def train_one_step(model: BaseTF2Model, train_x: np.ndarray,
train_y: np.ndarray, optimizer: tf.optimizers.Optimizer,
loss_fn: tf.losses.Loss):
"""
Perform one step gradient update
Later this function can be reused when training SoftmaxRegressionTF2 or LogisticRegressionTF2
"""
with tf.GradientTape() as tape:
# [TODO 1.12] Calculate model predictions and loss inside tf.GradientTape context
# HINT: Operations on trainable variables that executed inside tf.GradientTape context
# will be recorded automatically for autograd
# print(train_x.shape)
# print('model',model)
predictions = model(train_x)
# print(predictions.shape)
# print(train_y.shape)
train_loss = loss_fn(train_y, predictions)
# [TODO 1.13] Compute gradient of loss w.r.t model's parameters
# HINT : since BaseTF2Classifier extends tf.Module, you can access all
# model's parameters using `model.trainable_variables`
gradients = tape.gradient(train_loss, model.trainable_variables)
# [TODO 1.14] Perform one step weight update using optimizers.apply_gradients
# HINT: https://www.tensorflow.org/api_docs/python/tf/keras/optimizers/Optimizer
optimizer.apply_gradients(zip(gradients, model.trainable_variables))
# `train_loss`, `predictions` currently is a EegerTensor,
# you can access its numpy data by using
return train_loss.numpy()
def train_step(model: TBH, batch_data, bbn_dim, cbn_dim, batch_size,
actor_opt: tf.optimizers.Optimizer,
critic_opt: tf.optimizers.Optimizer):
random_binary = (tf.sign(tf.random.uniform([batch_size, bbn_dim]) - 0.5) +
1) / 2
random_cont = tf.random.uniform([batch_size, cbn_dim])
with tf.GradientTape() as actor_tape, tf.GradientTape() as critic_tape:
model_input = [batch_data, random_binary, random_cont]
model_output = model(model_input, training=True)
actor_loss = reconstruction_loss(model_output[1], batch_data[1]) - \
adv_loss(model_output[4], model_output[2]) - \
adv_loss(model_output[5], model_output[3])
critic_loss = adv_loss(model_output[4], model_output[2]) + adv_loss(
model_output[5], model_output[3])
actor_scope = model.encoder.trainable_variables + model.tbn.trainable_variables + \
model.decoder.trainable_variables
critic_scope = model.dis_1.trainable_variables + model.dis_2.trainable_variables
actor_gradient = actor_tape.gradient(actor_loss, sources=actor_scope)
critic_gradient = critic_tape.gradient(critic_loss,
sources=critic_scope)
actor_opt.apply_gradients(zip(actor_gradient, actor_scope))
critic_opt.apply_gradients(zip(critic_gradient, critic_scope))
return model_output[0].numpy(), actor_loss.numpy(), critic_loss.numpy()
def test_robust_optimizer_matches(
base_optimizer: tf.optimizers.Optimizer,
robust_optimizer: RobustOptimizer,
grads,
grads_with_nan,
vars,
other_vars,
):
keys = list(vars.keys())
vars_list = [vars[key] for key in keys]
grads_list = [grads[key] for key in keys]
for i in range(grads[keys[0]].shape[0]):
base_optimizer.apply_gradients(
zip([grad[i] for grad in grads_list], vars_list))
other_vars_list = [other_vars[key] for key in keys]
grads_with_nan_list = [grads_with_nan[key] for key in keys]
for i in range(grads_with_nan[keys[0]].shape[0]):
robust_optimizer.apply_gradients(
zip([grad[i] for grad in grads_with_nan_list], other_vars_list))
for key in keys:
assert np.all(np.isfinite(other_vars[key]))
assert_allclose(vars[key].numpy(), other_vars[key].numpy())
def step_train(model: Model, data, opt: tf.optimizers.Optimizer, t):
o_x, o_y, t_x, t_y = next(data)
with tf.GradientTape() as tape:
mean, var, mvn = model([o_x, o_y], t_x)
loss = model.obj(t_y, mvn)
gradient = tape.gradient(loss, sources=model.trainable_variables)
opt.apply_gradients(zip(gradient, model.trainable_variables))
tf.summary.scalar('train/loss', loss.numpy(), step=t)
