forked from tkopfMaster/Transfer_NN_Recoil
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Training.py
397 lines (336 loc) · 16.9 KB
/
Training.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
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
import numpy as np
np.random.seed(1234)
import tensorflow as tf
from tensorflow_derivative.inputs import Inputs as InputsDer
from tensorflow_derivative.outputs import Outputs as OutputsDer
from tensorflow_derivative.derivatives import Derivatives
import datetime
import h5py
import matplotlib.pyplot as plt
from matplotlib import cm
import sys
from sklearn.preprocessing import StandardScaler
from collections import OrderedDict
import pandas as pd
from pandas import Series, MultiIndex, DataFrame
import seaborn as sns
from matplotlib.lines import Line2D
import pickle
from matplotlib import ticker
formatter = ticker.ScalarFormatter(useMathText=True)
formatter.set_scientific(True)
formatter.set_powerlimits((-2,2))
reweighting = True
def loadAxesSettingsLoss(ax):
ax.set_facecolor('white')
ax.spines['top'].set_linewidth(1)
ax.spines['right'].set_linewidth(1)
ax.spines['bottom'].set_linewidth(1)
ax.spines['left'].set_linewidth(1)
ax.spines['top'].set_color('black')
ax.spines['right'].set_color('black')
ax.spines['bottom'].set_color('black')
ax.spines['left'].set_color('black')
plt.tick_params(axis='both', which='major', labelsize=18)
ax.yaxis.offsetText.set_fontsize(18)
ax.yaxis.offsetText.set_fontsize(18)
plt.grid(color='grey', linestyle='-', linewidth=0.4)
def loadInputsTargetsWeights(outputD, ll):
InputsTargets = h5py.File("%sNN_Input_apply_%s.h5" % (outputD,ll), "r")
norm = InputsTargets['Boson_Pt']
Target = InputsTargets['Target']
weight = InputsTargets['weights']
Input = np.row_stack((
InputsTargets['PF'],
InputsTargets['Track'],
InputsTargets['NoPU'],
InputsTargets['PUCorrected'],
InputsTargets['PU'],
InputsTargets['Puppi'],
InputsTargets['NVertex']
))
return (np.transpose(Input), np.transpose(Target), np.transpose(weight))
def loadInputsTargetsPVWeights(outputD, ll):
InputsTargets = h5py.File("%sNN_Input_apply_%s.h5" % (outputD,ll), "r")
norm = InputsTargets['Boson_Pt']
Target = InputsTargets['Target']
weight = InputsTargets['weights']
Input = np.row_stack((
InputsTargets['PF'],
InputsTargets['Track'],
InputsTargets['NoPU'],
InputsTargets['PUCorrected'],
InputsTargets['PU'],
InputsTargets['Puppi'],
InputsTargets['NVertex']
))
return (np.transpose(Input), np.transpose(Target), np.transpose(weight), np.transpose(InputsTargets['NVertex']))
def moving_average(data_set, periods):
weights = np.ones(periods) / periods
return np.convolve(data_set, weights, mode='valid')
def costExpectedRelAsypTRange(y_true,y_pred, weight, Ranges):
a_=tf.sqrt(tf.square(y_pred[:,0])+tf.square(y_pred[:,1]))
pZ = tf.sqrt(tf.square(y_true[:,0])+tf.square(y_true[:,1]))
alpha_a=tf.atan2(y_pred[:,1],y_pred[:,0])
alpha_Z=tf.atan2(y_true[:,1],y_true[:,0])
alpha_diff=tf.subtract(alpha_a,alpha_Z)
u_long = tf.cos(alpha_diff)*a_
cost = 0
for i in range(0,len(Ranges)-1):
