def main():
parser = create_train_parser()
args = parser.parse_args()
args.__dict__['grid_config'] = feature_psp.grid_config
if args.test_only:
with open(os.path.join(args.model_dir, 'config.json')) as f:
model_config = json.load(f)
args.num_conv = model_config['num_conv']
args.use_batch_norm = model_config['use_batch_norm']
if 'grid_config' in model_config:
args.__dict__['grid_config'] = util.dotdict(
model_config['grid_config'])
log_fmt = '%(asctime)s - %(name)s - %(levelname)s - %(message)s'
if args.debug:
logging.basicConfig(level=logging.DEBUG, format=log_fmt)
else:
logging.basicConfig(level=logging.INFO, format=log_fmt)
logging.info("Running 3D CNN PSP training...")
if args.unobserved:
args.output_dir = os.path.join(args.output_dir, 'None')
os.makedirs(args.output_dir, exist_ok=True)
else:
num = 0
while True:
dirpath = os.path.join(args.output_dir, str(num))
if os.path.exists(dirpath):
num += 1
else:
args.output_dir = dirpath
logging.info('Creating output directory {:}'.format(
args.output_dir))
os.mkdir(args.output_dir)
break
logging.info("\n" + str(json.dumps(args.__dict__, indent=4)) + "\n")
# Save config
with open(os.path.join(args.output_dir, 'config.json'), 'w') as f:
json.dump(args.__dict__, f, indent=4)
args.train_sharded = sh.Sharded.load(args.train_sharded)
args.val_sharded = sh.Sharded.load(args.val_sharded)
args.test_sharded = sh.Sharded.load(args.test_sharded)
logging.info("Writing all output to {:}".format(args.output_dir))
with tf.Session() as sess:
np.random.seed(args.random_seed)
tf.set_random_seed(args.random_seed)
train_model(sess, args)
import atom3d.shard.shard as sh
import examples.cnn3d.subgrid_gen as subgrid_gen
import examples.cnn3d.util as util
grid_config = util.dotdict({
# Mapping from elements to position in channel dimension.
'element_mapping': {
'C': 0,
'O': 1,
'N': 2,
'S': 3,
},
# Radius of the grids to generate, in angstroms.
'radius': 50.0,
# Resolution of each voxel, in angstroms.
'resolution': 1.0,
# Number of directions to apply for data augmentation.
'num_directions': 20,
# Number of rolls to apply for data augmentation.
'num_rolls': 20,
})
def df_to_feature(struct_df, grid_config, random_seed=None):
pos = struct_df[['x', 'y', 'z']].astype(np.float32)
center = util.get_center(pos)
rot_mat = subgrid_gen.gen_rot_matrix(grid_config, random_seed=random_seed)
import examples.cnn3d.util as util
grid_config = util.dotdict({
# Mapping from elements to position in channel dimension.
'element_mapping': {
'C': 0,
'O': 1,
'N': 2,
'S': 3
},
# Radius of the grids to generate, in angstroms.
'radius': 17.0,
# Resolution of each voxel, in angstroms.
'resolution': 1.0,
# Number of directions to apply for data augmentation.
'num_directions': 20,
# Number of rolls to apply for data augmentation.
'num_rolls': 20,
### PPI specific
# Number of negatives to sample per positive example. -1 means all.
'neg_to_pos_ratio': 1,
'neg_to_pos_ratio_testing': 1,
# Max number of positive regions to take from a structure. -1 means all.
'max_pos_regions_per_ensemble': 5,
'max_pos_regions_per_ensemble_testing': 5,
# Whether to use all negative at test time.
'full_test': False,
})
def main():
parser = create_train_parser()
args = parser.parse_args()
args.__dict__['grid_config'] = feature_qm9.grid_config
if args.test_only or args.resume_training:
test_only = args.test_only
resume_training = args.resume_training
batch_size = args.batch_size
unobserved = args.unobserved
num_epochs = args.num_epochs
model_dir = args.model_dir
use_ckpt_num = args.use_ckpt_num
with open(os.path.join(args.model_dir, 'config.json')) as f:
model_config = json.load(f)
args.__dict__ = model_config
if 'grid_config' in model_config:
args.__dict__['grid_config'] = util.dotdict(
model_config['grid_config'])
args.resume_training = resume_training
args.test_only = test_only
args.unobserved = unobserved
args.num_epochs = num_epochs
args.model_dir = model_dir
args.use_ckpt_num = use_ckpt_num
if args.test_only:
args.batch_size = batch_size
log_fmt = '%(asctime)s - %(name)s - %(levelname)s - %(message)s'
if args.debug:
logging.basicConfig(level=logging.DEBUG, format=log_fmt)
else:
logging.basicConfig(level=logging.INFO, format=log_fmt)
logging.info("Running 3D CNN QM9 training...")
if args.unobserved:
args.output_dir = os.path.join(args.output_dir, 'None')
os.makedirs(args.output_dir, exist_ok=True)
else:
num = 0
while True:
dirpath = os.path.join(args.output_dir, str(num))
if os.path.exists(dirpath):
num += 1
else:
try:
args.output_dir = dirpath
logging.info('Creating output directory {:}'.format(args.output_dir))
os.mkdir(args.output_dir)
break
except:
pass
logging.info("\n" + str(json.dumps(args.__dict__, indent=4)) + "\n")
# Save config
with open(os.path.join(args.output_dir, 'config.json'), 'w') as f:
json.dump(args.__dict__, f, indent=4)
logging.info("Writing all output to {:}".format(args.output_dir))
with tf.Session() as sess:
tf.set_random_seed(args.random_seed)
train_model(sess, args)
import examples.cnn3d.subgrid_gen as subgrid_gen
import examples.cnn3d.util as util
grid_config = util.dotdict({
# Mapping from elements to position in channel dimension.
'element_mapping': {
'H': 0,
'C': 1,
'O': 2,
'N': 3,
'S': 4,
'CL': 5,
'F': 6,
},
# Radius of the grids to generate, in angstroms.
'radius': 25.0,
# Resolution of each voxel, in angstroms.
'resolution': 1.0,
# Number of directions to apply for data augmentation.
'num_directions': 20,
# Number of rolls to apply for data augmentation.
'num_rolls': 20,
# Number of negatives to sample per positive example. -1 means all.
# positive = A (active), negative = I (inactive)
'neg_to_pos_ratio': 1,
})
def __get_subunit_name(subunits, mode='inactive'):
assert len(subunits) == 2
import examples.cnn3d.subgrid_gen as subgrid_gen
import examples.cnn3d.util as util
grid_config = util.dotdict({
# Mapping from elements to position in channel dimension.
'element_mapping': {
'H': 0,
'C': 1,
'O': 2,
'N': 3,
'S': 4,
},
# Radius of the grids to generate, in angstroms.
'radius': 25.0,
# Resolution of each voxel, in angstroms.
'resolution': 1.0,
# Number of directions to apply for data augmentation.
'num_directions': 20,
# Number of rolls to apply for data augmentation.
'num_rolls': 20,
# Number of negatives to sample per positive example. -1 means all.
'neg_to_pos_ratio': 1.0,
# Max number of positive regions to take from a structure. -1 means all.
'max_pos_per_shard': 200,
})
def __get_mutation_center(struct_df, label_info, center_at_mut=True):
if center_at_mut:
# Use CA position of the mutated residue as center for subgrid center