def save_activation(activation, L1):
#Parameters Of Experiments
cfg = {}
cfg['SGD_BATCHSIZE'] = 256
cfg['SGD_LEARNINGRATE'] = 0.0004
cfg['NUM_EPOCHS'] = 5000
cfg['FULL_MI'] = True
cfg['ACTIVATION'] = 'tanh'
cfg['LAYER_DIMS'] = [10, 7, 5, 4, 3]
ARCH_NAME = '-'.join(map(str, cfg['LAYER_DIMS']))
trn, tst = utils.get_IB_data('2017_12_21_16_51_3_275766')
cfg['SAVE_DIR'] = 'rawdata/' + cfg['ACTIVATION'] + '_' + ARCH_NAME
input_layer = keras.layers.Input((trn.X.shape[1], ))
clayer = input_layer
# clayer = keras.layers.Dense(cfg['LAYER_DIMS'][0],
# activation=cfg['ACTIVATION'],
# kernel_initializer=keras.initializers.truncated_normal(mean=0.0,
# stddev=1 / np.sqrt(float(cfg['LAYER_DIMS'][0]))),
# bias_initializer='zeros',
# activity_regularizer=keras.regularizers.l1(.01)
# )(clayer)
for n in cfg['LAYER_DIMS']:
clayer = keras.layers.Dense(
n,
activation=cfg['ACTIVATION'],
kernel_initializer=keras.initializers.truncated_normal(
mean=0.0, stddev=1 / np.sqrt(float(n))),
bias_initializer='zeros',
)(clayer)
output_layer = keras.layers.Dense(trn.nb_classes,
activation='softmax')(clayer)
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
optimizer = keras.optimizers.TFOptimizer(
tf.train.AdamOptimizer(learning_rate=cfg['SGD_LEARNINGRATE']))
model.compile(loss='categorical_crossentropy',
optimizer=optimizer,
metrics=['accuracy'])
def do_report(epoch):
if epoch < 20:
return True
elif epoch < 100:
return (epoch % 5 == 0)
elif epoch < 2000:
return (epoch % 20 == 0)
else:
return (epoch % 100 == 0)
reporter = loggingreporter.LoggingReporter(cfg=cfg,
trn=trn,
tst=tst,
do_save_func=do_report)
r = model.fit(x=trn.X,
y=trn.Y,
verbose=2,
batch_size=cfg['SGD_BATCHSIZE'],
epochs=cfg['NUM_EPOCHS'],
callbacks=[
reporter,
])
model = keras.models.Model(inputs=input_layer, outputs=output_layer)
optimizer = keras.optimizers.SGD(lr=cfg['SGD_LEARNINGRATE'])
model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['accuracy'])
# In[ ]:
def do_report(epoch):
# Only log activity for some epochs. Mainly this is to make things run faster.
if epoch < 20: # Log for all first 20 epochs
return True
elif epoch < 100: # Then for every 5th epoch
return (epoch % 5 == 0)
elif epoch < 200: # Then every 10th
return (epoch % 10 == 0)
else: # Then every 100th
return (epoch % 100 == 0)
reporter = loggingreporter.LoggingReporter(cfg=cfg,
trn=trn,
tst=tst,
do_save_func=do_report)
r = model.fit(x=trn.X, y=trn.Y,
verbose = 2,
batch_size = cfg['SGD_BATCHSIZE'],
epochs = cfg['NUM_EPOCHS'],
# validation_data=(tst.X, tst.Y),
callbacks = [reporter,])
def do_report(epoch):
# Only log activity for some epochs. Mainly this is to make things run faster.
if epoch < 20: # Log for all first 20 epochs
return True
elif epoch < 100: # Then for every 5th epoch
return (epoch % 5 == 0)
elif epoch < 200: # Then every 10th
return (epoch % 10 == 0)
else: # Then every 100th
return (epoch % 1000 == 0)
if not os.path.isdir(os.path.join(args.save_dir,'tensorboard')):
os.mkdir(os.path.join(args.save_dir,'tensorboard'))
reporter = loggingreporter.LoggingReporter(args=args,
trn=trn,
tst=tst,
do_save_func=do_report)
saver = keras.callbacks.ModelCheckpoint(os.path.join(args.save_dir,'chkpt.h5'),
monitor='val_loss',
verbose=0,
save_best_only=False,
save_weights_only=False,
mode='auto',
period=1000)
scheduler = keras.callbacks.ReduceLROnPlateau(monitor='val_loss',
factor=0.75,
patience=20,
verbose=0,