def _train_step(
self,
inputs,
targets,
batch: int,
batch_size: int,
optimizer: tf.keras.optimizers,
loss_fn: tf.keras.losses,
):
with tf.GradientTape() as tape:
outputs = self.model(inputs, training=True)
# Calculate the training prediction tokens
predictions = self.model.predictions(
outputs, self.train_dataloader.encoder_target
)
# Calculate the loss and update the parameters
loss = loss_fn(targets, outputs)
gradients = tape.gradient(loss, self.model.trainable_variables)
optimizer.apply_gradients(zip(gradients, self.model.trainable_variables))
metric = self.recorded_losses["train"]
metric(loss)
if base.Metrics.ABSOLUTE_ACC in self.metrics:
self._record_abs_acc(outputs, targets, batch, batch_size, "train")
logger.info(f"Batch #{batch} : training loss {metric.result()}")
return predictions
def trainer(
j_hat: JHat,
datasets: Tuple[tf.keras.utils.Sequence, tf.keras.utils.Sequence],
optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam,
learning_rate: float = 1e-3,
preprocess_fn: Callable = None,
epochs: int = 20,
reg_loss_fn: Callable = (lambda kernel: 0),
verbose: int = 1,
) -> None:
"""
Train the kernel to maximise an estimate of test power using minibatch gradient descent.
"""
ds_ref, ds_cur = datasets
optimizer = optimizer(learning_rate)
n_minibatch = min(len(ds_ref), len(ds_cur))
# iterate over epochs
loss_ma = 0.
for epoch in range(epochs):
if verbose:
pbar = tf.keras.utils.Progbar(n_minibatch, 1)
for step, (x_ref, x_cur) in enumerate(zip(ds_ref, ds_cur)):
if isinstance(preprocess_fn, Callable): # type: ignore
x_ref, x_cur = preprocess_fn(x_ref), preprocess_fn(x_cur)
with tf.GradientTape() as tape:
estimate = j_hat(x_ref, x_cur)
loss = -estimate + reg_loss_fn(j_hat.kernel) # ascent
grads = tape.gradient(loss, j_hat.trainable_weights)
optimizer.apply_gradients(zip(grads, j_hat.trainable_weights))
if verbose == 1:
loss_ma = loss_ma + (loss - loss_ma) / (step + 1)
pbar_values = [('loss', loss_ma)]
pbar.add(1, values=pbar_values)
def train_step(model: tf.keras.Model, x_batch: tf.Tensor, y_batch: tf.Tensor,
optimizer: tf.keras.optimizers) -> tf.keras.metrics.Mean:
with tf.GradientTape() as tape:
loss = loss_fn(model, x_batch, y_batch)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
return train_loss(loss)
def explainer_train_step(
index: tf.Tensor,
X: tf.Tensor,
y: tf.Tensor,
sample_weights: tf.Tensor,
explainer: tf.keras.Model,
optimizer: tf.keras.optimizers,
):
with tf.GradientTape() as tape:
loss = explainer_mse_loss(explainer(index, X), y, sample_weights)
gradients = tape.gradient(loss, explainer.trainable_variables)
optimizer.apply_gradients(zip(gradients, explainer.trainable_variables))
def train_iteration(initial_state: tf.Tensor, gamma: float,
max_iter_step: tf.int32, actor_model: tf.keras.Model,
critic_model: tf.keras.Model,
optimizer: tf.keras.optimizers) -> List[tf.Tensor]:
state_curr = initial_state
state_shape = initial_state.shape
step_reward = 0
avg_reward = 0.