mask = [tf.logical_and((pZ>Ranges[i]) , (pZ<=Ranges[i+1]))]
Response1 = tf.divide(tf.boolean_mask(u_long,tf.reshape(mask, [-1])), tf.boolean_mask(pZ,tf.reshape(mask,[-1])))
#print("tf shape Response1", tf.shape(Response1))
Response_Diff1 = tf.square(tf.reduce_sum(tf.nn.relu(Response1-1))-tf.reduce_sum(tf.nn.relu(1-Response1)))
cost1 = Response_Diff1*0.03
cost = cost + cost1
Response = tf.divide(u_long, pZ)
return cost+tf.sqrt(tf.reduce_sum(tf.square(Response-1)))
def costExpectedRelAsypTPVRange(y_true,y_pred, weight, pTRanges, PVRanges):
PV = weight
a_=tf.sqrt(tf.square(y_pred[:,0])+tf.square(y_pred[:,1]))
pZ = tf.sqrt(tf.square(y_true[:,0])+tf.square(y_true[:,1]))
alpha_a=tf.atan2(y_pred[:,1],y_pred[:,0])
alpha_Z=tf.atan2(y_true[:,1],y_true[:,0])
alpha_diff=tf.subtract(alpha_a,alpha_Z)
u_long = tf.cos(alpha_diff)*a_
cost = 0
for i in range(0,len(pTRanges)-1):
for j in range(0,len(PVRanges)-1):
maskpT = tf.logical_and((pZ>pTRanges[i]) , (pZ<=pTRanges[i+1]))
maskPV = tf.logical_and((PV>PVRanges[j]) , (PV<=PVRanges[j+1]))
mask = [tf.logical_and(tf.reshape(maskpT, [-1]),tf.reshape(maskPV, [-1]))]
Response1 = tf.divide(tf.boolean_mask(u_long,tf.reshape(mask, [-1])), tf.boolean_mask(pZ,tf.reshape(mask,[-1])))
Response_Diff1 = tf.square(tf.reduce_sum(tf.nn.relu(Response1-1))-tf.reduce_sum(tf.nn.relu(1-Response1)))
cost1 = Response_Diff1*0.03
cost = cost + cost1
Response = tf.divide(u_long, pZ)
return cost+tf.sqrt(tf.reduce_sum(tf.square(Response-1)))
def getpTRanges(pT):
pTRanges = np.linspace(np.floor(np.min(pT)),np.max(pT),10)
return pTRanges
def getPVRanges(PV):
PVbins = 5
pPV = 100/PVbins
PVRanges = [0]
PVRanges = np.append(PVRanges, [np.percentile(PV, i*pPV) for i in range(1,PVbins+1)])
return PVRanges
def NNmodel(x, reuse):
ndim = 128
with tf.variable_scope("model") as scope:
if reuse:
scope.reuse_variables()
w1 = tf.get_variable('w1', shape=(19,ndim), dtype=tf.float32,
initializer=tf.glorot_normal_initializer())
b1 = tf.get_variable('b1', shape=(ndim), dtype=tf.float32,
initializer=tf.constant_initializer(0.0))
w2 = tf.get_variable('w2', shape=(ndim, ndim), dtype=tf.float32,
initializer=tf.glorot_normal_initializer())
b2 = tf.get_variable('b2', shape=(ndim), dtype=tf.float32,
initializer=tf.constant_initializer(0.0))
w3 = tf.get_variable('w3', shape=(ndim, ndim), dtype=tf.float32,
initializer=tf.glorot_normal_initializer())
b3 = tf.get_variable('b3', shape=(ndim), dtype=tf.float32,
initializer=tf.constant_initializer(0.0))
w4 = tf.get_variable('w4', shape=(ndim, ndim), dtype=tf.float32,
initializer=tf.glorot_normal_initializer())
b4 = tf.get_variable('b4', shape=(ndim), dtype=tf.float32,
initializer=tf.constant_initializer(0.0))
w5 = tf.get_variable('w5', shape=(ndim, 2), dtype=tf.float32,
initializer=tf.glorot_normal_initializer())
b5 = tf.get_variable('b5', shape=(2), dtype=tf.float32,
initializer=tf.constant_initializer(0.0))
l1 = tf.nn.relu(tf.add(b1, tf.matmul(x, w1)))
l2 = tf.nn.relu(tf.add(b2, tf.matmul(l1, w2)))
l3 = tf.nn.relu(tf.add(b3, tf.matmul(l2, w3)))
l4 = tf.nn.relu(tf.add(b4, tf.matmul(l3, w4)))
