def _train(self, checkpoint_path, layout_optimizer=False, restore=False):
ops.reset_default_graph()
graph = ops.get_default_graph()
with session.Session(
config=get_config(layout_optimizer), graph=graph) as sess:
batch = 2
height = 6
width = 7
input_channels = 3
shape = [batch, height, width, input_channels]
image = array_ops.placeholder(dtype='float32', shape=shape)
conv1 = conv_layers.conv2d(image, 32, [3, 3])
conv2 = conv_layers.conv2d(conv1, 32, [3, 3])
optimizer = gradient_descent.GradientDescentOptimizer(0.01)
loss = math_ops.reduce_mean(conv2)
train_op = optimizer.minimize(loss)
saver = saver_lib.Saver(write_version=saver_pb2.SaverDef.V2)
if restore:
saver.restore(sess, checkpoint_path)
else:
sess.run(variables.global_variables_initializer())
np.random.seed(0)
for _ in range(2):
image_val = np.random.rand(*shape).astype(np.float32)
sess.run([loss, train_op], feed_dict={image: image_val})
if restore:
all_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
all_vars_values = [var.eval(session=sess) for var in all_vars]
return all_vars_values
else:
saver.save(sess, checkpoint_path)
def testGradient(self):
if not test.is_gpu_available(cuda_only=True):
self.skipTest('GPU required')
random_seed.set_random_seed(0)
x = random_ops.truncated_normal([1, 200, 200, 3], seed=0)
y = conv_layers.conv2d(x, 32, [3, 3])
z = conv_layers.conv2d(y, 32, [3, 3])
optimizer = gradient_descent.GradientDescentOptimizer(1e-4)
loss = math_ops.reduce_mean(z)
train_op = optimizer.minimize(loss)
graph = ops.get_default_graph()
graph.add_to_collection('train_op', train_op)
meta_graph = saver.export_meta_graph(graph_def=graph.as_graph_def())
rewrite_options = rewriter_config_pb2.RewriterConfig(
optimize_tensor_layout=True)
optimized_graph = tf_optimizer.OptimizeGraph(rewrite_options,
meta_graph)
found = 0
for node in optimized_graph.node:
if node.op in [
'Conv2D', 'Conv2DBackpropFilter', 'Conv2DBackpropInput'
]:
found += 1
self.assertEqual(node.attr['data_format'].s, 'NCHW')
self.assertEqual(found, 5)
def testFunctionalConv2DNoReuse(self): height, width = 7, 9 images = random_ops.random_uniform((5, height, width, 3), seed=1) conv_layers.conv2d(images, 32, [3, 3]) self.assertEqual(len(variables.trainable_variables()), 2) conv_layers.conv2d(images, 32, [3, 3]) self.assertEqual(len(variables.trainable_variables()), 4)
def _train(self, checkpoint_path, layout_optimizer=False, restore=False):
ops.reset_default_graph()
graph = ops.get_default_graph()
with session.Session(config=get_config(layout_optimizer),
graph=graph) as sess:
batch = 2
height = 6
width = 7
input_channels = 3
shape = [batch, height, width, input_channels]
image = array_ops.placeholder(dtype='float32', shape=shape)
conv1 = conv_layers.conv2d(image, 32, [3, 3])
conv2 = conv_layers.conv2d(conv1, 32, [3, 3])
optimizer = gradient_descent.GradientDescentOptimizer(0.01)
loss = math_ops.reduce_mean(conv2)
train_op = optimizer.minimize(loss)
saver = saver_lib.Saver(write_version=saver_pb2.SaverDef.V2)
if restore:
saver.restore(sess, checkpoint_path)
else:
sess.run(variables.global_variables_initializer())
np.random.seed(0)
for _ in range(2):
image_val = np.random.rand(*shape).astype(np.float32)
sess.run([loss, train_op], feed_dict={image: image_val})
if restore:
all_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
all_vars_values = [var.eval(session=sess) for var in all_vars]
return all_vars_values
else:
saver.save(sess, checkpoint_path)
def testFunctionalConv2DNoReuse(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
conv_layers.conv2d(images, 32, [3, 3])
self.assertEqual(len(variables.trainable_variables()), 2)
conv_layers.conv2d(images, 32, [3, 3])
self.assertEqual(len(variables.trainable_variables()), 4)
def testGradient(self):
if not test.is_gpu_available(cuda_only=True):
self.skipTest('GPU required')
random_seed.set_random_seed(0)
x = random_ops.truncated_normal([1, 200, 200, 3], seed=0)
y = conv_layers.conv2d(x, 32, [3, 3])
z = conv_layers.conv2d(y, 32, [3, 3])
optimizer = gradient_descent.GradientDescentOptimizer(1e-4)
loss = math_ops.reduce_mean(z)
train_op = optimizer.minimize(loss)
