raise Exception('cmd err in the argv[2]')
# summary
nz_rate_df = scimpute.nnzero_rate_df(df)
print('df.shape, [gene, cell]:', df.shape)
print('nz_rate: {}'.format(round(nz_rate_df, 3)))
print(df.ix[0:3, 0:3])
# exp
df = np.power(10, df) -1
print('after exp(value) - 1')
print(df.iloc[:3, :3])
# lib-size per million normalization
df = scimpute.df_normalization(df)
print('after normalization')
print(df.ix[0:3, 0:3])
read_per_gene = df.sum(axis=1)
read_per_cell = df.sum(axis=0)
print('sum_reads_per_gene:', read_per_gene[0:3])
print('sum_reads_per_cell:', read_per_cell[0:3])
# log(tpm+1) transformation
df = scimpute.df_log_transformation(df)
print('after log transformation')
print(df.ix[0:3, 0:3])
# save
print('saving output to:', outname)
scimpute.save_hd5(df, outname)
print('finished')
# summary
nz_rate_df = scimpute.nnzero_rate_df(df)
print('input matrix.shape:', df.shape)
print('nz_rate: {}'.format(round(nz_rate_df, 3)))
print(df.ix[0:3, 0:3])
# filter #
read_per_gene = df.sum(axis=1)
read_per_cell = df.sum(axis=0)
df_filtered = df.loc[(read_per_gene >= gene_min), (read_per_cell >= cell_min)]
nz_rate_filtered = scimpute.nnzero_rate_df(df_filtered)
print('filtered matrix : ', df_filtered.shape)
print('nz_rate:', nz_rate_filtered)
print(df_filtered.ix[0:3, 0:3])
scimpute.save_hd5(df_filtered, tag0 + '.hd5')
# histogram of filtered data
read_per_gene_filtered = df_filtered.sum(axis=1)
read_per_cell_filtered = df_filtered.sum(axis=0)
scimpute.hist_list(read_per_cell_filtered.values,
xlab='counts/cell',
title='Histogram of counts per cell' + tag)
scimpute.hist_list(read_per_gene_filtered.values,
xlab='counts/gene',
title='Histogram of counts per gene' + tag)
scimpute.hist_df(df_filtered,
xlab='counts',
title='Histogram of counts in expression matrix' + tag)
# histogram of log transformed filtered data
# Remove .x from ID
# df_big.index = df_big.index.to_series().astype(str).str.replace(r'\.[0-9]*','').astype(str)
# print('because the index is different, remove the appendix')
# print('big df after changing index', df_big.ix[0:5, 0:5])
print('df_big index is unique? {}'.format(df_big.index.is_unique))
print('df_small index is unique? {}'.format(df_small.index.is_unique))
# SELECT
print('selecting..')
df_selected = df_big.ix[df_small.index]
# Check null, fill zeros
null_gene_num = df_selected.ix[:, 1].isnull().sum()
print('there are {} genes from small-df not found in big-df'.format(
null_gene_num))
df_selected = df_selected.fillna(value=0)
print('those N.A. in selected-df has been filled with zeros')
null_gene_num2 = df_selected.ix[:, :].isnull().sum().sum()
print('Now, there are {} null values in selected-df'.format(null_gene_num2))
nz_selected = scimpute.nnzero_rate_df(df_selected)
print('nz_rate output: ', nz_selected)
# Finish
print('selected big df from small df, after fillna:\n', df_selected.ix[0:5,
0:5])
print('shape: ', df_selected.shape)
scimpute.save_hd5(df_selected, out_name)
print('Finished')
def late_main(p, log_dir, rand_state=3):
##0. read data and extract gene IDs and cell IDs
input_matrix, gene_ids, cell_ids = read_data(p)
##1. split data and save indexes
#input p, input_matrix, cell_ids
#return cell_ids_train, cell_ids_valid, cell_ids_test
m, n = input_matrix.shape
input_train, input_valid, input_test, train_idx, valid_idx, test_idx = \
scimpute.split__csr_matrix(input_matrix, a=p.a, b=p.b, c=p.c)
cell_ids_train = cell_ids[train_idx]
cell_ids_valid = cell_ids[valid_idx]
cell_ids_test = cell_ids[test_idx]
np.savetxt('{}/train.{}_index.txt'.format(p.stage, p.stage),
cell_ids_train,
fmt='%s')
np.savetxt('{}/valid.{}_index.txt'.format(p.stage, p.stage),
cell_ids_valid,
fmt='%s')
np.savetxt('{}/test.{}_index.txt'.format(p.stage, p.stage),
cell_ids_test,
fmt='%s')
print('RAM usage after splitting input data is: {} M'.format(usage()))
