-
Notifications
You must be signed in to change notification settings - Fork 2
/
regression.py
774 lines (642 loc) · 34.2 KB
/
regression.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
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
445
446
447
448
449
450
451
452
453
454
455
456
457
458
459
460
461
462
463
464
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
725
726
727
728
729
730
731
732
733
734
735
736
737
738
739
740
741
742
743
744
745
746
747
748
749
750
751
752
753
754
755
756
757
758
759
760
761
762
763
764
765
766
767
768
769
770
771
772
773
774
"""
Train a neural network to approximate a continous function
Generate/get training data methods:
size = number of training vectors
LJ: Lennard-Jones
SW: Stillinger-Weber
functionData: random 1-dim input and output data:
input: r = [size,1], output: E = [size,1] (LJ)
neighbourData: random N-dim input data (N neighbours)
and random 1-dim output data
output is total energy of N neighbours
input: r = [size,N], output: E = [size,1] (LJ)
if functionDerivative is specified (not None):
random 4N-dim input data (N neighoburs)
and random 4-dimloutput data
(x, y, z, r) of each neighbour is supplied
output is total energy and force (LJ)
input: r = [size,4N], output: [size,4] : (Fx, Fy ,Fz, E)
radialSymmetry: random M-dim input data and random 1-dim output data
use radial symmetry functions to transform radial/coordinates to
input: [size, M] where M is number of symmetry functions
output: [size, 1] (total energy)
angularSymmetry: same as for radialSymmetry, but using angular symmetry functions
to transform random data. SW potential is used to produce
output data
lammps: (x, y, z, r) of each neighbour and total E is sampled from lammps
input data is transformed with angular symmetry functions
same sizes of input and output as for radialSymmetry
and angularSymmetry
"""
import tensorflow as tf
import numpy as np
import sys
import datetime as time
import os
import shutil
import matplotlib.pyplot as plt
import DataGeneration.randomData as data
import DataGeneration.lammpsData as lammps
import DataGeneration.symmetries as symmetries
import neuralNetwork as nn
from Tools.inspect_checkpoint import print_tensors_in_checkpoint_file
from Tools.freeze_graph import freeze_graph
from time import clock as timer
import Tools.matplotlibParameters
loadFlag = False
loadDir = ''
loadFileName = ''
saveFlag = False
saveDirName = ''
summaryFlag = False
summaryDir = ''
saveGraphFlag = False
saveGraphName = ''
saveGraphProtoFlag = False
saveGraphProtoName = ''
saveMetaName = ''
saveMetaFlag = False
saveParametersFlag = False
plotFlag = False
plotErrorFlag = False
now = time.datetime.now().strftime("%d.%m-%H.%M.%S")
trainingDir = 'TrainingData' + '/' + now
# make directory for training data
if len(sys.argv) > 1:
i = 1
while i < len(sys.argv):
if sys.argv[i] == '--save' or sys.argv[i] == '--savegraph' or sys.argv[i] == '--savegraphproto' \
or sys.argv[i] == '--summary':
if os.path.exists(trainingDir):
print "Attempted to place data in existing directory, %s. Exiting." % trainingDir
exit(1)
else:
os.mkdir(trainingDir)
saveMetaName = trainingDir + '/' + 'meta.dat'
saveGraphName = trainingDir + '/' + 'graph.dat'
print "Making data directory: ", saveMetaName
break
i += 1
# process command-line input
if len(sys.argv) > 1:
i = 1
while i < len(sys.argv):
if sys.argv[i] == '--load':
i += 1
loadFlag = True
loadDir = sys.argv[i]
i += 1
# find latest checkpoint
loadDir = 'TrainingData/' + loadDir
checkpointFile = loadDir + "/Checkpoints/checkpoint_state"
with open(checkpointFile, 'r') as infile:
line = infile.readline()
words = line.split()
checkpoint = words[1][1:-1]
loadFileName = loadDir + "/Checkpoints/" + checkpoint
print
print "Information on trained graph: "
command = "cat " + loadDir + "/README.txt"
os.system(command)
print
elif sys.argv[i] == '--save':
i += 1
saveFlag = True
saveMetaFlag = True
print "Checkpoints will be saved"
# make new directory for checkpoints
saveDirName = trainingDir + '/Checkpoints'
os.mkdir(saveDirName)