# state_curr = tf.expand_dims(state_curr, 0)
for curr_step in tf.range(max_iter_step):
with tf.GradientTape(persistent=True) as tape:
action_net_out = actor_model(state_curr)
value_curr_out = critic_model(state_curr)
action_net_out = tf.reshape(action_net_out, action_shape)
action_prob_softmax = tf.nn.softmax(action_net_out)
action = tf.random.categorical(tf.math.log(action_prob_softmax), 1)
action = tf.squeeze(action)
state_next, step_reward = tf_env_step(action)
state_next.set_shape(state_shape)
state_next = tf.expand_dims(state_next, 0)
value_next_out = critic_model(state_next)
avg_reward = 0.99 * avg_reward + \
0.01 * tf.cast(step_reward, tf.float32)
TD_err = tf.cast(step_reward, dtype=tf.float32) - avg_reward + \
gamma * value_next_out - value_curr_out
loss_actor = -TD_err * tf.math.reduce_sum([
tf.math.log(action_prob_softmax[i, j])
for i, j in enumerate(action)
])
loss_critic = tf.square(TD_err)
grads_actor = tape.gradient(loss_actor,
actor_model.trainable_variables)
grads_critic = tape.gradient(loss_critic,
critic_model.trainable_variables)
optimizer.apply_gradients(
zip(grads_actor, actor_model.trainable_variables))
optimizer.apply_gradients(
zip(grads_critic, critic_model.trainable_variables))
state_curr = state_next
if curr_step % 100 == 1:
print(f'Step: {curr_step}, average rewards = {avg_reward}'
f' elite num = {len(env.elite_list)}\n')
return [avg_reward, len(env.elite_list)]
def trainer(
model: tf.keras.Model,
loss_fn: tf.keras.losses,
X_train: np.ndarray,
y_train: np.ndarray = None,
optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam(
learning_rate=1e-3),
loss_fn_kwargs: dict = None,
preprocess_fn: Callable = None,
epochs: int = 20,
batch_size: int = 64,
buffer_size: int = 1024,
verbose: bool = True,
log_metric: Tuple[str, "tf.keras.metrics"] = None,
callbacks: tf.keras.callbacks = None
) -> None: # TODO: incorporate callbacks + LR schedulers
"""
Train TensorFlow model.
Parameters
----------
model
Model to train.
loss_fn
Loss function used for training.
X_train
Training batch.
y_train
Training labels.
optimizer
Optimizer used for training.
loss_fn_kwargs
Kwargs for loss function.
preprocess_fn
Preprocessing function applied to each training batch.
epochs
Number of training epochs.
batch_size
Batch size used for training.
buffer_size
Maximum number of elements that will be buffered when prefetching.
verbose
Whether to print training progress.
log_metric
Additional metrics whose progress will be displayed if verbose equals True.
callbacks
Callbacks used during training.
"""
# create dataset
if y_train is None: # unsupervised model without teacher forcing
train_data = X_train
else:
train_data = (X_train, y_train)
train_data = tf.data.Dataset.from_tensor_slices(train_data)
train_data = train_data.shuffle(buffer_size=buffer_size).batch(batch_size)
n_minibatch = int(np.ceil(X_train.shape[0] / batch_size))
# iterate over epochs
for epoch in range(epochs):
if verbose:
pbar = tf.keras.utils.Progbar(n_minibatch, 1)
# iterate over the batches of the dataset
for step, train_batch in enumerate(train_data):
if y_train is None:
X_train_batch = train_batch
else:
X_train_batch, y_train_batch = train_batch
if isinstance(preprocess_fn, Callable): # type: ignore
X_train_batch = preprocess_fn(X_train_batch)
with tf.GradientTape() as tape:
preds = model(X_train_batch)
if y_train is None:
ground_truth = X_train_batch
else:
ground_truth = y_train_batch
# compute loss
if tf.is_tensor(preds):
args = [ground_truth, preds]
else:
args = [ground_truth] + list(preds)
if loss_fn_kwargs:
loss = loss_fn(*args, **loss_fn_kwargs)
else:
loss = loss_fn(*args)
if model.losses: # additional model losses
loss += sum(model.losses)
grads = tape.gradient(loss, model.trainable_weights)
optimizer.apply_gradients(zip(grads, model.trainable_weights))
if verbose:
loss_val = loss.numpy()
if loss_val.shape:
if loss_val.shape[0] != batch_size:
if len(loss_val.shape) == 1:
shape = (batch_size - loss_val.shape[0], )
elif len(loss_val.shape) == 2:
shape = (batch_size - loss_val.shape[0],
loss_val.shape[1]) # type: ignore
add_mean = np.ones(shape) * loss_val.mean()
loss_val = np.r_[loss_val, add_mean]
pbar_values = [('loss', loss_val)]
if log_metric is not None:
log_metric[1](ground_truth, preds)
pbar_values.append(
(log_metric[0], log_metric[1].result().numpy()))
pbar.add(1, values=pbar_values)
def trainer(model: tf.keras.Model,
loss_fn: tf.keras.losses,
X_train: np.ndarray,
y_train: np.ndarray = None,
optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam(
learning_rate=1e-3),
loss_fn_kwargs: dict = None,
epochs: int = 100,
batch_size: int = 1,
buffer_size: int = 1024,
shuffle: bool = False,
verbose: bool = True) -> None:
"""
Train TensorFlow model.
