-
Notifications
You must be signed in to change notification settings - Fork 9
/
model.py
executable file
·445 lines (376 loc) · 15.2 KB
/
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
#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Mon May 25 17:58:58 2020
@author: hsjomaa
"""
import tensorflow as tf
import pandas as pd
import json
import time
import os
from modules import FunctionF,FunctionH,FunctionG,PoolF,PoolG,PoolH
tf.random.set_seed(0)
class Model(object):
'''
Model
'''
def __init__(self,configuration,rootdir,for_eval=False,fine_tuning=False):
# data shape
self.batch_size = configuration['batch_size']
self.split = configuration['split']
self.searchspace = configuration['searchspace']
self.split = configuration['split']
self.nonlinearity_d2v = configuration['nonlinearity_d2v']
# Function F
self.units_f = configuration['units_f']
self.nhidden_f = configuration['nhidden_f']
self.architecture_f = configuration['architecture_f']
self.resblocks_f = configuration['resblocks_f']
# Function G
self.units_g = configuration['units_g']
self.nhidden_g = configuration['nhidden_g']
self.architecture_g = configuration['architecture_g']
# Function H
self.units_h = configuration['units_h']
self.nhidden_h = configuration['nhidden_h']
self.architecture_h = configuration['architecture_h']
self.resblocks_h = configuration['resblocks_h']
self.delta = configuration['delta']
self.gamma = configuration['gamma']
self.config_num = configuration["number"]
self.model =self.dataset2vecmodel(trainable=True)
self.trainable_count = int(sum([tf.keras.backend.count_params(p) for p in self.model.trainable_weights]))
configuration["trainable"] = self.trainable_count
# tracking
self.metrickeys = ['similarityloss','time',"roc"]
self.with_csv = True
# create a location if not evaluation model
if not for_eval:
self._create_metrics()
self.directory = self._create_dir(rootdir)
self._save_configuration(configuration)
def _create_dir(self,rootdir):
import datetime
# create directory
directory = os.path.join(rootdir, "checkpoints",f"searchspace-{self.searchspace}",f"split-{self.split}","dataset2vec",\
"vanilla",f"configuration-{self.config_num}",datetime.datetime.now().strftime('%Y-%m-%d-%H-%M-%S-%f'))
os.makedirs(directory)
return directory
@tf.function
def similarity(self,layer,positive_pair):
'''
Parameters
----------
layer : tf.Tensor
Extracted metafeatures; shape = [None,3,units_hh].
positive_pair : bool
indicator of similarity expected (between positive pair or negative pair).
Returns
-------
tf.Tensor
Similarity between metafeatures.
'''
# check if requires reshape
return tf.exp(-self.gamma*self.distance(layer,positive_pair))
@tf.function
def distance(self,layer,positive_pair):
'''
Return the cosine similarity between dataset metafeatures
Parameters
----------
layer : tf.Tensor
metafeatures.
Returns
-------
cos : tf.Tensor
Cosine similarity.
'''
# reshape metafeatures
layer = tf.reshape(layer,shape=(self.batch_size,3,self.units_h))
# average metafeatures of positive data
pos = tf.reduce_mean(layer[:,:2],axis=1)[:,None] if not positive_pair else layer[:,0][:,None]
# metafeatures of negative data
neg = layer[:,-1][:,None] if not positive_pair else layer[:,1][:,None]
# concatenate with negative meta-features
layer = tf.keras.layers.concatenate([pos,neg],axis=1)
dist = tf.norm(layer[:,0]-layer[:,1],axis=1)
return dist
@tf.function
def similarityloss(self,target_y,predicted_y):
'''
Compute the similarity log_loss between positive-pair metafeatures
and negative-pair metafeatures.
Parameters
----------
target_y : tf.Tensor
Similarity indicator.
predicted_y : tf.Tensor
Extracted metafeatures; shape = [None,3,units_hh].
Returns
-------
tf.Tensor
'''
negative_prob = self.similarity(predicted_y,positive_pair=False)
positive_prob = self.similarity(predicted_y,positive_pair=True)
logits = tf.concat([positive_prob,negative_prob],axis=0)
# create weight
similarityweight = tf.concat([tf.ones(shape=self.batch_size),self.delta*(tf.ones(shape=self.batch_size))],axis=0)
return tf.compat.v1.losses.log_loss(labels=target_y,predictions=logits, weights=similarityweight)
@tf.function
def loss(self,target_y,output,training=True):
'''
Compute the total loss of the network.