# copy the python source code used to run the training, to preserve
# the tf graph (which is not saved by tf.nn.Saver.save()).
shutil.copy2(sys.argv[0], saveDirName + '/')
elif sys.argv[i] == '--summary':
i += 1
summaryFlag = True
saveMetaFlag = True
print "Summaries will be saved"
# make new directory for summaries
summaryDir = trainingDir + '/Summaries'
os.mkdir(summaryDir)
elif sys.argv[i] == '--savegraph':
i += 1
saveGraphFlag = True
saveMetaFlag = True
print "Graph.txt will be saved"
elif sys.argv[i] == '--savegraphproto':
i += 1
saveGraphProtoFlag = True
saveMetaFlag = True
print "Binary graph will be saved"
elif sys.argv[i] == '--plot':
i += 1
plotFlag = True
elif sys.argv[i] == '--ploterror':
i += 1
plotErrorFlag = True
else:
i += 1
# copy readme file
if saveFlag and loadFlag:
os.system("cp " + loadDir + "/README.txt " + trainingDir)
class Regression:
def __init__(self, function, trainSize, batchSize, testSize, inputs, outputs,
functionDerivative=False, learningRate=0.001, RMSEtol=1e-10):
self.trainSize = trainSize
self.batchSize = batchSize
self.testSize = testSize
self.function = function
self.inputs = inputs
self.outputs = outputs
self.functionDerivative = functionDerivative
self.learningRate = learningRate
self.RMSEtol = RMSEtol
# save output to terminal
if saveFlag or saveGraphFlag:
filepath = trainingDir + '/output.txt'
self.outputFile = open(filepath, 'w')
sys.stdout = self.outputFile
def generateData(self, a, b, method, numberOfSymmFunc=10, neighbours=80, \
symmFuncType='G4', dataFolder='', batch=50,
varyingNeigh=True, forces=False, Behler=True,
klargerj=True, tags=False, atomType=0, nTypes=1, nAtoms=10,
normalize=False, shiftMean=False, standardize=False):
self.a, self.b = a, b
self.neighbours = neighbours
self.forces = forces
global saveParametersFlag
self.samplesDir = dataFolder
self.symmFuncType = symmFuncType
if method == 'twoBody':
print "method=twoBody: Generating random, radial 1-neighbour data..."
self.xTrain, self.yTrain, self.xTest, self.yTest = \
data.twoBodyEnergy(self.function, self.trainSize, self.testSize, a=a, b=b)
self.numberOfBatches = self.trainSize/self.batchSize
elif method == 'neighbourTwoBody':
if self.functionDerivative:
print "method=neighbourTwoBody: Generating random, radial N-neighbour data including force output..."
neighbours = self.inputs / 4
print neighbours
self.xTrain, self.yTrain = \
data.neighbourTwoBodyEnergyAndForce2(self.function, self.functionDerivative, \
self.trainSize, \
neighbours, self.outputs, a, b)
self.xTest, self.yTest = \
data.neighbourTwoBodyEnergyAndForce2(self.function, self.functionDerivative, \
self.testSize, \
neighbours, self.outputs, a, b)
else:
print "method=neighbourTwoBody: Generating random, radial N-neighbour data..."
self.xTrain, self.yTrain = \
data.neighbourData(self.function, self.trainSize, a, b, \
inputs=self.inputs, outputs=self.outputs)
self.xTest, self.yTest = \
data.neighbourData(self.function, self.testSize, a, b, \
inputs=self.inputs, outputs=self.outputs)
elif method == 'twoBodySymmetry':
print "method=twoBodySymmetry: Generating random, two-body N-neighbour data with symmetry functions..."