graph = ops.get_default_graph()
graph.add_to_collection('train_op', train_op)
meta_graph = saver_lib.export_meta_graph(graph_def=graph.as_graph_def())
rewrite_options = rewriter_config_pb2.RewriterConfig(
optimize_tensor_layout=True)
optimized_graph = tf_optimizer.OptimizeGraph(rewrite_options, meta_graph)
found = 0
for node in optimized_graph.node:
if node.op in ['Conv2D', 'Conv2DBackpropFilter', 'Conv2DBackpropInput']:
found += 1
self.assertEqual(node.attr['data_format'].s, 'NCHW')
self.assertEqual(found, 5)
def testInvalidStrides(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'strides'):
conv_layers.conv2d(images, 32, 3, strides=(1, 2, 3))
with self.assertRaisesRegexp(ValueError, 'strides'):
conv_layers.conv2d(images, 32, 3, strides=None)
def testInvalidKernelSize(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'kernel_size'):
conv_layers.conv2d(images, 32, (1, 2, 3))
with self.assertRaisesRegexp(ValueError, 'kernel_size'):
conv_layers.conv2d(images, 32, None)
def testInvalidStrides(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'strides'):
conv_layers.conv2d(images, 32, 3, strides=(1, 2, 3))
with self.assertRaisesRegexp(ValueError, 'strides'):
conv_layers.conv2d(images, 32, 3, strides=None)
def testInvalidKernelSize(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'kernel_size'):
conv_layers.conv2d(images, 32, (1, 2, 3))
with self.assertRaisesRegexp(ValueError, 'kernel_size'):
conv_layers.conv2d(images, 32, None)
def testGroupNormalizeInference(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
with variable_scope.variable_scope("vs", use_resource=True):
y = convolutional.conv2d(
x,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
gamma = constant_op.constant([0.5, 0.5], np.float32)
beta = constant_op.constant([0.5, 0.5], np.float32)
mean = constant_op.constant([0.5, 0.5], np.float32)
inv_std_dev = constant_op.constant([0.5, 0.5], np.float32)
y = gen_popnn_ops.popnn_group_norm_inference(
inputs=y,
gamma=gamma,
beta=beta,
mean=mean,
inv_std_dev=inv_std_dev,
data_format="NHWC",
epsilon=0.0015,
num_groups=2)
y = convolutional.conv2d(
y,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
y = gen_popnn_ops.popnn_group_norm_inference(
inputs=y,
gamma=gamma,
beta=beta,
mean=mean,
inv_std_dev=inv_std_dev,
data_format="NHWC",
epsilon=0.0015,
num_groups=2)
report = ReportJSON(self, sess)
sess.run(variables.global_variables_initializer())
report.reset()
sess.run(y, {x: np.zeros([1, 4, 4, 2])})
report.parse_log()
# Would fail if there were two batch norms in the graph
ok = [
'__seed*', 'Copy_',
'vs/conv2d/Conv2D/convolution.*/Conv_1x1/Convolve',
'vs/PopnnGroupNormInference/group-norm-inference*/'
]
report.assert_all_compute_sets_and_list(ok)
def testFunctionalConv2DReuseFromScope(self):
with variable_scope.variable_scope('scope'):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
conv_layers.conv2d(images, 32, [3, 3], name='conv1')
self.assertEqual(len(variables.trainable_variables()), 2)
with variable_scope.variable_scope('scope', reuse=True):
conv_layers.conv2d(images, 32, [3, 3], name='conv1')
self.assertEqual(len(variables.trainable_variables()), 2)
def testBatchNormAndGroupNormalizeMixedInference(self):
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
with variable_scope.variable_scope("vs", use_resource=True):
y = convolutional.conv2d(
x,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
gamma = constant_op.constant([0.5, 0.5], np.float32)
beta = constant_op.constant([0.5, 0.5], np.float32)
mean = constant_op.constant([0.5, 0.5], np.float32)
inv_std_dev = constant_op.constant([0.5, 0.5], np.float32)
y = gen_popnn_ops.popnn_group_norm_inference(
inputs=y,
gamma=gamma,
beta=beta,
mean=mean,
inv_std_dev=inv_std_dev,
data_format="NHWC",
epsilon=0.0015,
num_groups=2)
y = convolutional.conv2d(
y,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
y = layers_norm.batch_normalization(y, fused=True)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system(True, True, True)
with tu.ipu_session() as sess:
sess.run(variables.global_variables_initializer())
sess.run(report)
