def _test_image_producer(self, batch_group_size, put_slower_than_get):
# We use the variable x to simulate a staging area of images. x represents
# the number of batches in the staging area.
x = tf.Variable(0, dtype=tf.int32)
if put_slower_than_get:
put_dep = self._slow_tensorflow_op()
get_dep = tf.no_op()
else:
put_dep = tf.no_op()
get_dep = self._slow_tensorflow_op()
with tf.control_dependencies([put_dep]):
put_op = x.assign_add(batch_group_size, use_locking=True)
with tf.control_dependencies([get_dep]):
get_op = x.assign_sub(1, use_locking=True)
with self.test_session() as sess:
sess.run(tf.variables_initializer([x]))
image_producer = cnn_util.ImageProducer(sess, put_op, batch_group_size,
use_python32_barrier=False)
image_producer.start()
for _ in range(5 * batch_group_size):
sess.run(get_op)
# We assert x is nonnegative, to ensure image_producer never causes
# an unstage op to block. We assert x is at most 2 * batch_group_size,
# to ensure it doesn't use too much memory by storing too many batches
# in the staging area.
self.assertGreaterEqual(sess.run(x), 0)
self.assertLessEqual(sess.run(x), 2 * batch_group_size)
image_producer.notify_image_consumption()
self.assertGreaterEqual(sess.run(x), 0)
self.assertLessEqual(sess.run(x), 2 * batch_group_size)
image_producer.done()
time.sleep(0.1)
self.assertGreaterEqual(sess.run(x), 0)
self.assertLessEqual(sess.run(x), 2 * batch_group_size)
def model_fn(features, labels, mode):
models = []
logits = []
classes = []
init_op = [tf.train.get_or_create_global_step().initializer]
for (i, model_name) in enumerate(FLAGS.model_name.split(',')):
with tf.device("/gpu:%d" % i):
network_fn = getattr(nets, model_name)
models.append(network_fn(features, is_training=False))
logits.append(models[i].get_outputs()[-2])
classes.append(tf.argmax(logits[i], axis=1))
if FLAGS.checkpoint_path is None:
init_op.extend(models[i].pretrained())
scaffold = None
if FLAGS.checkpoint_path is None:
scaffold = tf.train.Scaffold(init_op=init_op)
loss = []
for i in range(len(models)):
cross_entropy = tf.losses.sparse_softmax_cross_entropy(
logits=logits[i], labels=labels)
loss.append(cross_entropy)
loss = tf.reduce_sum(loss)
metrics = None
if mode == tf.estimator.ModeKeys.EVAL:
metrics = {}
for i in range(len(models)):
top1 = tf.metrics.accuracy(labels=labels, predictions=classes[i])
top5 = contrib.metrics.streaming_sparse_recall_at_k(logits[i],
tf.cast(
labels,
tf.int64),
k=5)
size = sum(
[w.shape.num_elements() for w in models[i].get_weights()])
run_meta = tf.RunMetadata()
opts = tf.profiler.ProfileOptionBuilder.float_operation()
opts['output'] = 'none'
flops = tf.profiler.profile(tf.get_default_graph(),
run_meta=run_meta,
options=opts)
metrics.update({
"%dTop1" % i:
top1,
"%dTop5" % i:
top5,
"%dMAC" % i: (tf.constant(flops.total_float_ops), tf.no_op()),
"%dSize" % i: (tf.constant(size), tf.no_op())
})
return tf.estimator.EstimatorSpec(mode=mode,
scaffold=scaffold,
predictions=None,
loss=loss,
train_op=None,
eval_metric_ops=metrics,
export_outputs=None)
def train_operation(self, loss, varaiables):
"""
Training operation
Parameters
----------
loss : tensor
current loss of the model
varaiables : list
list of the variables that should be trained
Returns
-------
tensorflow operation
the training operation
"""
if len(varaiables) > 0:
gradients = self.optimizer.get_gradients(loss, varaiables)
grads = list(zip(gradients, varaiables))
clipped_grads = []
