-
Notifications
You must be signed in to change notification settings - Fork 0
/
mlp_modifyoutput.py
357 lines (287 loc) · 14.8 KB
/
mlp_modifyoutput.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
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
from eagle.io import *
from eagle.plot import *
from eagle.nn import *
import numpy as np
from numpy.lib.recfunctions import structured_to_unstructured
import tensorflow as tf
import shap
from keras import backend
from argparse import ArgumentParser
parser = ArgumentParser()
parser.add_argument("input", help = "Choose nocolors, subsetcolors or allcolors")
parser.add_argument("sampling", help = "Choose None, random or uniform")
args = parser.parse_args()
if args.sampling == 'None': args.sampling = None
assert args.input == 'nocolors' or args.input == 'subsetcolors' or args.input == 'allcolors'
assert args.sampling == None or args.sampling == 'random' or args.sampling == 'uniform'
seed = 7
np.random.seed(seed) #fix seed for reproducibility
tf.set_random_seed(seed)
plotting = True
inp = args.input #'nocolors'#'subsetcolors'#'allcolors'#
#EAGLE settings
fol = ''
sim = 'RefL0100N1504' # simulation
cat = 'sdss' # catalogue to build model to
dust = 'dusty' # '' if no dust
snap = 27 # snapshot to use
redshift = 1.0063854e-01#0.1# redshift of snapshot
#redshift_array = np.flip(np.array([2.2204460e-16, 1.0063854e-01, 1.8270987e-01, 2.7090108e-01, 3.6566857e-01, 5.0310730e-01, 6.1518980e-01, 7.3562960e-01, 8.6505055e-01, 1.0041217e+00, 1.2593315e+00, 1.4867073e+00, 1.7369658e+00, 2.0124102e+00, 2.2370370e+00, 2.4784133e+00, 3.0165045e+00, 3.5279765e+00, 3.9836636e+00, 4.4852138e+00, 5.0372367e+00, 5.4874153e+00, 5.9711623e+00, 7.0495663e+00, 8.0746160e+00, 8.9878750e+00, 9.9930330e+00]))
#redshift = redshift_array[snap - 2]
fluxes = ('u', 'g', 'r', 'i', 'z')
if inp == 'nocolors':
colors = ()
elif inp == 'subsetcolors':
colors = ('ug', 'gr', 'ri', 'iz')
elif inp == 'allcolors':
colors = ('ug', 'ur', 'ui', 'uz', 'gr', 'gi', 'gz', 'ri', 'rz', 'iz')
xcols = fluxes + colors
xcols_string = ','.join(xcols)
#eagle
eagle_xtype = 'flux'
eagle_xcols = [dust + eagle_xtype + '_' + cat + '_' + f for f in xcols]
eagle_ycols = ['m_star']#['sfr']#
#sdss
sdss_xtype = 'flux'
sdss_xcols = [sdss_xtype + '_' + f for f in xcols]
sdss_ycols = ['Mstellar_median']#['SFR_median']#
#input output plot
#xnames=[dust + ' ' + cat + ' ' + i for i in list(fluxes) + [c[0] + '-' + c[1] for c in colors]]
xnames= list(fluxes) + [c[0] + '-' + c[1] for c in colors]
ynames=['$\log_{10} M_{*} (M_{\odot})$']#['$\log_{10}$ SFR $(M_{\odot} yr^{-1})$']#
#preprocessing
cs = False ; cs_mask = None #central satellite analysis (only change cs)
sampling = args.sampling#'uniform'
N=625
#bins = 10 ; count = 50 ; N = 625 ; scaling = True if sampling == None else False #both same uniform mass distribution
#bins=10 ; count=125 ; N=625
#bins=200 ; count=5 ; N=625
"""
#ML settings GridSearch
h_nodes = [20, 50, 50]
activation = ['tanh', 'linear', 'tanh', 'linear']
optimizer = 'Adamax'
"""
if inp == 'nocolors':
#Hyperas no colors
h_nodes = [40, 50, 30]
activation = ['linear', 'tanh', 'tanh', 'linear']
elif inp == 'subsetcolors':
#Hyperas subset colors
h_nodes = [50, 50, 20]
activation = ['relu', 'tanh', 'linear', 'linear']
elif inp == 'allcolors':
#Hyperas all colors
h_nodes = [30, 30, 50]
activation = ['linear', 'tanh', 'relu', 'linear']
optimizer = 'Adam'
dropout = [0., 0., 0.]
perc_train = .8 #todo: this is not used anywhere, give it to io
loss = 'mean_absolute_error'
epochs = 15
batch_size = 128
def cross_validation(X, Y, K, seed):
kf = KFold(n_splits=K, random_state=seed)
X_train = [] ; Y_train = []; X_val = []; Y_val = []
for idx_train, idx_val in kf.split(X):
X_train.append(X[idx_train])
Y_train.append(Y[idx_train])
X_val.append(X[idx_val])
Y_val.append(Y[idx_val])
return X_train, Y_train, X_val, Y_val
def add_scores(score, x, y, y_pred):
y.reshape(len(y),) ; y_pred.reshape(len(y_pred),)
score[1] = r2_score(y, y_pred)
return np.append(score, [adjusted_R_squared(score[1], x.shape[1], len(y)), np.mean(y_pred - y), np.var(y_pred - y)])
"""
Note that we perform everything in logscale since the errors are also given in log scale!