sess.run(y, {x: np.zeros([1, 4, 4, 2])})
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
# Would fail if there were two batch norms in the graph
ok = [
'__seed*', 'host-exchange-local-copy', 'Copy_',
'vs/conv2d/Conv2D/convolution.*/Conv_1x1/Convolve',
'vs/PopnnGroupNormInference/custom-call*/',
'vs/batch_normalization/FusedBatchNorm/batch-norm-inference.*/'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testFunctionalConv2DNoReuse(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
conv_layers.conv2d(images, 32, [3, 3])
self.assertEqual(
len(ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)), 2)
conv_layers.conv2d(images, 32, [3, 3])
self.assertEqual(
len(ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES)), 4)
def testFunctionalConv2DReuseFromScope(self):
with variable_scope.variable_scope('scope'):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
conv_layers.conv2d(images, 32, [3, 3], name='conv1')
self.assertEqual(len(variables.trainable_variables()), 2)
with variable_scope.variable_scope('scope', reuse=True):
conv_layers.conv2d(images, 32, [3, 3], name='conv1')
self.assertEqual(len(variables.trainable_variables()), 2)
def _bottleneck(self, inputs, size=None):
conv1 = conv2d(inputs, filters=size, kernel_size=1, padding='same')
ac2 = self._add_common_layers(conv1)
conv2 = conv2d(ac2, filters=size, kernel_size=3, padding='same')
ac3 = self._add_common_layers(conv2)
conv3 = conv2d(ac3, filters=size * 4, kernel_size=1, padding='same')
# This 1x1 conv is used to match the dimension of x and F(x)
hack_conv = conv2d(inputs, filters=size * 4, kernel_size=1, padding='same')
return tf.add(hack_conv, conv3)
def scconv_Qd(self, scc_type, bn_type, use_relu, input_layer=None):
if input_layer is None:
input_layer = self.top_layer
name = 'scconv' + str(self.counts['scconv'])
self.counts['scconv'] += 1
strides = [1, 1]
channel_pos = 'channels_first'
print('input of scconv: ', input_layer.name, input_layer.get_shape())
with tf.variable_scope(name + '_h'):
x = tf.pad(input_layer, [[0, 0], [0, 0], [1, 1], [1, 1]])
x = conv_layers.conv2d(x, filters=input_layer.get_shape()[1].value / 2, kernel_size=3, strides=strides,
padding='VALID', data_format=channel_pos,
kernel_initializer=None,
use_bias=(bn_type == 'NONE'))
if bn_type != 'NONE':
x = tf.contrib.layers.batch_norm(
x,
decay=self.batch_norm_config.get('decay', 0.9),
scale=True,
epsilon=self.batch_norm_config.get('epsilon', 0.001),
is_training=self.phase_train,
fused=True,
data_format=bn_type)
if use_relu:
x = tf.nn.relu(x)
print('height compressed: ', x.name, x.get_shape())
with tf.variable_scope(name + '_w'):
x = tf.transpose(x, [0, 2, 1, 3])
x = tf.pad(x, [[0, 0], [0, 0], [1, 1], [1, 1]])
x = conv_layers.conv2d(x, filters=input_layer.get_shape()[2].value / 2, kernel_size=3, strides=strides,
padding='VALID', data_format=channel_pos,
kernel_initializer=None,
use_bias=(bn_type == 'NONE'))
if bn_type != 'NONE':
x = tf.contrib.layers.batch_norm(
x,
decay=self.batch_norm_config.get('decay', 0.9),
scale=True,
epsilon=self.batch_norm_config.get('epsilon', 0.001),
is_training=self.phase_train,
fused=True,
data_format=bn_type)
if use_relu:
x = tf.nn.relu(x)
x = tf.transpose(x, [0, 2, 1, 3])
print('width compressed: ', x.name, x.get_shape())
self.top_layer = x
return x
def flow(self, source, reference):
ds=1
activate=tf.nn.relu#tf.nn.leaky_relu
ki=tf.contrib.layers.xavier_initializer()
n,h,w,c=reference.shape
with tf.variable_scope('flow',reuse=tf.AUTO_REUSE) as scope:
x=tf.concat([reference,source],-1)
x=conv2d(x,32,9,strides=ds, padding='same', activation=activate, kernel_initializer=ki,name='conv0')
x=tf.nn.max_pool(x, [1,2,2,1], [1,2,2,1], padding='SAME', name='pool0')
x=conv2d(x,32,9,strides=ds, padding='same', activation=activate, kernel_initializer=ki,name='conv1')
x=tf.nn.max_pool(x, [1,2,2,1], [1,2,2,1], padding='SAME', name='pool1')
x=tf.image.resize_images(x,[h,w],method=0)
uv=conv2d(x,2,3,strides=ds, padding='same', activation=tf.nn.tanh, kernel_initializer=ki,name='conv2')
return uv
def cnn_2d(images, is_training):
"""
Build the model for 2D-CNN.