return loss.numpy()
def train_step(model: JMLH, batch_data, opt: tf.optimizers.Optimizer):
with tf.GradientTape() as tape:
model_input = batch_data
code, prob, cls_prob = model(model_input, training=True)
loss = jmlh_loss(prob, cls_prob, label=batch_data[2])
gradient = tape.gradient(loss, sources=model.trainable_variables)
opt.apply_gradients(zip(gradient, model.trainable_variables))
return code.numpy(), loss.numpy()
def train_op(self, x_batch: tf.Tensor, optimizer: tf.optimizers.Optimizer):
with tf.GradientTape() as tape:
loss = self.loss_op(x_batch)
variables = tape.watched_variables()
gradient = tape.gradient(loss, variables)
optimizer.apply_gradients(
zip(gradient,
variables)
)
return loss
def train_single_epoch(model: tf.keras.Model,
anchors: tf.Tensor,
dataset: tf.data.Dataset,
optimizer: tf.optimizers.Optimizer,
grad_accum_steps: int,
loss_fn: LossFn,
epoch: int,
num_classes: int,
print_every: int = 10):
acc_gradients = []
running_loss = tf.metrics.Mean()
running_clf_loss = tf.metrics.Mean()
running_reg_loss = tf.metrics.Mean()
for i, (images, (labels, bbs)) in enumerate(dataset):
target_reg, target_clf = utils.anchors.anchor_targets_bbox(
anchors, images, bbs, labels, num_classes)
reg_loss, clf_loss, grads = _train_step(model=model,
optimizer=optimizer,
loss_fn=loss_fn,
images=images,
regress_targets=target_reg,
labels=target_clf)
if len(acc_gradients) == 0:
acc_gradients = grads
else:
acc_gradients = [g1 + g2 for g1, g2 in zip(acc_gradients, grads)]
if (i + 1) % grad_accum_steps == 0:
optimizer.apply_gradients(
zip(acc_gradients, model.trainable_variables))
acc_gradients = []
running_loss(reg_loss + clf_loss)
running_clf_loss(clf_loss)
running_reg_loss(reg_loss)
if (i + 1) % print_every == 0:
lr = get_lr(optimizer)
print(f'Epoch[{epoch}] '
f'loss: {running_loss.result():.6f} '
f'clf. loss: {running_clf_loss.result():.6f} '
f'reg. loss: {running_reg_loss.result():.6f} '
f'learning rate: {lr:.6f}')
if len(acc_gradients) > 0:
optimizer.apply_gradients(zip(acc_gradients,
model.trainable_variables))
def execute(
bath_of_experiences: List[Experience],
target_net: Model,
policy_net: Model,
gamma: float,
number_of_actions: int,
optimizer: tf.optimizers.Optimizer,
) -> float:
(
states,
actions,
rewards,
next_states,
dones,
) = ExtractExperiencesService.execute(bath_of_experiences)
formatted_next_states = np.atleast_2d(next_states).astype(np.float32)
formatted_rewards = np.atleast_2d(rewards).astype(np.float32)
output_from_policy_net = policy_net(formatted_next_states)
argmax_from_policy_net_output = np.argmax(output_from_policy_net,
axis=1)
output_from_target_net = target_net(formatted_next_states)
actions_in_one_hot = tf.one_hot(
np.squeeze(argmax_from_policy_net_output), number_of_actions)
actions_gathered = tf.math.reduce_sum(output_from_target_net *
actions_in_one_hot,
axis=1,
keepdims=True)
q_s_a_prime = formatted_rewards + gamma * actions_gathered * (1 -
dones)
q_s_a_prime_tensor = tf.convert_to_tensor(q_s_a_prime,
dtype=tf.float32)
with tf.GradientTape() as tape:
formatted_states = np.atleast_2d(states).astype(np.float32)
formatted_actions = np.squeeze(actions).astype(np.float32)
output_from_policy_net = policy_net(formatted_states)
actions_in_one_hot = tf.one_hot(formatted_actions,
number_of_actions)
q_s_a = tf.math.reduce_sum(output_from_policy_net *
actions_in_one_hot,
axis=1,
keepdims=True)
loss = tf.math.reduce_mean(tf.square(q_s_a_prime_tensor - q_s_a))
policy_net_variables = policy_net.trainable_variables
gradients = tape.gradient(loss, policy_net_variables)
optimizer.apply_gradients(zip(gradients, policy_net_variables))
return loss.numpy()
def _train_step(model: tf.keras.Model, optimizer: tf.optimizers.Optimizer,