# todo: for backward support for older parameter files only
# sample_size is 1000 in default; if sample_size is less than the number of cells (m),
# we reconstruct the training and validation sets by randomly sampling.
try:
p.sample_size
sample_size = p.sample_size
except:
sample_size = int(9e4)
if sample_size < m:
np.random.seed(1)
rand_idx = np.random.choice(range(len(cell_ids_train)),
min(sample_size, len(cell_ids_train)))
sample_train = input_train[rand_idx, :].todense()
sample_train_cell_ids = cell_ids_train[rand_idx]
rand_idx = np.random.choice(range(len(cell_ids_valid)),
min(sample_size, len(cell_ids_valid)))
sample_valid = input_valid[rand_idx, :].todense()
sample_valid_cell_ids = cell_ids_valid[rand_idx]
#?? the following sample_input is a matrix sampled randomly, and should it be a matrix containing
# sample_training and sample_valid
rand_idx = np.random.choice(range(m), min(sample_size, m))
sample_input = input_matrix[rand_idx, :].todense()
sample_input_cell_ids = cell_ids[rand_idx]
del rand_idx
gc.collect()
np.random.seed()
else:
sample_input = input_matrix.todense()
sample_train = input_train.todense()
sample_valid = input_valid.todense()
sample_input_cell_ids = cell_ids
sample_train_cell_ids = cell_ids_train
sample_valid_cell_ids = cell_ids_valid
print('len of sample_train: {}, sample_valid: {}, sample_input {}'.format(
len(sample_train_cell_ids), len(sample_valid_cell_ids),
len(sample_input_cell_ids)))
##2. model training and validation
#2.1 init --> keep this in the main
tf.reset_default_graph()
# define placeholders and variables
X = tf.placeholder(tf.float32, [None, n], name='X_input') # input
pIn_holder = tf.placeholder(tf.float32,
name='p.pIn') #keep_prob for dropout
pHidden_holder = tf.placeholder(tf.float32,
name='p.pHidden') #keep_prob for dropout
#2.2 define layers and variables
# input p, X, pIn_holder, pHidden_holder, n
# return a_bottleneck, h(d_a1)
a_bottleneck, h = build_late(X,
pHidden_holder,
pIn_holder,
p,
n,
rand_state=3)
#2.3 define loss
# input X, h, p
# return mse_nz, mse, reg_term
mse_nz, mse, reg_term = build_metrics(X, h, p.reg_coef)
#2.4 costruct the trainer --> keep this section in the main
optimizer = tf.train.AdamOptimizer(p.learning_rate)
if p.mse_mode in ('mse_omega', 'mse_nz'):
print('training on mse_nz')
trainer = optimizer.minimize(mse_nz + reg_term)
elif p.mse_mode == 'mse':
print('training on mse')
trainer = optimizer.minimize(mse + reg_term)
else:
raise Exception('mse_mode spelled wrong')
#2.5 Init a session accoding to the run_flag
sess = tf.Session()
# restore variables
saver = tf.train.Saver()
if p.run_flag == 'load_saved':
print('*** In TL Mode')
saver.restore(sess, "./step1/step1.ckpt")
elif p.run_flag == 'rand_init':
print('*** In Rand Init Mode')
init = tf.global_variables_initializer()
sess.run(init)
elif p.run_flag == 'impute':
print('*** In impute mode loading "step2.ckpt"..')