Parameters
----------
target_y : tuple(tf.Tensor)
(Similarty Indicator,targetresponse)
output : tuple
Output of the model.
training : bool
important to specify keys of metrics dict.
Returns
-------
loss : tf.Tensor
'''
# add prefix
prefix = '' if training else 'vld'
# parse target output
similaritytarget = target_y['similaritytarget']
# create metrics placeholder
metrics = {}
# split the output
metafeatures = output['metafeatures']
losses = {}
l = None
# Compute Similarity Loss
losses.update({'similarity':self.similarityloss(similaritytarget,predicted_y=metafeatures)})
l = losses['similarity']
metrics.update({f'{prefix}similarityloss':losses['similarity']})
return l,metrics
def train_step(self,x,y,optimizer,clip=True,no_metrics=False):
starttime = time.time()
with tf.GradientTape() as tape:
# training=True is only needed if there are layers with different
# behavior during training versus inference (e.g. Dropout).
output = self.model(x, training=True)
loss,metrics = self.loss(y, output)
gradients = tape.gradient(loss, self.model.trainable_variables)
# clip gradients
if clip:
gradients = [tf.clip_by_value(t,clip_value_min=-0.5,clip_value_max=0.5) for t in gradients]
optimizer.apply_gradients(zip(gradients, self.model.trainable_variables ))
# add time required to run backprop
metrics.update({'time':tf.constant(time.time()-starttime)})
# update metrics
if no_metrics:
pass
else:
self._update_metrics(metrics)
# return metrics
return metrics
def update_tracker(self,training,metrics=None):
'''
update csv tracker
Parameters
----------
training : bool
which tracker to update.
metrics : dict
dictionary of tf.keras.metrics, default None
Returns
-------
None.
'''
# use existing metric or passed down ?
metrics = self.metrics if metrics is None else metrics
key = 'train' if training else 'valid'
# iterate over metrics
for metric,_ in metrics.items():
# update cell in dataframe
value = self.metrics[metric]
if metric != 'time':
# check if passed down
if hasattr(value,'result'):
value = value.result().numpy()
elif hasattr(value,'numpy'):
value = value.numpy()
else:
value = value
self.csv[key].at[self.update_counter[key],metric] = value
else:
self.csv[key].at[self.update_counter[key],metric] = self.metrics[metric].result().numpy()
# update key counter
self.update_counter[key] += 1
def _update_metrics(self,metrics):
'''
Update metrics trackers
Parameters
----------
metrics : dict
Returns
-------
None.
'''
# iterate over metrics
for metric,value in metrics.items():
# update dictionaries
# check if metric in keys
if metric in list(self.metrics.keys()):
# check if metric is a tf.keras function
if hasattr(self.metrics[metric],'call'):
# update metric
self.metrics[metric](value)
# if the metric is not
else:
self.metrics[metric] = value
else:
# add dictionary
self.metrics.update({metric:value})
def reset_states(self):
'''
Reset tracking metrics
Returns
-------
None.
'''
for metric in self.metrics.keys():
# check if metric is tf.keras function
if hasattr(self.metrics[metric],'reset_states'):
# do not reset time
if metric != 'time':
self.metrics[metric].reset_states()
else:
self.metrics[metric] = None
def dump(self):
'''
Save csv progress
'''
for key in ['train','valid']:
self.csv[key].to_csv(f'{self.directory}/{key}-progress.csv')
def _create_metrics(self):
'''
Create tracking metrics
Returns
-------
None.
