/
run.py
192 lines (146 loc) · 5.87 KB
/
run.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
import matplotlib
matplotlib.use('Agg')
import numpy as np
import os
import yaml
import pandas as pd
from scvis.model import SCVIS
from scvis import plot
from scvis import data
import matplotlib.pyplot as plt
try:
import cPickle as pickle
except ImportError:
import pickle
CURR_PATH = os.path.dirname(os.path.abspath(__file__))
def train(args):
x, y, architecture, hyperparameter, train_data, model, normalizer, out_dir, name = \
_init_model(args, 'train')
iter_per_epoch = round(x.shape[0] / hyperparameter['batch_size'])
max_iter = int(iter_per_epoch * hyperparameter['max_epoch'])
if max_iter < 3000:
max_iter = 3000
elif max_iter > 30000:
max_iter = np.max([30000, iter_per_epoch * 2])
name += '_iter_' + str(max_iter)
res = model.train(data=train_data,
batch_size=hyperparameter['batch_size'],
verbose=args.verbose,
verbose_interval=args.verbose_interval,
show_plot=args.show_plot,
plot_dir=os.path.join(out_dir, (name+"_intermediate_result")),
max_iter=max_iter,
pretrained_model=args.pretrained_model_file)
model.set_normalizer(normalizer)
# Save the trained model
out_dir = args.out_dir
if not os.path.isdir(out_dir):
os.mkdir(out_dir)
model_dir = os.path.join(out_dir, "model")
if not os.path.isdir(model_dir):
os.mkdir(model_dir)
model_name = name + ".ckpt"
model_name = os.path.join(model_dir, model_name)
model.save_sess(model_name)
# The objective function trace plot
elbo = res['elbo']
tsne_cost = res['tsne_cost']
iteration = len(elbo)
avg_elbo = elbo - tsne_cost
for i in range(iteration)[1:]:
avg_elbo[i] = (elbo[i] - tsne_cost[i]) / i + \
avg_elbo[i-1] * (i-1) / i
plot.plot_trace([range(iteration)] * 3,
[avg_elbo, elbo, tsne_cost],
['avg_cost', 'elbo', 'tsne_cost'])
fig_file = name + '_obj.png'
fig_file = os.path.join(out_dir, fig_file)
plt.savefig(fig_file)
obj_file = name + '_obj.tsv'
obj_file = os.path.join(out_dir, obj_file)
res = pd.DataFrame(np.column_stack((elbo, tsne_cost)),
columns=['elbo', 'tsne_cost'])
res.to_csv(obj_file, sep='\t', index=True, header=True)
# Save the mapping results
_save_result(x, y, model, out_dir, name)
return()
def map(args):
x, y, architecture, hyperparameter, train_data, model, _, out_dir, name = \
_init_model(args, 'map')
name = "_".join([name, "map"])
_save_result(x, y, model, out_dir, name)
return()
def _init_model(args, mode):
x = pd.read_csv(args.data_matrix_file, sep='\t').values
config = {}
config_file = CURR_PATH + '/config/model_config.yaml'
config_file = args.config_file or config_file
try:
config_file_yaml = open(config_file, 'r')
config = yaml.load(config_file_yaml)
config_file_yaml.close()
except yaml.YAMLError as exc:
print('Error in the configuration file: {}'.format(exc))
architecture = config['architecture']
architecture.update({'input_dimension': x.shape[1]})
hyperparameter = config['hyperparameter']
if hyperparameter['batch_size'] > x.shape[0]:
hyperparameter.update({'batch_size': x.shape[0]})
model = SCVIS(architecture, hyperparameter)
normalizer = 1.0
if args.pretrained_model_file is not None:
model.load_sess(args.pretrained_model_file)
normalizer = model.get_normalizer()
if mode == 'train':
if args.normalize is not None:
normalizer = float(args.normalize)
else:
normalizer = np.max(np.abs(x))
else:
if args.normalize is not None:
normalizer = float(args.normalize)
x /= normalizer
y = None
if args.data_label_file is not None:
label = pd.read_csv(args.data_label_file, sep='\t').values
label = pd.Categorical(label[:, 0])
y = label.codes
# fixed random seed
np.random.seed(0)
train_data = data.DataSet(x, y)
out_dir = args.out_dir
if not os.path.isdir(out_dir):
os.makedirs(out_dir)
name = '_'.join(['perplexity', str(hyperparameter['perplexity']),
'regularizer', str(hyperparameter['regularizer_l2']),
'batch_size', str(hyperparameter['batch_size']),
'learning_rate', str(hyperparameter['optimization']['learning_rate']),
'latent_dimension', str(architecture['latent_dimension']),
'activation', str(architecture['activation']),
'seed', str(hyperparameter['seed'])])
return x, y, architecture, hyperparameter, train_data, model, normalizer, out_dir, name
def _save_result(x, y, model, out_dir, name):
z_mu, _ = model.encode(x)
plt.figure(figsize=(12, 8))
plt.scatter(z_mu[:, 0], z_mu[:, 1], c=y, s=10)
if y is not None:
plt.colorbar()
fig_name = name + '.png'
fig_name = os.path.join(out_dir, fig_name)
plt.savefig(fig_name)
z_mu = pd.DataFrame(z_mu, columns=['z_coordinate_'+str(i) for i in range(z_mu.shape[1])])
map_name = name + '.tsv'
map_name = os.path.join(out_dir, map_name)
z_mu.to_csv(map_name, sep='\t', index=True, header=True)
##
log_likelihood = model.get_log_likelihood(x)
plt.figure(figsize=(12, 8))
plt.scatter(z_mu.iloc[:, 0], z_mu.iloc[:, 1], c=log_likelihood, s=10)
plt.colorbar()
fig_name = name + '_log_likelihood' + '.png'
fig_name = os.path.join(out_dir, fig_name)
plt.savefig(fig_name)
log_likelihood = pd.DataFrame(log_likelihood, columns=['log_likelihood'])
map_name = name + '_log_likelihood' + '.tsv'
map_name = os.path.join(out_dir, map_name)
log_likelihood.to_csv(map_name, sep='\t', index=True, header=True)