def main(argv):
if len(argv) > 1:
raise app.UsageError('Too many command-line arguments.')
# Set random seed
np.random.seed(FLAGS.seed)
torch.manual_seed(FLAGS.seed)
if torch.cuda.is_available(): # cuda device
device = 'cuda'
torch.cuda.set_device(FLAGS.gpu)
else:
device = 'cpu'
adata, trainloader, testloader = load_dataset(
FLAGS.input_path,
batch_name='batch',
min_genes=FLAGS.min_peaks,
min_cells=FLAGS.min_cells,
batch_size=FLAGS.batch_size,
n_top_genes=FLAGS.n_feature,
log=None,
)
input_dim = adata.shape[1]
tf.io.gfile.makedirs(FLAGS.output_path)
dims = [input_dim, FLAGS.latent, FLAGS.encode_dim, FLAGS.decode_dim]
model = SCALE(dims, n_centroids=FLAGS.n_centroids)
print(model)
print('\n## Training Model ##')
model.init_gmm_params(testloader)
model.fit(trainloader,
lr=FLAGS.lr,
weight_decay=FLAGS.weight_decay,
verbose=FLAGS.verbose,
device=device,
max_iter=FLAGS.max_iter,
outdir=FLAGS.output_path)
adata.obsm['latent'] = model.encodeBatch(testloader,
device=device,
out='z')
dr = pd.DataFrame(adata.obsm['latent'], index=adata.obs_names)
with tf.io.gfile.GFile(os.path.join(FLAGS.output_path, 'SCALE.csv'),
'w') as f:
dr.to_csv(f)
device=device,
max_iter=args.max_iter,
# name=name,
outdir=outdir)
torch.save(model.state_dict(), os.path.join(outdir,
'model.pt')) # save model
else:
print('\n## Loading Model: {}\n'.format(args.pretrain))
model.load_model(args.pretrain)
model.to(device)
### output ###
print('outdir: {}'.format(outdir))
# 1. latent feature
adata.obsm['latent'] = model.encodeBatch(testloader,
device=device,
out='z')
# 2. cluster
sc.pp.neighbors(adata, n_neighbors=30, use_rep='latent')
if args.cluster_method == 'leiden':
sc.tl.leiden(adata)
elif args.cluster_method == 'kmeans':
kmeans = KMeans(n_clusters=k, n_init=20, random_state=0)
adata.obs['kmeans'] = kmeans.fit_predict(
adata.obsm['latent']).astype(str)
# if args.reference in adata.obs:
# cluster_report(adata.obs[args.reference].cat.codes, adata.obs[args.cluster_method].astype(int))
sc.settings.figdir = outdir
weight_decay=args.weight_decay,
verbose=args.verbose,
device=device,
max_iter=args.max_iter,
name=name,
outdir=outdir)
# torch.save(model.to('cpu').state_dict(), os.path.join(outdir, 'model.pt')) # save model
else:
print('\n## Loading Model: {}\n'.format(args.pretrain))
model.load_model(args.pretrain)
model.to(device)
### output ###
print('outdir: {}'.format(outdir))
# 1. latent feature
feature = model.encodeBatch(testloader, device=device, out='z')
pd.DataFrame(feature).to_csv(os.path.join(outdir, 'feature.txt'),
sep='\t',
header=False)
# 2. cluster assignments
pred = model.predict(testloader, device)
pd.Series(pred, index=dataset.barcode).to_csv(os.path.join(
outdir, 'cluster_assignments.txt'),
sep='\t',
header=False)
# 3. imputed data
if args.impute or args.binary:
recon_x = model.encodeBatch(testloader,
device,