saver.restore(sess, './step2/step2.ckpt')
p.max_training_epochs = 0
p.learning_rate = 0.0
## save_whole_imputation
save_whole_imputation(sess, X, h, a_bottleneck, pIn_holder,
pHidden_holder, input_matrix, gene_ids, cell_ids,
p, m)
print('imputation finished')
#toc_stop = time.time()
#print("reading took {:.1f} seconds".format(toc_stop - tic_start))
exit()
else:
raise Exception('run_flag err')
# define tensor_board writer
batch_writer = tf.summary.FileWriter(log_dir + '/batch', sess.graph)
valid_writer = tf.summary.FileWriter(log_dir + '/valid', sess.graph)
# prep mini-batch, and reporter vectors
num_batch = int(math.floor(len(train_idx) // p.batch_size)) # floor
epoch_log = []
mse_nz_batch_vec, mse_nz_valid_vec = [], [
] #, mse_nz_train_vec = [], [], []
mse_batch_vec, mse_valid_vec = [], [] # mse = MSE(X, h)
#msej_batch_vec, msej_valid_vec = [], [] # msej = MSE(X, h), for genej, nz_cells
print('RAM usage after building the model is: {} M'.format(usage()))
epoch = 0
#2.6. pre-training epoch (0)
#save imputation results before training steps
print("Evaluation: epoch{}".format(epoch))
epoch_log.append(epoch)
mse_train, mse_nz_train = sess.run([mse, mse_nz],
feed_dict={
X: sample_train,
pHidden_holder: 1.0,
pIn_holder: 1.0
})
mse_valid, mse_nz_valid = sess.run([mse, mse_nz],
feed_dict={
X: sample_valid,
pHidden_holder: 1.0,
pIn_holder: 1.0
})
print("mse_nz_train=", round(mse_nz_train, 3), "mse_nz_valid=",
round(mse_nz_valid, 3))
print("mse_train=", round(mse_train, 3), "mse_valid=", round(mse_valid, 3))
mse_batch_vec.append(mse_train)
mse_valid_vec.append(mse_valid)
mse_nz_batch_vec.append(mse_nz_train)
mse_nz_valid_vec.append(mse_nz_valid)
#2.7. training epochs (1-)
for epoch in range(1, p.max_training_epochs + 1):
tic_cpu, tic_wall = time.clock(), time.time()
ridx_full = np.random.choice(len(train_idx),
len(train_idx),
replace=False)
#2.7.1 training model on mini-batches
for i in range(num_batch):
# x_batch
indices = np.arange(p.batch_size * i, p.batch_size * (i + 1))
ridx_batch = ridx_full[indices]
# x_batch = df1_train.ix[ridx_batch, :]
x_batch = input_train[ridx_batch, :].todense()
sess.run(trainer,
feed_dict={
X: x_batch,
pIn_holder: p.pIn,
pHidden_holder: p.pHidden
})
toc_cpu, toc_wall = time.clock(), time.time()
#2.7.2 save the results of epoch 1 and all display steps (epochs)
if (epoch == 1) or (epoch % p.display_step == 0):
tic_log = time.time()
print(
'#Epoch {} took: {} CPU seconds; {} Wall seconds'.format(
epoch, round(toc_cpu - tic_cpu, 2),
round(toc_wall - tic_wall, 2)))
print('num-mini-batch per epoch: {}, till now: {}'.format(
i + 1, epoch * (i + 1)))
print('RAM usage: {:0.1f} M'.format(usage()))
# debug
# print('d_w1', sess.run(d_w1[1, 0:4])) # verified when GradDescent used
# training mse and mse_nz of the last batch
mse_batch, mse_nz_batch, h_batch = sess.run([mse, mse_nz, h],
feed_dict={
X: x_batch,
pHidden_holder:
1.0,
pIn_holder: 1.0
})
# validation mse and mse_nz of the sample validation set (1000)
mse_valid, mse_nz_valid, Y_valid = sess.run([mse, mse_nz, h],
