# return tf.reduce_mean(tf.pow(y_true - x_out, 2))+tf.abs(var-0.5)*1e-4/(var+1e-5)
# return tf.losses.softmax_cross_entropy(y_true,x_out)
return tf.losses.sigmoid_cross_entropy(y_true, x_out) #loss1
# mean, var = tf.nn.moments(tf.reshape(x_out, [-1]), axes=[0])
# return tf.losses.sigmoid_cross_entropy(y_true,x_out) + 1e-4/(var+1e-5)
# return tf.reduce_mean(tf.keras.losses.categorical_crossentropy(y_true,x_out))
# return tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(logits = x_out , labels = y_true))
conv = ng.Model(data_provider=dp,
optimizer=tf.train.AdamOptimizer,
loss=loss,
restore=restore,
model_add='./model/1',
arch=arch)
#for i in range(1):
# print( 'set ' + str(i+1) + '/1')
# conv.train(data_provider=dp,training_epochs=2,
# iterations=100, n_s=100,
# learning_rate = 0.001, verbose=1)
# for _ in range(5):
# x,y = dp(1)
# pred = conv.predict(x)
# print(np.mean(x),np.argmax(y),np.argmax(pred))
#
#conv.train(data_provider=dp,training_epochs=2,
return ccg.filters(x, edd_method='sch')
dp = DataProvider(x_files,
y_files,
alpha,
nx=nx,
ny=ny,
n_buffer=2,
reload_rate=10000,
filt=filtf)
model_add = './models/' + str(n_layers) + '_layers_f' + filt + '/'
res_dir = './results/' + str(n_layers) + '_layers_f' + filt + '/'
ccg.ch_mkdir(res_dir + 'plots')
model = ng.Model(dp, restore=0, model_add=model_add + str(0), arch=arch)
print('# of variables:', model.n_variables)
if os.path.exists(res_dir + 'info.npy'):
i, dp.alpha, dalpha, learning_rate = np.load(res_dir + 'info.npy')
i = int(i)
model.model_add = model_add + str(i)
print('Loading model ' + str(i) + ' ...')
model.restore()
else:
i = 0
for _ in range(ntry):
alphas.append(dp.alpha)
#learning_rate = 5e-6
if qq0 == 0:
model_add_now = model0_add
else:
model_add_now = model_add
exec('from arch_' + arch + ' import architecture')
def arch(x_in):
return architecture(x_in, trainable=0)
restore = 1
conv = ng.Model(data_provider=dpt,
restore=restore,
model_add=model_add_now,
arch=arch)
print(conv.model_add)
conv.model_add = model_add
print(conv.model_add)
print('')
print('Number of trainable variables: {:d}'.format(conv.n_variables))
#nqq = qq0+1
for qq in range(qq0, nqq):
print(qq)
the_print('ROUND: ' + str(qq) + ', learning rate=' + str(learning_rate),
bgc='blue')
conv.train(data_provider=dpt,
try:
qq0, learning_rate = np.load(res_file + '_prop.npy')
except:
qq0 = np.load(res_file + '_prop.npy')
learning_rate = learning_rate / (decay)**qq0
qq0 = int(qq0)
qq0 = qq0 + 1
print('The learning will begin from {} q number.'.format(qq0))
#exit()
ns = 10
#learning_rate = 5e-6
restore = qq0 != 0
conv = ng.Model(data_provider=dpt,
restore=restore,
model_add=model_add,
arch='arch_' + args.arch)
print('')
print('Number of trainable variables: {:d}'.format(conv.n_variables))
nqq = qq0 + 1
for qq in range(qq0, nqq):
print(qq)
the_print('ROUND: ' + str(qq) + ', learning rate=' + str(learning_rate),
bgc='blue')
conv.train(data_provider=dpt,
training_epochs=5,
iterations=20,
n_s=ns,
learning_rate=learning_rate,
#fig, (ax1,ax2)= plt.subplots(ncols=2, nrows=1, figsize=(20, 10))
#ax1.imshow(x[0,:,:,0])
#ax1.axis('off')
#ax2.imshow(y[0,:,:,0])
#ax2.axis('off')
def arch(x_in):
x_out = architecture(x_in=x_in, n_layers=5, res=2)
return x_out
model = ng.Model(nx=nside,
ny=nside,
n_channel=1,
n_class=1,
restore=0,
model_add='./model/' + str(0) + '_' + str(dp.alpha),
arch=arch)
print('# of variables:', model.n_variables)
alphas = []
dalpha = 0.05
p_move = 0
for i in range(50):
model.train(data_provider=dp,
training_epochs=10,