-
Notifications
You must be signed in to change notification settings - Fork 1
/
sps.py
328 lines (291 loc) · 12.7 KB
/
sps.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
from utils.app_sps import parse_args
import sys
import os
import tensorflow as tf
from utils.misc import Unbuffered, print_arguments, invoke_with_log
from models.base_models_mtl import convNet2SPS, convNet2CrossStitch, norm_conv3_crossstitch, norm_conv3_sps
import shared
from evaluation import Evaluator, EvaluationCallback
import shutil
from models.tensor_trace_norm import TensorTraceNorm
import time
from utils.deap import load_deap_data, select_channels_deap
import json
class SPS:
def __init__(self, args, outpath):
self.args = args
self.data = args.dataset
self.data_path = self._get_opportunity_data_path()
self.out_path = outpath
def _get_opportunity_data_path(self):
if self.data == 'opportunity':
nbSensors = self.args.opportunity_num_sensors
if nbSensors == 107:
pathExtension = os.path.join(
'all_sensors', str(self.args.opportunity_time_window))
else:
pathExtension = f'{nbSensors}_highest_var_sensors'
channels = self.args.labels
dataPath = os.path.join(self.args.opportunity_path, pathExtension)
return dataPath
elif self.data == 'deap':
return args.deap_path
def build_model_sps(self):
print(
f"Building model of type {self.args.model} for {len(self.label_names)} tasks")
if self.args.model == "sps-opportunity":
model, ttn_layers = convNet2SPS(
input_shape=self.input_shape,
label_names=self.label_names,
nbClasses=self.num_classes,
args=self.args)
self.model = model
self.ttn_layers = ttn_layers
elif self.args.model == 'sps-deap':
model, ttn_layers = norm_conv3_sps(
input_shape=self.input_shape,
label_names=self.label_names,
nb_classes=self.num_classes
)
self.model = model
self.ttn_layers = ttn_layers
elif self.args.model == 'cross-stitch':
if self.args.csmodel == 'normConv3':
self.model = norm_conv3_crossstitch(
input_shape=self.input_shape,
nb_classes=self.num_classes,
label_names=self.label_names
)
else:
self.model = convNet2CrossStitch(
input_shape=self.input_shape,
label_names=self.label_names,
nbClasses=self.num_classes,
args=self.args)
else:
raise AttributeError()
self.model.summary()
jsonname = os.path.join(
self.out_path, f"{self.args.model}_{self.args.tag}.json")
with open(jsonname, "w") as jf:
jf.write(self.model.to_json(indent=2))
model_plot = os.path.join(
self.out_path, f"{self.args.model}_{self.args.tag}.png")
tf.keras.utils.plot_model(self.model, to_file=model_plot,
show_shapes=True, show_layer_names=True,
dpi=320)
print(f"Saved model img to {model_plot} and json to {jsonname}")
def set_data(self, train_data, test_data, val_data,
input_shape, label_names, num_classes, deap_config):
self.train_data = train_data
self.test_data = test_data
self.val_data = val_data
self.input_shape = input_shape
self.label_names = label_names
self.num_classes = num_classes
self.deap_config = deap_config
if self.data == 'deap':
deap_available = self.deap_config["files"]["available"]
train_available = deap_available["train"]
validation_available = deap_available["test"]
train_batches = int(train_available / args.batch)
validation_batches = int(validation_available / args.batch)
print(
f"DEAP: training on {train_available} samples, {train_batches} batches")
print(
f" testing on {validation_available} samples, {validation_batches} batches")
self.args.steps = train_batches
self.args.validation_steps = validation_batches
elif self.data == 'opportunity':
self.args.validation_steps = None
def generate_callbacks(self):
callbacks = []
tbpath = os.path.join(self.out_path, "tensorboard")
symtbpath = os.path.join(args.output, "tensorboard", args.tag)
if not os.path.exists(tbpath):
os.makedirs(tbpath)
if not os.path.exists(symtbpath):
os.symlink(tbpath, symtbpath)
print(f"Symlinked {tbpath} -> {symtbpath}")
log_files_list = os.listdir(tbpath)
if log_files_list != []:
for fn in log_files_list:
print(f"Deleting {os.path.join(tbpath, fn)}")
shutil.rmtree(os.path.join(tbpath, fn))
checkpath = os.path.join(self.out_path, 'checkpoint/')
if not os.path.exists(checkpath):
os.makedirs(checkpath)
tb_callback = tf.keras.callbacks.TensorBoard(log_dir=tbpath,
update_freq='epoch',
write_graph=True,
write_images=True)
callbacks.append(tb_callback)
check_name = os.path.join(checkpath, f'{args.model}_{args.tag}.hdf5')
if self.data == 'opportunity':
monitorname = f"out_{self.label_names[0]}_fmeasure"
if len(self.label_names) == 1:
monitorname = 'fmeasure'
elif self.data == 'deap':
monitorname = f"val_out_{self.label_names[0]}_accuracy"
check_callback = tf.keras.callbacks.\
ModelCheckpoint(check_name,
monitor=monitorname,
save_best_only=True,
mode='max',
save_freq='epoch',
save_weights_only=False)
callbacks.append(check_callback)
if self.data == 'opportunity':
evaluator = Evaluator(self.label_names)
eval_dir = os.path.join(outpath, 'evaluation')