"""
#-----------------EAGLE preprocessing-------------------
eagle = EAGLE(fol, sim, cat, dust, snap, redshift, seed, eagle_xcols, eagle_ycols, eagle_xtype)
eagle.preprocess(colors)
#-----------------SDSS preprocessing-------------------
sdss = SDSS(sdss_xcols, sdss_ycols, sdss_xtype, redshift)
#datacols
prefix = sdss.ycols[0].replace('median', '')
sdss.datacols += [prefix + '16th', prefix + '84th']
#read data
sdss.read_data()
#select only galaxies of given redshift
sdss.select_redshift(frac=5e-3)
#to magnitude
if sdss.xtype == 'flux':
to_magnitude(sdss, sdss.xtype, luminosity_distance_sdss)
#add colors
add_colors(sdss, sdss.xtype+'_', colors)
#remove galaxies with nan mass
sdss.data = sdss.data[~np.isnan(structured_to_unstructured(sdss.data[sdss.ycols]))]
#-----------------Add variance to mass-------------------
#array of mass errors
mass_sigma = (sdss.data[prefix + '84th'] - sdss.data[prefix + '16th'])/2
#mass arrays
eagle_mass = structured_to_unstructured(eagle.data[eagle.ycols])
sdss_mass = structured_to_unstructured(sdss.data[sdss.ycols])
#mass_bins = 20 ; mass_count = 25
mass_bins = 300; mass_count = 5
#mass errors arrays
mass_errors = np.empty((mass_bins))
#histogram of sdss logarithmic stellar mass
hist_sdss, mass_edges = np.histogram(sdss_mass, bins=mass_bins)
#histogram of eagle logarithmic stellar mass
hist_eagle = np.histogram(eagle_mass, bins=mass_edges)[0]
#remaining bins
remain = (hist_sdss > mass_count) & (hist_eagle > mass_count)
remain = consecutive_trues(remain)
print("Number of remaining bins: ", len(np.where(remain)[0]))
for j in range(mass_bins):
mask_eagle = (eagle_mass >= mass_edges[j]) & (eagle_mass < mass_edges[j+1])
mask_sdss = (sdss_mass >= mass_edges[j]) & (sdss_mass < mass_edges[j+1])
if remain[j]:
mass_errors[j] = np.average(mass_sigma[mask_sdss])
eagle.data[eagle.ycols][mask_eagle] = np.random.normal(structured_to_unstructured(eagle.data[eagle.ycols][mask_eagle]), mass_errors[j])
else:
eagle.data[eagle.ycols][mask_eagle] = np.nan
sdss.data[sdss.ycols][mask_sdss] = np.nan
#remove nans
eagle.data = eagle.data[~np.isnan(structured_to_unstructured(eagle.data[eagle.ycols]))]
sdss.data = sdss.data[~np.isnan(structured_to_unstructured(sdss.data[sdss.ycols]))]
#eagle remaining preprocessing steps
eagle.x, eagle.y = divide_input_output(eagle)
#sdss remaining preprocessing steps
sdss.datacols = sdss.xcols + sdss.ycols
select_cols(sdss)
sdss.x, sdss.y = divide_input_output(sdss)
sdss.x, sdss.y = remove_zero(sdss.x, sdss.y)
sdss.x, sdss.y = remove_minone(sdss.x, sdss.y)
sdss.y, sdss.y = remove_nan(sdss.x, sdss.y)
sdss.x, sdss.y = remove_inf(sdss.x, sdss.y)
#determine number of bins for sampling
#binsize = 0.01046
#range_mass = np.amax(eagle.y) - np.amin(eagle.y)
#bins = int(range_mass / binsize)
#count = int(np.ceil(N / bins / 5)*5)+5
#print("number of bins of size ", binsize, " dex: ", bins)
edges = np.load('/disks/strw9/vanweenen/mrp2/plots/edges_remain.npy')
bins=10 ; count=125
sample(eagle, sdss, sampling, edges, count, N)
eagle.scaling()
sdss.scaling(eagle)
x_train, x_test, y_train, y_test = train_test_split(eagle.x, eagle.y, test_size=0.2, random_state=eagle.seed, shuffle=True)
print("Total size of data: %s; size of training set: %s ; size of test set: %s"%(len(eagle.x), len(x_train), len(x_test)))
print("EAGLE: len(eagle.x) = %s ; len(x_train) = %s ; len(x_test) = %s ; len(sdss.x) = %s"%(len(eagle.x), len(x_train), len(x_test), len(sdss.x)))
if plotting:
if sampling == 'uniform':
edges = eagle.yscaler.transform(eagle.edges.reshape(-1,1)).T[0]
else:
edges = 7
#edges = 10
