def test_dense_output(self):
dense_inputs = tf.convert_to_tensor(
np.random.uniform(size=(10, 10)).astype('f'))
# Create some sparse data where multiple rows and columns are missing.
sparse_inputs = tf.SparseTensor(
indices=np.random.randint(low=0, high=10, size=(5, 2)),
values=np.random.uniform(size=(5,)).astype('f'),
dense_shape=[10, 10])
sparse_inputs = tf.sparse.reorder(sparse_inputs)
layer = keras.layers.Dense(
5,
kernel_initializer=keras.initializers.RandomUniform(),
bias_initializer=keras.initializers.RandomUniform(),
dtype='float32')
dense_outputs = layer(dense_inputs)
sparse_outpus = layer(sparse_inputs)
expected_dense = tf.add(
tf.matmul(dense_inputs, keras.backend.get_value(layer.kernel)),
keras.backend.get_value(layer.bias))
expected_sparse = tf.add(
tf.matmul(
tf.sparse.to_dense(sparse_inputs),
keras.backend.get_value(layer.kernel)),
keras.backend.get_value(layer.bias))
self.assertAllClose(dense_outputs, expected_dense)
self.assertAllClose(sparse_outpus, expected_sparse)
def mlp(self, x):
layer_1 = tf.nn.relu(
tf.add(tf.matmul(x, self.h1_weights), self.h1_bias))
layer_2 = tf.nn.relu(
tf.add(tf.matmul(layer_1, self.h2_weights), self.h2_bias))
return tf.nn.relu(
tf.add(tf.matmul(layer_2, self.out_weights), self.out_bias))
def _apply_divergence_concrete(self,scale_factor, name):
divergence_fn = (lambda pl, pr: (tf.reduce_sum(tf.add(
tf.multiply(pl,tf.subtract(tf.math.log(
tf.add(pl,tfk.backend.epsilon())), tf.math.log(pr))),
tf.multiply( tf.subtract(tfk.backend.constant(1),pl),
tf.subtract(tf.math.log(
tf.add(tf.subtract(tfk.backend.constant(1),pl),tfk.backend.epsilon())),
tf.math.log(pr)))) )
/tf.cast(scale_factor, dtype=tf.float32)))
divergence = tf.identity(
divergence_fn(self.p_post, self.p_prior),
name=name)
self.add_loss(divergence)
def test_dense_output(self):
dense_inputs = tf.convert_to_tensor(
np.random.uniform(size=(10, 10)).astype("f"))
# Create some sparse data where multiple rows and columns are missing.
sparse_inputs = tf.SparseTensor(
indices=np.random.randint(low=0, high=10, size=(5, 2)),
values=np.random.uniform(size=(5, )).astype("f"),
dense_shape=[10, 10],
)
sparse_inputs = tf.sparse.reorder(sparse_inputs)
# Create some ragged data.
ragged_inputs = tf.RaggedTensor.from_row_splits(
np.random.uniform(size=(10, 10)).astype("f"),
row_splits=[0, 4, 6, 6, 9, 10],
)
layer = keras.layers.Dense(
5,
kernel_initializer=keras.initializers.RandomUniform(),
bias_initializer=keras.initializers.RandomUniform(),
dtype="float32",
)
dense_outputs = layer(dense_inputs)
sparse_outpus = layer(sparse_inputs)
ragged_outputs = layer(ragged_inputs)
expected_dense = tf.add(
tf.matmul(dense_inputs, keras.backend.get_value(layer.kernel)),
keras.backend.get_value(layer.bias),
)
expected_sparse = tf.add(
tf.matmul(
tf.sparse.to_dense(sparse_inputs),
keras.backend.get_value(layer.kernel),
),
keras.backend.get_value(layer.bias),
)
expected_ragged_values = tf.add(
tf.matmul(ragged_inputs.flat_values,
keras.backend.get_value(layer.kernel)),
keras.backend.get_value(layer.bias),
)
expected_ragged = tf.RaggedTensor.from_row_splits(
expected_ragged_values, row_splits=[0, 4, 6, 6, 9, 10])
self.assertAllClose(dense_outputs, expected_dense)
self.assertAllClose(sparse_outpus, expected_sparse)
self.assertAllClose(ragged_outputs, expected_ragged)
def _hash_values_to_bins(self, values):
"""Converts a non-sparse tensor of values to bin indices."""
