-
Notifications
You must be signed in to change notification settings - Fork 1
/
surrogate_model.py
489 lines (362 loc) · 17.6 KB
/
surrogate_model.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
import random
import h5py
import copy
import numpy as np
import tensorflow as tf
from tensorflow import keras
from abc import ABC, abstractmethod
from tensorflow.keras.models import load_model
from lume_model.models import SurrogateModel
from lume_model.utils import load_variables
from lume_model.variables import (
ScalarInputVariable,
ScalarOutputVariable,
ImageInputVariable,
ImageOutputVariable,
)
class BaseKerasModel(SurrogateModel, ABC):
def __init__(self, *, model_file, input_variables, output_variables):
# Save init
self.model_file = model_file
self.input_variables = input_variables
self.output_variables = output_variables
# Load attributes from file
with h5py.File(self.model_file, "r") as h5:
self.attrs = dict(h5.attrs)
# Load model etc.
self.json_string = self.attrs["JSON"]
# load model in thread safe manner
self.thread_graph = tf.Graph()
with self.thread_graph.as_default():
self.model = tf.keras.models.model_from_json(
self.json_string.decode("utf-8")
)
self.model.load_weights(model_file)
# SUBCLASSING THE SURROGATE MODEL SUBCLASS ENFORCES THAT THIS IS DEFINED
def evaluate(self, input_variables):
input_dictionary = {} # maps variable_name -> value
# convert list of input variables to dictionary
input_variables = {input_variable.name: input_variable for input_variable in input_variables}
# prepare input dictionary, accounting for any missing values using defaults
for variable_name in self.input_ordering:
if variable_name in input_variables:
if input_variables[variable_name].value is not None:
input_dictionary[variable_name] = input_variables[variable_name].value
else:
input_dictionary[variable_name] = input_variables[variable_name].default
else:
input_dictionary[variable_name] = self.input_variables[variable_name].default
# MUST IMPLEMENT A format_input METHOD TO CONVERT FROM DICT -> MODEL INPUT
formatted_input = self.format_input(input_dictionary)
# call prediction in threadsafe manner
with self.thread_graph.as_default():
model_output = self.model.predict(formatted_input)
# MUST IMPLEMENT AN OUTPUT -> DICT METHOD
output = self.parse_output(model_output)
# PREPARE OUTPUTS WILL FORMAT RETURN VARIABLES (DICT-> VARIABLES)
return self.prepare_outputs(output)
def random_evaluate(self):
random_input = copy.deepcopy(self.input_variables)
for variable in self.input_ordering:
if self.input_variables[variable].variable_type == "scalar":
random_input[variable].value = np.random.uniform(
self.input_variables[variable].value_range[0],
self.input_variables[variable].value_range[1],
)
else:
random_input[variable].value = self.input_variables[variable].default
pred = self.evaluate(list(random_input.values()))
return pred
def prepare_outputs(self, predicted_output):
"""
Prepares the model outputs to be served so no additional
manipulation happens in the OnlineSurrogateModel class
Args:
model_outputs (dict): Dictionary of output variables to np.ndarrays of outputs
Returns:
dict: Dictionary of output variables to respective scalars
"""
for variable in self.output_variables.values():
if variable.variable_type == "scalar":
self.output_variables[variable.name].value = predicted_output[
variable.name
]
elif variable.variable_type == "image":
self.output_variables[variable.name].value = predicted_output[
variable.name
].reshape(variable.shape)
# update limits
if self.output_variables[variable.name].x_min_variable:
self.output_variables[variable.name].x_min = predicted_output[
self.output_variables[variable.name].x_min_variable
]
if self.output_variables[variable.name].x_max_variable:
self.output_variables[variable.name].x_max = predicted_output[
self.output_variables[variable.name].x_max_variable
]
if self.output_variables[variable.name].y_min_variable:
self.output_variables[variable.name].y_min = predicted_output[
self.output_variables[variable.name].y_min_variable
]
if self.output_variables[variable.name].y_max_variable:
self.output_variables[variable.name].y_max = predicted_output[
self.output_variables[variable.name].y_max_variable
]
return list(self.output_variables.values())
@abstractmethod
def format_input(self, input_dictionary):
# MUST IMPLEMENT A METHOD TO CONVERT INPUT DICTIONARY TO MODEL INPUT
pass
@abstractmethod
def parse_output(self, model_output):
# MUST IMPLEMENT A METHOD TO CONVERT MODEL OUTPUT TO A DICTIONARY OF VARIABLE NAME -> VALUE
pass
class ScaledModel(BaseKerasModel):
"""
Example Usage:
Load model and use a dictionary of inputs to evaluate the NN.