plot_data = PLOT_DATA(xnames+ynames, sim=sim, snap=snap, N=len(sdss.y), inp=inp, sampling=str(sampling))
plot_data.hist_data(('eagle-train', 'eagle-test', 'sdss-total'), [np.hstack((x_train, y_train)), np.hstack((x_test, y_test)), np.hstack((sdss.x, sdss.y))], edges, xlim=[-1.5,1.5], ylim=[-1.1,1.1])
plot_data.datanames = xnames+[ynames[0].split('(')[0]+'$']
plot_data.statistical_matrix(x_test, y_test, ['eagle'], simple=True)
plot_data.statistical_matrix(sdss.x, sdss.y, ['sdss'], simple=True)
input_size = len(x_train[0])
output_size = len(y_train[0])
nodes = [input_size] + h_nodes + [output_size]
print("nodes in the network: ", nodes)
#read hyperparameters
nn = NN(input_size, output_size, h_nodes, activation, dropout, loss, epochs, batch_size, optimizer)
new_result = AdditionalValidationSets([(sdss.x, sdss.y, 'val2')])
callbacks = [new_result]
#------------------------------------------Cross-validate on EAGLE--------------------------------------------------------------
K = 5
#crossvalidation data
X_train, Y_train, X_val, Y_val = cross_validation(x_train, y_train, K, seed)
cv_score = np.empty((K, 5))
for i in range(K):
#make architecture of the network
nn.MLP_model()
result = nn.model.fit(X_train[i], Y_train[i], batch_size=batch_size, epochs=epochs, verbose=0, validation_data=(X_val[i], Y_val[i]), callbacks=callbacks)
cv_score[i,:2] = nn.model.evaluate(X_val[i], Y_val[i], verbose=0)
y_pred = nn.model.predict(X_val[i]) #predict
backend.clear_session()
cv_score[i] = add_scores(cv_score[i,:2], X_val[i], Y_val[i], y_pred) #add scores
avg_cv_score = np.average(cv_score, axis=0)
print("Cross-validated errors (EAGLE): MAE = %.3e ; R^2 = %.3f ; adjusted R^2 = %.3f ; error_mean = %.3e ; error_sigma = %.3e"%tuple(avg_cv_score))
#------------------------------------------Train and test on EAGLE--------------------------------------------------------------
#train, evaluate, predict, postprocess
nn.MLP_model()
result = nn.model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=0, validation_data=(x_test, y_test), callbacks=callbacks) #train
score = nn.model.evaluate(x_test, y_test, verbose=0) #evaluate
y_pred = nn.model.predict(x_test) #predict
score = add_scores(score, x_test, y_test, y_pred) #add scores
print("Errors (EAGLE): MAE = %.3e ; R^2 = %.3f ; adjusted R^2 = %.3f ; error_mean = %.3e ; error_sigma = %.3e"%tuple(score))
#central satellites analysis
if cs == True:
cs_mask = eagle.cs_test
score_c = nn.model.evaluate(x_test[cs_mask], y_test[cs_mask], verbose=0)
score_s = nn.model.evaluate(x_test[~cs_mask], y_test[~cs_mask], verbose=0)
score_c = add_scores(score_c, x_test[cs_mask], y_test[cs_mask], y_pred[cs_mask])
score_s = add_scores(score_s, x_test[~cs_mask], y_test[~cs_mask], y_pred[~cs_mask])
print("Errors (EAGLE) centrals: MAE = %.3e ; R^2 = %.3g ; adjusted R^2 = %.3g ; error_mean = %.3e ; error_sigma = %.3e"%tuple(score_c))
print("Errors (EAGLE) satellites: MAE = %.3e ; R^2 = %.3g ; adjusted R^2 = %.3g ; error_mean = %.3e ; error_sigma = %.3e"%tuple(score_s))
#------------------------------------------Compare with SDSS--------------------------------------------------------------
#evaluate, predict, postprocess
sdss_score = nn.model.evaluate(sdss.x, sdss.y, verbose=0) #evaluate
sdss.ypred = nn.model.predict(sdss.x) #predict
sdss_score = add_scores(sdss_score, sdss.x, sdss.y, sdss.ypred)
print("Errors (SDSS): MAE = %.3e ; R^2 = %.3g ; adjusted R^2 = %.3g ; error_mean = %.3e ; error_sigma = %.3e"%tuple(sdss_score))
np.save('/disks/strw9/vanweenen/mrp2/plots/score_'+xcols_string+'_'+eagle_ycols[0]+'_inp='+inp+'_sampling='+str(sampling)+'.pdf', np.vstack((avg_cv_score, score, sdss_score)))