self.xTrain, self.yTrain, self.parameters = \
data.neighbourTwoBodySymmetry(self.function, self.trainSize, \
neighbours, numberOfSymmFunc, symmFuncType, a, b,
varyingNeigh=varyingNeigh)
self.xTest, self.yTest, _ = \
data.neighbourTwoBodySymmetry(self.function, self.testSize, \
neighbours, numberOfSymmFunc, symmFuncType, a, b,
varyingNeigh=varyingNeigh)
if saveMetaFlag:
saveParametersFlag = True
elif method == 'threeBodySymmetry':
print "method=threeBodySymmetry: Generating random, three-body N-neighbour data with symmetry functions..."
self.xTrain, self.yTrain, self.parameters = \
data.neighbourThreeBodySymmetry(self.function, self.trainSize, \
neighbours, numberOfSymmFunc, symmFuncType, a, b,
varyingNeigh=varyingNeigh)
self.xTest, self.yTest, _ = \
data.neighbourThreeBodySymmetry(self.function, self.testSize, \
neighbours, numberOfSymmFunc, symmFuncType, a, b,
varyingNeigh=varyingNeigh)
self.inputs = len(self.parameters)
if saveMetaFlag:
saveParametersFlag = True
elif method == 'lammpsSi' or method == 'lammpsSiO2':
if self.function == None:
print "method=lammps: Reading data from lammps simulations, including energies..."
else:
print "method=lammps: Reading data from lammps simulations, not including energies..."
if not dataFolder:
print "Path to folder where data is stored must be supplied"
self.samplesDir = dataFolder
# write content of README file to terminal
print
print "Content of lammps data file: "
command = "cat " + dataFolder + "README.txt"
os.system(command)
print
if method == 'lammpsSi':
print 'Training Si'
self.nTypes = nTypes
# save means of all symm funcs if shifting and saving
if saveFlag:
saveFolder = trainingDir
else:
saveFolder = ''
self.xTrain, self.yTrain, self.xTest, self.yTest, self.inputs, self.outputs, self.parameters, \
self.Ftrain, self.Ftest = \
lammps.SiTrainingData(dataFolder, symmFuncType, function=self.function, forces=forces, Behler=Behler,
klargerj=klargerj, tags=tags, normalize=normalize, shiftMean=shiftMean,
standardize=standardize, trainingDir=saveFolder)
else:
print 'Training SiO2'
self.atomType = atomType
self.nTypes = nTypes
self.xTrain, self.yTrain, self.xTest, self.yTest, self.inputs, self.outputs, self.parameters, \
self.elem2param = \
lammps.SiO2TrainingData(dataFolder, symmFuncType, atomType, forces=forces, nAtoms=nAtoms)
# set different sizes based on lammps data
self.trainSize = self.xTrain.shape[0]
self.testSize = self.xTest.shape[0]
############# EDIT EDIT EDIT change set ###############
#self.xTrain = self.xTrain[:300]
#self.trainSize = self.xTrain.shape[0]
if batch is None:
print
print "Doing offline learning"
batch = self.trainSize
elif batch > 1:
print
print "Doing online learning with", batch, "batches"
else:
print "Batch has to be 1 or above, exiting"
exit(1)
# set batch size, ensure that train size is a multiple of batch size
rest = self.trainSize % batch
if rest != 0:
self.trainSize -= rest
indicies = np.random.choice(self.trainSize, rest)
self.xTrain = np.delete(self.xTrain, indicies, axis=0)
self.yTrain = np.delete(self.yTrain, indicies, axis=0)
self.batchSize = batch
self.numberOfBatches = self.trainSize / batch
if saveMetaFlag:
saveParametersFlag = True
else:
print "Invalid data generation method chosen. Exiting..."