logits = tf.add(b5, tf.matmul(l4, w5), name='output')
return logits, logits
def getNNModel(outputDir, loss_fct, ll, plotsD):
if loss_fct == 'relResponseAsypTPVRange':
Inputs, Targets, Weights, PV = loadInputsTargetsPVWeights(outputDir, ll)
else:
Inputs, Targets, Weights = loadInputsTargetsWeights(outputDir, ll)
Boson_Pt = np.sqrt(np.square(Targets[:,0])+np.square(Targets[:,1]))
num_events = Inputs.shape[0]
train_test_splitter = 0.5
training_idx = np.random.choice(np.arange(Inputs.shape[0]), int(Inputs.shape[0]*train_test_splitter), replace=False)
#Write Test Idxs
test_idx = np.setdiff1d( np.arange(Inputs.shape[0]), training_idx)
dset = Test_Idx.create_dataset("Test_Idx", dtype='f', data=test_idx)
#Get Datasets for training and validation
Inputs_train, Inputs_test = Inputs[training_idx,:], Inputs[test_idx,:]
Targets_train, Targets_test = Targets[training_idx,:], Targets[test_idx,:]
train_val_splitter = 0.9
train_train_idx_idx = np.random.choice(np.arange(training_idx.shape[0]), int(training_idx.shape[0]*train_val_splitter), replace=False)
train_train_idx = training_idx[train_train_idx_idx]
train_val_idx = training_idx[ np.setdiff1d( np.arange(training_idx.shape[0]), train_train_idx_idx)]
Inputs_train_train, Inputs_train_val = Inputs[train_train_idx,:], Inputs[train_val_idx,:]
Targets_train, Targets_test = Targets[train_train_idx,:], Targets[train_val_idx,:]
if reweighting and not (loss_fct == 'relResponseAsypTPVRange'):
weights_train_, weights_val_ = Weights[train_train_idx,:], Weights[train_val_idx,:]
elif reweighting and loss_fct == 'relResponseAsypTPVRange':
prob_train_, prob_val_ = Weights[train_train_idx,:], Weights[train_val_idx,:]
weights_train_, weights_val_ = PV[train_train_idx,:], PV[train_val_idx,:]
else:
print("No reweighting")
weights_train_, weights_val_ = np.repeat(1., len(train_train_idx)) , np.repeat(1., len(train_val_idx))
weights_train_.shape = (len(train_train_idx),1)
weights_val_.shape = (len(train_val_idx),1)
#Assignment of datasets used in the training
data_train = Inputs_train_train
labels_train = Targets_train
data_val = Inputs_train_val
labels_val = Targets_test
weights_train = weights_train_
weights_val = weights_val_
#Batchsizes
batchsize = 4500
batchsize_val = 10000
#Get stings of inputs
MET_definitions = ['PF', 'Track', 'NoPU', 'PUCorrected', 'PU', 'Puppi']
Variables = ['x','y','SumEt']
Variables = [Variables[:] for _ in range(len(MET_definitions))]
MET_definitions = np.repeat(MET_definitions,3)
Variables = [item for sublist in Variables for item in sublist]
Inputstring = [MET+'_'+Variable for MET,Variable in zip(MET_definitions,Variables)]
Inputstring = np.append(Inputstring, 'NVertex')
#Placeholders
xDer = InputsDer(Inputstring)
x = xDer.placeholders
y = tf.placeholder(tf.float32, shape=[batchsize, labels_train.shape[1]])
w = tf.placeholder(tf.float32, shape=[batchsize, weights_train.shape[1]])
x_ = tf.placeholder(tf.float32)
y_ = tf.placeholder(tf.float32)
w_ = tf.placeholder(tf.float32)
#GPU configs
print("tf.test.gpu_device_name()", tf.test.gpu_device_name())
