exp_counter = 1
num_exps = len(n_src) * reps * len(percs) * len(methods)
# Run jobs
for s in range(len(n_src)):
accs = np.zeros((len(acc_funcs), reps, len(percs), len(methods)))
accs_desc = list()
opt_mix_ind = np.zeros((reps, len(percs)))
opt_mix_aris = np.zeros((reps, len(percs)))
num_strat = np.zeros((reps, len(percs), len(methods)))
res_desc = []
r = 0
while r < reps:
# Split data in source and target randomly (mode =1) or randomly stratified (mode = 2)
src, trg, src_labels, trg_labels = split_source_target(
data, labels, mode=1, target_ncells=n_trg, source_ncells=n_src[s])
trg_labels = np.array(trg_labels, dtype=np.int)
src_labels = np.array(src_labels, dtype=np.int)
# 3.a. Subsampling order for target
inds = np.random.permutation(trg_labels.size)
# 3.b. Use perfect number of latent states for nmf and sc3
src_lbl_set = np.unique(src_labels)
n_trg_cluster = np.unique(trg_labels).size
n_src_cluster = src_lbl_set.size
# 3.c. train source once per repetition
print "Train source data of rep {0}".format(r + 1)
source_nmf = None
if len(target_ncells_range) == 0:
target_ncells_range = [args.target_ncells]
for sidx, source_ncells in enumerate(source_ncells_range):
for tidx, target_ncells in enumerate(target_ncells_range):
print(
'\nSplit artificial single-cell RNA-seq data in target and source data.'
)
data_source, data_target, true_labels_source, true_labels_target = \
split_source_target(
data,
labels,
target_ncells = target_ncells,
source_ncells = source_ncells,
source_clusters = source_clusters,
noise_target = args.noise_target,
noise_sd = args.noise_sd,
mode = args.splitting_mode,
common = args.common
)
print 'Target data dimension: ', data_target.shape
print 'Source data dimension: ', data_source.shape
# 3. GENERATE GENE AND CELL NAMES
gene_ids = np.arange(args.num_genes)
# 4. SAVE RESULTS
print('Saving target data to \'{0}\'.'.format(args.fout_target_data))
np.savetxt(os.path.splitext(args.fout_target_data)[0] + "_T" +
str(tidx + 1) + "_" + str(target_ncells) + "_S" +
if len(source_ncells_range) == 0:
source_ncells_range = [args.source_ncells]
if len(target_ncells_range) == 0:
target_ncells_range = [args.target_ncells]
for sidx, source_ncells in enumerate(source_ncells_range):
for tidx, target_ncells in enumerate(target_ncells_range):
print('\nSplit artificial single-cell RNA-seq data in target and source data.')
data_source, data_target, true_labels_source, true_labels_target = \
split_source_target(
data,
labels,
target_ncells = target_ncells,
source_ncells = source_ncells,
source_clusters = source_clusters,
noise_target = args.noise_target,
noise_sd = args.noise_sd,
mode = args.splitting_mode
)
print 'Target data dimension: ', data_target.shape
print 'Source data dimension: ', data_source.shape
# 3. GENERATE GENE AND CELL NAMES
gene_ids = np.arange(args.num_genes)
# 4. SAVE RESULTS
print('Saving target data to \'{0}\'.'.format(args.fout_target_data))
np.savetxt(
os.path.splitext(args.fout_target_data)[0] +
"_T" + str(tidx+1) + "_" + str(target_ncells) +
def experiment_loop(fname, methods, acc_funcs, n_src=800, n_trg=800, n_genes=1000, mode=2, reps=10,
cluster_mode=False, cluster_spec=[1, 2, 3, [4, 5], [6, [7, 8]]], percs=[0.1, 0.4, 0.8]):
