def main():
tf.set_random_seed(10)
with tf.Session() as sess:
rnn_cell = tf.nn.rnn_cell.LSTMCell(10)
# defining initial state
initial_state = rnn_cell.zero_state(4, dtype=tf.float32)
inputs = tf.Variable(tf.random_uniform(shape = (4, 30, 100)), name='input')
inputs = tf.identity(inputs, "input_node")
# 'state' is a tensor of shape [batch_size, cell_state_size]
outputs, state = tf.nn.dynamic_rnn(rnn_cell, inputs, initial_state=initial_state, dtype=tf.float32)
y1 = tf.identity(outputs, 'outputs')
y2 = tf.identity(state, 'state')
t1 = tf.ones([4, 30, 10])
t2 = tf.ones([4, 10])
loss = tf.reduce_sum((y1 - t1) * (y1 - t1)) + tf.reduce_sum((y2 - t2) * (y2 - t2))
tf.identity(loss, name = "lstm_loss")
# tf.summary.FileWriter('/tmp/log', tf.get_default_graph())
net_outputs = map(lambda x: tf.get_default_graph().get_tensor_by_name(x), argv[2].split(','))
run_model(net_outputs, argv[1], None, argv[3] == 'True')
def testGradient(self):
# Set graph seed for determinism.
random_seed = 42
tf.set_random_seed(random_seed)
with self.test_session():
for test_case in self._TEST_CASES:
np.random.seed(random_seed)
in_shape = test_case['in_shape']
in_val = tf.constant(np.random.random(in_shape),
dtype=tf.float32)
for padding in ['VALID', 'SAME']:
out_val = tf.extract_image_patches(in_val,
test_case['ksizes'],
test_case['strides'],
test_case['rates'],
padding)
out_shape = out_val.get_shape().as_list()
err = tf.test.compute_gradient_error(
in_val, in_shape, out_val, out_shape
)
print('extract_image_patches gradient err: %.4e' % err)
self.assertLess(err, 1e-4)
def testAtrousFullyConvolutionalValues(self):
"""Verify dense feature extraction with atrous convolution."""
nominal_stride = 32
for output_stride in [4, 8, 16, 32, None]:
with slim.arg_scope(resnet_utils.resnet_arg_scope()):
with tf.Graph().as_default():
with self.test_session() as sess:
tf.set_random_seed(0)
inputs = create_test_input(2, 81, 81, 3)
# Dense feature extraction followed by subsampling.
output, _ = self._resnet_small(inputs, None,
is_training=False,
global_pool=False,
output_stride=output_stride)
if output_stride is None:
factor = 1
else:
factor = nominal_stride // output_stride
output = resnet_utils.subsample(output, factor)
# Make the two networks use the same weights.
tf.get_variable_scope().reuse_variables()
# Feature extraction at the nominal network rate.
expected, _ = self._resnet_small(inputs, None,
is_training=False,
global_pool=False)
sess.run(tf.global_variables_initializer())
self.assertAllClose(output.eval(), expected.eval(),
atol=1e-4, rtol=1e-4)
def testResumeTrainAchievesRoughlyTheSameLoss(self):
logdir = os.path.join(tempfile.mkdtemp(prefix=self.get_temp_dir()),
'tmp_logs')
number_of_steps = [300, 301, 305]
for i in range(len(number_of_steps)):
with tf.Graph().as_default():
tf.set_random_seed(i)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
tf_predictions = LogisticClassifier(tf_inputs)
slim.losses.log_loss(tf_predictions, tf_labels)
total_loss = slim.losses.get_total_loss()
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
train_op = slim.learning.create_train_op(
total_loss, optimizer)
loss = slim.learning.train(
train_op, logdir, number_of_steps=number_of_steps[i],
log_every_n_steps=10)
self.assertIsNotNone(loss)
self.assertLess(loss, .015)
def gradient_memory_mbs():
"""Evaluates gradient, prints peak memory."""
