def main():
args = get_arguments()
data = pandas.read_hdf(args.infile, 'table')
keys = data[args.smiles_column].map(len) < 121
if args.length <= len(keys):
data = data[keys].sample(n=args.length)
else:
data = data[keys]
structures = data[args.smiles_column].map(lambda x: list(x.ljust(120)))
if args.property_column:
properties = data[args.property_column][keys]
del data
train_idx, test_idx = map(
np.array, train_test_split(structures.index, test_size=0.20))
charset = list(reduce(lambda x, y: set(y) | x, structures, set()))
one_hot_encoded_fn = lambda row: map(
lambda x: one_hot_array(x, len(charset)), one_hot_index(row, charset))
h5f = h5py.File(args.outfile, 'w')
h5f.create_dataset('charset', data=charset)
def create_chunk_dataset(h5file,
dataset_name,
dataset,
dataset_shape,
chunk_size=1000,
apply_fn=None):
new_data = h5file.create_dataset(dataset_name,
dataset_shape,
chunks=tuple([chunk_size] +
list(dataset_shape[1:])))
for (chunk_ixs, chunk) in chunk_iterator(dataset):
if not apply_fn:
new_data[chunk_ixs, ...] = chunk
else:
new_data[chunk_ixs, ...] = apply_fn(chunk)
create_chunk_dataset(
h5f,
'data_train',
train_idx, (len(train_idx), 120, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn, structures[ch])))
create_chunk_dataset(
h5f,
'data_test',
test_idx, (len(test_idx), 120, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn, structures[ch])))
if args.property_column:
h5f.create_dataset('property_train', data=properties[train_idx])
h5f.create_dataset('property_test', data=properties[test_idx])
h5f.close()
def main():
args = get_arguments()
data1 = open(args.infile, 'r')
data1 = data1.readlines()
data2 = open(args.infile2, 'r')
data2 = data2.readlines()
data3 = open(args.infile3, 'r')
data3 = data3.readlines()
len_train = len(data1)
len_test = len(data2) + len(data3)
data = np.array(data1 + data2 + data3)
train_idx = np.arange(len_train)
np.random.shuffle(train_idx)
test_idx = np.arange(len_train, len_train + len_test)
np.random.shuffle(test_idx)
if args.length < len(data):
data = np.random.choice(data, args.length)
charset = list(reduce(lambda x, y: set(y) | x, data, set()))
one_hot_encoded_fn = lambda row: map(
lambda x: one_hot_array(x, len(charset)), one_hot_index(row, charset))
h5f = h5py.File(args.outfile, 'w')
h5f.create_dataset('charset', data=charset)
def create_chunck_dataset(h5file,
dataset_name,
dataset,
dataset_shape,
chunk_size=1000,
apply_fn=None):
new_data = h5file.create_dataset(dataset_name,
dataset_shape,
chunks=tuple([chunk_size] +
list(dataset_shape[1:])))
for (chunk_ixs, chunk) in chunk_iterator(dataset):
if not apply_fn:
new_data[chunk_ixs, ...] = chunk
else:
#print(chunk)
#print(type(chunk))
new_data[chunk_ixs, ...] = apply_fn(chunk)
create_chunck_dataset(
h5f,
'data_train',
train_idx, (len(train_idx), MAX_LEN_STR, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn, data[ch])))
create_chunck_dataset(
h5f,
'data_test',
test_idx, (len(test_idx), MAX_LEN_STR, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn, data[ch])))
h5f.close()
def main():
args = get_arguments()
data = pandas.read_hdf(args.infile, 'table')
keys = data[args.smiles_column].map(len) < 121
if args.length <= len(keys):
data = data[keys].sample(n = args.length)
else:
data = data[keys]
structures = data[args.smiles_column].map(lambda x: list(x.ljust(120)))
if args.property_column:
properties = data[args.property_column][keys]
del data
train_idx, test_idx = map(np.array,
train_test_split(structures.index, test_size = 0.20))
charset = list(reduce(lambda x, y: set(y) | x, structures, set()))
one_hot_encoded_fn = lambda row: map(lambda x: one_hot_array(x, len(charset)),
one_hot_index(row, charset))
h5f = h5py.File(args.outfile, 'w')
h5f.create_dataset('charset', data = charset)
def create_chunk_dataset(h5file, dataset_name, dataset, dataset_shape,
chunk_size=1000, apply_fn=None):
new_data = h5file.create_dataset(dataset_name, dataset_shape,