Inputs:
-- images: Images placeholder
-- is_training: bool placeholder, training or not
Output:
-- Logits: Return the output of the model
"""
# Build the CNN model
l_conv1 = conv2d(images,
CONV1_FILTERS,
KERNEL_SIZE1,
strides=STRIDE_CONV1,
activation=relu,
name='Conv1')
l_maxpool1 = max_pooling2d(l_conv1,
POOL_SIZE1,
POOL_SIZE1,
padding='same',
name='Maxpool1')
l_conv2 = conv2d(l_maxpool1,
CONV2_FILTERS,
KERNEL_SIZE2,
strides=STRIDE_CONV2,
activation=relu,
name='Conv2')
l_maxpool2 = max_pooling2d(l_conv2,
POOL_SIZE2,
POOL_SIZE2,
padding='same',
name='Maxpool2')
l_flatten = flatten(l_maxpool2, scope='Flatten')
l_fc1 = dense(l_flatten, FC1, activation=relu, name='Fc1')
l_drop = dropout(l_fc1, DROP_RATE, training=is_training, name='Dropout')
l_fc2 = dense(l_drop, FC2, activation=relu, name='Fc2')
logits = dense(l_fc2, NUM_CLASSES, name='Output')
return logits
def _conv_with_bn(self, idx, bottom):
# initializer_dict = {'gamma': init_ops.glorot_normal_initializer()}
conved = conv2d(
inputs=bottom,
filters=self.deps[idx],
kernel_size=[3, 3],
strides=[1, 1],
padding='same',
activation=None,
use_bias=True,
kernel_initializer=tf.contrib.layers.xavier_initializer(),
name=self.conv_names[idx])
bn = batch_norm(inputs=conved,
decay=0.99,
center=True,
scale=True,
activation_fn=tf.nn.relu,
is_training=self.training,
scope=self.conv_names[idx])
# initializer_dict = {'gamma': init_ops.glorot_normal_initializer()}
# conved = conv2d(inputs=bottom, filters=self.deps[idx], kernel_size=[3,3], strides=[1,1], padding='same',
# activation=tf.nn.relu, use_bias=True, kernel_initializer=tf.contrib.layers.xavier_initializer(), name=self.conv_names[idx])
# bn = batch_norm(inputs=conved, center=True, scale=True, activation_fn=None, is_training=self.training, scope=self.conv_names[idx],
# param_initializers=initializer_dict)
return bn
def _conv2d_impl(self, input_layer, num_channels_in, filters, kernel_size,
strides, padding, kernel_initializer):
if self.use_tf_layers:
return conv_layers.conv2d(input_layer,
filters,
kernel_size,
strides,
padding,
self.channel_pos,
kernel_initializer=kernel_initializer,
use_bias=False)
else:
weights_shape = [
kernel_size[0], kernel_size[1], num_channels_in, filters
]
weights = self.get_variable("conv2d/kernel",
weights_shape,
self.variable_dtype,
self.dtype,
initializer=kernel_initializer)
if self.data_format == "NHWC":
strides = [1] + strides + [1]
else:
strides = [1, 1] + strides
return tf.nn.conv2d(input_layer,
weights,
strides,
padding,
data_format=self.data_format)
def _conv2d_impl(self, input_layer, num_channels_in, filters, kernel_size,
strides, padding, kernel_initializer):
"""Construct a custom convolution layer."""