start_time0 = time.perf_counter()
start_time = start_time0
tf.reset_default_graph()
tf.set_random_seed(1)
train_op, loss = create_train_op_and_loss()
print("Graph construction: %.2f ms" %(1000*(time.perf_counter()-start_time)))
g = tf.get_default_graph()
ops = g.get_operations()
for op in ge.filter_ops_from_regex(ops, "block_layer"):
tf.add_to_collection("checkpoints", op.outputs[0])
sess = create_session()
sessrun(tf.global_variables_initializer())
start_time = time.perf_counter()
sessrun(train_op)
start_time = time.perf_counter()
print("loss %f"%(sess.run(loss),))
print("Compute time: %.2f ms" %(1000*(time.perf_counter()-start_time)))
mem_use = mem_util.peak_memory(run_metadata)['/gpu:0']/1e6
print("Memory used: %.2f MB "%(mem_use))
total_time = time.perf_counter()-start_time0
assert total_time < 100
return mem_use
def testSplitApplyMerge(self):
# Repeatability. SGD has a tendency to jump around, even here.
tf.set_random_seed(1)
with self.test_session() as sess:
# Use sampling to train REINFORCE
with st.value_type(st.SampleAndReshapeValue(n=1)):
(route_selection,
routing_loss,
final_loss) = build_split_apply_merge_model()
sgd = tf.train.GradientDescentOptimizer(1.0).minimize(final_loss)
tf.global_variables_initializer().run()
for i in range(10):
# Run loss and inference step. This toy problem converges VERY quickly.
(routing_loss_v, final_loss_v, route_selection_v, _) = sess.run(
[routing_loss, final_loss, tf.identity(route_selection), sgd])
print(
"Iteration %d, routing loss: %s, final_loss: %s, "
"route selection: %s"
% (i, routing_loss_v, final_loss_v, route_selection_v))
self.assertAllEqual([0, 0, 1, 1], route_selection_v)
self.assertAllClose([0.0, 0.0, 0.0, 0.0], routing_loss_v)
self.assertAllClose(0.0, final_loss_v)
def testNoneGlobalStep(self):
with tf.Graph().as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
tf_predictions = BatchNormClassifier(tf_inputs)
slim.losses.log_loss(tf_predictions, tf_labels)
total_loss = slim.losses.get_total_loss()
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
train_op = slim.learning.create_train_op(total_loss,
optimizer,
global_step=None)
global_step = slim.get_or_create_global_step()
with tf.Session() as sess:
# Initialize all variables
sess.run(tf.global_variables_initializer())
for _ in range(10):
sess.run([train_op])
global_step = global_step.eval()
# Since train_op don't use global_step it shouldn't change.
self.assertAllClose(global_step, 0)
def testEmptyUpdateOps(self):
with tf.Graph().as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
tf_predictions = BatchNormClassifier(tf_inputs)
slim.losses.log_loss(tf_predictions, tf_labels)
total_loss = slim.losses.get_total_loss()
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
train_op = slim.learning.create_train_op(total_loss, optimizer,
update_ops=[])
moving_mean = tf.contrib.framework.get_variables_by_name('moving_mean')[0]
moving_variance = tf.contrib.framework.get_variables_by_name(
'moving_variance')[0]
with tf.Session() as sess:
# Initialize all variables
sess.run(tf.global_variables_initializer())
mean, variance = sess.run([moving_mean, moving_variance])
# After initialization moving_mean == 0 and moving_variance == 1.
self.assertAllClose(mean, [0] * 4)
self.assertAllClose(variance, [1] * 4)
for _ in range(10):
sess.run([train_op])
mean = moving_mean.eval()
variance = moving_variance.eval()