# Add histograms for gradients.
for grad, var in grads:
if grad is not None:
# try catching nans
grad = tf.where(tf.math.is_finite(grad), grad,
tf.zeros_like(grad))
# limit the maximum change to any variable to 0.5
grad = tf.clip_by_value(grad, -0.5/self.initial_lr,
0.5/self.initial_lr)
clipped_grads += [(grad, var)]
tf.summary.histogram(var.op.name + '/gradients', grad)
# safe plotting
var_plot = tf.where(tf.math.is_finite(var), var,
tf.zeros_like(var))
tf.summary.histogram(var.op.name, var_plot)
# process the updates of the batch norm layers
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
update_ops = []
for lay in self.context.update_ops:
update_ops += lay.updates
if len(clipped_grads) > 0:
apply_gradient_op = \
self.optimizer.apply_gradients(clipped_grads)
with tf.control_dependencies([apply_gradient_op] + update_ops):
train_parameters_op = tf.no_op(name='train')
else:
with tf.control_dependencies(update_ops):
train_parameters_op = tf.no_op(name='train')
return train_parameters_op
else:
return tf.no_op(name='train')
def old_old_weight_update_op(self):
with tf.name_scope(self._spec.name):
if self._spec.prune_option != 'second_order_gradient':
return tf.no_op('gradient_update_no_op')
if not self._assign_old_old_weight_ops:
self._get_assign_old_old_weight_ops()
with tf.control_dependencies(self._assign_old_old_weight_ops):
tf.logging.info('Updating old old weights.')
return tf.no_op('old_old_weight_update')
def gradient_update_op(self):
with tf.name_scope(self._spec.name):
if self._spec.prune_option not in ('first_order_gradient',
'second_order_gradient'):
return tf.no_op('gradient_update_no_op')
if not self._assign_gradient_ops:
self._get_assign_gradient_ops()
with tf.control_dependencies([
tf.assign(self._last_gradient_update_step,
self._global_step,
name='last_gradient_update_step_assign')
]):
with tf.control_dependencies(self._assign_gradient_ops):
tf.logging.info('Updating gradients.')
return tf.no_op('gradient_update')
def randomize_op(self):
"""Randomizes the augmentation according to the `random_config`."""
if self.children:
return tf.group([child.randomize_op() for child in self.children])
if self.random_config is None:
return tf.no_op()
config = self.random_config
assign_rotate = self.rotate.assign_bool(config.rotation_probability)
assign_smooth = self.smooth.assign_bool(config.smooth_probability)
assign_contrast = self.contrast.assign_bool(
config.contrast_probability)
assign_negate = self.negate.assign_bool(config.negate_probability)
assign_resize = self.resize.assign_bool(config.resize_probability)
assign_roll = self.roll.assign_bool(config.roll_probability)
angle_range = config.angle_range / 180.0 * np.pi
assign_angle = self.angle.assign_uniform(scale=angle_range)
assign_size = self.size.assign_uniform(scale=1.0)
assign_alpha = self.alpha.assign_uniform(scale=1.0)
assign_roll_x = self.roll_x.assign_uniform(scale=config.roll_range)
assign_roll_y = self.roll_y.assign_uniform(scale=config.roll_range)
assign_rotate_90 = self.rotate_90_times.assign_uniform(scale=2.0)
return tf.group(assign_rotate, assign_smooth, assign_contrast,
assign_angle, assign_negate, assign_resize,
assign_alpha, assign_size, assign_roll, assign_roll_x,
assign_roll_y, assign_rotate_90)
def _add_train_graph(self):
"""Define the training operation."""
mc = self.mc
self.global_step = tf.Variable(0, name='global_step', trainable=False)
lr = tf.train.exponential_decay(mc.LEARNING_RATE,
self.global_step,
mc.DECAY_STEPS,
mc.LR_DECAY_FACTOR,
staircase=True)
tf.summary.scalar('learning_rate', lr)
_add_loss_summaries(self.loss)
opt = tf.train.MomentumOptimizer(learning_rate=lr,
momentum=mc.MOMENTUM)
grads_vars = opt.compute_gradients(self.loss, tf.trainable_variables())
with tf.variable_scope('clip_gradient') as scope:
for i, (grad, var) in enumerate(grads_vars):
grads_vars[i] = (tf.clip_by_norm(grad, mc.MAX_GRAD_NORM), var)
apply_gradient_op = opt.apply_gradients(grads_vars,
global_step=self.global_step)
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)
for grad, var in grads_vars:
if grad is not None:
tf.summary.histogram(var.op.name + '/gradients', grad)
with tf.control_dependencies([apply_gradient_op]):
self.train_op = tf.no_op(name='train')
def LogAndSummarizeMetrics(metrics, use_streaming_mean=True):
"""Logs and summarizes metrics.
