if os.path.exists(save_dir + 'Test_RMSE_ll.txt'):
os.remove(save_dir + 'Test_RMSE_ll.txt')
if os.path.exists(save_dir + 'best_arc_scores.txt'):
os.remove(save_dir + 'best_arc_scores.txt')
while iteration < BO_rounds:
if args.predictor:
pred = model.predictor(torch.FloatTensor(X_test).to(device))
pred = pred.detach().cpu().numpy()
pred = (-pred - mean_y_train) / std_y_train
uncert = np.zeros_like(pred)
else:
# We fit the GP
M = 500
sgp = SparseGP(X_train, 0 * X_train, y_train, M)
sgp.train_via_ADAM(X_train, 0 * X_train, y_train, X_test, X_test * 0, \
y_test, minibatch_size = 2 * M, max_iterations = max_iter, learning_rate = lr)
pred, uncert = sgp.predict(X_test, 0 * X_test)
print("predictions: ", pred.reshape(-1))
print("real values: ", y_test.reshape(-1))
error = np.sqrt(np.mean((pred - y_test)**2))
testll = np.mean(sps.norm.logpdf(pred - y_test, scale=np.sqrt(uncert)))
print('Test RMSE: ', error)
print('Test ll: ', testll)
pearson = float(pearsonr(pred.reshape(-1), y_test.reshape(-1))[0])
print('Pearson r: ', pearson)
with open(save_dir + 'Test_RMSE_ll.txt', 'a') as test_file:
test_file.write(
'Test RMSE: {:.4f}, ll: {:.4f}, Pearson r: {:.4f}\n'.format(
error, testll, pearson))
y_train = y[permutation][0:np.int(np.round(0.9 * n))]
y_test = y[permutation][np.int(np.round(0.9 * n)):]
import os.path
np.random.seed(random_seed)
iteration = 0
while iteration < 5:
# We fit the GP
np.random.seed(iteration * random_seed)
M = 500
sgp = SparseGP(X_train, 0 * X_train, y_train, M)
sgp.train_via_ADAM(X_train, 0 * X_train, y_train, X_test, X_test * 0, \
y_test, minibatch_size = 10 * M, max_iterations = 50, learning_rate = 0.0005)
pred, uncert = sgp.predict(X_test, 0 * X_test)
error = np.sqrt(np.mean((pred - y_test)**2))
testll = np.mean(sps.norm.logpdf(pred - y_test, scale=np.sqrt(uncert)))
print 'Test RMSE: ', error
print 'Test ll: ', testll
pred, uncert = sgp.predict(X_train, 0 * X_train)
error = np.sqrt(np.mean((pred - y_train)**2))
trainll = np.mean(sps.norm.logpdf(pred - y_train, scale=np.sqrt(uncert)))
print 'Train RMSE: ', error
print 'Train ll: ', trainll
# We load the decoder to obtain the molecules
n = X.shape[0]
permutation = np.random.choice(n, n, replace=False)
X_train = X[permutation, :][0:np.int(np.round(0.9 * n)), :]
X_test = X[permutation, :][np.int(np.round(0.9 * n)):, :]
y_train = y[permutation][0:np.int(np.round(0.9 * n))]
y_test = y[permutation][np.int(np.round(0.9 * n)):]
iteration = 0
while iteration < 5:
np.random.seed(iteration * args.seed)
M = 500
sgp = SparseGP(X_train, 0 * X_train, y_train, M)
sgp.train_via_ADAM(X_train, 0 * X_train, y_train, X_test, X_test * 0, \
y_test, minibatch_size = 10 * M, max_iterations = cmd_args.num_epochs, learning_rate = args.gp_lr)
# pred, uncert = sgp.predict(X_test, 0 * X_test)
# error = np.sqrt(np.mean((pred - y_test)**2))
# testll = np.mean(sps.norm.logpdf(pred - y_test, scale = np.sqrt(uncert)))
# print 'Test RMSE: ', error
# print 'Test ll: ', testll
# pred, uncert = sgp.predict(X_train, 0 * X_train)
# error = np.sqrt(np.mean((pred - y_train)**2))
# trainll = np.mean(sps.norm.logpdf(pred - y_train, scale = np.sqrt(uncert)))
# print 'Train RMSE: ', error
# print 'Train ll: ', trainll
next_inputs = sgp.batched_greedy_ei(50, np.min(X_train, 0),
np.max(X_train, 0))
# y /= np.max(y)
assert X.shape[0] == y.shape[0]
n = X.shape[ 0 ]
permutation = np.random.choice(n, n, replace = False)
