def computate_train_and_eval(classes, batch, epochs_without_learning,
model_name, kfold, pretrain):
# Cria a arquitetura densene1t pre-treinada
print('Inicializando a rede...')
print('Estrutura da {}: '.format(model_name))
model_ft, model_input_size = new_model(classes, model_name, pretrain)
print(model_ft)
print("Pré-Treinada = %s" % pretrain)
#Inicializa objeto para criacao dos graficos
if (kfold < 2):
grafico = statistics_plot(model_name, 'train_eval')
else:
grafico = statistics_plot(model_name, 'kfold')
model_ft = train(model_ft, model_name, model_input_size, grafico, classes,
batch, epochs_without_learning, kfold, pretrain)
#Testa a rede sobre o dataset de test
eval(grafico,
model_input_size,
model_name,
model_ft,
inside_trainning=True)
save_dir = "./modelo_treinado/" + model_name + "-best" + ".pth"
save_dir_grafico = "./estatisticas/" + model_name + ".html"
#Salva o modelo treinado em disco
print('Salvando modelo em ', save_dir)
torch.save(model_ft.state_dict(), save_dir)
print('Modelo salvo com sucesso')
#Exibe o grafico no navegador
grafico.show()
grafico.write_html(save_dir_grafico)
def computate_eval_only(classes, batch, epochs_without_learning, model_name,
kfold, pretrained):
save_dir_grafico = "./modelo_treinado/" + model_name + "eval-only" + ".html"
model_ft, model_input_size = new_model(classes, model_name, pretrained)
grafico = statistics_plot(model_name, 'eval')
eval(grafico,
model_input_size,
model_name,
model_ft,
inside_trainning=False)
grafico.show()
grafico.write_html(save_dir_grafico)
def train_bj_model(y_train, y_test, x1_train, x1_test, x2_train, x2_test,
x3_train, x3_test):
model = new_model()
model.fit(
{
'tem_input': x1_train,
'spa_input': x2_train,
'meo_input': x3_train
}, [y_train],
epochs=80,
validation_split=0.2,
batch_size=128,
shuffle=True,
verbose=1)
score = model.evaluate(x={
'tem_input': x1_test,
'spa_input': x2_test,
'meo_input': x3_test
},
y=y_test,
batch_size=30,
verbose=1,
sample_weight=None,
steps=None)
print(score)
pre = (model.predict({
'tem_input': x1_test,
'spa_input': x2_test,
'meo_input': x3_test
}))
a = smape(y_test, pre)
print(smape(y_test, pre))
s = '0509//' '1st' + '.h5'
model.save(s)
np.save('result', a)
return a, score
def compile_model():
return model.new_model(INPUT_SIZE, OUTPUT_SIZE)
def model_vacio():
return md.new_model()
y_train, y_test = g48[:6300], g48[6300:]
x1_train, x1_test = tem[:6300], tem[6300:]
x2_train, x2_test = spc[:6300], spc[6300:]
x3_train, x3_test = me[:6300], me[6300:]
#x1_train, x1_test ,\
#x2_train,x2_test, \
#x3_train , x3_test,\
#y_train, y_test =train_test_split(tem,spc,me,g48,test_size=0.33, random_state=42)
from model import new_model
model = new_model()
model.fit({
'tem_input': x1_train,
'spa_input': x2_train,
'meo_input': x3_train
}, [y_train],
epochs=100,
validation_split=0.2,
batch_size=128,
shuffle=True,
verbose=1)
score = model.evaluate(x={
'tem_input': x1_test,
'spa_input': x2_test,