Metrics are added to the LOGGING_OUTPUTS collection.
Args:
metrics: A dictionary of scalar metrics.
use_streaming_mean: If true, the metrics will be averaged using a running
mean.
Returns:
If use_streaming_mean is true, then this will be the op that you need to
regularly call to update the running mean. Otherwise, this is a no-op.
"""
prefix = tf.get_default_graph().get_name_scope()
if prefix:
prefix += "/"
logging_collection = tf.get_collection_ref(LOGGING_OUTPUTS)
update_ops = [tf.no_op()]
for name, value in metrics.items():
if use_streaming_mean:
value, update_op = tf.metrics.mean(value)
update_ops.append(update_op)
logging_collection.append((prefix + name, value))
tf.summary.scalar(name, value)
return tf.group(*update_ops)
def build_g_loss(self, labels, outputs):
self.g_log_losses = []
update_ops = []
loss_key = 'GeneratorLoss'
with tf.variable_scope(loss_key):
# L1 loss
l1_loss = tf.losses.absolute_difference(labels, outputs, 1.0,
loss_collection=None)
tf.losses.add_loss(l1_loss)
update_ops.append(self.loss_summary('l1_loss', l1_loss, self.g_log_losses))
# SSIM loss
labelsY = layers.RGB2Y(labels, self.data_format)
outputsY = layers.RGB2Y(outputs, self.data_format)
ssim_loss = 1 - layers.MS_SSIM2(labelsY, outputsY, sigma=[1.5, 4.0, 10.0],
L=1, norm=False, data_format=self.data_format)
tf.losses.add_loss(ssim_loss * 0.1)
update_ops.append(self.loss_summary('ssim_loss', ssim_loss, self.g_log_losses))
# regularization loss
reg_losses = tf.losses.get_regularization_losses('Generator')
reg_loss = tf.add_n(reg_losses)
# tf.losses.add_loss(reg_loss)
update_ops.append(self.loss_summary('reg_loss', reg_loss))
# final loss
losses = tf.losses.get_losses(loss_key)
self.g_loss = tf.add_n(losses, 'total_loss')
update_ops.append(self.loss_summary('loss', self.g_loss))
# accumulate operator
with tf.control_dependencies(update_ops):
self.g_losses_acc = tf.no_op('accumulator')
def backward():
x = tf.placeholder(tf.float32,
shape=[
BATCH_SIZE, lenet5_forward.IMAGE_SIZE,
lenet5_forward.IMAGE_SIZE,
lenet5_forward.NUM_CHANNELS
])
y_ = tf.placeholder(tf.float32, [None, lenet5_forward.OUTPUT_NODE])
y = lenet5_forward.forward(x, True, REGULARIZER)
global_step = tf.Variable(0, trainable=False)
ce = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=y,
labels=tf.argmax(
y_, 1))
cem = tf.reduce_mean(ce)
loss = cem + tf.add_n(tf.get_collection('losses'))
learning_rate = tf.train.exponential_decay(LEARNING_RATE_BASE,
global_step,
train_num_examples / BATCH_SIZE,
LEARNING_RETE_DECAY,
staircase=True)
train_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(
loss, global_step=global_step)
ema = tf.train.ExponentialMovingAverage(MOVING_AVERAGE_DECAY, global_step)
ema_op = ema.apply(tf.trainable_variables())
with tf.control_dependencies([train_step, ema_op]):
train_op = tf.no_op(name="train")
saver = tf.train.Saver()
img_batch, label_batch = lenet5_generateds.get_tfRecord(BATCH_SIZE, True)
with tf.Session() as sess:
init_op = tf.global_variables_initializer()
sess.run(init_op)
ckpt = tf.train.get_checkpoint_state(MODEL_SAVE_PATH)
if ckpt and ckpt.model_checkpoint_path:
saver.restore(sess, ckpt.model_checkpoint_path)
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(sess=sess, coord=coord)
for i in range(STEPS):
xs, ys = sess.run([img_batch, label_batch])
_, loss_val, step = sess.run([train_op, loss, global_step],
feed_dict={
x: xs,
y_: ys
})
if i % 100 == 0:
print(
"After %d training step(s), loss on training batch is %g."