"""
def __init__(self, *, model_file=None, input_variables=None, output_variables=None):
super().__init__(model_file=model_file, input_variables=input_variables, output_variables=output_variables)
# Collect attributes
self.input_ordering = self.attrs["input_ordering"]
self.output_ordering = self.attrs["output_ordering"]
self.input_scales = {
self.input_ordering[i]: self.attrs["input_scales"][i]
for i in range(len(self.input_ordering))
}
self.output_scales = {
self.output_ordering[i]: self.attrs["output_scales"][i]
for i in range(len(self.output_ordering))
}
self.input_offsets = {
self.input_ordering[i]: self.attrs["input_offsets"][i]
for i in range(len(self.input_ordering))
}
self.output_offsets = {
self.output_ordering[i]: self.attrs["output_offsets"][i]
for i in range(len(self.output_ordering))
}
## Set basic values needed for input and output scaling
self.model_value_max = self.attrs["upper"]
self.model_value_min = self.attrs["lower"]
def scale_inputs(self, input_values):
data_scaled = {}
for i, input_variable in enumerate(self.input_ordering):
if self.input_variables[input_variable].variable_type == "scalar":
data_scaled[input_variable] = self.model_value_min + (
(input_values[input_variable] - self.input_offsets[input_variable])
* (self.model_value_max - self.model_value_min)
/ self.input_scales[input_variable]
)
elif self.input_variables[input_variable].variable_type == "image":
data_scaled[input_variable] = (self.model_value_max - self.model_value_min) * (
(input_values[input_variable] / self.input_scales[input_variable])
- self.input_offsets[input_variable]
)
data_scaled[input_variable] = data_scaled[input_variable].reshape(
self.input_variables[input_variable].shape
)
return data_scaled
def unscale_outputs(self, output_values):
unscaled_outputs = {}
for output_variable in output_values:
# Scale scalar variable
if self.output_variables[output_variable].variable_type == "scalar":
unscaled_outputs[output_variable] = (
output_values[output_variable]
* self.output_scales[output_variable]
/ (self.model_value_max - self.model_value_min)
+ self.output_offsets[output_variable]
)
# Scale image variable
elif self.output_variables[output_variable].variable_type == "image":
unscaled_image = (
output_values[output_variable]
+ self.output_offsets[output_variable]
) * self.output_scales[output_variable]
# Reshape image
unscaled_outputs[output_variable] = unscaled_image.reshape(
self.output_variables[output_variable].shape
)
return unscaled_outputs
class Model(ScaledModel):
def format_input(self, input_dictionary):
# scale inputs
input_dictionary = self.scale_inputs(input_dictionary)
#image = input_dictionary["input_image"].reshape(1, 50, 50, 1)
scalar_inputs = np.array([
input_dictionary['distgen:r_dist:sigma_xy:value'],
input_dictionary['distgen:t_dist:length:value'],
input_dictionary['distgen:total_charge:value'],
input_dictionary['SOL1:solenoid_field_scale'],
input_dictionary['CQ01:b1_gradient'],
input_dictionary['SQ01:b1_gradient'],
input_dictionary['L0A_phase:dtheta0_deg'],
input_dictionary['L0A_scale:voltage'],
input_dictionary['end_mean_z']
]).reshape((1,9))
model_input = [scalar_inputs]
return model_input
def parse_output(self, model_output):
parsed_output = {}
image_output = model_output[0][0]
parsed_output["out_xmin"] = image_output[0]
parsed_output["out_xmax"] = image_output[1]
parsed_output["out_ymin"] = image_output[2]
parsed_output["out_ymax"] = image_output[3]
parsed_output["x:y"] = image_output[4:].reshape((50,50))
parsed_output.update(dict(zip(self.output_variables.keys(), model_output[1][0].T)))
# unscale
parsed_output = self.unscale_outputs(parsed_output)
return parsed_output
# to lume-keras
class ScaleLayer(keras.layers.Layer):
trainable = False
def __init__(self, offset, scale, lower,upper, **kwargs):
super(ScaleLayer, self).__init__(**kwargs)
self.scale = scale
self.offset = offset
self.lower = lower
self.upper = upper
def call(self, inputs):
return self.lower+((inputs-self.offset)*(self.upper-self.lower)/self.scale)
# MUST OVERRIDE IN ORDER TO SAVE + LOAD W/KERAS
# STORE SALE, etc.
def get_config(self):
return {'scale': self.scale,'offset': self.offset,'lower': self.lower,'upper': self.upper}
# to lume-keras
class UnScaleLayer(keras.layers.Layer):
trainable = False
def __init__(self, offset, scale, lower,upper, **kwargs):
super(UnScaleLayer, self).__init__(**kwargs)
self.scale = scale
self.offset = offset
self.lower = lower
self.upper = upper
def call(self, inputs):
return (((inputs-self.lower)*self.scale)/(self.upper-self.lower)) + self.offset
# MUST OVERRIDE IN ORDER TO SAVE + LOAD W/KERAS
# STORE SALE, etc.
def get_config(self):
return {'scale': self.scale,'offset': self.offset,'lower': self.lower,'upper': self.upper}
# To lume-keras
class UnScaleImg(keras.layers.Layer):
trainable = False
def __init__(self, img_offset, img_scale, **kwargs):
super(UnScaleImg, self).__init__(**kwargs)
self.img_scale = img_scale
self.img_offset = img_offset
def call(self, inputs):