hash_bins = self.num_bins
mask = None
# If mask_value is set, the zeroth bin is reserved for it.
if self.mask_value is not None and hash_bins > 1:
hash_bins -= 1
mask = tf.equal(values, self.mask_value)
# Convert all values to strings before hashing.
if values.dtype.is_integer:
values = tf.as_string(values)
# Hash the strings.
if self.strong_hash:
values = tf.strings.to_hash_bucket_strong(values,
hash_bins,
name='hash',
key=self.salt)
else:
values = tf.strings.to_hash_bucket_fast(values,
hash_bins,
name='hash')
if mask is not None:
values = tf.add(values, tf.ones_like(values))
values = tf.where(mask, tf.zeros_like(values), values)
return values
def call(self, inputs, training=True, survival_prob=None):
"""Implementation of call().
Args:
inputs: the inputs tensor.
training: boolean, whether the model is constructed for training.
survival_prob: float, between 0 to 1, drop connect rate.
Returns:
A output tensor.
"""
x = inputs
if self._block_args.expand_ratio != 1:
x = self._relu_fn(self._bn0(self._expand_conv(x), training=training))
x = self._relu_fn(self._bn1(self._depthwise_conv(x), training=training))
if self._has_se:
se_tensor = tf.reduce_mean(
x, self._spatial_dims, keepdims=True)
se_tensor = self._se_expand(self._relu_fn(self._se_reduce(se_tensor)))
x = tf.sigmoid(se_tensor) * x
x = self._bn2(self._project_conv(x), training=training)
# Add identity so that quantization-aware training can insert quantization
# ops correctly.
x = tf.identity(x)
if self._clip_projection_output:
x = tf.clip_by_value(x, -6, 6)
if all(
s == 1 for s in self._block_args.strides
) and self._block_args.input_filters == self._block_args.output_filters:
if survival_prob:
x = utils.drop_connect(x, training, survival_prob)
x = tf.add(x, inputs)
return x
def __call__(self, step):
with tf.name_scope(self.name or "PolynomialDecay") as name:
initial_learning_rate = tf.convert_to_tensor(
self.initial_learning_rate, name="initial_learning_rate")
dtype = initial_learning_rate.dtype
end_learning_rate = tf.cast(self.end_learning_rate, dtype)
power = tf.cast(self.power, dtype)
global_step_recomp = tf.cast(step, dtype)
decay_steps_recomp = tf.cast(self.decay_steps, dtype)
if self.cycle:
# Find the first multiple of decay_steps that is bigger than
# global_step. If global_step is zero set the multiplier to 1
multiplier = tf.where(
tf.equal(global_step_recomp, 0), 1.0,
tf.math.ceil(global_step_recomp / self.decay_steps))
decay_steps_recomp = tf.multiply(decay_steps_recomp,
multiplier)
else:
# Make sure that the global_step used is not bigger than decay_steps.
global_step_recomp = tf.minimum(global_step_recomp,
decay_steps_recomp)
p = tf.divide(global_step_recomp, decay_steps_recomp)
return tf.add(tf.multiply(
initial_learning_rate - end_learning_rate,
tf.pow(1 - p, power)),
end_learning_rate,
name=name)
def _apply_variational_kernel(self, inputs):
if not isinstance(
self.kernel_posterior, independent_lib.Independent
) or not isinstance(self.kernel_posterior.distribution, normal_lib.Normal):
self.kernel_posterior_tensor = self.kernel_posterior_tensor_fn(
self.kernel_posterior
)
self.kernel_posterior_affine = None
self.kernel_posterior_affine_tensor = None
outputs = self._convolution_op(inputs, self.kernel_posterior_tensor)
return outputs
else:
self.kernel_posterior_affine = normal_lib.Normal(
loc=tf.zeros_like(self.kernel_posterior.distribution.loc),
scale=self.kernel_posterior.distribution.scale,
)
self.kernel_posterior_affine_tensor = self.kernel_posterior_tensor_fn(
self.kernel_posterior_affine
)
self.kernel_posterior_tensor = None
outputs_m = self._convolution_op(
inputs, self.kernel_posterior.distribution.loc
)
outputs_v = self._convolution_op(
tf.square(inputs),
tf.square(self.kernel_posterior.distribution.stddev()),
)
k_size = tf_layers_util.normalize_tuple(self.kernel_size, 3, "k_size")
g_shape = [1 for i in k_size] + [1, self.filters]
outputs_e = tf.random.normal(shape=g_shape, dtype=self.dtype)
if self.is_mc:
err = tf.sqrt(tf.add(outputs_v, tf.keras.backend.epsilon())) * outputs_e
return outputs_m + err
else:
return outputs_m
def _set_values_using_indicator(self, x, indicator, val):
"""Set the indicated fields of x to val.