loss_fn: LossFn, images: tf.Tensor, regress_targets: tf.Tensor,
labels: tf.Tensor) -> Tuple[float, float]:
with tf.GradientTape() as tape:
regressors, clf_probas = model(images)
reg_loss, clf_loss = loss_fn(labels, clf_probas, regress_targets,
regressors)
loss = reg_loss + clf_loss
grads = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(grads, model.trainable_variables))
return reg_loss, clf_loss
def step_train(model: Model, data, opt: tf.optimizers.Optimizer, t):
batch_data = next(data)
x = tf.cast(batch_data['x'], tf.float32)
summary_step = -1 if t % 50 > 0 else t
with tf.GradientTape() as tape:
encoded, vq_feat, context_ind, decoded, _ = model(x)
loss = model.loss(x, encoded, context_ind, decoded, step=summary_step)
gradient = tape.gradient(loss, sources=model.trainable_variables)
opt.apply_gradients(zip(gradient, model.trainable_variables))
if summary_step >= 0:
gt = tf.reshape(x, [-1, 28, 28, 1])
rec = tf.reshape(decoded, [-1, 28, 28, 1])
tf.summary.image('train/gt', gt, summary_step, max_outputs=1)
tf.summary.image('train/rec', rec, summary_step, max_outputs=1)
return loss.numpy()
def train_step(model: TBH, batch_data, bbn_dim, cbn_dim, batch_size,
actor_opt: tf.optimizers.Optimizer,
critic_opt: tf.optimizers.Optimizer):
random_binary = (tf.sign(tf.random.uniform([batch_size, bbn_dim]) - 0.5) +
1) / 2
random_cont = tf.random.uniform([batch_size, cbn_dim])
# tf.GradientTape: give exposure and record sequence of gradients
# https://stackoverflow.com/questions/53953099/what-is-the-purpose-of-the-tensorflow-gradient-tape
with tf.GradientTape() as actor_tape, tf.GradientTape() as critic_tape:
model_input = [batch_data, random_binary, random_cont]
model_output = model(model_input, training=True)
# Original Code
# actor_loss = reconstruction_loss(model_output[1], batch_data[1]) - \
# adv_loss(model_output[4], model_output[2]) - \
# adv_loss(model_output[5], model_output[3])
# critic_loss = adv_loss(model_output[4], model_output[2]) + adv_loss(model_output[5], model_output[3])
# Testing Code: critic_loss = adv_loss
actor_loss = reconstruction_loss(model_output[1], batch_data[1]) \
- tf.reduce_mean(tf.keras.losses.binary_crossentropy(tf.ones_like(model_output[2]), model_output[2]))\
- tf.reduce_mean(tf.keras.losses.binary_crossentropy(tf.ones_like(model_output[3]), model_output[3]))
# log(d(x'))
#+ adv_loss(model_output[4], model_output[2]) \
#+ adv_loss(model_output[5], model_output[3])
# adv_loss
critic_loss = -adv_loss(model_output[4], model_output[2]) - adv_loss(
model_output[5], model_output[3])
actor_scope = model.encoder.trainable_variables + model.tbn.trainable_variables + \
model.decoder.trainable_variables
critic_scope = model.dis_1.trainable_variables + model.dis_2.trainable_variables
actor_gradient = actor_tape.gradient(actor_loss, sources=actor_scope)
critic_gradient = critic_tape.gradient(critic_loss,
sources=critic_scope)
actor_opt.apply_gradients(zip(actor_gradient, actor_scope))
critic_opt.apply_gradients(zip(critic_gradient, critic_scope))
return model_output[0].numpy(), actor_loss.numpy(), critic_loss.numpy()
def train_step(model: Model, batch_data, opt: tf.optimizers.Optimizer, step):
feat = batch_data[1]
label = batch_data[2]
summary_step = -1 if step % 50 > 0 else step
with tf.GradientTape() as tape:
net_out = model(feat, training=True)
loss = model.loss(feat, net_out, step=summary_step)
gradient = tape.gradient(loss, sources=model.trainable_variables)
opt.apply_gradients(zip(gradient, model.trainable_variables))
code, prob, fc_cls = net_out['decoder']
if summary_step >= 0:
sim_gt = tf.expand_dims(tf.expand_dims(label_relevance(label), 0), -1)
batch_map = eval_cls_map(code.numpy(), code.numpy(), label.numpy(),