# Since we skip update_ops the moving_vars are not updated.
self.assertAllClose(mean, [0] * 4)
self.assertAllClose(variance, [1] * 4)
def __init__(self):
# Import data
error = None
for _ in range(10):
try:
self.mnist = input_data.read_data_sets(
"/tmp/tensorflow/mnist/input_data", one_hot=True)
error = None
break
except Exception as e:
error = e
time.sleep(5)
if error:
raise ValueError("Failed to import data", error)
# Set seed and build layers
tf.set_random_seed(0)
self.x = tf.placeholder(tf.float32, [None, 784], name="x")
self.y_ = tf.placeholder(tf.float32, [None, 10], name="y_")
y_conv, self.keep_prob = deepnn(self.x)
# Need to define loss and optimizer attributes
self.loss = tf.reduce_mean(
tf.nn.softmax_cross_entropy_with_logits(
labels=self.y_, logits=y_conv))
self.optimizer = tf.train.AdamOptimizer(1e-4)
self.variables = ray_tf_utils.TensorFlowVariables(
self.loss, tf.get_default_session())
# For evaluating test accuracy
correct_prediction = tf.equal(
tf.argmax(y_conv, 1), tf.argmax(self.y_, 1))
self.accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
def initialize_parameters():
"""
Initializes parameters to build a neural network with tensorflow. The shapes are:
W1 : [25, 12288]
b1 : [25, 1]
W2 : [12, 25]
b2 : [12, 1]
W3 : [6, 12]
b3 : [6, 1]
Returns:
parameters -- a dictionary of tensors containing W1, b1, W2, b2, W3, b3
"""
tf.set_random_seed(1) # so that your "random" numbers match ours
### START CODE HERE ### (approx. 6 lines of code)
W1 = tf.get_variable("W1", [25,12288], initializer = tf.contrib.layers.xavier_initializer(seed = 1))
b1 = tf.get_variable("b1", [25,1], initializer = tf.zeros_initializer())
W2 = tf.get_variable("W2", [12,25], initializer = tf.contrib.layers.xavier_initializer(seed = 1))
b2 = tf.get_variable("b2", [12,1], initializer = tf.zeros_initializer())
W3 = tf.get_variable("W3", [6,12], initializer = tf.contrib.layers.xavier_initializer(seed = 1))
b3 = tf.get_variable("b3", [6,1], initializer = tf.zeros_initializer())
### END CODE HERE ###
parameters = {"W1": W1,
"b1": b1,
"W2": W2,
"b2": b2,
"W3": W3,
"b3": b3}
return parameters
def construct_graph(self, training, seed):
"""Returns a TensorflowGraph object."""
graph = tf.Graph()
# Lazily created by _get_shared_session().
shared_session = None
# Cache of TensorFlow scopes, to prevent '_1' appended scope names
# when subclass-overridden methods use the same scopes.
name_scopes = {}
# Setup graph
with graph.as_default():
if seed is not None:
tf.set_random_seed(seed)
(output, labels, weights) = self.build(graph, name_scopes, training)
if training:
loss = self.add_training_cost(graph, name_scopes, output, labels, weights)
else:
loss = None
output = self.add_output_ops(graph, output) # add softmax heads
return TensorflowGraph(
graph=graph,
session=shared_session,
name_scopes=name_scopes,
output=output,
labels=labels,
weights=weights,
loss=loss)
def wide_model(numeric_input, category_input, vocabs):
transpose_category_input = tf.transpose(category_input)
category_sum = None
# Append embadding category to numeric_sum
for i in range(0, len(vocabs)):
embedding = tf.get_variable("wideem" + str(i), [vocabs[i], 8],
initializer=tf.contrib.layers.xavier_initializer()
#partitioner=tf.fixed_size_partitioner(n_pss))
#partitioner=tf.min_max_variable_partitioner(n_pss, 0, 2 << 10)
)
# Pick one column from category input
col = tf.gather(transpose_category_input, [i])[0]
#col = tf.nn.embedding_lookup(transpose_category_input, [i])[0]
# Same as make [0001]*[w1,w2,w3,w4] = lookup w4
#embedded_col = embedding_lookup(tf.identity(embedding), col) # number * embedding output number
embedded_col = embedding_ops.embedding_lookup_unique(embedding, col)