X_train = X[ permutation, : ][ 0 : np.int(np.round(0.9 * n)), : ]
X_test = X[ permutation, : ][ np.int(np.round(0.9 * n)) :, : ]
y_train = y[ permutation ][ 0 : np.int(np.round(0.9 * n)) ]
y_test = y[ permutation ][ np.int(np.round(0.9 * n)) : ]
np.random.seed(0)
M = 500
sgp = SparseGP(X_train, 0 * X_train, y_train, M)
sgp.train_via_ADAM(X_train, 0 * X_train, y_train, X_test, X_test * 0, \
y_test, minibatch_size = 10 * M, max_iterations = cmd_args.num_epochs, learning_rate = args.gp_lr)
with open('%s/sgp-e-%d-seed-%d-lr-%.4f.txt' % (cmd_args.save_dir, cmd_args.num_epochs, args.seed, args.gp_lr), 'w') as f:
pred, uncert = sgp.predict(X_test, 0 * X_test)
error = np.sqrt(np.mean((pred - y_test)**2))
testll = np.mean(sps.norm.logpdf(pred - y_test, scale = np.sqrt(uncert)))
f.write('Test RMSE: %.10f\n' % error)
f.write('Test ll: %.10f\n' % testll)
print 'Test RMSE: ', error
print 'Test ll: ', testll
pred, uncert = sgp.predict(X_train, 0 * X_train)
error = np.sqrt(np.mean((pred - y_train)**2))
trainll = np.mean(sps.norm.logpdf(pred - y_train, scale = np.sqrt(uncert)))
f.write('Train RMSE: %.10f\n' % error)
f.write('Train ll: %.10f\n' % trainll)
logP_values_normalized = (np.array(logP_values) -
np.mean(logP_values)) / np.std(logP_values)
cycle_scores_normalized = (np.array(cycle_scores) -
np.mean(cycle_scores)) / np.std(cycle_scores)
iteration = 0
while iteration < 2:
# We fit the GP
np.random.seed(iteration * random_seed)
M = 1
sgp = SparseGP(X_train, 0 * X_train, y_train, M)
sgp.train_via_ADAM(X_train,
0 * X_train,
y_train,
X_test,
X_test * 0,
y_test,
minibatch_size=2,
max_iterations=100,
learning_rate=0.001)
pred, uncert = sgp.predict(X_test, 0 * X_test)
error = np.sqrt(np.mean((pred - y_test)**2))
testll = np.mean(sps.norm.logpdf(pred - y_test, scale=np.sqrt(uncert)))
print 'Test RMSE: ', error
print 'Test ll: ', testll
pred, uncert = sgp.predict(X_train, 0 * X_train)
error = np.sqrt(np.mean((pred - y_train)**2))
trainll = np.mean(sps.norm.logpdf(pred - y_train, scale=np.sqrt(uncert)))
print 'Train RMSE: ', error
def run_bo_demo():
import sys
sys.path.append('/home/icml18-jtnn')
import pickle
import gzip
from sparse_gp import SparseGP
import scipy.stats as sps
import numpy as np
import os.path
import time
import rdkit
from rdkit.Chem import MolFromSmiles, MolToSmiles
from rdkit.Chem import Descriptors
from rdkit.Chem import PandasTools
import torch
import torch.nn as nn
from jtnn import create_var, JTNNVAE, Vocab
start_time = time.time()
lg = rdkit.RDLogger.logger()
lg.setLevel(rdkit.RDLogger.CRITICAL)
# We define the functions used to load and save objects
def save_object(obj, filename):
result = pickle.dumps(obj)
with gzip.GzipFile(filename, 'wb') as dest:
dest.write(result)
dest.close()
def load_object(filename):
with gzip.GzipFile(filename, 'rb') as source:
result = source.read()
ret = pickle.loads(result)
source.close()
return ret
vocab_path = '../data/vocab.txt'
#save_dir=save_dir
vocab = [x.strip("\r\n ") for x in open(vocab_path)]
vocab = Vocab(vocab)
# print(opts.save_dir)
hidden_size = 450
latent_size = 56
depth = 3
random_seed = 1
model = JTNNVAE(vocab, hidden_size, latent_size, depth)
model.load_state_dict(
torch.load('../molvae/MPNVAE-h450-L56-d3-beta0.005/model.iter-4',
map_location=lambda storage, loc: storage))
#model = model.cuda()
# We load the random seed
np.random.seed(random_seed)