% (step, loss_val))
saver.save(sess,
os.path.join(MODEL_SAVE_PATH, MODEL_NAME),
global_step=global_step)
coord.request_stop()
coord.join(threads)
def batch_norm(x, is_training, bn_decay):
input_dims = x.get_shape()[-1].value
moment_dims = list(range(len(x.get_shape()) - 1))
beta = tf.Variable(tf.zeros_initializer()(shape=[input_dims]),
dtype=tf.float32,
trainable=True,
name='beta')
gamma = tf.Variable(tf.ones_initializer()(shape=[input_dims]),
dtype=tf.float32,
trainable=True,
name='gamma')
batch_mean, batch_var = tf.nn.moments(x, moment_dims, name='moments')
decay = bn_decay if bn_decay is not None else 0.9
ema = tf.train.ExponentialMovingAverage(decay=decay)
# Operator that maintains moving averages of variables.
ema_apply_op = tf.cond(is_training,
lambda: ema.apply([batch_mean, batch_var]),
lambda: tf.no_op())
# Update moving average and return current batch's avg and var.
def mean_var_with_update():
with tf.control_dependencies([ema_apply_op]):
return tf.identity(batch_mean), tf.identity(batch_var)
# ema.average returns the Variable holding the average of var.
mean, var = tf.cond(
is_training, mean_var_with_update, lambda:
(ema.average(batch_mean), ema.average(batch_var)))
x = tf.nn.batch_normalization(x, mean, var, beta, gamma, 1e-3)
return x
def _decay_weights_op(self, var, learning_rate, apply_state):
do_decay = self._do_use_weight_decay(var.name)
if do_decay:
return var.assign_sub(learning_rate * var *
apply_state['weight_decay_rate'],
use_locking=self._use_locking)
return tf.no_op()
def summary(images, name):
"""As a hack, saves image summaries by adding to `eval_metric_ops`."""
images = tf.saturate_cast(images * 255 + 0.5, tf.uint8)
eval_metric_ops[name] = (tf.summary.image(name,
images,
max_outputs=2),
tf.no_op())
def _set_up_cache(self):
"""Replace fields with cached versions.
Returns:
TensorFlow op to update the cache.
"""
return tf.no_op() # By default, don't cache.
def testReuseVars(self):
height, width = 3, 3
with self.test_session() as sess:
image_shape = (10, height, width, 3)
image_values = np.random.rand(*image_shape)
expected_mean = np.mean(image_values, axis=(0, 1, 2))
expected_var = np.var(image_values, axis=(0, 1, 2))
images = tf.constant(image_values, shape=image_shape, dtype=tf.float32)
output = ops.batch_norm(images, decay=0.1, is_training=False)
update_ops = tf.get_collection(ops.UPDATE_OPS_COLLECTION)
with tf.control_dependencies(update_ops):
barrier = tf.no_op(name='gradient_barrier')
output = control_flow_ops.with_dependencies([barrier], output)
# Initialize all variables
sess.run(tf.global_variables_initializer())
moving_mean = variables.get_variables('BatchNorm/moving_mean')[0]
moving_variance = variables.get_variables('BatchNorm/moving_variance')[0]
mean, variance = sess.run([moving_mean, moving_variance])
# After initialization moving_mean == 0 and moving_variance == 1.
self.assertAllClose(mean, [0] * 3)
self.assertAllClose(variance, [1] * 3)
# Simulate assigment from saver restore.
init_assigns = [tf.assign(moving_mean, expected_mean),
tf.assign(moving_variance, expected_var)]
sess.run(init_assigns)
for _ in range(10):
sess.run([output], {images: np.random.rand(*image_shape)})
mean = moving_mean.eval()
variance = moving_variance.eval()
# Although we feed different images, the moving_mean and moving_variance
# shouldn't change.
self.assertAllClose(mean, expected_mean)
self.assertAllClose(variance, expected_var)
def testComputeMovingVars(self):
height, width = 3, 3
with self.test_session() as sess:
image_shape = (10, height, width, 3)
image_values = np.random.rand(*image_shape)
expected_mean = np.mean(image_values, axis=(0, 1, 2))
expected_var = np.var(image_values, axis=(0, 1, 2))
images = tf.constant(image_values, shape=image_shape, dtype=tf.float32)
output = ops.batch_norm(images, decay=0.1)
update_ops = tf.get_collection(ops.UPDATE_OPS_COLLECTION)
with tf.control_dependencies(update_ops):
barrier = tf.no_op(name='gradient_barrier')
output = control_flow_ops.with_dependencies([barrier], output)
# Initialize all variables
sess.run(tf.global_variables_initializer())
moving_mean = variables.get_variables('BatchNorm/moving_mean')[0]
moving_variance = variables.get_variables('BatchNorm/moving_variance')[0]
mean, variance = sess.run([moving_mean, moving_variance])