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
#Define the neural network architecture
batch_train = [data_train, labels_train, weights_train]
batch_val = [data_val, labels_val, weights_val]
logits_train, f_train= NNmodel(x, reuse=False)
yDer = OutputsDer(logits_train, ['x','y'])
logits_val, f_val= NNmodel(x_, reuse=True)
#Initialize derivates
derivatives = Derivatives(xDer, yDer)
d={}
for i in range(0,len(Inputstring)):
Variable = sorted(Inputstring)[i]
d["1dxd"+Variable]=derivatives.get('x', [Variable])
d["1dyd"+Variable]=derivatives.get('y', [Variable])
list_derivatestensor = d.values()
print("loss fct", loss_fct)
if (loss_fct=="mean_squared_error"):
loss_train = tf.reduce_mean(
tf.losses.mean_squared_error(labels=y, predictions=logits_train))
loss_val = tf.reduce_mean(
tf.losses.mean_squared_error(labels=y_, predictions=logits_val))
minimize_loss = tf.train.AdamOptimizer().minimize(loss_train)
elif (loss_fct=="relResponseAsypTRange"):
pTRanges = []
pTRanges = getpTRanges(Boson_Pt)
loss_train = costExpectedRelAsypTRange(y, logits_train, w, pTRanges)
loss_val = costExpectedRelAsypTRange(y_, logits_val, w_, pTRanges)
minimize_loss = tf.train.AdamOptimizer().minimize(loss_train)
elif (loss_fct=="relResponseAsypTPVRange"):
pTRanges, PVRanges = [], []
pTRanges = getpTRanges(Boson_Pt)
PVRanges = getPVRanges(PV)
loss_train = costExpectedRelAsypTPVRange(y, logits_train, w, pTRanges, PVRanges)
loss_val = costExpectedRelAsypTPVRange(y_, logits_val, w_, pTRanges, PVRanges)
minimize_loss = tf.train.AdamOptimizer().minimize(loss_train)
else:
sys.exit("Error, no suitable loss declared")
# ## Run the training
sess.run(tf.global_variables_initializer())
losses_train = []
losses_val = []
#Initialize summary
summary_train = tf.summary.scalar("loss_train", loss_train)
summary_val = tf.summary.scalar("loss_val", loss_val)
writer = tf.summary.FileWriter("./logs/{}".format(
datetime.datetime.now().strftime("%Y-%m-%d %H:%M:%S")), sess.graph)
saver = tf.train.Saver()
#Initialize Training steps
min_valloss = [1000000000000]
max_steps = 300000
saveStep = 100
early_stopping = 0
best_model = 0
if loss_fct == 'relResponseAsypTPVRange':
batch_prob = prob_train_.flatten() * 1 / (np.sum(prob_train_.flatten()))
batch_prob_val = prob_val_.flatten() * 1 / (np.sum( prob_val_.flatten()))
else:
batch_prob = weights_train.flatten() * 1 / (np.sum(weights_train.flatten()))
batch_prob_val = weights_val.flatten() * 1 / (np.sum( weights_val.flatten()))
pT = np.sqrt(np.square(labels_train[:,0]) + np.square(labels_train[:,1]))
#Preprocessing
preprocessing_input = StandardScaler()
preprocessing_output = StandardScaler()
preprocessing_i = preprocessing_input.fit(Inputs)
preprocessing_o = preprocessing_output.fit(Targets)
#Save preprocessing in pickle
pickle.dump(preprocessing_i, open("preprocessing_input.pickle", "wb"))
pickle.dump(preprocessing_o, open("preprocessing_output.pickle", "wb"))
for i_step in range(max_steps):
batch_train_idx = np.random.choice(np.arange(data_train.shape[0]), batchsize, p=batch_prob, replace=False)
