def train(self, batch_size, epochs, xtrain, ytrain, validation_data=None):
train_dataset = tf.data.Dataset.from_tensor_slices(
(xtrain.astype(np.float32), ytrain.astype(np.float32)))
train_dataset = train_dataset.shuffle(
buffer_size=xtrain.shape[0],
reshuffle_each_iteration=True).batch(batch_size)
#grad_vars = self.model.trainable_weights
#zero_grads = [tf.zeros_like(w) for w in grad_vars]
#self.optimizer.apply_gradients(zip(zero_grads, grad_vars))
red_lr = ReduceLROnPlateau(self.optimizer, 0.8, 10, 1e-5)
train_loss = []
valid_loss = []
epoch_loss_avg = tf.keras.metrics.Mean()
for i in range(epochs):
epoch_loss_avg.reset_states()
for x, y in train_dataset:
loss = self.train_step(self.model, x, y)
epoch_loss_avg.update_state(loss)
train_loss.append(epoch_loss_avg.result())
red_lr.on_epoch_end(train_loss[-1], i)
if (validation_data):
yvalid = self.model(validation_data[0], training=True)
valid_loss.append(
np.mean(self.loss_fn(yvalid, validation_data[1]).numpy()))
print("Step {} loss {} valid_loss {}".format(
i, epoch_loss_avg.result(), valid_loss[i]))
else:
print("Step {} loss {}".format(i, epoch_loss_avg.result()))
return train_loss, valid_loss
def train(self,
model,
batch_size,
epochs,
xtrain,
ytrain,
indx,
validation_data=None):
train_dataset = tf.data.Dataset.from_tensor_slices(
(xtrain.astype(np.float32), ytrain.astype(np.float32)))
train_dataset = train_dataset.shuffle(
buffer_size=xtrain.shape[0],
reshuffle_each_iteration=True).batch(batch_size)
#self.optimizer = tf.keras.optimizers.RMSprop(self.lr)
red_lr = ReduceLROnPlateau(self.optimizers[indx], 0.8, 10, 1e-5)
if (validation_data):
validation_data[0] = validation_data[0].astype(np.float32)
validation_data[1] = validation_data[1].astype(np.float32)
train_loss = []
valid_loss = []
mse_loss = []
anchor_loss = []
epoch_loss_avg = tf.keras.metrics.Mean()
mse_loss_avg = tf.keras.metrics.Mean()
anchor_loss_avg = tf.keras.metrics.Mean()
for i in range(epochs):
epoch_loss_avg.reset_states()
mse_loss_avg.reset_states()
anchor_loss_avg.reset_states()
for x, y in train_dataset:
total_loss, mse, anchor = self.train_step(
model, x, y, self.optimizers[indx], indx)
epoch_loss_avg.update_state(total_loss)
mse_loss_avg.update_state(mse)
anchor_loss_avg.update_state(anchor)
train_loss.append(epoch_loss_avg.result().numpy())
mse_loss.append(mse_loss_avg.result().numpy())
anchor_loss.append(anchor_loss_avg.result().numpy())
red_lr.on_epoch_end(train_loss[-1], i)
if (validation_data):
valid_loss.append(
np.mean(
self.loss_fn(model, validation_data[0],
validation_data[1], False).numpy()))
print("Step {} loss {} valid_loss {}".format(
i, epoch_loss_avg.result(), valid_loss[i]))
else:
print("Step {} loss {}".format(i, epoch_loss_avg.result()))
return [train_loss, mse_loss, anchor_loss], valid_loss
def train(self,
batch_size,
epochs,
xtrain,
ytrain,
kl_weight=1e-3,
validation_data=None):
train_dataset = tf.data.Dataset.from_tensor_slices(
(xtrain.astype(np.float32), ytrain.astype(np.float32)))
train_dataset = train_dataset.shuffle(
buffer_size=xtrain.shape[0],
reshuffle_each_iteration=True).batch(batch_size)
train_loss = []
NLL_loss = []
KL_loss = []
valid_loss = []
red_lr = ReduceLROnPlateau(self.optimizer, 0.8, 10, 1e-5)
epoch_loss_avg = tf.keras.metrics.Mean()
NLL_loss_avg = tf.keras.metrics.Mean()
KL_loss_avg = tf.keras.metrics.Mean()
for i in range(epochs):
epoch_loss_avg.reset_states()
NLL_loss_avg.reset_states()
KL_loss_avg.reset_states()
for x, y in train_dataset:
loss, NLL, KL = self.train_step(self.model, x, y, kl_weight)
epoch_loss_avg.update_state(loss)
NLL_loss_avg.update_state(NLL)
KL_loss_avg.update_state(KL)
train_loss.append(epoch_loss_avg.result().numpy())
NLL_loss.append(NLL_loss_avg.result().numpy())
KL_loss.append(KL_loss_avg.result().numpy())
red_lr.on_epoch_end(train_loss[-1], i)
if (validation_data):
valid_loss.append(
np.mean(
self.loss_fn(self.model, validation_data[0],
validation_data[1], kl_weight,
False).numpy()))
print("Step {} loss {} valid_loss {}".format(
i, epoch_loss_avg.result(), valid_loss[i]))
else:
print("Step {} loss {}".format(i, epoch_loss_avg.result()))
return [train_loss, NLL_loss, KL_loss], valid_loss
test_loss, test_accuracy = model.test_on_batch(x_batch_test,
y_batch_test)
testing_acc.append(test_accuracy)
testing_loss.append(test_loss)
train_logs_dict = get_logs(train_logs_dict, epoch, model, x_train, y_train)
test_logs_dict = get_logs(test_logs_dict, epoch, model, x_test, y_test)
logs = {
'acc': np.mean(training_acc),
'loss': np.mean(training_loss),
'val_loss': np.mean(testing_loss),
'val_acc': np.mean(testing_acc)
}
modelcheckpoint.on_epoch_end(epoch, logs)
earlystop.on_epoch_end(epoch, logs)
reduce_lr.on_epoch_end(epoch, logs)
tensorboard.on_epoch_end(epoch, logs)
print(
"accuracy: {}, loss: {}, validation accuracy: {}, validation loss: {}".
format(np.mean(training_acc), np.mean(training_loss),
np.mean(testing_acc), np.mean(testing_loss)))
if model.stop_training:
break
earlystop.on_train_end()
modelcheckpoint.on_train_end()
reduce_lr.on_train_end()
tensorboard.on_train_end()
# confusion metric for training
y_train_pred = model.predict(x_train).argmax(axis=1)