model_name, model_name, iter_)
net_param = caffe_util.NetParameter.from_prototxt(model_file)
net_param.set_molgrid_data_source(data_file, '')
data_param = net_param.get_molgrid_data_param(caffe.TEST)
data_param.random_rotation = True
net = caffe_util.Net.from_param(net_param, weights_file, phase=caffe.TEST)
latent_name = '{}_latent_fc'.format(dict(r='rec', l='lig')[encode])
after_latent_name = 'lig_latent_defc'
net.forward(end=latent_name)
n_samples = n = net.blobs[latent_name].shape[0]
v0 = np.array(net.blobs[latent_name].data[0])
v1 = np.array(net.blobs[latent_name].data[1])
net.blobs[latent_name].data[0:n // 2] = lerp(v0, v1, np.linspace(0, 1, n // 2))
net.blobs[latent_name].data[n // 2:n] = lerp(v1, v0, np.linspace(0, 1, n // 2))
net.forward(start=after_latent_name)
dx_groups = {}
blob_names = ['lig_gen']
for j in range(n_samples):
for blob_name in blob_names:
grid_name = '{}_{}k_{}_{}'.format(name, iter_, blob_name, j)
grid = np.array(net.blobs[blob_name].data[j])
dx_groups[grid_name] = g.write_grids_to_dx_files(
grid_name, grid, channels, center, 0.5)
pymol_file = '{}.pymol'.format(name)
other_files = [rec_file, lig_file]
g.write_pymol_script(pymol_file, dx_groups, other_files, [center, center])
for name, blob in latent_blobs.items():
latent_vecs['zero'][name] = np.zeros(blob.shape[1:])
blob.data[...] = 0.0
for target in latent_vecs:
for name in latent_vecs[target]:
latent_vec = latent_vecs[target][name]
print name, target, latent_vec.shape, np.linalg.norm(latent_vec)
grid_names = []
for rec_target in latent_vecs:
for lig_target in latent_vecs:
idx = len(grid_names)
latent_blobs['rec'].data[idx, ...] = latent_vecs[rec_target]['rec']
latent_blobs['lig'].data[idx, ...] = latent_vecs[lig_target]['lig']
grid_names.append('REC_{}_LIG_{}'.format(rec_target, lig_target))
print grid_names[-1]
net.forward(start='latent_concat')
dx_groups = {}
for i, grid_name in enumerate(grid_names):
grid = net.blobs['lig_gen'].data[i]
dx_groups[grid_name] = generate.write_grids_to_dx_files(
grid_name, grid, channels, np.zeros(3), 0.5)
generate.write_pymol_script('REC_COND.pymol', dx_groups, rec_files + lig_files,
2 * centers)
print 'done'
net.forward(end='latent_concat')
net.blobs['lig_latent_mean'] = 0.0
net.blobs['lig_latent_std'] = 1.0
net.forward(start='lig_latent_noise')
for other_file in rec_files + lig_files:
shutil.copyfile(other_file, os.path.basename(other_file))
rec_files = [os.path.basename(f) for f in rec_files]
lig_files = [os.path.basename(f) for f in lig_files]
dx_groups = []
target_grids = dict()
for i, (rec_file, lig_file) in enumerate(zip(rec_files, lig_files)):
target = os.path.splitext(os.path.basename(lig_file))[0]
grid = 8.0 * np.array(net.blobs['lig_gen'].data[i])
generate.write_grids_to_dx_files(target, grid, channels, centers[i], 0.5)
dx_groups.append(target)
target_grids[target] = grid
fit_algo = dict(greedy=True,
bonded=True,
max_init_bond_E=0.2,
lambda_E=0.2,
radius_factor=1.33)
def do_fit(args):
lig_file, center = args
target = os.path.splitext(os.path.basename(lig_file))[0]
grid = target_grids[target]
xyz, c, bonds, loss = generate.fit_atoms_to_grids(grid,