exit(1)
# print out sizes
print
print "##### Training parameters #####"
print "Training set size: ", self.trainSize
print "Test set size: ", self.testSize
print "Batch size: ", self.batchSize
print "Number of batches: ", self.numberOfBatches
print "Learning rate: ", self.learningRate
def constructNetwork(self, nLayers, nNodes, activation=tf.nn.sigmoid, \
wInit='normal', bInit='normal', stdDev=1.0, constantValue=0.1):
self.nLayers = nLayers
self.nNodes = nNodes
self.activation = activation
self.wInit = wInit
self.bInit = bInit
# print out...
print
print "##### network parameters #####"
print "Inputs: ", self.inputs
print "Outputs: ", self.outputs
print "Number of layers: ", nLayers
print "Number of nodes: ", nNodes
print "Activation function: ", activation.__name__
print "Weight initialization: ", wInit
print "Bias initialization: ", bInit
print "Setting up NN..."
# input placeholders
with tf.name_scope('input'):
self.x = tf.placeholder('float', [None, self.inputs], name='x-input')
self.y = tf.placeholder('float', [None, self.outputs], name='y-input')
self.neuralNetwork = nn.neuralNetwork(nNodes, nLayers, activation,
weightsInit=wInit, biasesInit=bInit,
stdDev=stdDev, inputs=self.inputs, outputs=self.outputs,
constantValue=constantValue)
self.makeNetwork = lambda data : self.neuralNetwork.model(self.x)
def train(self, numberOfEpochs):
trainSize = self.trainSize
batchSize = self.batchSize
testSize = self.testSize
xTrain = self.xTrain
yTrain = self.yTrain
xTest = self.xTest
yTest = self.yTest
numberOfBatches = self.numberOfBatches
x = self.x
y = self.y
nNodes = self.nNodes
nLayers = self.nLayers
learningRate = self.learningRate
testRMSELow = 1.0
epochLow = 0
timeElapsed = 0
# begin session
with tf.Session() as sess:
# pass data to network and receive output
prediction = self.makeNetwork(x)
with tf.name_scope('L2Norm'):
# HAVE CHANGED HERE!!!!
trainCost = tf.div( tf.nn.l2_loss( tf.subtract(prediction, y) ), batchSize, name='/trainCost')
testCost = tf.div( tf.nn.l2_loss( tf.subtract(prediction, y) ), testSize, name='/testCost')
tf.summary.scalar('L2Norm', trainCost/batchSize)
with tf.name_scope('MAD'):
MAD = tf.reduce_sum( tf.abs( tf.subtract(prediction, y) ) )
with tf.name_scope('train'):
trainStep = tf.train.AdamOptimizer(learning_rate=learningRate).minimize(trainCost)
with tf.name_scope('networkGradient'):
networkGradient = tf.gradients(self.neuralNetwork.allActivations[-1], x)
#with tf.name_scope('L2Force'):
# CFDATrain = tf.nn.l2_loss( tf.subtract(networkGradient, xTrain) )
# initialize variables or restore from file
saver = tf.train.Saver(keep_checkpoint_every_n_hours=1)
sess.run(tf.global_variables_initializer())
if loadFlag:
saver.restore(sess, loadFileName)
print 'Model %s restored' % loadFileName
# merge all the summaries and write them out to training directory
if summaryFlag:
merged = tf.merge_all_summaries()
train_writer = tf.train.SummaryWriter(summaryDir + '/train', sess.graph)
test_writer = tf.train.SummaryWriter(summaryDir + '/test')
# decide how often to print and store things
every = 1000/self.numberOfBatches
if loadFlag and (plotFlag or plotErrorFlag) and not saveFlag:
numberOfEpochs = -1
# train
print
print "##### Starting training session #####"