# We load the data (y is minued!)
X = np.loadtxt('latent_features_demo.txt')
y = -np.loadtxt('targets_demo.txt')
y = y.reshape((-1, 1))
n = X.shape[0]
# print(X.shape[1])
permutation = np.random.choice(n, n, replace=False)
# print(n)
X_train = X[permutation, :][0:np.int(np.round(0.8 * n)), :]
X_test = X[permutation, :][np.int(np.round(0.8 * n)):, :]
# print(X_train.shape)
y_train = y[permutation][0:np.int(np.round(0.8 * n))]
y_test = y[permutation][np.int(np.round(0.8 * n)):]
np.random.seed(random_seed)
logP_values = np.loadtxt('logP_values_demo.txt')
SA_scores = np.loadtxt('SA_scores_demo.txt')
cycle_scores = np.loadtxt('cycle_scores_demo.txt')
SA_scores_normalized = (np.array(SA_scores) -
np.mean(SA_scores)) / np.std(SA_scores)
logP_values_normalized = (np.array(logP_values) -
np.mean(logP_values)) / np.std(logP_values)
cycle_scores_normalized = (np.array(cycle_scores) -
np.mean(cycle_scores)) / np.std(cycle_scores)
iteration = 0
while iteration < 1:
# We fit the GP
np.random.seed(iteration * random_seed)
M = 1
sgp = SparseGP(X_train, 0 * X_train, y_train, M)
sgp.train_via_ADAM(X_train,
0 * X_train,
y_train,
X_test,
X_test * 0,
y_test,
minibatch_size=2,
max_iterations=100,
learning_rate=0.001)
pred, uncert = sgp.predict(X_test, 0 * X_test)
error = np.sqrt(np.mean((pred - y_test)**2))
testll = np.mean(sps.norm.logpdf(pred - y_test, scale=np.sqrt(uncert)))
# print 'Test RMSE: ', error
# print 'Test ll: ', testll
pred, uncert = sgp.predict(X_train, 0 * X_train)
error = np.sqrt(np.mean((pred - y_train)**2))
trainll = np.mean(
sps.norm.logpdf(pred - y_train, scale=np.sqrt(uncert)))
# print 'Train RMSE: ', error
# print 'Train ll: ', trainll
# We pick the next 60 inputs
next_inputs = sgp.batched_greedy_ei(60, np.min(X_train, 0),
np.max(X_train, 0))
valid_smiles = []
valid_mols = []
new_features = []
for i in xrange(60):
all_vec = next_inputs[i].reshape((1, -1))
tree_vec, mol_vec = np.hsplit(all_vec, 2)
tree_vec = create_var(torch.from_numpy(tree_vec).float())
mol_vec = create_var(torch.from_numpy(mol_vec).float())
s = model.decode(tree_vec, mol_vec, prob_decode=False)
if s is not None:
valid_smiles.append(s)
# print(MolFromSmiles(s))
valid_mols.append(str(MolFromSmiles(s)))
new_features.append(all_vec)
print len(valid_smiles), "molecules are found"
valid_smiles = valid_smiles[:50]
valid_mols = valid_mols[:50]
new_features = next_inputs[:50]
new_features = np.vstack(new_features)
# save_object(valid_smiles, save_dir + "/valid_smiles{}.dat".format(iteration))
# save_object(valid_mols,save_dir + '/valid_mols{}.png'.format(iteration))
# save_object(mol1
import sascorer
import networkx as nx
from rdkit.Chem import rdmolops
scores = []
for i in range(len(valid_smiles)):
current_log_P_value = Descriptors.MolLogP(
MolFromSmiles(valid_smiles[i]))
current_SA_score = -sascorer.calculateScore(
MolFromSmiles(valid_smiles[i]))
cycle_list = nx.cycle_basis(
nx.Graph(