label.numpy())
tf.summary.image('sim/gt', sim_gt, step=summary_step, max_outputs=1)
tf.summary.scalar('map/train', batch_map, step=summary_step)
return code, loss
def step_train(model: BasicMAB, data: BasicData, opt: tf.optimizers.Optimizer,
t):
x = data.next()
step = t if t % 10 == 0 else -1
with tf.GradientTape() as tape:
cls, att = model(x)
loss = model.obj(cls, x['label'])
gradient = tape.gradient(loss, sources=model.trainable_variables)
opt.apply_gradients(zip(gradient, model.trainable_variables))
if t % 10 == 0:
img = tf.squeeze(att, 2)[0, :, :100]
tf.summary.scalar('train/loss', loss, step=t)
tf.summary.image('train/att',
img[tf.newaxis, :, :, tf.newaxis],
step=t)
step_test(model, data, t)
return loss.numpy()
def optimizer_minimize(self, loss: tf.Tensor, tape: tf.GradientTape,
optimizer: tf.optimizers.Optimizer, model: k.Model):
"""apply gradients
Args:
loss (tf.Tensor):
tape (tf.GradientTape):
optimizer (tf.optimizers.Optimizer):
model (k.Model):
"""
with tape:
scaled_loss = loss / self.strategy.num_replicas_in_sync
if isinstance(optimizer,
tf.keras.mixed_precision.experimental.LossScaleOptimizer):
scaled_loss = optimizer.get_scaled_loss(loss)
grad = tape.gradient(scaled_loss, model.trainable_variables)
if isinstance(optimizer,
tf.keras.mixed_precision.experimental.LossScaleOptimizer):
grad = optimizer.get_unscaled_gradients(grad)
optimizer.apply_gradients(zip(grad, model.trainable_variables))
return scaled_loss
def minimize_variables(this_loss, vars: tf.Tensor,
optimizer: tf.optimizers.Optimizer,
tape: tf.GradientTape):
grads = tape.gradient(this_loss, vars)
optimizer.apply_gradients(zip(grads, vars))
def train_single_epoch(model: tf.keras.Model,
anchors: tf.Tensor,
dataset: tf.data.Dataset,
optimizer: tf.optimizers.Optimizer,
grad_accum_steps: int,
loss_fn: LossFn,
steps: int,
epoch: int,
num_classes: int,
print_every: int = 10):
@tf.function(input_signature=[
tf.TensorSpec(shape=[None, None, None, 3], dtype=tf.float32),
tf.TensorSpec(shape=[None, None, 5], dtype=tf.float32),
tf.TensorSpec(shape=[None, None, num_classes + 1], dtype=tf.float32)
])
def train_step(images, r_targets, c_targets):
return _train_step(model=model,
optimizer=optimizer,
loss_fn=loss_fn,
images=images,
regress_targets=r_targets,
labels=c_targets)
acc_gradients = []
running_loss = tf.metrics.Mean()
running_clf_loss = tf.metrics.Mean()
running_reg_loss = tf.metrics.Mean()
for i, (images, (labels, bbs)) in enumerate(dataset):
target_reg, target_clf = utils.anchors.anchor_targets_bbox(
anchors, images, bbs, labels, num_classes)
reg_loss, clf_loss, grads = train_step(images=images,
r_targets=target_reg,
c_targets=target_clf)
if tf.math.is_nan(reg_loss) or tf.math.is_nan(clf_loss):
print('Loss NaN, skipping training step')
if len(acc_gradients) == 0:
acc_gradients = grads
else:
acc_gradients = [g1 + g2 for g1, g2 in zip(acc_gradients, grads)]
if (i + 1) % grad_accum_steps == 0:
optimizer.apply_gradients(
zip(acc_gradients, model.trainable_variables))
acc_gradients = []
running_loss(reg_loss + clf_loss)
running_clf_loss(clf_loss)
running_reg_loss(reg_loss)
if (i + 1) % print_every == 0:
lr = get_lr(optimizer)
print(f'Epoch[{epoch}] [{i}/{steps}] '
f'loss: {running_loss.result():.6f} '
f'clf. loss: {running_clf_loss.result():.6f} '
f'reg. loss: {running_reg_loss.result():.6f} '
f'learning rate: {lr:.6f}')
if len(acc_gradients) > 0:
optimizer.apply_gradients(zip(acc_gradients,
model.trainable_variables))
def _train_step(model: VAE, x: tf.Tensor,
optimizer: tf.optimizers.Optimizer):
with tf.GradientTape() as tape:
loss = compute_loss(model, x)
gradients = tape.gradient(loss, model.trainable_variables)
optimizer.apply_gradients(zip(gradients, model.trainable_variables))