start = timer()
for epoch in xrange(numberOfEpochs+1):
# for shuffling the training set
indicies = np.random.choice(trainSize, trainSize, replace=False)
# offline learning
if batchSize == trainSize:
# pick whole set in random order
xBatch = xTrain[indicies]
yBatch = yTrain[indicies]
# train
sess.run(trainStep, feed_dict={x: xBatch, y: yBatch})
# online learning
else:
# loop through whole set, train each iteration
for b in xrange(numberOfBatches):
batch = indicies[b*batchSize:(b+1)*batchSize]
xBatch = xTrain[batch]
yBatch = yTrain[batch]
# train
sess.run(trainStep, feed_dict={x: xBatch, y: yBatch})
if summaryFlag:
if not epoch % every:
summary = sess.run(merged, feed_dict={x: xBatch, y: yBatch})
train_writer.add_summary(summary, epoch)
# calculate cost every every epoch
if not epoch % every or epoch == numberOfEpochs:
trainError, absErrorTrain = sess.run([trainCost, MAD], feed_dict={x: xBatch, y: yBatch})
testError, absErrorTest = sess.run([testCost, MAD], feed_dict={x: xTest, y: yTest})
trainRMSE = np.sqrt(2*trainError)
testRMSE = np.sqrt(2*testError)
if testRMSE < testRMSELow:
testRMSELow = testRMSE
epochLow = epoch
end = timer()
timeElapsed = end - start
print 'Cost/N train test at epoch %4d: TF: %g %g, RMSE: %g %g, MAD: %g %g' % \
( epoch, trainError, testError, \
trainRMSE, \
testRMSE, \
absErrorTrain/float(batchSize), \
absErrorTest/float(testSize) )
sys.stdout.flush()
if testRMSE / trainRMSE > 10:
print 'Overfitting is occuring, training ends'
break
if summaryFlag:
summary = sess.run(merged, feed_dict={x: xTest, y: yTest})
test_writer.add_summary(summary, epoch)
# if an argument is passed, save the graph variables ('w', 'b') and dump
# some info about the training so far to TrainingData/<this run>/meta.dat
if saveMetaFlag:
if epoch == 0:
saveEpochNumber = 0
with open(saveMetaName, 'w') as outFile:
outStr = '# epochs: %d train: %d, test: %d, batch: %d, nodes: %d, layers: %d \n' \
% (numberOfEpochs, trainSize, testSize, batchSize, nNodes, nLayers)
outStr += 'a: %1.1f, b: %1.1f, activation: %s, wInit: %s, bInit: %s, learnRate: %g, symm: %s' % \
(self.a, self.b, self.activation.__name__, self.wInit, self.bInit, self.learningRate, self.symmFuncType)
outFile.write(outStr + '\n')
outStr = 'Inputs: %d, outputs: %d, loaded: %s, sampled: %s \n' % \
(self.inputs, self.outputs, loadDir, self.samplesDir)
outFile.write(outStr)
outStr = '%d %g %g' % \
(epoch, trainRMSE, testRMSE)
outFile.write(outStr + '\n')
else:
if not epoch % every:
with open(saveMetaName, 'a') as outFile :
outStr = '%d %g %g' % \
(epoch, trainRMSE, testRMSE)
outFile.write(outStr + '\n')
if saveFlag or saveGraphProtoFlag:
if not epoch % every:
saveFileName = saveDirName + '/' 'ckpt'
saver.save(sess, saveFileName, global_step=epoch,
latest_filename="checkpoint_state")
# finish training if RMSE of test set is below tolerance
if testRMSE < self.RMSEtol:
print "Reached RMSE tolerance"
break
if numberOfEpochs == -1:
print sess.run(trainCost, feed_dict={x: xTrain[0].reshape([1,self.inputs]), y: yTrain[0].reshape([1,1])})
print xTrain[0]
print yTrain[0]
print '%.16g' % sess.run(prediction, feed_dict={x: xTrain[0].reshape([1,self.inputs])})
print sess.run(networkGradient, feed_dict={x: xTrain})
# elapsed time
end = timer();