Args:
x: tensor.
indicator: boolean with same shape as x.
val: scalar with value to set.
Returns:
modified tensor.
"""
indicator = tf.cast(indicator, x.dtype)
return tf.add(tf.multiply(x, 1 - indicator), val * indicator)
def __call__(self, step):
with tf.name_scope(self.name or "InverseTimeDecay") as name:
initial_learning_rate = tf.convert_to_tensor(
self.initial_learning_rate, name="initial_learning_rate")
dtype = initial_learning_rate.dtype
decay_steps = tf.cast(self.decay_steps, dtype)
decay_rate = tf.cast(self.decay_rate, dtype)
global_step_recomp = tf.cast(step, dtype)
p = global_step_recomp / decay_steps
if self.staircase:
p = tf.floor(p)
const = tf.cast(tf.constant(1), dtype)
denom = tf.add(const, tf.multiply(decay_rate, p))
return tf.divide(initial_learning_rate, denom, name=name)
def _hash_values_to_bins(self, values):
"""Converts a non-sparse tensor of values to bin indices."""
str_to_hash_bucket = self._get_string_to_hash_bucket_fn()
num_available_bins = self.num_bins
mask = None
# If mask_value is set, the zeroth bin is reserved for it.
if self.mask_value is not None and num_available_bins > 1:
num_available_bins -= 1
mask = tf.equal(values, self.mask_value)
# Convert all values to strings before hashing.
if values.dtype.is_integer:
values = tf.as_string(values)
values = str_to_hash_bucket(values, num_available_bins, name='hash')
if mask is not None:
values = tf.add(values, tf.compat.v1.ones_like(values))
values = tf.compat.v1.where(mask, tf.compat.v1.zeros_like(values), values)
return values
def _create_ensemble_logits_helper(self,
weighted_subnetworks,
bias,
summary,
key=None,
index=None):
"""Returns the AdaNet ensemble logits and regularization term for key."""
subnetwork_logits = []
for weighted_subnetwork in weighted_subnetworks:
subnetwork_logits.append(
_lookup_if_dict(weighted_subnetwork.logits, key))
with tf_compat.v1.variable_scope(
"logits_{}".format(index) if index else "logits"):
ensemble_logits = _lookup_if_dict(bias, key)
for logits in subnetwork_logits:
ensemble_logits = tf.add(ensemble_logits, logits)
return ensemble_logits
def _dft_magnitude(self, signal):
"""Compute DFT and then its magnitude.
It is avoiding tflite incompatible ops.
Args:
signal: has dims [..., frame_size]
Returns:
magnitude_spectrogram: with dims [..., fft_size]
"""
real_spectrum = tf.matmul(signal, self.real_dft_tensor)
imag_spectrum = tf.matmul(signal, self.imag_dft_tensor)
magnitude_spectrum = tf.add(real_spectrum * real_spectrum,
imag_spectrum * imag_spectrum)
# if self.magnitude_squared:
# return magnitude_spectrum
# else:
return tf.sqrt(magnitude_spectrum)
def cal_longest_subsequence(softmaxed_logits):
int_logits = tf.dtypes.cast(tf.round(softmaxed_logits), dtype=tf.int32)
index_tensor = tf.range(softmaxed_logits.shape[1], dtype=tf.int32)
t_index = tf.reshape(index_tensor, [softmaxed_logits.shape[1], 1])
new_seq = tf.transpose(tf.matmul(int_logits, t_index))[0].numpy().tolist()
# new_seq = [3,2,4,5,6,5,5,6,7]
# print(new_seq)
subseq = []
indexseq = []
for i in range(len(new_seq)):
if i == 0:
subseq.append(new_seq[i])
indexseq.append(i)
else:
if new_seq[i] > subseq[-1]:
subseq.append(new_seq[i])
indexseq.append(i)
elif new_seq[i] < subseq[0]:
subseq[0] = new_seq[i]
indexseq[0] = i
else:
index = binarySearch(subseq, 0, len(subseq) - 1, new_seq[i])
if index != -1:
subseq[index] = new_seq[i]
indexseq[index] = i
# print(subseq)
# print(indexseq)
subseq_tensor = tf.reshape(subseq, [1, -1])
index_tensor = tf.reshape(indexseq, [1, -1])
# print(subseq_tensor,index_tensor)
te = tf.subtract(subseq_tensor, index_tensor)
# print(te)
minus_result = tf.square(tf.subtract(subseq_tensor, index_tensor))
one_tensor = tf.ones([1, len(subseq)], tf.int32)
result = tf.divide(one_tensor, tf.add(one_tensor, minus_result))
# return tf.reduce_sum(result)
return subseq
def main(_):