if self.use_tf_layers:
assert not (self.params.tanh_weight_transform
or self.params.quant_weight)
return conv_layers.conv2d(input_layer,
filters,
kernel_size,
strides,
padding,
self.channel_pos,
kernel_initializer=kernel_initializer,
use_bias=False)
else:
weights_shape = [
kernel_size[0], kernel_size[1], num_channels_in, filters
]
# We use the name 'conv2d/kernel' so the variable has the same name as its
# tf.layers equivalent. This way, if a checkpoint is written when
# self.use_tf_layers == True, it can be loaded when
# self.use_tf_layers == False, and vice versa.
weights = self.get_variable('conv2d/kernel',
weights_shape,
self.variable_dtype,
self.dtype,
initializer=kernel_initializer)
if self.params.tanh_weight_transform:
if not (self.params.first_weight_name in weights.name
or self.params.last_weight_name in weights.name):
print('Dorefa quantizing weight %s' % weights.name)
weights = self.dorefa_weight_quantize(
weights, self.params.quant_weight,
self.params.quant_weight_bits,
self.params.quant_weight_per_channel,
self.params.quant_weight_delay)
elif self.params.quant_weight:
if not (self.params.first_weight_name in weights.name
or self.params.last_weight_name in weights.name):
print('Quantizing weight %s' % weights.name)
weights = self.last_value_quantize(
weights,
per_channel=self.params.quant_weight_per_channel,
is_training=self.phase_train,
num_bits=self.params.quant_weight_bits,
narrow_range=self.params.quant_weight_narrow_range,
relative_quantile=self.params.
quant_weight_relative_quantile,
freeze=self.params.freeze_weight_range,
quant_delay=self.params.quant_weight_delay)
if self.data_format == 'NHWC':
strides = [1] + strides + [1]
else:
strides = [1, 1] + strides
return tf.nn.conv2d(input_layer,
weights,
strides,
padding,
data_format=self.data_format)
def testConv2DFloat16(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 4),
dtype='float16')
output = conv_layers.conv2d(images, 32, [3, 3], activation=nn_ops.relu)
self.assertListEqual(output.get_shape().as_list(),
[5, height - 2, width - 2, 32])
def _conv2d_impl(self, input_layer, num_channels_in, filters, kernel_size,
strides, padding, kernel_initializer):
if self.use_tf_layers:
return conv_layers.conv2d(input_layer,
filters,
kernel_size,
strides,
padding,
self.channel_pos,
kernel_initializer=kernel_initializer,
use_bias=False)
else:
weights_shape = [
kernel_size[0], kernel_size[1], num_channels_in, filters
]