#------------------------------------------Calculate shapley values--------------------------------------------------------------
explainer = shap.DeepExplainer(nn.model, x_train)
shap_eagle = explainer.shap_values(x_test)[0]
shap_sdss = explainer.shap_values(sdss.x)[0]
shapdir = '/disks/strw9/vanweenen/mrp2/plots/mlp_dustyflux_mstar/feature_importance/shap/tests_linearfunc/one_coeff/'
#np.save(shapdir + 'shap_eagle_' + inp + '.npy', shap_eagle)
#np.save(shapdir + 'shap_sdss_' + inp + '.npy', shap_sdss)
#shap_eagle_ref = np.load(shapdir + 'shap_eagle_' + inp + '.npy')
#shap_sdss_ref = np.load(shapdir + 'shap_sdss_' + inp + '.npy')
#shap_eagle -= shap_eagle_ref
#shap_sdss -= shap_sdss_ref
plot_shap = PLOT_SHAP(xnames, xcols_string, inp, str(sampling))
plot_shap.summary_plot(x_test, shap_eagle, 'eagle', shap_sdss, 'sdss', plottype = 'bar')
plot_shap.summary_plot(x_test, shap_eagle, 'eagle', plottype = 'dot', dotsize=10)
#------------------------------------------Plotting--------------------------------------------------------------
x_test, y_test, y_pred = eagle.postprocess(x_test, y_test, y_pred) #postprocessing
#plot
if plotting:
plot = PLOT_NN(eagle, nn, xnames, ynames, dataname='eagle', N=len(y_test), xcols_string=xcols_string, inp=inp, sampling=str(sampling), score=score)#[mse[19], r2[19]]
plot.plot_learning_curve(new_result)
plot.plot_input_output(x_test, y_test, y_pred, 'scatter') #scatter, contour
plot.plot_true_predict(y_test, y_pred, 'scatter', cs_mask) #scatter, hexbin, contour, contourf
plot.plot_output_error(y_test, y_pred, 'contourf', cs_mask) #scatter, hexbin, contour
sdss.x, sdss.y, sdss.ypred = sdss.postprocess(eagle, sdss.x, sdss.y, sdss.ypred) #postprocessing
#plot
if plotting:
plot = PLOT_NN(eagle, nn, xnames, ynames, dataname='sdss', N=len(sdss.y), xcols_string=xcols_string, inp=inp, sampling=str(sampling), score=sdss_score)#[mse[19], r2[19]]
plot.plot_input_output(sdss.x, sdss.y, sdss.ypred) #scatter, contour
plot.plot_true_predict(sdss.y, sdss.ypred, 'scatter') #scatter, hexbin, contour, contourf
plot.plot_output_error(sdss.y, sdss.ypred, 'contourf', ylim=(-0.15, 0.3)) #scatter, hexbin, contour
"""
#------------------------------------------CREATE GIF--------------------------------------------------------------
#read data
eagle = EAGLE(fol, sim, cat, snap, redshift, eagle_xcols, eagle_ycols, perc_train)
x_train, y_train, x_test, y_test = eagle.preprocess(eagle_xtype)
input_size = len(x_train[0])
output_size = len(y_train[0])
nodes = [input_size] + h_nodes + [output_size]
print("nodes in the network: ", nodes)
#read hyperparameters and make network
nn = NN(input_size, output_size, h_nodes, activation, dropout, loss, epochs, batch_size, optimizer)
nn.MLP_model()
Y_pred, mse, r2 = [], [], []
for i in range(epochs):
#train, evaluate, predict, postprocess
result = nn.model.fit(x_train, y_train, batch_size=batch_size, epochs=1, verbose=0, validation_data=(x_test, y_test)) #train
score = nn.model.evaluate(x_test, y_test, verbose=0) #evaluate
y_pred = nn.model.predict(x_test) #predict
x_test_log, y_test_log, y_pred = eagle.postprocess(x_test, y_test, y_pred) #postprocess
#plotting
plot = PLOT_NN(eagle, nn, xnames, ynames, MLtype='MLP', score=score, epochs=i)
#plot.plot_input_output(x_test_log, y_test_log, y_pred)
#plot.plot_true_predict(y_test_log, y_pred)
#save values for gif
Y_pred.append(y_pred) ; mse.append(score[0]) ; r2.append(score[1])
#plot gif
plot.gif_input_output(x_test_log, y_test_log, Y_pred=Y_pred, mse=mse, r2=r2)
"""