# After initialization moving_mean == 0 and moving_variance == 1.
self.assertAllClose(mean, [0] * 3)
self.assertAllClose(variance, [1] * 3)
for _ in range(10):
sess.run([output])
mean = moving_mean.eval()
variance = moving_variance.eval()
# After 10 updates with decay 0.1 moving_mean == expected_mean and
# moving_variance == expected_var.
self.assertAllClose(mean, expected_mean)
self.assertAllClose(variance, expected_var)
def train(total_loss,
global_step,
optimizer,
learning_rate,
moving_average_decay,
update_gradient_vars,
log_histograms=True):
# Generate moving averages of all losses and associated summaries.
loss_averages_op = _add_loss_summaries(total_loss)
# Compute gradients.
with tf.control_dependencies([loss_averages_op]):
if optimizer == 'ADAGRAD':
opt = tf.train.AdagradOptimizer(learning_rate)
elif optimizer == 'ADADELTA':
opt = tf.train.AdadeltaOptimizer(learning_rate,
rho=0.9,
epsilon=1e-6)
elif optimizer == 'ADAM':
opt = tf.train.AdamOptimizer(learning_rate,
beta1=0.9,
beta2=0.999,
epsilon=0.1)
elif optimizer == 'RMSPROP':
opt = tf.train.RMSPropOptimizer(learning_rate,
decay=0.9,
momentum=0.9,
epsilon=1.0)
elif optimizer == 'MOM':
opt = tf.train.MomentumOptimizer(learning_rate,
0.9,
use_nesterov=True)
else:
raise ValueError('Invalid optimization algorithm')
grads = opt.compute_gradients(total_loss, update_gradient_vars)
# Apply gradients.
apply_gradient_op = opt.apply_gradients(grads, global_step=global_step)
# Add histograms for trainable variables.
if log_histograms:
for var in tf.trainable_variables():
tf.summary.histogram(var.op.name, var)
# Add histograms for gradients.
if log_histograms:
for grad, var in grads:
if grad is not None:
tf.summary.histogram(var.op.name + '/gradients', grad)
# Track the moving averages of all trainable variables.
variable_averages = tf.train.ExponentialMovingAverage(
moving_average_decay, global_step)
variables_averages_op = variable_averages.apply(tf.trainable_variables())
with tf.control_dependencies([apply_gradient_op, variables_averages_op]):
train_op = tf.no_op(name='train')
return train_op
def construct_model(input_tensors, encoder_w0, decoder0, prefix=None):
"""Construct model."""
facto = tf.placeholder_with_default(1.0, ())
context_xs = input_tensors['inputa']
context_ys = input_tensors['labela']
target_xs = input_tensors['inputb']
target_ys = input_tensors['labelb']
# sample ws ~ w|(x_all,a), rs = T(ws, ys), r = mean(rs), z = T(r)
# x_all = tf.concat([context_xs, target_xs], axis=1) #n_task * 20 * (128*128)
# y_all = tf.concat([context_ys, target_ys], axis=1)
x_all = context_xs
y_all = context_ys
# n_task * [n_im] * d_z
if 'train' in prefix:
z_samples = xy_to_z(x_all, y_all, encoder_w0) * facto
else:
z_samples = xy_to_z(context_xs, context_ys, encoder_w0) * facto
target_ws = encoder_w(target_xs, encoder_w0)
input_zxs = tf.concat([z_samples, target_ws], axis=-1)
# sample y_hat ~ y|(w,z)
with tf.variable_scope('decoder'):
target_yhat_mu = decoder0(input_zxs) # n_task * n_im * dim_y
# when var of p(y | x,z) is fixed, neg-loglik <=> MSE
mse_loss = mse(target_yhat_mu, target_ys)
tf.summary.scalar(prefix + 'mse', mse_loss)
optimizer1 = tf.train.AdamOptimizer(FLAGS.update_lr)
optimizer2 = tf.train.AdamOptimizer(FLAGS.update_lr)
if 'train' in prefix:
if FLAGS.weight_decay:
loss = mse_loss
optimizer = contrib_opt.AdamWOptimizer(
weight_decay=FLAGS.beta, learning_rate=FLAGS.update_lr)
gvs = optimizer.compute_gradients(loss)
train_op = optimizer.apply_gradients(gvs)
else:
THETA = ( # pylint: disable=invalid-name
tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='decoder') +
tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='encoder_w'))
all_var = tf.trainable_variables()
PHI = [v for v in all_var if v not in THETA] # pylint: disable=invalid-name
loss = mse_loss
gvs_theta = optimizer1.compute_gradients(loss, THETA)
train_theta_op = optimizer1.apply_gradients(gvs_theta)
gvs_phi = optimizer2.compute_gradients(loss, PHI)
train_phi_op = optimizer2.apply_gradients(gvs_phi)
with tf.control_dependencies([train_theta_op, train_phi_op]):
train_op = tf.no_op()
return mse_loss, train_op, facto
else:
return mse_loss
def Lock(self):
"""Used to lock in the personalization."""