# We use the name 'conv2d/kernel' so the variable has the same name as its
# tf.layers equivalent. This way, if a checkpoint is written when
# self.use_tf_layers == True, it can be loaded when
# self.use_tf_layers == False, and vice versa.
weights = self.get_variable('conv2d/kernel',
weights_shape,
self.variable_dtype,
self.dtype,
initializer=kernel_initializer)
if self.data_format == 'NHWC':
strides = [1] + strides + [1]
else:
strides = [1, 1] + strides
return tf.nn.conv2d(input_layer,
weights,
strides,
padding,
data_format=self.data_format)
def _conv2d_impl(self, input_layer, num_channels_in, filters, kernel_size,
strides, padding, kernel_initializer):
if self.use_tf_layers:
return conv_layers.conv2d(
input_layer,
filters,
kernel_size,
strides,
padding,
self.channel_pos,
kernel_initializer=kernel_initializer,
use_bias=False)
else:
weights_shape = [
kernel_size[0], kernel_size[1], num_channels_in, filters
]
weights = self.get_variable(
'conv2d/kernel',
weights_shape,
self.variable_dtype,
self.dtype,
initializer=kernel_initializer)
if self.data_format == 'NHWC':
strides = [1] + strides + [1]
else:
strides = [1, 1] + strides
return tf.nn.conv2d(
input_layer,
weights,
strides,
padding,
data_format=self.data_format)
def testBatchNormalizeInferenceMatchWithSharding(self):
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
with variable_scope.variable_scope("vs", use_resource=True):
with tu.ipu_shard(0):
a = convolutional.conv2d(
x,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
b = layers_norm.batch_normalization(a, fused=True)
with tu.ipu_shard(0):
c = convolutional.conv2d(
b,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
d = layers_norm.batch_normalization(c, fused=True)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system(True, True, True, sharded=True)
with tu.ipu_session() as sess:
sess.run(variables.global_variables_initializer())
sess.run(report)
sess.run(d, {x: np.zeros([1, 4, 4, 2])})
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
# Would fail if there were two batch norms in the graph
ok = [
'__seed*', '*OnTileCopy*',
'vs/conv2d/Conv2D/convolution.*/Conv_1x1/Convolve',
'vs/batch_normalization/FusedBatchNorm/batch-norm-inference.*/'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testBatchNormalizeInferenceDontMatchDifferentTypes(self):
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
with variable_scope.variable_scope("vs", use_resource=True):
y = convolutional.conv2d(
x,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
y = layers_norm.batch_normalization(y, fused=True)
y = math_ops.cast(y, np.float16)
y = convolutional.conv2d(
y,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
y = layers_norm.batch_normalization(y, fused=True)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
tu.configure_ipu_system(True, True, True)
with tu.ipu_session() as sess:
sess.run(variables.global_variables_initializer())
sess.run(report)
sess.run(y, {x: np.zeros([1, 4, 4, 2])})
result = sess.run(report)
s = tu.extract_all_strings_from_event_trace(result)
cs_list = tu.get_compute_sets_from_report(s)
# Matches two convolutions
ok = [
'__seed*', 'host-exchange-local-copy-', 'Copy_',
'vs/conv2d/Conv2D/convolution.*/Conv_1x1',
'vs/batch_normalization/FusedBatchNorm/batch-norm-inference.*/',
'vs/Cast/convert.*/Cast',
'vs/conv2d_1/Conv2D/convolution.*/Conv_1x1',
'vs/batch_normalization_1/FusedBatchNormV2/batch-norm-inference.*/'
]
self.assertTrue(tu.check_all_compute_sets_and_list(cs_list, ok))
def testFunctionalConv2DInitializerFromScope(self):
with self.test_session() as sess:
with variable_scope.variable_scope(
'scope', initializer=init_ops.ones_initializer()):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
conv_layers.conv2d(images, 32, [3, 3], name='conv1')
weights = variables.trainable_variables()
# Check the names of weights in order.
self.assertTrue('kernel' in weights[0].name)
self.assertTrue('bias' in weights[1].name)
sess.run(variables.global_variables_initializer())
weights = sess.run(weights)
# Check that the kernel weights got initialized to ones (from scope)
self.assertAllClose(weights[0], np.ones((3, 3, 3, 32)))
# Check that the bias still got initialized to zeros.
self.assertAllClose(weights[1], np.zeros((32)))
def testCloneBatchNorm(self):
g = ops.Graph()
with g.as_default():
np.random.seed(1234)
x_value = np.random.random([2, 5, 5, 3])
w_value = np.random.random([3, 3, 3, 2])
is_training_t = array_ops.placeholder(dtypes.bool,
name='is_training_t')
x_t = array_ops.constant(x_value, dtype=dtypes.float32, name='x_t')
y_t = conv2d(
x_t,
2, [3, 3],
kernel_initializer=init_ops.constant_initializer(w_value))
y_t = batch_norm(y_t, training=is_training_t)
optimizer_t = train.AdamOptimizer()
optimize_t = optimizer_t.minimize(math_ops.reduce_sum(y_t))
with self.test_session(use_gpu=True) as sess:
sess.run(variables.global_variables_initializer())
y_test_1 = sess.run(y_t, feed_dict={is_training_t: False})
sess.run(optimize_t, feed_dict={is_training_t: True})
y_test_2 = sess.run(y_t, feed_dict={is_training_t: False})
is_training = array_ops.placeholder(dtypes.bool,
name='is_training')
x = array_ops.constant(np.zeros([2, 5, 5, 3]),
dtype=dtypes.float32,
name='x')
y = conv2d(
x,
2, [3, 3],
kernel_initializer=init_ops.constant_initializer(w_value))
y = batch_norm(y, training=is_training)
x_new = array_ops.constant(x_value, dtype=dtypes.float32, name='x')
y_out = meta_graph.clone(y, "copy", replace={x: x_new})
optimizer = train.AdamOptimizer()
optimize = optimizer.minimize(math_ops.reduce_sum(y_out))
with self.test_session(use_gpu=True) as sess:
sess.run(variables.global_variables_initializer())
y_out_1 = sess.run(y_out, feed_dict={is_training: False})
y_out_2 = sess.run(y_out, feed_dict={is_training: False})
sess.run(optimize, feed_dict={is_training: True})
y_out_3 = sess.run(y_out, feed_dict={is_training: False})
self.assertAllClose(y_out_1, y_out_2)
self.assertTrue(np.abs(y_out_1 - y_out_3).max() > 1e-6)
self.assertAllClose(y_test_1, y_out_1)
self.assertAllClose(y_test_2, y_out_3)
def testFunctionalConv2DInitializerFromScope(self):
with self.test_session() as sess:
with variable_scope.variable_scope(
'scope', initializer=init_ops.ones_initializer()):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
conv_layers.conv2d(images, 32, [3, 3], name='conv1')
weights = variables.trainable_variables()