# Wrap the TensorFlow Session object for debugging.
# TODO(anthonyjliu): Enable debugger from flags
if FLAGS.debug and FLAGS.tensorboard_debug_address:
raise ValueError(
"The --debug and --tensorboard_debug_address flags are mutually "
"exclusive.")
if FLAGS.debug:
raise NotImplementedError(
"tfdbg v2 support for debug_fibonacci is not implemented yet")
elif FLAGS.tensorboard_debug_address:
raise NotImplementedError(
"tfdbg v2 support for debug_fibonacci is not implemented yet")
# Construct the TensorFlow network.
n0 = tf.constant(np.ones([FLAGS.tensor_size] * 2), dtype=tf.int32)
n1 = tf.constant(np.ones([FLAGS.tensor_size] * 2), dtype=tf.int32)
for _ in xrange(2, FLAGS.length):
n0, n1 = n1, tf.add(n0, n1)
print("Fibonacci number at position %d:\n%s" % (FLAGS.length, n1.numpy()))
def maybe_update_alpha():
"""Maybe update the alpha param.
Checks if global_step is between begin_compression_step and
end_compression_step, and if the current training step is a
compression step.
Returns:
Boolean tensor whether the training step is a compression step.
"""
is_step_within_compression_range = tf.logical_and(
tf.greater_equal(tf.cast(self._global_step, tf.int32),
self._spec.begin_compression_step),
tf.logical_or(
tf.less_equal(tf.cast(self._global_step, tf.int32),
self._spec.end_compression_step),
tf.less(self._spec.end_compression_step, 0)))
is_compression_step = tf.less_equal(
tf.add(self.last_alpha_update_step,
self._spec.compression_frequency),
tf.cast(self._global_step, tf.int32))
return tf.logical_and(is_step_within_compression_range,
is_compression_step)
def from_importance_weights(log_rhos,
discounts,
rewards,
values,
bootstrap_value,
clip_rho_threshold=1.0,
clip_pg_rho_threshold=1.0,
name='vtrace_from_importance_weights'):
r"""V-trace from log importance weights.
Calculates V-trace actor critic targets as described in
"IMPALA: Scalable Distributed Deep-RL with
Importance Weighted Actor-Learner Architectures"
by Espeholt, Soyer, Munos et al.
In the notation used throughout documentation and comments, T refers to the
time dimension ranging from 0 to T-1. B refers to the batch size and
NUM_ACTIONS refers to the number of actions. This code also supports the
case where all tensors have the same number of additional dimensions, e.g.,
`rewards` is [T, B, C], `values` is [T, B, C], `bootstrap_value` is [B, C].
Args:
log_rhos: A float32 tensor of shape [T, B, NUM_ACTIONS] representing the log
importance sampling weights, i.e. log(target_policy(a) /
behaviour_policy(a)). V-trace performs operations on rhos in log-space for
numerical stability.
discounts: A float32 tensor of shape [T, B] with discounts encountered when
following the behaviour policy.
rewards: A float32 tensor of shape [T, B] containing rewards generated by
following the behaviour policy.
values: A float32 tensor of shape [T, B] with the value function estimates
wrt. the target policy.
bootstrap_value: A float32 of shape [B] with the value function estimate at
time T.
clip_rho_threshold: A scalar float32 tensor with the clipping threshold for
importance weights (rho) when calculating the baseline targets (vs).
rho^bar in the paper. If None, no clipping is applied.
clip_pg_rho_threshold: A scalar float32 tensor with the clipping threshold
on rho_s in \rho_s \delta log \pi(a|x) (r + \gamma v_{s+1} - V(x_s)). If
None, no clipping is applied.
name: The name scope that all V-trace operations will be created in.