if category_sum is None:
category_sum = embedded_col
else:
category_sum = tf.concat([category_sum, embedded_col], 1)
tf.set_random_seed(1)
w = tf.get_variable("W", [numeric_input.shape[1] + category_sum.shape[1], 1], initializer=tf.contrib.layers.xavier_initializer())
wmodel_logits_sum = tf.matmul(tf.concat([numeric_input, category_sum], 1), w)
return wmodel_logits_sum
def testGradientWithZeroWeight(self):
with tf.Graph().as_default():
tf.set_random_seed(0)
inputs = tf.ones((2, 3))
weights = tf.get_variable('weights',
shape=[3, 4],
initializer=tf.truncated_normal_initializer())
predictions = tf.matmul(inputs, weights)
optimizer = tf.train.MomentumOptimizer(learning_rate=0.001, momentum=0.9)
loss = tf.contrib.losses.mean_pairwise_squared_error(
predictions,
predictions,
0)
gradients_to_variables = optimizer.compute_gradients(loss)
init_op = tf.initialize_all_variables()
with self.test_session() as sess:
sess.run(init_op)
for grad, _ in gradients_to_variables:
np_grad = sess.run(grad)
self.assertFalse(np.isnan(np_grad).any())
def _do_sampling(self, logits, num_samples, sampler):
"""Samples using the supplied sampler and inputs.
Args:
logits: Numpy ndarray of shape [batch_size, num_classes].
num_samples: Int; number of samples to draw.
sampler: A sampler function that takes (1) a [batch_size, num_classes]
Tensor, (2) num_samples and returns a [batch_size, num_samples] Tensor.
Returns:
Frequencies from sampled classes; shape [batch_size, num_classes].
"""
with self.test_session() as sess:
tf.set_random_seed(1618)
op = sampler(tf.constant(logits), num_samples)
d = sess.run(op)
batch_size, num_classes = logits.shape
freqs_mat = []
for i in range(batch_size):
cnts = dict(collections.Counter(d[i, :]))
freqs = [(cnts[k] * 1. / num_samples if k in cnts else 0)
for k in range(num_classes)]
freqs_mat.append(freqs)
return freqs_mat
def main(_):
with tf.Session() as sess:
env = gym.make(ENV_NAME)
np.random.seed(RANDOM_SEED)
tf.set_random_seed(RANDOM_SEED)
env.seed(RANDOM_SEED)
state_dim = env.observation_space.shape[0]
action_dim = env.action_space.shape[0]
action_bound = env.action_space.high
# Ensure action bound is symmetric
assert (env.action_space.high == -env.action_space.low)
actor = ActorNetwork(sess, state_dim, action_dim, action_bound,
ACTOR_LEARNING_RATE, TAU)
critic = CriticNetwork(sess, state_dim, action_dim,
CRITIC_LEARNING_RATE, TAU, actor.get_num_trainable_vars())
if GYM_MONITOR_EN:
if not RENDER_ENV:
env = wrappers.Monitor(
env, MONITOR_DIR, video_callable=False, force=True)
else:
env = wrappers.Monitor(env, MONITOR_DIR, force=True)
train(sess, env, actor, critic)
if GYM_MONITOR_EN:
env.monitor.close()
def __init__(self, input_dim=None, output_dim=1, init_path=None, opt_algo='gd', learning_rate=1e-2, l2_weight=0,
random_seed=None):
Model.__init__(self)
init_vars = [('w', [input_dim, output_dim], 'xavier', dtype),
('b', [output_dim], 'zero', dtype)]
self.graph = tf.Graph()
with self.graph.as_default():
if random_seed is not None:
tf.set_random_seed(random_seed)
self.X = tf.sparse_placeholder(dtype)
self.y = tf.placeholder(dtype)
self.vars = utils.init_var_map(init_vars, init_path) # 初始化变量w, b
w = self.vars['w']
b = self.vars['b']
xw = tf.sparse_tensor_dense_matmul(self.X, w)
logits = tf.reshape(xw + b, [-1])
self.y_prob = tf.sigmoid(logits)
self.loss = tf.reduce_mean(
tf.nn.sigmoid_cross_entropy_with_logits(labels=self.y, logits=logits)) + \
l2_weight * tf.nn.l2_loss(xw)
self.optimizer = utils.get_optimizer(opt_algo, learning_rate, self.loss)
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
self.sess = tf.Session(config=config)
tf.global_variables_initializer().run(session=self.sess)
def testSampleMultipleTimes(self):