return (inputs+self.img_offset)*self.img_scale
# MUST OVERRIDE IN ORDER TO SAVE + LOAD W/KERAS
# STORE SALE, etc.
def get_config(self):
return {'img_scale': self.img_scale,'img_offset': self.img_offset}
class CustomBaseKerasModel(SurrogateModel, ABC):
def __init__(self, *, model_file, input_variables, output_variables):
# Save init
self.model_file = model_file
self.input_variables = input_variables
self.output_variables = output_variables
# Load attributes from file
with h5py.File(self.model_file, "r") as h5:
self.attrs = dict(h5.attrs)
# load model in thread safe manner
self.thread_graph = tf.Graph()
with self.thread_graph.as_default():
self.model = load_model(model_file ,custom_objects={'ScaleLayer': ScaleLayer,'UnScaleLayer': UnScaleLayer,'UnScaleImg': UnScaleImg})
# SUBCLASSING THE SURROGATE MODEL SUBCLASS ENFORCES THAT THIS IS DEFINED
def evaluate(self, input_variables):
# convert list of input variables to dictionary
input_dictionary = {input_variable.name: input_variable.value for input_variable in input_variables}
# MUST IMPLEMENT A format_input METHOD TO CONVERT FROM DICT -> MODEL INPUT
formatted_input = self.format_input(input_dictionary)
# call prediction in threadsafe manner
with self.thread_graph.as_default():
model_output = self.model.predict(formatted_input)
# MUST IMPLEMENT AN OUTPUT -> DICT METHOD
output = self.parse_output(model_output)
# PREPARE OUTPUTS WILL FORMAT RETURN VARIABLES (DICT-> VARIABLES)
return self.prepare_outputs(output)
def random_evaluate(self):
random_input = copy.deepcopy(self.input_variables)
for variable in self.input_variables:
if self.input_variables[variable].variable_type == "scalar":
random_input[variable].value = np.random.uniform(
self.input_variables[variable].value_range[0],
self.input_variables[variable].value_range[1],
)
else:
random_input[variable].value = self.input_variables[variable].default
pred = self.evaluate(list(random_input.values()))
return pred
def prepare_outputs(self, predicted_output):
"""
Prepares the model outputs to be served so no additional
manipulation happens in the OnlineSurrogateModel class
Args:
model_outputs (dict): Dictionary of output variables to np.ndarrays of outputs
Returns:
dict: Dictionary of output variables to respective scalars
"""
for variable in self.output_variables.values():
if variable.variable_type == "scalar":
self.output_variables[variable.name].value = predicted_output[
variable.name
]
elif variable.variable_type == "image":
self.output_variables[variable.name].value = predicted_output[
variable.name
].reshape(variable.shape)
# update limits
if self.output_variables[variable.name].x_min_variable:
self.output_variables[variable.name].x_min = predicted_output[
self.output_variables[variable.name].x_min_variable
]
if self.output_variables[variable.name].x_max_variable:
self.output_variables[variable.name].x_max = predicted_output[
self.output_variables[variable.name].x_max_variable
]
if self.output_variables[variable.name].y_min_variable:
self.output_variables[variable.name].y_min = predicted_output[
self.output_variables[variable.name].y_min_variable
]
if self.output_variables[variable.name].y_max_variable:
self.output_variables[variable.name].y_max = predicted_output[
self.output_variables[variable.name].y_max_variable
]
return list(self.output_variables.values())
@abstractmethod
def format_input(self, input_dictionary):
# MUST IMPLEMENT A METHOD TO CONVERT INPUT DICTIONARY TO MODEL INPUT
pass
@abstractmethod
def parse_output(self, model_output):
# MUST IMPLEMENT A METHOD TO CONVERT MODEL OUTPUT TO A DICTIONARY OF VARIABLE NAME -> VALUE
pass
class AutoScaledModel(CustomBaseKerasModel):
def format_input(self, input_dictionary):
scalar_inputs = np.array([
input_dictionary['distgen:r_dist:sigma_xy:value'],
input_dictionary['distgen:t_dist:length:value'],
input_dictionary['distgen:total_charge:value'],
input_dictionary['SOL1:solenoid_field_scale'],
input_dictionary['CQ01:b1_gradient'],
input_dictionary['SQ01:b1_gradient'],
input_dictionary['L0A_phase:dtheta0_deg'],
input_dictionary['L0A_scale:voltage'],
input_dictionary['end_mean_z']
]).reshape((1,9))
model_input = [scalar_inputs]
return model_input
def parse_output(self, model_output):
parsed_output = {}
parsed_output["x:y"] = model_output[0][0].reshape((50,50))
# NTND array attributes MUST BE FLOAT 64!!!! np.float() should be moved to lume-epics
parsed_output["out_xmin"] = np.float64(model_output[1][0][0])
parsed_output["out_xmax"] = np.float64(model_output[1][0][1])
parsed_output["out_ymin"] = np.float64(model_output[1][0][2])
parsed_output["out_ymax"] = np.float64(model_output[1][0][3])
parsed_output.update(dict(zip(self.output_variables.keys(), model_output[2][0].T)))
return parsed_output