lock_ops = [tf.no_op()] # 什么都不做，仅做为点位符使用控制边界。
if self.lowrank_adaptation:
# compute the new W
left_adapt = tf.squeeze(self.left_adapt)
right_adapt = tf.squeeze(self.right_adapt)
# final_w = self.W + tf.matmul(left_adapt, right_adapt)
final_w = tf.add(tf.matmul(left_adapt, right_adapt),self.W)
#self.lockedW = final_w
#lock_ops.append(self.lockedW)
# self.lockedW.assign(final_w)
#lock = tf.assign(self.lockedW, final_w) #您可以使用tf.identity取消引用_ref类型
self.lockedW = final_w
lock2 = tf.identity(self.lockedW)
lock_ops.append(lock2) #
else:
# self.lockedW = self.W
# lock_ops.append(self.lockedW)
lock3 = tf.assign(self.lockedW, self.W)
lock4 = tf.identity(self.lockedW)
lock_ops.append(lock4)
# lock_ops.append(self.lockedW.assign(self.W))
if self.mikolov_adapt:
final_bias = tf.squeeze(self.bias + self.delta)
lock_ops.append(self.lockedBias.assign(final_bias))
else:
lock_ops.append(self.lockedBias.assign(self.bias))
# self.lock_op = tf.group(*lock_ops,name="lock_op") # 没有返回值，只是个op
self.lock_op = tf.tuple(lock_ops,name="lock_op") #返回的是list of tensor。
def reconstruction_loss(self, x_input, x_target, x_length, z=None,
c_input=None):
# Split output for each core model.
split_x_input = tf.split(x_input, self._output_depths, axis=-1)
split_x_target = tf.split(x_target, self._output_depths, axis=-1)
loss_outputs = []
# Compute reconstruction losses for the split output.
for i, cd in enumerate(self._core_decoders):
with tf.variable_scope('core_decoder_%d' % i):
# TODO(adarob): Sample initial inputs when using scheduled sampling.
loss_outputs.append(
cd.reconstruction_loss(
split_x_input[i], split_x_target[i], x_length, z, c_input))
r_losses, metric_maps, decode_results = list(zip(*loss_outputs))
# Merge the metric maps by passing through renamed values and taking the
# mean across the splits.
merged_metric_map = {}
for metric_name in metric_maps[0]:
metric_values = []
for i, m in enumerate(metric_maps):
merged_metric_map['%s/output_%d' % (metric_name, i)] = m[metric_name]
metric_values.append(m[metric_name][0])
merged_metric_map[metric_name] = (
tf.reduce_mean(metric_values), tf.no_op())
return (tf.reduce_sum(r_losses, axis=0),
merged_metric_map,
self._merge_decode_results(decode_results))
def decompress_blocks(self, sess, blocks, x_shape, debug=False):
"""Uses the decompression model to decompress a point cloud"""
dec_blocks = []
debug_t_list = []
for i, (strings, best_threshold_idx) in enumerate(blocks):
logger.info(f'Decompress block {i}/{len(blocks)}: start')
strings = [[s] for s in strings]
threshold = self.thresholds[best_threshold_idx]
logger.info(f'Decompress block {i}/{len(blocks)}: run session')
fetches = [self.x_hat, self.debug_tensors if debug else tf.no_op()]
x_hat, debug_tensors = sess.run(
fetches,
feed_dict={
self.x_shape_t: x_shape,
**dict(zip(self.strings_t, strings))
})
logger.info(f'Decompress block {i}/{len(blocks)}: session done')
x_hat = x_hat[
0,
0, :, :, :] if self.data_format == 'channels_first' else x_hat[
0, :, :, :, 0]
x_hat = x_hat > threshold
pa = np.argwhere(x_hat).astype('float32')
logger.info(f'Decompress block {i}/{len(blocks)}: done')
dec_blocks.append(pa)
debug_t_list.append(debug_tensors)
return dec_blocks, debug_t_list
def assert_shape_equal_along_first_dimension(shape_a, shape_b):
"""Asserts that shape_a and shape_b are the same along the 0th-dimension.