batch_val_idx = np.random.choice(np.arange(data_val.shape[0]), batchsize, p=batch_prob_val, replace=False)
summary_, loss_, _ = sess.run([summary_train, loss_train, minimize_loss], feed_dict={x: preprocessing_input.transform(data_train[batch_train_idx,:]), y: labels_train[batch_train_idx,:], w: weights_train[batch_train_idx,:]})
losses_train.append(loss_)
writer.add_summary(summary_, i_step)
summary_, loss_ = sess.run([summary_val, loss_val], feed_dict={x_: preprocessing_input.transform(data_val[batch_val_idx,:]), y_: labels_val[batch_val_idx,:], w_: weights_val[batch_val_idx,:]})
losses_val.append(loss_)
writer.add_summary(summary_, i_step)
if i_step % saveStep == 0:
batch_val_idx_100 = np.random.choice(np.arange(data_val.shape[0]), batchsize_val, p=batch_prob_val, replace=False)
loss_ = sess.run(loss_val, feed_dict={x_: preprocessing_input.transform(data_val[batch_val_idx_100,:]), y_: labels_val[batch_val_idx_100,:], w_: weights_val[batch_val_idx_100,:]})
if loss_<min(min_valloss):
best_model = i_step
saver.save(sess, "%sNNmodel"%outputDir, global_step=i_step)
outputs = ["output"]
constant_graph = tf.graph_util.convert_variables_to_constants(
sess, sess.graph.as_graph_def(), outputs)
tf.train.write_graph(constant_graph, outputDir, "constantgraph.pb", as_text=False)
early_stopping = 0
print("better val loss found at ", i_step)
list_derivatestensor = d.values()
else:
early_stopping += 1
print("increased early stopping to ", early_stopping)
if early_stopping == 120:
break
min_valloss.append(loss_)
print('gradient step No ', i_step)
print("validation loss", loss_)
#Plot loss functions
fig=plt.figure(figsize=(10,6))
fig.patch.set_facecolor('white')
ax = plt.subplot(111)
plt.plot(range(1, len(moving_average(np.asarray(losses_train[0:(best_model+800)]), 800))+1), moving_average(np.asarray(losses_train[0:(best_model+800)]), 800), lw=3, label="Training batches", color='#ED2024')
plt.plot(range(1, len(moving_average(np.asarray(losses_val[0:(best_model+800)]), 800))+1), moving_average(np.asarray(losses_val[0:(best_model+800)]), 800), lw=3, label="Validation batches", color='#2C6AA8')
plt.xlabel("Training step", fontsize=22), plt.ylabel("$\\langle \mathscr{L} \\rangle_{800}$", fontsize=22)
plt.yscale('log')
plt.xscale('log')
legend = plt.legend()
plt.setp(plt.gca().get_legend().get_texts(), fontsize=18)
plt.tick_params(axis='both', which='major', labelsize=18)
plt.grid(color='black', linestyle='-', linewidth=0.1)
ax.set_facecolor('white')
ax.yaxis.offsetText.set_fontsize(18)
loadAxesSettingsLoss(ax)
plt.savefig("%sLoss_ValLoss.png"%(plotsD), bbox_inches="tight")
plt.close()
dset = NN_Output.create_dataset("loss", dtype='f', data=losses_train)
dset2 = NN_Output.create_dataset("val_loss", dtype='f', data=losses_val)
NN_Output.close()
if __name__ == "__main__":
outputDir = sys.argv[1]
loss_fct = str(sys.argv[2])
ll = sys.argv[3]
plotsD = sys.argv[4]
print(outputDir)
NN_Output = h5py.File("%sNN_Output_%s.h5"%(outputDir,ll), "w")
Test_Idx = h5py.File("%sTest_Idx_%s.h5" % (outputDir, ll), "w")
getNNModel(outputDir, loss_fct, ll, plotsD)