Returns:
A VTraceReturns namedtuple (vs, pg_advantages) where:
vs: A float32 tensor of shape [T, B]. Can be used as target to
train a baseline (V(x_t) - vs_t)^2.
pg_advantages: A float32 tensor of shape [T, B]. Can be used as the
advantage in the calculation of policy gradients.
"""
log_rhos = tf.convert_to_tensor(log_rhos, dtype=tf.float32)
discounts = tf.convert_to_tensor(discounts, dtype=tf.float32)
rewards = tf.convert_to_tensor(rewards, dtype=tf.float32)
values = tf.convert_to_tensor(values, dtype=tf.float32)
bootstrap_value = tf.convert_to_tensor(bootstrap_value, dtype=tf.float32)
if clip_rho_threshold is not None:
clip_rho_threshold = tf.convert_to_tensor(clip_rho_threshold,
dtype=tf.float32)
if clip_pg_rho_threshold is not None:
clip_pg_rho_threshold = tf.convert_to_tensor(clip_pg_rho_threshold,
dtype=tf.float32)
# Make sure tensor ranks are consistent.
rho_rank = log_rhos.shape.ndims # Usually 2.
values.shape.assert_has_rank(rho_rank)
bootstrap_value.shape.assert_has_rank(rho_rank - 1)
discounts.shape.assert_has_rank(rho_rank)
rewards.shape.assert_has_rank(rho_rank)
if clip_rho_threshold is not None:
clip_rho_threshold.shape.assert_has_rank(0)
if clip_pg_rho_threshold is not None:
clip_pg_rho_threshold.shape.assert_has_rank(0)
with tf.name_scope(name):
rhos = tf.exp(log_rhos)
if clip_rho_threshold is not None:
clipped_rhos = tf.minimum(clip_rho_threshold,
rhos,
name='clipped_rhos')
else:
clipped_rhos = rhos
cs = tf.minimum(1.0, rhos, name='cs')
# Append bootstrapped value to get [v1, ..., v_t+1]
values_t_plus_1 = tf.concat(
[values[1:], tf.expand_dims(bootstrap_value, 0)], axis=0)
deltas = clipped_rhos * (rewards + discounts * values_t_plus_1 -
values)
acc = tf.zeros_like(bootstrap_value)
vs_minus_v_xs = []
for i in range(int(discounts.shape[0]) - 1, -1, -1):
discount, c, delta = discounts[i], cs[i], deltas[i]
acc = delta + discount * c * acc
vs_minus_v_xs.append(acc)
# Reversing
vs_minus_v_xs = vs_minus_v_xs[::-1]
# Add V(x_s) to get v_s.
vs = tf.add(vs_minus_v_xs, values, name='vs')
# Advantage for policy gradient.
vs_t_plus_1 = tf.concat(
[vs[1:], tf.expand_dims(bootstrap_value, 0)], axis=0)
if clip_pg_rho_threshold is not None:
clipped_pg_rhos = tf.minimum(clip_pg_rho_threshold,
rhos,
name='clipped_pg_rhos')
else:
clipped_pg_rhos = rhos
pg_advantages = (clipped_pg_rhos *
(rewards + discounts * vs_t_plus_1 - values))
# Make sure no gradients backpropagated through the returned values.
return VTraceReturns(vs=tf.stop_gradient(vs),
pg_advantages=tf.stop_gradient(pg_advantages))
def add_or_or(x1, x2):
if x1.dtype == tf.bool:
assert x2.dtype == tf.bool
return tf.logical_or(x1, x2)
return tf.add(x1, x2)
def compute_loss_and_metrics(mu,
log_sigma_sq,
regression_targets,
labels,
task_type,
model_uncertainty,
loss_config,
regularization_loss=0.,
confidence_interval=95,
mode='train'):
"""Computes loss statistics and other metrics."""