If the shapes are static, raises a ValueError when the shapes
mismatch.
If the shapes are dynamic, raises a tf InvalidArgumentError when the shapes
mismatch.
Args:
shape_a: a list containing shape of the first tensor.
shape_b: a list containing shape of the second tensor.
Returns:
Either a tf.no_op() when shapes are all static and a tf.assert_equal() op
when the shapes are dynamic.
Raises:
ValueError: When shapes are both static and unequal.
"""
if isinstance(shape_a[0], int) and isinstance(shape_b[0], int):
if shape_a[0] != shape_b[0]:
raise ValueError('Unequal first dimension {}, {}'.format(
shape_a[0], shape_b[0]))
else:
return tf.no_op()
else:
return tf.assert_equal(shape_a[0], shape_b[0])
def __run_test(self, sess):
# Var is initialized by var_init
var = tf.get_variable('var', [1], dtype=tf.int32)
var_init = tf.constant([0])
var_initialize = var.assign(var_init)
# Computation of mean
input1 = tf.constant(
[[[1.0, 2.0], [3.0, 4.0]], [[1.0, 2.0], [3.0, 4.0]]], name='input1')
mean = tf.reduce_mean(input1, var)
# For updating the Var
const_var = tf.constant([1])
var_add = tf.add(var, const_var)
var_update = var.assign(var_add)
# update to happen after mean computation
with tf.control_dependencies([mean]):
var_update = var.assign(var_add)
with tf.control_dependencies([var_update]):
update_op = tf.no_op('train_op')
# Initialize Var
var_init_value = sess.run((var_initialize))
# Compute mean and var updates
mean_values = []
for i in range(3):
(result_mean, result_up) = sess.run((mean, update_op))
mean_values.append(result_mean)
# Compute Final Values
var_final_val = var.eval(sess)
return var_init_value, mean_values, var_final_val
def _compressor_op(self, matrix_compressor, a_matrix_tfvar):
"""Creates compressor op based on simhash matrix_compressor.
Meant to create the factors once at begin_compression_step tailored
for simhash (which has only one output b_matrix).
Args:
matrix_compressor: specifies the matrix compressor object.
a_matrix_tfvar: the tf tensor to be compressed.
Returns:
a tf.no_op object with assign ops as control dependencies.
"""
# py_func is not supported on TPU so need non py_func implementation
# The following line seems to be needed otherwise it says machines with tpu
# don't support pyfunc.
use_tpu = self._spec.use_tpu
if use_tpu:
[b_matrix_out] = matrix_compressor.tpu_matrix_compressor(a_matrix_tfvar)
else:
[b_matrix_out
] = tf.compat.v1.py_func(matrix_compressor.static_matrix_compressor,
[a_matrix_tfvar], [tf.float32])
b_matrix_assign_op = tf.compat.v1.assign(
self.b_matrix_tfvar, b_matrix_out, name='_b_matrix_assign_op')
with tf.control_dependencies([b_matrix_assign_op]):
return tf.no_op('compresor_b_matrix_update')
def assert_shape_equal(shape_a, shape_b):
"""Asserts that shape_a and shape_b are equal.
If the shapes are static, raises a ValueError when the shapes
mismatch.
If the shapes are dynamic, raises a tf InvalidArgumentError when the shapes
mismatch.
Args:
shape_a: a list containing shape of the first tensor.
shape_b: a list containing shape of the second tensor.
Returns:
Either a tf.no_op() when shapes are all static and a tf.assert_equal() op
when the shapes are dynamic.
Raises:
ValueError: When shapes are both static and unequal.