flatten = lambda l: flatten(l[0]) + (flatten(l[1:]) if len(l) > 1 else []) if type(l) is list else [l]
n_cluster = len(flatten(cluster_spec))
print 'Number of cluster is ', n_cluster
accs = np.zeros((len(acc_funcs), reps, len(percs), len(methods)))
accs_desc = [''] * len(acc_funcs)
num_strat = np.zeros((reps, len(percs), len(methods)))
res_desc = []
for r in range(reps):
# 1. Generate scRNA data
data, labels = generate_toy_data(num_genes=n_genes,
num_cells=2.*(n_trg + n_src), # generate more data
cluster_spec=cluster_spec)
# 2. Split source and target according to specified mode/setting
src, trg, src_labels, trg_labels = split_source_target(data, labels,
target_ncells=n_trg, source_ncells=n_src,
mode=mode, source_clusters=None,
noise_target=False, noise_sd=0.1)
trg_labels = np.array(trg_labels, dtype=np.int)
src_labels = np.array(src_labels, dtype=np.int)
# 3.a. Subsampling order for target
inds = np.random.permutation(trg_labels.size)
# 3.b. Use perfect number of latent states for nmf and sc3
src_lbl_set = np.unique(src_labels)
n_trg_cluster = np.unique(trg_labels).size
n_src_cluster = src_lbl_set.size
# 3.c. Target data subsampling loop
for i in range(len(percs)):
if cluster_mode:
n_trg_cluster = percs[i]
p_trg = trg[:, inds[:n_trg]].copy()
p_trg_labels = trg_labels[inds[:n_trg]].copy()
else:
n_trg_perc = np.int(n_trg * percs[i])
p_trg = trg[:, inds[:n_trg_perc]].copy()
p_trg_labels = trg_labels[inds[:n_trg_perc]].copy()
# 4. MTL/DA mixing parameter loop
res_desc = list()
for m in range(len(methods)):
desc, lbls, reject = methods[m](src.copy(), src_labels.copy(),
p_trg.copy(), p_trg_labels.copy(),
n_src_cluster=n_src_cluster, n_trg_cluster=n_trg_cluster)
res_desc.append(desc)
# evaluation
strat_lbl_inds = []
for n in range(p_trg_labels.size):
if p_trg_labels[n] in src_lbl_set:
strat_lbl_inds.append(n)
for f in range(len(acc_funcs)):
accs[f, r, i, m], desc = acc_funcs[f](p_trg.copy(), p_trg_labels.copy(),
lbls.copy(), reject, strat_lbl_inds)
accs_desc[f] = desc
# save the result and then plot
np.savez(fname, methods=methods, acc_funcs=acc_funcs, accs=accs, accs_desc=accs_desc,
percs=percs, reps=reps, n_genes=n_genes, n_src=n_src, n_trg=n_trg,
desc=res_desc, mode=mode, num_strat=num_strat, cluster_mode=cluster_mode)
print('Done.')
num_strat = np.zeros((reps, len(percs), len(methods)))
res_desc = []
r = 0
while r < reps:
# 1. Generate scRNA data
data, labels = generate_toy_data(num_genes=genes[g],
num_cells=10. *
(n_trg + n_src[s]),
cluster_spec=cluster_spec)
# 2. Split source and target according to specified mode/setting
src, trg, src_labels, trg_labels = split_source_target(
data,
labels,
target_ncells=n_trg,
source_ncells=n_src[s],
mode=7,
source_clusters=None,
noise_target=False,
noise_sd=0.1,
common=common[c],
cluster_spec=cluster_spec)
trg_labels = np.array(trg_labels, dtype=np.int)
src_labels = np.array(src_labels, dtype=np.int)
# 3.a. Subsampling order for target
inds = np.random.permutation(trg_labels.size)
# 3.b. Use perfect number of latent states for nmf and sc3
src_lbl_set = np.unique(src_labels)
n_trg_cluster = np.unique(trg_labels).size
n_src_cluster = src_lbl_set.size
# 3.c. train source once per repetition
source_nmf = NmfClustering(src,