if not os.path.isdir(eval_dir):
os.makedirs(eval_dir)
eval_callback = EvaluationCallback(
self.val_data,
self.label_names,
self.num_classes,
eval_dir)
callbacks.append(eval_callback)
return callbacks
def loss_trace_norm(self, ttn_layers, base_fun, ttn_weight=None):
def loss(y_true, y_pred):
def _ttn(layer_idx, ttn_layer):
weights = [self.model.get_layer(
name=conv_name).kernel for conv_name in ttn_layer]
shapeX = weights[0].get_shape().as_list()
dimX = len(shapeX)
# weights = [l.weights[0] for l in layers]
stack = tf.stack(weights, axis=dimX)
t = TensorTraceNorm(stack)
tr = tf.reduce_sum(t)
return tr
cce = base_fun(y_true, y_pred)
ttn = [_ttn(layer_idx, ttn_layer)
for layer_idx, ttn_layer
in enumerate(ttn_layers)]
sttn = tf.reduce_sum(ttn)
if ttn_weight is not None:
sttn = tf.math.scalar_mul(ttn_weight,
sttn)
final_loss = cce + sttn
return final_loss
if ttn_layers is None:
return base_fun
return loss
def get_loss_fun(self):
if "sps" in self.args.model:
if self.data == 'deap':
if self.args.deap_one_hot:
base_fun = tf.keras.losses.categorical_crossentropy
else:
base_fun = tf.keras.losses.sparse_categorical_crossentropy
else:
base_fun = tf.keras.losses.categorical_crossentropy
print(
f"loss will be {base_fun.__name__} with a trace loss weight of {args.traceloss}")
return self.loss_trace_norm(self.ttn_layers,
base_fun,
self.args.traceloss)
elif self.args.model == "cross-stitch":
# return tf.keras.losses.categorical_crossentropy
if self.args.deap_one_hot:
return tf.keras.losses.categorical_crossentropy
else:
return tf.keras.losses.sparse_categorical_crossentropy
else:
raise ValueError()
def train(self):
self.callbacks = self.generate_callbacks()
self.loss_fun = self.get_loss_fun()
self.optimizer = shared.build_optimizer(self.args.optimizer_args)
optimizer_string = "default Adagrad(1.0)" if args.optimizer_args is None else f"{args.optimizer_args['name']} with kwargs {json.dumps(args.optimizer_args['kwargs'])}"
self.metrics = shared.generate_metrics_dict(
self.label_names,
self.num_classes,
self.args,
self.data,
out_format_string="out_{ln}"
)
self.model.compile(
optimizer=self.optimizer,
loss=self.loss_fun,
loss_weights=self.args.loss_weights,
metrics=self.metrics
)
print('Initiating the training phase ...')
print("Hyperparameter summary:")
print(f" Epochs: {self.args.epochs}")
print(f" Batch size: {self.args.batch}")
if self.data == 'opportunity':
print(f" Window size: {self.args.opportunity_time_window}")
print(f" Sensors: {self.args.opportunity_num_sensors}")
print(f" Tasks: {', '.join(self.label_names)}")
print(f" Loss weights: \
[{', '.join(str(lw) for lw in self.args.loss_weights)}]")
print(f" Optimizer: {optimizer_string}")
start = time.time()
print(f"\n Current Unix time: {start}")
self.model.fit(
self.train_data,
verbose=1,
epochs=self.args.epochs,
steps_per_epoch=self.args.steps,
validation_data=self.test_data,
validation_steps=self.args.validation_steps,
callbacks=self.callbacks)
print('##############################################')
end = time.time()
print('Total time used: %.2f seconds' % (end-start))
# Save the weights of the network
model_json = self.model.to_json()
jsonname = os.path.join(
self.out_path,
f"model{self.args.model}_{self.args.tag}.json")
with open(jsonname, "w") as json_file:
json_file.write(model_json)
# serialize weights to HDF5
hdfname = os.path.join(
self.out_path,
f"model{self.args.model}_{self.args.tag}_weights.h5")
self.model.save_weights(hdfname)
print("Saved model to folder:" + self.out_path)
final_name = os.path.join(
outpath, f'{self.args.model}_{self.args.tag}_model+weights.hdf5')
self.model.save(final_name)
def load_data(args, sps):
if args.dataset == 'opportunity':
train_data, val_data, input_shape, label_names, num_classes = shared.load_opportunity_data_mtl_tf(
sps.data_path, args)
sps.set_data(train_data, None, val_data,
input_shape, label_names, num_classes, None)
elif args.dataset == 'deap':
label_names, num_classes, all_names = select_channels_deap(
args.labels)
data_train, data_test, shape, config = load_deap_data(
args,
label_names)
data_test_test, _ = data_test # do not need the val dataset which is unbatches
sps.set_data(data_train, data_test_test, None,
shape, label_names, num_classes, config)
def training_loop(args, outpath):
print_arguments(args)
sps = SPS(args, outpath)
load_data(args, sps)
sps.build_model_sps()
if not args.dry_run:
sps.train()
if __name__ == "__main__":
args = parse_args()
if args.dataset is not None:
args.output = args.output.replace("$dataset$", args.dataset.upper())
else:
raise ValueError()
outpath = os.path.join(args.output, args.model, args.tag)
logfile = os.path.join(outpath, f"{args.model}_{args.tag}.log")
if not os.path.isdir(outpath):
os.makedirs(outpath)
with open(logfile, "w") as log:
sys.stdout = Unbuffered(sys.stdout, log)
print(f"Logging to {logfile}")
training_loop(args=args, outpath=outpath)
print("\nDone.", flush=True)