scalars_to_log = dict()
vectors_to_log = dict()
scalars_to_log['regularization_loss'] = regularization_loss
vectors_to_log['mu'] = mu
if task_type == TASK_CLASSIFICATION:
cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(
logits=mu, labels=labels, name='cross_entropy')
classification_loss = tf.reduce_mean(cross_entropy, name='class_loss')
total_loss = classification_loss
sigma = None
scalars_to_log['classification_loss'] = classification_loss
predicted_labels = tf.argmax(mu, axis=1)
correct_predictions = equal32(predicted_labels, labels)
else:
regression_loss = mse_loss(mu, regression_targets)
if 'mse_normalize' in loss_config and loss_config['mse_normalize']:
assert task_type in [
TASK_GROUNDED_UNNORMALIZED_REGRESSION,
TASK_NORMALIZED_REGRESSION
]
regression_loss = normalize_regression_loss(regression_loss, mu)
avg_regression_loss = tf.reduce_mean(regression_loss)
vectors_to_log['regression_loss'] = regression_loss
scalars_to_log['regression_loss'] = avg_regression_loss
scalars_to_log['avg_mu'] = tf.reduce_mean(mu)
scalars_to_log['var_mu'] = tf.reduce_mean(
mse_loss(mu, tf.reduce_mean(mu)))
predicted_labels = tf.cast(mu > 0, tf.int64)
correct_predictions = equal32(predicted_labels, labels)
if model_uncertainty:
# This implements Eq. (1) in https://arxiv.org/pdf/1612.01474.pdf
inv_sigma_sq = tf.math.exp(-log_sigma_sq)
scaled_regression_loss = regression_loss * inv_sigma_sq
scaled_regression_loss = tf.reduce_mean(scaled_regression_loss)
uncertainty_loss = tf.reduce_mean(log_sigma_sq)
total_loss = uncertainty_loss + scaled_regression_loss
scalars_to_log['uncertainty_loss'] = uncertainty_loss
scalars_to_log['scaled_regression_loss'] = scaled_regression_loss
scalars_to_log['uncertainty_plus_scaled_regression'] = total_loss
sigma = tf.math.exp(log_sigma_sq / 2.)
vectors_to_log['sigma'] = sigma
scalars_to_log['avg_sigma'] = tf.reduce_mean(sigma)
var_sigma = tf.reduce_mean(mse_loss(sigma, tf.reduce_mean(sigma)))
scalars_to_log['var_sigma'] = var_sigma
# Compute # of labels that fall into the confidence interval.
std_factor = get_std_factor_from_confidence_percent(
confidence_interval)
lower_bound = mu - std_factor * sigma
upper_bound = mu + std_factor * sigma
preds = tf.logical_and(tf.greater(regression_targets, lower_bound),
tf.less(regression_targets, upper_bound))
percent_in_conf_interval = tf.reduce_mean(
tf.cast(preds, tf.float32))
scalars_to_log[
'percent_in_conf_interval'] = percent_in_conf_interval * 100
error_sigma_corr = tfp.stats.correlation(x=regression_loss,
y=sigma,
event_axis=None)
scalars_to_log['error_sigma_correlation'] = error_sigma_corr
dists = tfp.distributions.Normal(mu, sigma)
probs = dists.prob(regression_targets)
scalars_to_log['avg_prob'] = tf.reduce_mean(probs)
else:
total_loss = avg_regression_loss
loss_name = str(mode) + '_loss'
total_loss = tf.add(total_loss, regularization_loss, name=loss_name)
scalars_to_log[loss_name] = total_loss
vectors_to_log['correct_predictions'] = correct_predictions
scalars_to_log['prediction_accuracy'] = tf.reduce_mean(correct_predictions)
# Validate that metrics outputted are exactly what is expected
expected = get_all_metric_names(task_type, model_uncertainty, loss_config,
mode, False)
assert set(expected) == set(scalars_to_log.keys())
return scalars_to_log, vectors_to_log
def train_loop_body(step):
train_op = optimizer.minimize(
build_loss_fn if tf.executing_eagerly() else build_loss_fn())
return tf.tuple(tensors=[tf.add(step, 1)], control_inputs=[train_op])