'''
# create epoch counter
self.update_counter = {'train':0,'valid':0}
# create empty dictionary
self.metrics = {}
# fill dictionary with keys and values
[self.metrics.update({_:tf.keras.metrics.Mean(name=_)}) for _ in self.metrickeys if _ !='time']
# fix time metrics
self.metrics['time'] = tf.keras.metrics.Sum(name='time')
# check if csv required
if self.with_csv:
# create csv dictionary
self.csv = {}
# add training csv tracker
self.csv.update({'train': pd.DataFrame(data=None,columns=[_ for _ in self.metrickeys if 'vld' not in _])})
# add training csv tracker
self.csv.update({'valid': pd.DataFrame(data=None,columns=[_ for _ in self.metrickeys if 'vld' in _])})
def report(self):
template = 'Similarity: {:.5f}, ROC: {:.5f}, Time: {:.2f} s '
print(template.format(self.metrics['similarityloss'].result(),
self.metrics['roc'].result(),
self.metrics['time'].result()))
def _save_configuration(self,configuration):
configuration.update({"savedir":self.directory})
filepath = os.path.join(self.directory,"configuration.txt")
with open(filepath, 'w') as json_file:
json.dump(configuration, json_file)
def save_weights(self,iteration=None):
'''
Save weights of model with provided description
Parameters
----------
description: str
name of weights to save.
Returns
-------
None.
'''
# define filepath
iteration = f"-{iteration}" if iteration is not None else ''
filepath = os.path.join(self.directory,f"iteration{iteration}","weights")
os.makedirs(filepath,exist_ok=True)
# save internal model weights
self.model.save_weights(filepath=os.path.join(filepath,"weights"))
def set_weights(self,weights=None):
'''
Update the weights of the internal model with backend model
weights or with provided weights.
Parameters
----------
weights : List[tf.Variable], optional
Weights of the trainable variables. The default is None.
'''
self.model.set_weights(weights=weights)
def get_weights(self,internal=True):
'''
Return weights of the (internal) model
Parameters
----------
internal : bool, optional
indicator of type of model for which we want
to get weights. The default is True.
Returns
-------
weights : list(tf.Tensor)
'''
# get weights
weights = self.model.get_weights()
return weights
def getmetafeatures(self,x):
output = self.model(x,training=False)
layer = PoolH(self.batch_size,self.units_h)(output['metafeatures'],ignore_negative=True)
return layer
# @tf.function
def predict(self,x,y):
'''
Return the distribution of the target task
Parameters
----------
x : tuple(tf.Tensor)
input.
y : tuple(tf.Tensor)
output.
Returns
-------
y_mean : TYPE
DESCRIPTION.
y_logvar : TYPE
DESCRIPTION.
'''
# predict
output = self.model(x,training=False)
phi = output['metafeatures']
posprob = self.similarity(phi,positive_pair=True)
negprob = self.similarity(phi,positive_pair=False)
proba = tf.concat([posprob,negprob],axis=0)
return proba,y["similaritytarget"]
def dataset2vecmodel(self,trainable):
# input two dataset2vec shape = [None,2], i.e. flattened tabular batch
x = tf.keras.Input(shape=(2),dtype=tf.float32)
# Number of sampled classes from triplets
nclasses = tf.keras.Input(shape=(self.batch_size*3),dtype=tf.int32,batch_size=1)
# Number of sampled features from triplets
nfeature = tf.keras.Input(shape=(self.batch_size*3),dtype=tf.int32,batch_size=1)
# Number of sampled instances from triplets
ninstanc = tf.keras.Input(shape=(self.batch_size*3),dtype=tf.int32,batch_size=1)
# Encode the predictor target relationship across all instances
layer = FunctionF(units = self.units_f,nhidden = self.nhidden_f,nonlinearity = self.nonlinearity_d2v,architecture=self.architecture_f,resblocks=self.resblocks_f,trainable=trainable)(x)
# Average over instances
layer = PoolF(units=self.units_f)(layer,nclasses[0],nfeature[0],ninstanc[0])
# Encode the interaction between features and classes across the latent space
layer = FunctionG(units = self.units_g,nhidden = self.nhidden_g,nonlinearity = self.nonlinearity_d2v,architecture = self.architecture_g,trainable=trainable)(layer)
# Average across all instances
layer = PoolG(units=self.units_g)(layer,nclasses[0],nfeature[0])
# Extract the metafeatures
metafeatures = FunctionH(units = self.units_h,nhidden = self.nhidden_h, nonlinearity = self.nonlinearity_d2v,architecture=self.architecture_h,trainable=trainable,resblocks=self.resblocks_h)(layer)
# define hierarchical dataset representation model
output = {'metafeatures':metafeatures}
dataset2vec = tf.keras.Model(inputs=[x,nclasses,nfeature,ninstanc], outputs=output)
return dataset2vec