"""
if (all(isinstance(dim, int) for dim in shape_a)
and all(isinstance(dim, int) for dim in shape_b)):
if shape_a != shape_b:
raise ValueError('Unequal shapes {}, {}'.format(shape_a, shape_b))
else:
return tf.no_op()
else:
return tf.assert_equal(shape_a, shape_b)
def _train_discriminator(self, nodes: Nodes, optimizer, create_summaries):
"""Creates a train_op for the discriminator.
Args:
nodes: Instance of Nodes, the nodes of the model to feed to D.
optimizer: Discriminator optimizer. Passed in because it will be re-used
in the different discriminator steps.
create_summaries: If True, create summaries.
Returns:
A training op if training, else no_op.
"""
d_out = self._compute_discriminator_out(
nodes,
create_summaries,
gradients_to_generator=False) # Only train discriminator!
d_loss = self._create_gan_loss(d_out, create_summaries, mode="d_loss")
if not self.training:
return tf.no_op()
self._add_hook(tf.train.NanTensorHook(d_loss))
# Getting the variables here because they don't exist before calling
# _compute_discriminator_out for the first time!
disc_vars = self._discriminator.trainable_variables
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
with tf.name_scope("min_d"):
train_op_d = optimizer.minimize(
d_loss, self._global_step_disc, disc_vars)
return train_op_d
def _compressor_op(self, matrix_compressor, a_matrix_tfvar):
"""Creates compressor op based on matrix_compressor.
Meant to create the factors once at begin_compression_step.
Args:
matrix_compressor: specifies the matrix compressor object.
a_matrix_tfvar: the tf tensor to be compressed.
"""
# py_func is not supported on TPU so need non py_func implementation
use_tpu = self._spec.use_tpu
# Seeing some tf.py_func error because of which the
# following may be needed, so enforcing TF operation updates.
if use_tpu:
[b_matrix_out, c_matrix_out
] = matrix_compressor.tpu_matrix_compressor(a_matrix_tfvar)
else:
[b_matrix_out, c_matrix_out] = tf.compat.v1.py_func(
matrix_compressor.static_matrix_compressor, [a_matrix_tfvar],
[tf.float32, tf.float32])
b_matrix_assign_op = tf.compat.v1.assign(self.b_matrix_tfvar,
b_matrix_out,
name='b_matrix_assign_op')
c_matrix_assign_op = tf.compat.v1.assign(self.c_matrix_tfvar,
c_matrix_out,
name='c_matrix_assign_op')
with tf.control_dependencies([b_matrix_assign_op, c_matrix_assign_op]):
logging.info('Updating b_matrix,c_matrix.')
return tf.no_op('compresor_b_matrix_and_c_matrix_update')
def _apply_dense(self, grad, var):
# We actually apply grads in _finish. This function is used
# to record intermediate variables related to the individual gradients
# which we eventually combine in _finish to obtain global statistics
# (e.g. the L1 norm of the full gradient).
self.grads[var] = grad
betting_fraction = self.get_slot(var, OUTER_BETTING_FRACTION)
self.betting_fraction_dot_product_deltas[var] = tf.reduce_sum(
betting_fraction * grad)
# Wealth increases by -g \cdot w where w is the parameter value.
# Since w = Wealth * v with betting fraction v, we can write
# the wealth increment as -(g \cdot v) Wealth.
# TODO(cutkosky): at one point there was a bug in which epsilon
# was not added here. It seemed performance may have degraded
# somewhat after fixing this. Find out why this would be.
wealth_delta = -self.betting_fraction_dot_product_deltas[
var] * self._get_non_slot(OUTER_WEALTH)
self.wealth_deltas[var] = wealth_delta
self.grad_norms[var] = tf.norm(grad, 1)
return tf.no_op()
def compute_gradients(self, loss):
grads = self.opt.compute_gradients(loss, self.var_list)
updates = [self.accum_vars[v].assign_add(g) for (g, v) in grads]
updates.append(self.total_loss.assign_add(loss))
updates.append(self.count_loss.assign_add(1.0))
with tf.control_dependencies(updates):
return tf.no_op()
def all_update_op(self):
"""Returns the combine update tf OP."""
# TODO(nishanthd): implement all_update_op logic inside the wrapper
with tf.compat.v1.name_scope(self._scope):
with tf.control_dependencies(self._update_ops):
logging.info('Updating all compression_ops.')
self._all_update_op = tf.no_op('update_all_compression_ops')
return self._all_update_op