# Check the names of weights in order.
self.assertTrue('kernel' in weights[0].name)
self.assertTrue('bias' in weights[1].name)
sess.run(variables.global_variables_initializer())
weights = sess.run(weights)
# Check that the kernel weights got initialized to ones (from scope)
self.assertAllClose(weights[0], np.ones((3, 3, 3, 32)))
# Check that the bias still got initialized to zeros.
self.assertAllClose(weights[1], np.zeros((32)))
def training_step(inputs, scale):
outputs = convolutional.conv2d(
inputs,
filters=16,
kernel_size=(3, 3),
data_format="channels_first",
kernel_regularizer=make_regularizer(scale))
loss = math_ops.reduce_mean(math_ops.square(outputs))
return loss.op
def _conv(self, idx, bottom):
return conv2d(inputs=bottom,
filters=self.deps[idx],
kernel_size=[3, 3],
strides=[1, 1],
padding='same',
activation=nn.relu,
use_bias=True,
name=self.conv_names[idx])
def training_step(inputs, scale):
outputs = convolutional.conv2d(
inputs,
filters=16,
kernel_size=(3, 3),
data_format="channels_first",
kernel_regularizer=make_regularizer(scale))
loss = math_ops.reduce_mean(math_ops.square(outputs))
return loss.op
def conv(input, filters, stride, name):
return conv2d(input,
filters, [3, 3],
strides=[stride, stride],
name=name,
padding='same',
activation=None,
use_bias=False,
kernel_initializer=tf.random_normal_initializer(
stddev=np.sqrt(2.0 / 9 / filters)))
def add_transition(name, input, nb_filters):
# shape = input.get_shape().as_list()
# in_channel = shape[3]
with tf.variable_scope(name) as scope:
l = self._batch_norm_default(input, name)
l = tf.nn.relu(l)
l = conv2d(l, nb_filters, [1,1], strides=[1,1], name='conv1',
padding='same', activation=None, use_bias=False)
l = tf.nn.relu(l)
l = self._avgpool(l, 2)
return l
def testBatchNormalizeInferenceDontMatchDifferentTypes(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
with variable_scope.variable_scope("vs", use_resource=True):
y = convolutional.conv2d(
x,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
y = layers_norm.batch_normalization(y, fused=True)
y = math_ops.cast(y, np.float16)
y = convolutional.conv2d(
y,
2,
1,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
y = layers_norm.batch_normalization(y, fused=True)
report = ReportJSON(self, sess)
sess.run(variables.global_variables_initializer())
report.reset()
sess.run(y, {x: np.zeros([1, 4, 4, 2])})
report.parse_log()
# Matches two convolutions
ok = [
'__seed*', 'Copy_', 'vs/conv2d/Conv2D/convolution.*/Conv_1x1',
'vs/batch_normalization/FusedBatchNorm*/batch-norm-inference.*/',
'vs/Cast/convert.*/Cast',
'vs/conv2d_1/Conv2D/convolution.*/Conv_1x1',
'vs/batch_normalization_1/FusedBatchNorm*/batch-norm-inference.*/'
]
report.assert_all_compute_sets_and_list(ok)
def cnn_test(images, is_training):
# Build the CNN model
l_conv1 = conv2d(images,
10, [7, 7],
strides=[1, 1],
activation=relu,
name='Conv1')
l_flatten = flatten(l_conv1, scope='Flatten')
logits = dense(l_flatten, NUM_CLASSES, name='Output')
return logits
def testCloneConvolution(self):
g = ops.Graph()
with g.as_default():
x = array_ops.ones([2, 5, 5, 3], name='x')
y = conv2d(x, 2, [3, 3], padding='same')
x_new = array_ops.zeros([2, 5, 5, 3], name='x_new')