chunks=tuple([chunk_size]+list(dataset_shape[1:])))
for (chunk_ixs, chunk) in chunk_iterator(dataset):
if not apply_fn:
new_data[chunk_ixs, ...] = chunk
else:
new_data[chunk_ixs, ...] = apply_fn(chunk)
create_chunk_dataset(h5f, 'data_train', train_idx,
(len(train_idx), 120, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn,
structures[ch])))
create_chunk_dataset(h5f, 'data_test', test_idx,
(len(test_idx), 120, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn,
structures[ch])))
if args.property_column:
h5f.create_dataset('property_train', data = properties[train_idx])
h5f.create_dataset('property_test', data = properties[test_idx])
h5f.close()
def main():
#args = get_arguments()
data = pandas.read_hdf('data/smiles_50k.h5', 'table')
keys = data[SMILES_COL_NAME].map(len) < 121
if MAX_NUM_ROWS <= len(keys):
data = data[keys].sample(n = MAX_NUM_ROWS)
else:
data = data[keys]
structures = data[SMILES_COL_NAME].map(lambda x: list(x.ljust(120)))
del data
train_idx, test_idx = map(np.array,
train_test_split(structures.index, test_size = 0.20))
charset = list(reduce(lambda x, y: set(y) | x, structures, set()))
one_hot_encoded_fn = lambda row: map(lambda x: one_hot_array(x, len(charset)),
one_hot_index(row, charset))
h5f = h5py.File('data/processed.h5', 'w')
h5f.create_dataset('charset', data = charset)
def create_chunk_dataset(h5file, dataset_name, dataset, dataset_shape,
chunk_size=1000, apply_fn=None):
new_data = h5file.create_dataset(dataset_name, dataset_shape,
chunks=tuple([chunk_size]+list(dataset_shape[1:])))
for (chunk_ixs, chunk) in chunk_iterator(dataset):
if not apply_fn:
new_data[chunk_ixs, ...] = chunk
else:
new_data[chunk_ixs, ...] = apply_fn(chunk)
create_chunk_dataset(h5f, 'data_train', train_idx,
(len(train_idx), 120, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn,
structures[ch])))
create_chunk_dataset(h5f, 'data_test', test_idx,
(len(test_idx), 120, len(charset)),
apply_fn=lambda ch: np.array(map(one_hot_encoded_fn,
structures[ch])))
h5f.close()
def interpolate(source, dest, steps, charset, model, latent_dim, width):
source_just = source.ljust(width)
dest_just = dest.ljust(width)
one_hot_encoded_fn = lambda row: map(lambda x: one_hot_array(x, len(charset)),
one_hot_index(row, charset))
source_encoded = numpy.array(map(one_hot_encoded_fn, source_just))
source_x_latent = model.encoder.predict(source_encoded.reshape(1, width, len(charset)))
dest_encoded = numpy.array(map(one_hot_encoded_fn, dest_just))
dest_x_latent = model.encoder.predict(dest_encoded.reshape(1, width, len(charset)))
step = (dest_x_latent - source_x_latent)/float(steps)
results = []
for i in range(steps):
item = source_x_latent + (step * i)
sampled = model.decoder.predict(item.reshape(1, latent_dim)).argmax(axis=2)[0]
sampled = decode_smiles_from_indexes(sampled, charset)
results.append( (i, item, sampled) )
return results
def interpolate(source, dest, steps, charset, model, latent_dim, width):
source_just = source.ljust(width)
dest_just = dest.ljust(width)
one_hot_encoded_fn = lambda row: list(map(lambda x: one_hot_array(x, len(charset)),
one_hot_index(row, charset)))
source_encoded = numpy.array(list(map(one_hot_encoded_fn, source_just)))
source_x_latent = model.encoder.predict(source_encoded.reshape(1, width, len(charset)))
dest_encoded = numpy.array(list(map(one_hot_encoded_fn, dest_just)))
dest_x_latent = model.encoder.predict(dest_encoded.reshape(1, width, len(charset)))
step = (dest_x_latent - source_x_latent)/float(steps)
results = []
for i in range(steps):
item = source_x_latent + (step * i)
sampled = model.decoder.predict(item.reshape(1, latent_dim)).argmax(axis=2)[0]
sampled = decode_smiles_from_indexes(sampled, charset)
results.append( (i, item, sampled) )
return results
def one_hot_encoded_fn(row):
return map(lambda x: one_hot_array(x, len(charset)), one_hot_index(row, charset))