Parameters
----------
model
Model to train.
loss_fn
Loss function used for training.
X_train
Training batch.
y_train
Training labels.
optimizer
Optimizer used for training.
loss_fn_kwargs
Kwargs for loss function.
epochs
Number of training epochs.
batch_size
Batch size used for training.
buffer_size
Maximum number of elements that will be buffered when prefetching.
shuffle
Whether to shuffle training data.
verbose
Whether to print training progress.
"""
# create dataset
if y_train is None: # unsupervised model
train_data = X_train
else:
train_data = (X_train, y_train)
train_data = tf.data.Dataset.from_tensor_slices(train_data)
if shuffle:
train_data = train_data.shuffle(
buffer_size=buffer_size).batch(batch_size)
n_minibatch = int(np.ceil(X_train.shape[0] / batch_size))
# iterate over epochs
for epoch in range(epochs):
if verbose:
pbar = tf.keras.utils.Progbar(n_minibatch, 1)
# iterate over the batches of the dataset
for step, train_batch in enumerate(train_data):
if y_train is None:
X_train_batch = train_batch
else:
X_train_batch, y_train_batch = train_batch
with tf.GradientTape() as tape:
preds = model(X_train_batch)
if y_train is None:
ground_truth = X_train_batch
else:
ground_truth = y_train_batch
# compute loss
if tf.is_tensor(preds):
args = [ground_truth, preds]
else:
args = [ground_truth] + list(preds)
if loss_fn_kwargs:
loss = loss_fn(*args, **loss_fn_kwargs)
else:
loss = loss_fn(*args)
if model.losses: # additional model losses
loss += sum(model.losses)
grads = tape.gradient(loss, model.trainable_weights)
optimizer.apply_gradients(zip(grads, model.trainable_weights))
if verbose:
loss_val = loss.numpy().mean()
pbar_values = [('loss', loss_val)]
pbar.add(1, values=pbar_values)
def trainer(model: tf.keras.Model,
loss_fn: tf.keras.losses,
x_train: np.ndarray,
y_train: np.ndarray = None,
dataset: tf.keras.utils.Sequence = None,
optimizer: tf.keras.optimizers = tf.keras.optimizers.Adam(
learning_rate=1e-3),
loss_fn_kwargs: dict = None,
preprocess_fn: Callable = None,
epochs: int = 20,
reg_loss_fn: Callable = (lambda model: 0),
batch_size: int = 64,
buffer_size: int = 1024,
verbose: bool = True,
log_metric: Tuple[str, "tf.keras.metrics"] = None,
callbacks: tf.keras.callbacks = None) -> None:
"""
Train TensorFlow model.
Parameters
----------
model
Model to train.
loss_fn
Loss function used for training.
x_train
Training data.
y_train
Training labels.
dataset
Training dataset which returns (x, y).
optimizer
Optimizer used for training.
loss_fn_kwargs
Kwargs for loss function.
preprocess_fn
Preprocessing function applied to each training batch.
epochs
Number of training epochs.
reg_loss_fn
Allows an additional regularisation term to be defined as reg_loss_fn(model)
batch_size
Batch size used for training.
buffer_size
Maximum number of elements that will be buffered when prefetching.
verbose
Whether to print training progress.
log_metric
Additional metrics whose progress will be displayed if verbose equals True.
callbacks
Callbacks used during training.