rdmolops.GetAdjacencyMatrix(MolFromSmiles(
valid_smiles[i]))))
if len(cycle_list) == 0:
cycle_length = 0
else:
cycle_length = max([len(j) for j in cycle_list])
if cycle_length <= 6:
cycle_length = 0
else:
cycle_length = cycle_length - 6
current_cycle_score = -cycle_length
current_SA_score_normalized = (
current_SA_score - np.mean(SA_scores)) / np.std(SA_scores)
current_log_P_value_normalized = (
current_log_P_value -
np.mean(logP_values)) / np.std(logP_values)
current_cycle_score_normalized = (
current_cycle_score -
np.mean(cycle_scores)) / np.std(cycle_scores)
score = current_SA_score_normalized + current_log_P_value_normalized + current_cycle_score_normalized
scores.append(-score) #target is always minused
# print valid_smiles
# print scores
# save_object(scores, save_dir + "/scores{}.dat".format(iteration))
if len(new_features) > 0:
X_train = np.concatenate([X_train, new_features], 0)
y_train = np.concatenate([y_train, np.array(scores)[:, None]], 0)
iteration += 1
# print('Seconds taken: %s' %(time.time() -start_time))
all_smiles = []
#all_smiles=valid_smiles+scores+valid_mols
all_smiles.extend(zip(valid_smiles, scores, valid_mols))
all_smiles = [(x, -y, z) for x, y, z in all_smiles]
all_smiles = sorted(all_smiles, key=lambda x: x[1], reverse=True)
# return valid_smiles[0:5],scores[0:5],valid_mols[0:5]
return all_smiles[0:3]
def main(input_directory, output_directory):
"""
:param input_directory: directory to which the output of Branin_Sampler.py was saved.
:param output_directory: directory in which to save the plots.
"""
np.random.seed(2)
# Load the dataset
X_bran = genfromtxt(input_directory + '/inputs.csv',
delimiter=',',
dtype='float32')
y_con = genfromtxt(input_directory + '/constraint_targets.csv',
delimiter=',',
dtype='int')
y_reg = genfromtxt(input_directory + '/branin_targets.csv',
delimiter=',',
dtype='float32')
y_reg = y_reg.reshape((-1, 1))
# We convert constraint targets from one-hot to categorical.
y_con_cat = np.zeros(len(y_con), dtype=int)
i = 0
for element in y_con:
if element[0] == 1:
y_con_cat[i] = 1
else:
y_con_cat[i] = 0
i += 1
y_con = y_con_cat
n_bran = X_bran.shape[0] # number of examples
permutation = np.random.choice(n_bran, n_bran,
replace=False) # We shuffle the data
X_tr_bran = X_bran[permutation, :][40:np.int(np.round(
0.9 * n_bran)), :] # 50/10 train/test split.
X_te_bran = X_bran[permutation, :][
np.int(np.round(0.8 * n_bran)):np.int(np.round(0.9 * n_bran)), :]
y_tr_reg = y_reg[permutation][40:np.int(
np.round(0.9 * n_bran)
)] # 10:20 have balanced class split after the permutation is applied with random seed = 1
y_te_reg = y_reg[permutation][np.int(np.round(0.8 * n_bran)):np.
int(np.round(0.9 * n_bran))]
y_tr_con = y_con[permutation][40:np.int(
np.round(0.9 * n_bran)
)] # no test set for constraint as traning subroutine for BNN doesn't require it
y_te_con = y_con[permutation][np.int(np.round(0.8 * n_bran)):np.