print "Time elapsed: %g" % (end-start)
# write network to file when training is finished
if saveGraphFlag:
with open(saveGraphName, 'w') as outFile:
outStr = "%1d %1d %s %d %d" % (nLayers, nNodes, self.activation.__name__, \
self.inputs, self.outputs)
outFile.write(outStr + '\n')
size = len(self.neuralNetwork.allWeights)
for i in range(size):
weights = sess.run(self.neuralNetwork.allWeights[i])
if i < size-1:
for j in range(len(weights)):
for k in range(len(weights[0])):
outFile.write("%.16g" % weights[j][k])
outFile.write(" ")
outFile.write("\n")
else:
for j in range(len(weights[0])):
for k in range(len(weights)):
outFile.write("%.16g" % weights[k][j])
outFile.write(" ")
outFile.write("\n")
outFile.write("\n")
for biasVariable in self.neuralNetwork.allBiases:
biases = sess.run(biasVariable)
for j in range(len(biases)):
outFile.write("%.16g" % biases[j])
outFile.write(" ")
outFile.write("\n")
# save parameters to file
if saveParametersFlag:
parameters = self.parameters
numberOfParameters = len(parameters[0])
saveParametersName = trainingDir + '/' + 'parameters.dat'
if self.nTypes > 1:
with open(saveParametersName, 'w') as outFile:
outStr = "%d %d" % (len(parameters), self.atomType)
outFile.write(outStr + '\n')
# G2
for jtype in xrange(self.nTypes):
key = (self.atomType, jtype)
if key in self.elem2param:
interval = self.elem2param[(self.atomType,jtype)]
for s, p in enumerate(parameters[interval[0]:interval[1]], interval[0]):
for param in p:
outFile.write("%g " % param)
outFile.write("\n")
outFile.write("\n")
# G4/G5
for jtype in xrange(self.nTypes):
for ktype in xrange(self.nTypes):
key = (self.atomType, jtype, ktype)
if key in self.elem2param:
interval = self.elem2param[(self.atomType,jtype,ktype)]
for s, p in enumerate(parameters[interval[0]:interval[1]], interval[0]):
for param in p:
outFile.write("%g " % param)
outFile.write("\n")
outFile.write("\n")
else:
with open(saveParametersName, 'w') as outFile:
# write number of symmfuncs and number of unique parameters
outStr = "%d" % len(parameters)
outFile.write(outStr + '\n')
for symmFunc in range(len(parameters)):
for param in parameters[symmFunc]:
outFile.write("%g" % param)
if numberOfParameters > 1:
outFile.write(" ")
outFile.write("\n")
# freeze graph
if saveGraphProtoFlag:
tf.train.write_graph(sess.graph_def, trainingDir, 'graph.pb')
input_graph_path = trainingDir + '/graph.pb'
input_saver_def_path = ""
input_binary = False
input_checkpoint_path = saveFileName + '-' + str(saveEpochNumber-1)
output_node_names = "outputLayer/activation"
restore_op_name = "save/restore_all"
filename_tensor_name = "save/Const:0"
output_graph_path = trainingDir + '/frozen_graph.pb'
clear_devices = False
freeze_graph(input_graph_path, input_saver_def_path,
input_binary, input_checkpoint_path,
output_node_names, restore_op_name,
filename_tensor_name, output_graph_path,
clear_devices, "")
if saveFlag or saveGraphFlag:
self.outputFile.close()
# plot RMSE as function of epoch
if plotFlag:
if loadFlag and not saveFlag:
location = loadDir + '/meta.dat'
else:
location = saveMetaName
with open(location) as infile:
# skip headers
infile.readline(); infile.readline(); infile.readline()
# read RMSE of train and test
epoch = []; trainError = []; testError = [];