feed_dict={
X: sample_valid,
pHidden_holder:
1.0,
pIn_holder: 1.0
})
toc_log = time.time()
print('mse_nz_batch:{}; mse_omage_valid: {}'.format(
mse_nz_batch, mse_nz_valid))
print('mse_batch:', mse_batch, '; mse_valid:', mse_valid)
print('log time for each epoch: {}\n'.format(
round(toc_log - tic_log, 1)))
mse_batch_vec.append(mse_batch)
mse_valid_vec.append(mse_valid)
mse_nz_batch_vec.append(mse_nz_batch)
mse_nz_valid_vec.append(mse_nz_valid)
epoch_log.append(epoch)
#2.7.3 save snapshot step
if (epoch % p.snapshot_step == 0) or (epoch == p.max_training_epochs):
tic_log2 = time.time()
#1.save imputation results
#if the input matrix is large (m > p.large_size), only save the
#imputation results of a small sample set (sample_input)
print("> Impute and save.. ")
if m > p.large_size:
Y_input_df = fast_imputation(sess, h, X, pIn_holder,
pHidden_holder, sample_input,
gene_ids, sample_input_cell_ids)
scimpute.save_hd5(
Y_input_df,
"{}/sample_imputation.{}.hd5".format(p.stage, p.stage))
else:
Y_input_df = fast_imputation(sess, h, X, pIn_holder,
pHidden_holder,
input_matrix.todense(), gene_ids,
cell_ids)
scimpute.save_hd5(
Y_input_df,
"{}/imputation.{}.hd5".format(p.stage, p.stage))
#2.save model
print('> Saving model..')
save_path = saver.save(sess, log_dir + "/{}.ckpt".format(p.stage))
print("Model saved in: %s" % save_path)
#3.save the training and test curve
if p.mse_mode in ('mse_nz', 'mse_omega'):
#learning_curve_mse_nz(skip=math.floor(epoch / 5 / p.display_step))
learning_curve_mse_nz(epoch_log,
mse_nz_batch_vec,
mse_nz_valid_vec,
p.stage,
skip=math.floor(epoch / 5 /
p.display_step))
elif p.mse_mode == 'mse':
#learning_curve_mse(skip=math.floor(epoch / 5 / p.display_step))
learning_curve_mse(epoch_log,
mse_batch_vec,
mse_valid_vec,
p.stage,
skip=math.floor(epoch / 5 / p.display_step))
#4.save save_bottleneck_representation
print("> save bottleneck_representation")
code_bottleneck_input = sess.run(a_bottleneck,
feed_dict={
X: sample_input,
pIn_holder: 1,
pHidden_holder: 1
})
np.save('{}/code_neck_valid.{}.npy'.format(p.stage, p.stage),
code_bottleneck_input)
#save_weights()
save_weights(sess, p.stage, en_de_layers=p.l)
#visualize_weights()
visualize_weights(sess, p.stage, en_de_layers=p.l)
toc_log2 = time.time()
log2_time = round(toc_log2 - tic_log2, 1)
min_mse_valid = min(mse_nz_valid_vec)
# os.system(
# '''for file in {0}/*npy
# do python -u weight_clustmap.py $file {0}
# done'''.format(p.stage)
# )
print('min_mse_nz_valid till now: {}'.format(min_mse_valid))
print('snapshot_step: {}s'.format(log2_time))
batch_writer.close()
valid_writer.close()
sess.close()
def save_whole_imputation(sess, X, h, a_bottleneck, pIn_holder, pHidden_holder,
input_matrix, gene_ids, cell_ids, p, m):
''' calculate and save imputation results for an input matrix at the 'impute' mode. If the number
of cells is larger than a threshold (large_size: 1e5), save results of m//p.sample_size 'folds'.