int(np.round(0.9 * n_bran))]
# We plot the data used to initialise the surrogate model
X1 = X_tr_bran[:, 0]
X2 = X_tr_bran[:, 1]
save_object(X1, output_directory + "/X1.dat")
save_object(X2, output_directory + "/X2.dat")
# We store the best feasible value found in the training set for reference
feasible_vals = []
for i in range(X_tr_bran.shape[0]):
if y_tr_con[i] == 0:
continue
feasible_vals.append([branin(tuple(X_tr_bran[i]))])
best_tr = min(feasible_vals)
best_tr = best_tr[0]
save_object(best_tr,
output_directory + "/best_feasible_training_point.dat")
# We set the number of data colletion iterations
num_iters = 4
for iteration in range(num_iters):
# We train the regression model
# We fit the GP
# M = np.int(np.maximum(10,np.round(0.1 * n_bran)))
M = 20
sgp = SparseGP(X_tr_bran, 0 * X_tr_bran, y_tr_reg, M)
sgp.train_via_ADAM(X_tr_bran,
0 * X_tr_bran,
y_tr_reg,
X_te_bran,
X_te_bran * 0,
y_te_reg,
minibatch_size=M,
max_iterations=400,
learning_rate=0.005)
save_object(sgp, output_directory + "/sgp{}.dat".format(iteration))
# We load the saved gp
sgp = load_object(output_directory + "/sgp{}.dat".format(iteration))
# We load some previous trained gp
pred, uncert = sgp.predict(X_te_bran, 0 * X_te_bran)
error = np.sqrt(np.mean((pred - y_te_reg)**2))
testll = np.mean(
sps.norm.logpdf(pred - y_te_reg, scale=np.sqrt(uncert)))
print('Test RMSE: ', error)
print('Test ll: ', testll)
pred, uncert = sgp.predict(X_tr_bran, 0 * X_tr_bran)
error = np.sqrt(np.mean((pred - y_tr_reg)**2))
trainll = np.mean(
sps.norm.logpdf(pred - y_tr_reg, scale=np.sqrt(uncert)))
print('Train RMSE: ', error)
print('Train ll: ', trainll)
# we train the constraint network
# We load the random seed
seed = 1
np.random.seed(seed)
# We load the data
datasets, n, d, n_labels = load_data(X_tr_bran, y_tr_con, X_te_bran,
y_te_con)
train_set_x, train_set_y = datasets[0]
test_set_x, test_set_y = datasets[1]
N_train = train_set_x.get_value(borrow=True).shape[0]
N_test = test_set_x.get_value(borrow=True).shape[0]
layer_sizes = [d, 50, n_labels]
n_samples = 50
alpha = 0.5
learning_rate = 0.001
v_prior = 1.0
batch_size = 10
print('... building model')
sys.stdout.flush()
bb_alpha = BB_alpha(layer_sizes, n_samples, alpha, learning_rate,
v_prior, batch_size, train_set_x, train_set_y,
N_train, test_set_x, test_set_y, N_test)
print('... training')
sys.stdout.flush()
test_error, test_ll = bb_alpha.train(400)
# We save the trained BNN
sys.setrecursionlimit(4000) # Required to save the BNN
save_object(bb_alpha,
output_directory + "/bb_alpha{}.dat".format(iteration))
# We pick the next 5 inputs based on random sampling
np.random.seed()
num_inputs = 1
x1 = np.random.uniform(-5, 10, size=num_inputs)
x2 = np.random.uniform(0, 15, size=num_inputs)
random_inputs = np.zeros([num_inputs, 2])
random_inputs[:, 0] = x1
random_inputs[:, 1] = x2
reg_scores = [] # collect y-values for Branin-Hoo function
con_scores = [] # collect y-values for Constraint function
probs = [] # collect the probabilities of satisfying the constraint
log_probs = [
] # collect the log probabilities of satisfying the constraint
for i in range(random_inputs.shape[0]):
reg_scores.append([branin(tuple(random_inputs[i]))])
if (random_inputs[i][0] - 2.5)**2 + (random_inputs[i][1] -
7.5)**2 <= 50:
con_scores.append(np.int64(1))
else:
con_scores.append(np.int64(0))
probs.append(
bb_alpha.prediction_probs(random_inputs[i].reshape(
1, d))[0][0][1])
log_probs.append(
bb_alpha.pred_log_probs(random_inputs[i].reshape(1,
d))[0][0][1])
print(i)
# print the value of the Branin-Hoo function at the data points we have acquired
print(reg_scores)
# save y-values and (x1,x2)-coordinates of locations chosen for evaluation
save_object(reg_scores,
output_directory + "/scores{}.dat".format(iteration))
save_object(random_inputs,
output_directory + "/next_inputs{}.dat".format(iteration))
save_object(con_scores,
output_directory + "/con_scores{}.dat".format(iteration))
save_object(probs, output_directory + "/probs{}.dat".format(iteration))
save_object(log_probs,
output_directory + "/log_probs{}.dat".format(iteration))
# extend labelled training data for next cycle
X_tr_bran = np.concatenate([X_tr_bran, random_inputs], 0)
y_tr_reg = np.concatenate([y_tr_reg, np.array(reg_scores)], 0)
y_tr_con = np.concatenate([y_tr_con, np.array(con_scores)], 0)
best_so_far(
output_directory, num_iters
) # Plot the best point as a function of the data collection iteration number
GP_contours(output_directory,
num_iters) # Plot the contours of the GP regression model
BNN_contours(output_directory,
num_iters) # Plot the contours of the BNN constraint model
initial_data(
output_directory) # Plot the data used to initialise the model