def custom_fit(model,
callbacks,
x_train,
y_train,
x_test,
y_test,
epochs,
batch_size,
dir_name,
compare_title,
draw_step=10,
verbose=1):
epochs_step = int(epochs / draw_step)
if dir_name != None:
os.mkdir(dir_name)
full_loss_history = np.empty(0)
for init_epoch in np.arange(0, epochs, step=epochs_step):
save = False if dir_name == None else True
if save:
save_path = dir_name + "/" + "val_loss.png"
else:
save_path = None
history = model.fit(x=x_train,
y=y_train,
batch_size=batch_size,
epochs=init_epoch + epochs_step,
verbose=verbose,
callbacks=callbacks,
validation_data=(x_test, y_test),
initial_epoch=init_epoch)
full_loss_history = np.append(full_loss_history,
history.history['val_loss'])
plt.plot(np.transpose(x_test)[0], y_test, '.')
plt.plot(np.transpose(x_test)[0], model.predict(x_test), '.')
plt.legend(('function', 'approximation'),
loc='upper left',
shadow=True)
plt.title(compare_title + "\nval_loss = %.4f\nepoch = %d" %
(history.history["val_loss"][history.epoch.__len__() - 1],
init_epoch + epochs_step))
if dir_name != None:
plt.savefig(
dir_name + "/" + "%.d_compare_%.4f.png" %
(init_epoch + epochs_step,
history.history["val_loss"][history.epoch.__len__() - 1]),
dpi=200)
plt.show()
plt.close()
if (history.epoch.__len__() - 1 != epochs_step):
epochs = init_epoch + history.epoch.__len__()
break
gr.plot_graphic(x=np.arange(1, epochs + 1),
y=full_loss_history,
x_label='epochs',
y_label='val_loss',
title="val_loss" + ' history',
save_path=save_path,
save=save,
show=True)
return model
# 4 model fitting
model.compile(optimizer=optimizer, loss='mse', metrics=['mse'])
history = model.fit(x=x_train,
y=y_train,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
validation_data=(x_test, y_test))
gr.plot_graphic(x=history.epoch,
y=np.array(history.history["val_loss"]),
x_label='epochs',
y_label='val_loss',
title="val_loss" + ' history',
save=False,
show=True)
plt_x_zero = np.empty(0)
plt_y_zero = np.empty(0)
plt_x_one = np.empty(0)
plt_y_one = np.empty(0)
plt_x_two = np.empty(0)
plt_y_two = np.empty(0)
plt_x_three = np.empty(0)
plt_y_three = np.empty(0)
model = Sequential()
model.add(Dense(first_layer_nur, input_dim=1, kernel_initializer='glorot_normal', activation='linear'))
model.compile(loss='mean_squared_error', optimizer=SGD(lr=lr), metrics=['mae'])
# 2.1.3-------------------------------------------------------
history = model.fit(x_train, y_train, batch_size=batch_size, epochs=epochs, verbose=verbose)
score = model.evaluate(x_test, y_test, verbose=verbose)
print("\nabsolute_error on train data\t %.f%%" % (history.history['mean_absolute_error'][epochs - 1] * 100))
print("\nabsolute_error on testing data %.f%%" % (score[1] * 100))
print("loss on train data %.f%%" % (history.history['loss'][epochs - 1] * 100))
gr.plot_graphic(x=history.epoch, y=np.array(history.history['loss']),
x_label='epochs', y_label='loss', title='mean_squared_error history', show=True)
gr.plot_graphic(x=history.epoch, y=np.array(history.history['mean_absolute_error']),
x_label='epochs', y_label='accuracy', title='mean_absolute_error history', show=True)
plt.plot(np.append(x_train, x_test), model.predict(np.append(x_train, x_test)), '.')
plt.plot(np.append(x_train, x_test), np.append(y_train, y_test), '.')
plt.legend(('approximation', 'function'), loc='upper left', shadow=True)
plt.show()
plt.close()
# 2.2-------------------------------------------------------
# Выбор скорости обучения