y_out = meta_graph.clone(y,
"conv2d/copy1",
from_scope="conv2d",
replace={x: x_new})
with self.test_session(use_gpu=True) as sess:
sess.run(variables.global_variables_initializer())
y_out_ = sess.run(y_out)
self.assertEqual(y_out.op.inputs[0].op.inputs[1].name,
"conv2d/kernel/read:0")
self.assertAllClose(y_out_, np.zeros((2, 5, 5, 2)))
def conv_bn(x,
feature_channel,
kernel_size,
block_name,
stride=(1, 1),
padding='same',
dilation_rate=(1, 1),
activation=None):
with tf.name_scope(block_name):
x = conv2d(x,
feature_channel,
kernel_size,
stride=stride,
padding=padding,
dilation_rate=dilation_rate,
activation=activation,
name=block_name + 'conv')
x = batch_normalization(x, name=block_name + 'bn')
return x
def _simple_model(self, image, fused, freeze_mode):
output_channels, kernel_size = 2, 3
conv = conv_layers.conv2d(
image,
output_channels,
kernel_size,
use_bias=False,
kernel_initializer=init_ops.ones_initializer())
bn_layer = normalization_layers.BatchNormalization(fused=fused)
bn_layer._bessels_correction_test_only = False
training = not freeze_mode
bn = bn_layer.apply(conv, training=training)
loss = math_ops.reduce_sum(math_ops.abs(bn))
optimizer = gradient_descent.GradientDescentOptimizer(0.01)
if not freeze_mode:
update_ops = ops.get_collection(ops.GraphKeys.UPDATE_OPS)
with ops.control_dependencies(update_ops):
train_op = optimizer.minimize(loss)
else:
train_op = optimizer.minimize(loss)
saver = saver_lib.Saver(write_version=saver_pb2.SaverDef.V2)
return loss, train_op, saver
def _conv2d_impl(self, input_layer, num_channels_in, filters, kernel_size,
strides, padding, kernel_initializer):
if self.use_tf_layers:
return conv_layers.conv2d(input_layer, filters, kernel_size, strides,
padding, self.channel_pos,
kernel_initializer=kernel_initializer,
use_bias=False)
else:
weights_shape = [kernel_size[0], kernel_size[1], num_channels_in, filters]
# We use the name 'conv2d/kernel' so the variable has the same name as its
# tf.layers equivalent. This way, if a checkpoint is written when
# self.use_tf_layers == True, it can be loaded when
# self.use_tf_layers == False, and vice versa.
weights = self.get_variable('conv2d/kernel', weights_shape,
self.variable_dtype, self.dtype,
initializer=kernel_initializer)
if self.data_format == 'NHWC':
strides = [1] + strides + [1]
else:
strides = [1, 1] + strides
return tf.nn.conv2d(input_layer, weights, strides, padding,
data_format=self.data_format)
def test_invalid_shape(self):
inputs = random_ops.random_uniform((10, 100, 100, 3), seed=1)
graph = conv_layers.conv2d(inputs, 3, 10, strides=(1, 1))
with self.assertRaisesRegexp(ValueError, 'number of features'):
graph = maxout.maxout(graph, num_units=2)
def testInvalidDataFormat(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 3), seed=1)
with self.assertRaisesRegexp(ValueError, 'data_format'):
conv_layers.conv2d(images, 32, 3, data_format='invalid')
def testConv2DFloat16(self):
height, width = 7, 9
images = random_ops.random_uniform((5, height, width, 4), dtype='float16')
output = conv_layers.conv2d(images, 32, [3, 3], activation=nn_ops.relu)
self.assertListEqual(output.get_shape().as_list(),
[5, height - 2, width - 2, 32])