Parameters
----------
'''
if m > p.large_size:
#impute on small data blocks to avoid high memory cost
n_out_batches = m // p.sample_size
print('num_out_batches:', n_out_batches)
handle2 = open('./{}/latent_code.{}.csv'.format(p.stage, p.stage), 'w')
with open('./{}/imputation.{}.csv'.format(p.stage, p.stage),
'w') as handle:
for i_ in range(n_out_batches + 1):
start_idx = i_ * p.sample_size
end_idx = min((i_ + 1) * p.sample_size, m)
print('saving:', start_idx, end_idx)
x_out_batch = input_matrix[start_idx:end_idx, :].todense()
y_out_batch = sess.run(h,
feed_dict={
X: x_out_batch,
pIn_holder: 1,
pHidden_holder: 1
})
df_out_batch = pd.DataFrame(data=y_out_batch,
columns=gene_ids,
index=cell_ids[range(
start_idx, end_idx)])
latent_code = sess.run(a_bottleneck,
feed_dict={
X: x_out_batch,
pIn_holder: 1,
pHidden_holder: 1
})
latent_code_df = pd.DataFrame(data=latent_code,
index=cell_ids[range(
start_idx, end_idx)])
if i_ == 0:
df_out_batch.to_csv(handle, float_format='%.6f')
latent_code_df.to_csv(handle2, float_format='%.6f')
print(
'RAM usage during mini-batch imputation and saving output: ',
'{} M'.format(usage()))
else:
df_out_batch.to_csv(handle, header=None)
latent_code_df.to_csv(handle2, header=None)
handle2.close()
else: # if m the # of cells is less than large_size (1e5))
Y_input_arr = sess.run(h,
feed_dict={
X: input_matrix.todense(),
pIn_holder: 1,
pHidden_holder: 1
})
# save sample imputation
Y_input_df = pd.DataFrame(data=Y_input_arr,
columns=gene_ids,
index=cell_ids)
latent_code = sess.run(a_bottleneck,
feed_dict={
X: input_matrix.todense(),
pIn_holder: 1,
pHidden_holder: 1
})
latent_code_df = pd.DataFrame(data=latent_code, index=cell_ids)
print('RAM usage during whole data imputation and saving output: ',
'{} M'.format(usage()))
scimpute.save_hd5(Y_input_df,
"{}/imputation.{}.hd5".format(p.stage, p.stage))
scimpute.save_hd5(latent_code_df,
"{}/latent_code.{}.hd5".format(p.stage, p.stage))
# read data (so that output matrix is [sample, gene])
if matrix_type == 'cell_row':
df = pd.read_hdf(file_name).transpose()
elif matrix_type == 'gene_row':
df = pd.read_hdf(file_name)
# summary
print('input shape [genes, samples]:', df.shape, df.ix[0:3, 0:2])
nz_rate_in = scimpute.nnzero_rate_df(df)
print('nz_rate_in: {}'.format(nz_rate_in))
# read list
if list_name.endswith('.csv'):
list_df = pd.read_csv(list_name, index_col=0, sep='\t', header=None)
elif list_name.endswith('.hd5'):
list_df = scimpute.read_hd5(list_name)
print('list:', list_df.shape, list_df.index)
# filter
df_yes = df.ix[list_df.index]
overlap = df.index.isin(list_df.index)
df_no = df.ix[~overlap]
print('matrix yes: ', df_yes.shape, df_yes.ix[0:3, 0:2])
print('matrix no: ', df_no.shape, df_no.ix[0:3, 0:2])
# output result dataframe
scimpute.save_hd5(df_yes, out_prefix + '_yes.hd5')
scimpute.save_hd5(df_no, out_prefix + '_no.hd5')
