60],
'mixing_layers': [60]
}
for k in range(len(use_entropy_cost)):
mdnmp = MDNMP(n_comps=nmodel,
d_input=6,
d_output=knum,
nn_structure=nn_structure,
scaling=1)
if use_entropy_cost[k]:
mdnmp.lratio['entropy'] = 1
else:
mdnmp.lratio['entropy'] = 0
csrates = np.zeros(shape=(MAX_EXPNUM, len(nsamples)))
for expId in range(MAX_EXPNUM):
trdata, tdata, trvmps, tvmps = train_test_split(
data, vmps, test_size=0.3, random_state=rstates[expId])
trdata, _, trvmps, _ = train_test_split(
trdata, trvmps, test_size=0.3, random_state=rstates[expId])
print("use {} data for training and {} data for testing".format(
np.shape(trdata)[0],
np.shape(tdata)[0]))
print("======== Exp: {} with {} ========".format(
expId, model_names[k]))
weights = np.ones(shape=(np.shape(trvmps)[0], 1))
train_weights = np.copy(weights)
trqueries = trdata[:, 0:6]
mdnmp.build_mdn(learning_rate=0.00003)
mdnmp.init_train()
mdnmp.train(trqueries,
trvmps,
train_weights,
max_epochs=20000,
is_load=False,
is_save=False)
tqueries = tdata[:, 0:6]
for i in range(len(nsamples)):
wout, _ = mdnmp.predict(tqueries, nsamples[i])
starts = tdata[:, 6:8]
goals = tdata[:, 8:10]
srate, _ = evaluate_docking(wout, tqueries, starts, goals)
csrates[expId, i] = srate
srates[model_names[k]] = np.mean(csrates, axis=0)
return srates