for line in infile:
words = line.split()
epoch.append(float(words[0]))
trainError.append(float(words[1]))
testError.append(float(words[2]))
plt.plot(epoch, trainError, 'b-', epoch, testError, 'g-')
plt.xlabel('Epoch')
plt.ylabel('RMSE')
plt.legend(['Training set', 'Test set'], prop={'size':20})
plt.axis([0, 5000, 0, max(trainError + testError)])
plt.tight_layout()
#plt.savefig('../Oppgaven/Figures/Implementation/overfitting.pdf')
plt.show()
# plot error of 1-dim function on interval [a,b]
if plotErrorFlag:
# make interval [a,b]
N = 1000
#interval = np.linspace(self.a, self.b, N)
interval = np.sort(self.xTest, axis=0)
# evaluate trained network on this interval
energiesNN = sess.run(prediction, feed_dict={x: interval})
energiesLJ = self.function(interval)
# read file where NN is evaluated and differentiated in C++
Cfile = 'Tests/TrainLennardJones/energyAndDerivativeC.dat'
energiesC = []; derivativesC = []
with open(Cfile, 'r') as infile:
for line in infile:
words = line.split()
energiesC.append(float(words[0]))
derivativesC.append(float(words[1]))
# plot NN energy and LJ energy together
plt.plot(interval, energiesLJ, 'b-', interval, energiesNN, 'g-')
plt.legend(['LJ energy', 'NN energy'])
plt.show()
# plot energy error
energyError = energiesLJ - energiesNN
print "RMSE: ", np.sqrt(np.sum(energyError**2)/N)
print "RMSE relative: ", np.sqrt(np.sum((energyError/energiesLJ)**2)/N)
print "Min E:", np.min(np.abs(energiesLJ))
plt.figure()
plt.plot(interval, energyError/energiesLJ)
plt.xlabel(r'$r_{ij} \, [\mathrm{\AA{}}]$')
plt.ylabel(r'$\mathrm{Absolute} \; \mathrm{error} \, [\mathrm{eV}]$')
plt.legend(['$E_{\mathrm{LJ}} - E_{\mathrm{NN}}$'], prop={'size':20})
plt.tight_layout()
#plt.savefig('../Oppgaven/Figures/Implementation/LJError.pdf')
#plt.show()
# plot NN energy in Python and C++ together to check that they are the same
#plt.figure()
#plt.plot(energiesC, energiesNN)
#plt.xlabel('C++ manual NN energy')
#plt.ylabel('Python TF API NN energy')
#plt.show()
##### derivatives #####
# calculate derivative LJ and derivative of NN in Python
derivativeLJ = self.functionDerivative(interval)
derivativeNN = sess.run(networkGradient, feed_dict={x: interval} )[0]
#derivativeNN = derivativeNN
# plot NN derivative and LJ derivative together
plt.figure()
plt.plot(interval, derivativeLJ, 'g-', interval, derivativeNN, 'b-')
plt.legend(['Derivative LJ', 'Derivative NN'])
#plt.show()
# plot derivative error
derivativeError = derivativeLJ - derivativeNN
print 'Derivative RMSE', np.sqrt(np.sum(derivativeError**2)/N)
plt.figure()
plt.plot(interval, derivativeError)
plt.xlabel(r'$r_{ij} \, [\mathrm{\AA{}}]$')
plt.ylabel(r'$\mathrm{Absolute} \; \mathrm{error} \, [\mathrm{eV}/\mathrm{\AA{}}]$')
plt.legend(['$dE_{\mathrm{LJ}}/dr_{ij} - dE_{\mathrm{NN}}/dr_{ij}$'], prop={'size':20}, loc=4)
plt.tight_layout()
plt.savefig('../Oppgaven/Figures/Implementation/LJErrorDerivative.pdf')
#plt.show()
# plot NN derivative in Python and C++ together to check that they are the same
#plt.figure()
#plt.plot(derivativesC, derivativeNN)
#plt.xlabel('C++ manual NN derivative')
#plt.ylabel('Python TF API derivative')
plt.show()
return testRMSELow, epochLow, timeElapsed
if __name__ == '__main__':
pass