"""
return_xy = False if not isinstance(
dataset, tf.keras.utils.Sequence) and y_train is None else True
if not isinstance(dataset, tf.keras.utils.Sequence): # create dataset
train_data = x_train if y_train is None else (x_train, y_train)
dataset = tf.data.Dataset.from_tensor_slices(train_data)
dataset = dataset.shuffle(buffer_size=buffer_size).batch(batch_size)
n_minibatch = len(dataset)
if loss_fn_kwargs:
loss_fn = partial(loss_fn, **loss_fn_kwargs)
# iterate over epochs
for epoch in range(epochs):
if verbose:
pbar = tf.keras.utils.Progbar(n_minibatch, 1)
if hasattr(dataset, 'on_epoch_end'):
dataset.on_epoch_end()
loss_val_ma = 0.
for step, data in enumerate(dataset):
x, y = data if return_xy else (data, None)
if isinstance(preprocess_fn, Callable): # type: ignore
x = preprocess_fn(x)
with tf.GradientTape() as tape:
y_hat = model(x)
y = x if y is None else y
if isinstance(loss_fn, Callable): # type: ignore
args = [y, y_hat
] if tf.is_tensor(y_hat) else [y] + list(y_hat)
loss = loss_fn(*args)
else:
loss = 0.
if model.losses: # additional model losses
loss += sum(model.losses)
loss += reg_loss_fn(
model) # alternative way they might be specified
grads = tape.gradient(loss, model.trainable_weights)
optimizer.apply_gradients(zip(grads, model.trainable_weights))
if verbose:
loss_val = loss.numpy()
if loss_val.shape:
if loss_val.shape[0] != batch_size:
if len(loss_val.shape) == 1:
shape = (batch_size - loss_val.shape[0], )
elif len(loss_val.shape) == 2:
shape = (batch_size - loss_val.shape[0],
loss_val.shape[1]) # type: ignore
add_mean = np.ones(shape) * loss_val.mean()
loss_val = np.r_[loss_val, add_mean]
loss_val_ma = loss_val_ma + (loss_val - loss_val_ma) / (step +
1)
pbar_values = [('loss_ma', loss_val_ma)]
if log_metric is not None:
log_metric[1](y, y_hat)
pbar_values.append(
(log_metric[0], log_metric[1].result().numpy()))
pbar.add(1, values=pbar_values)
def run(
self,
loss_fn: tf.keras.losses,
optimizer: tf.keras.optimizers,
batch_size: int,
num_epoch: int,
checkpoint=None,
):
"""Pretraining session."""
logger.info("Creating datasets...")
train_dataset = self.train_dataloader.create_dataset()
valid_dataset = self.valid_dataloader.create_dataset()
train_dataset = self.model.preprocessing(train_dataset)
valid_dataset = self.model.preprocessing(valid_dataset)
logger.info("Creating results directory...")
directory = os.path.join(
"results/" + self.model.title + "_mono",
datetime.now().strftime("%Y-%m-%d %H:%M:%S"),
)
os.makedirs(directory, exist_ok=True)
if checkpoint is not None:
logger.info(f"Loading model {checkpoint}")
self.model.load(str(checkpoint))
else:
checkpoint = 0
logger.info("Beginning pretraining session...")
for epoch in range(checkpoint + 1, num_epoch + 1):
logger.debug(f"Epoch {epoch}...")
for i, minibatch in enumerate(
train_dataset.padded_batch(
batch_size, padded_shapes=self.model.padded_shapes)):
logger.debug(minibatch)
with tf.GradientTape() as tape:
loss = self._step(minibatch,
i,
loss_fn,
"train",
training=True)
gradients = tape.gradient(loss,
self.model.trainable_variables)
optimizer.apply_gradients(
zip(gradients, self.model.trainable_variables))
logger.debug("Saving training loss")
self.history.record("train_loss", self.losses["train"].result())
for i, minibatch in enumerate(
valid_dataset.padded_batch(
batch_size, padded_shapes=self.model.padded_shapes)):
loss = self._step(minibatch,
i,
loss_fn,
"valid",
training=False)
logger.debug("Saving validation loss")
self.history.record("valid_loss", self.losses["valid"].result())
self.model.save(epoch)
self.losses["train"].reset_states()
self.losses["valid"].reset_states()
self.history.save(directory + f"/history-{epoch}")
def _calculate_and_apply_gradients(model: tf.keras.Sequential,
optimizer: tf.keras.optimizers,
gradient_tape: tf.GradientTape,
loss: [float]):
gradients = gradient_tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
def apply_gradients(optimizer: tf.keras.optimizers, gradients: tf.Tensor,
variables):
optimizer.apply_gradients(zip(gradients, variables))