def test_mnist():
(train_x, train_y), (test_x, test_y) = mnist.load_data()
val_x = train_x[50000:]
val_y = train_y[50000:]
train_x = train_x[:50000]
train_y = train_y[:50000]
batch_size = 200
modle = models.Sequential()
modle.add(layers.Linear(28, input_shape=(None, train_x.shape[1])))
modle.add(layers.ReLU())
modle.add(layers.Linear(10))
modle.add(layers.ReLU())
modle.add(layers.Linear(10))
modle.add(layers.Softmax())
acc = losses.categorical_accuracy.__name__
modle.compile(losses.CrossEntropy(),
optimizers.SGD(lr=0.001),
metrics=[losses.categorical_accuracy])
modle.summary()
history = modle.train(train_x,
train_y,
batch_size,
epochs=32,
validation_data=(val_x, val_y))
epochs = range(1, len(history["loss"]) + 1)
plt.plot(epochs, history["loss"], 'ro', label="Traning loss")
plt.plot(epochs, history["val_loss"], 'go', label="Validating loss")
plt.plot(epochs, history[acc], 'r', label="Traning accuracy")
plt.plot(epochs, history["val_" + acc], 'g', label="Validating accuracy")
plt.title('Training/Validating loss/accuracy')
plt.xlabel('Epochs')
plt.ylabel('Loss/Accuracy')
plt.legend()
plt.show(block=True)
if not "sgd_momentum" in parameters["training"]:
parameters["training"]['sgd_momentum']=0.9
optimizer = optim.SGD(net.parameters(), lr=parameters["training"]['learning_rate'], momentum=parameters["training"]['sgd_momentum'])
else:
optimizer = optim.Adam(net.parameters(), lr=parameters["training"]['learning_rate'])
# LOSS
if parameters["training"]["loss_function"]=="dice":
if (not "dice_smooth" in parameters["training"]):
parameters["training"]['dice_smooth']=0.001
loss_function = losses.Dice(smooth=parameters["training"]['dice_smooth'])
else:
loss_function = losses.CrossEntropy()
# LR SCHEDULE
if parameters["training"]["lr_schedule"]=="cosine":
scheduler = optim.lr_scheduler.CosineAnnealingLR(optimizer, parameters["training"]["nb_epochs"])
elif parameters["training"]["lr_schedule"]=="poly":
if not "poly_schedule_p" in parameters["training"]:
parameters["training"]['poly_schedule_p']=0.9
lr_lambda = lambda epoch: (1-float(epoch)/parameters["training"]["nb_epochs"])**parameters["training"]["poly_schedule_p"]
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lr_lambda)
else:
scheduler = optim.lr_scheduler.LambdaLR(optimizer, lambda epoch: 1)