# 5 component mixture.
logits = [-10.0, -5.0, 0.0, 5.0, 10.0]
mus = [-5.0, 0.0, 5.0, 4.0, 20.0]
sigmas = [0.1, 5.0, 3.0, 0.2, 4.0]
n = 100
tf.set_random_seed(654321)
components = [
tfd.Normal(loc=mu, scale=sigma) for mu, sigma in zip(mus, sigmas)
]
cat = tfd.Categorical(logits, dtype=tf.int32, name="cat1")
dist1 = tfd.Mixture(
cat,
components,
name="mixture1",
use_static_graph=self.use_static_graph)
samples1 = self.evaluate(dist1.sample(n, seed=123456))
tf.set_random_seed(654321)
components2 = [
tfd.Normal(loc=mu, scale=sigma) for mu, sigma in zip(mus, sigmas)
]
cat2 = tfd.Categorical(logits, dtype=tf.int32, name="cat2")
dist2 = tfd.Mixture(
cat2,
components2,
name="mixture2",
use_static_graph=self.use_static_graph)
samples2 = self.evaluate(dist2.sample(n, seed=123456))
self.assertAllClose(samples1, samples2)
def _runSamplingBenchmark(self, name, create_distribution, use_gpu,
num_components, batch_size, num_features,
sample_size):
config = tf.ConfigProto()
config.allow_soft_placement = True
np.random.seed(127)
with tf.Session(config=config, graph=tf.Graph()) as sess:
tf.set_random_seed(0)
with tf.device("/device:GPU:0" if use_gpu else "/cpu:0"):
mixture = create_distribution(
num_components=num_components,
batch_size=batch_size,
num_features=num_features)
sample_op = mixture.sample(sample_size).op
sess.run(tf.global_variables_initializer())
reported = self.run_op_benchmark(
sess,
sample_op,
min_iters=10,
name=("%s_%s_components_%d_batch_%d_features_%d_sample_%d" %
(name, use_gpu, num_components, batch_size, num_features,
sample_size)))
tf.logging.vlog(2, "\t".join(["%s", "%d", "%d", "%d", "%d", "%g"]) % (
use_gpu, num_components, batch_size, num_features, sample_size,
reported["wall_time"]))
def main(_):
# Fixed seed for repeatability
seed = 8964
tf.set_random_seed(seed)
np.random.seed(seed)
random.seed(seed)
if FLAGS.legacy_mode and FLAGS.seq_length < 3:
raise ValueError('Legacy mode supports sequence length > 2 only.')
if not gfile.Exists(FLAGS.checkpoint_dir):
gfile.MakeDirs(FLAGS.checkpoint_dir)
train_model = model.Model(data_dir=FLAGS.data_dir,
is_training=True,
learning_rate=FLAGS.learning_rate,
beta1=FLAGS.beta1,
reconstr_weight=FLAGS.reconstr_weight,
smooth_weight=FLAGS.smooth_weight,
ssim_weight=FLAGS.ssim_weight,
icp_weight=FLAGS.icp_weight,
batch_size=FLAGS.batch_size,
img_height=FLAGS.img_height,
img_width=FLAGS.img_width,
seq_length=FLAGS.seq_length,
legacy_mode=FLAGS.legacy_mode)
train(train_model, FLAGS.pretrained_ckpt, FLAGS.checkpoint_dir,
FLAGS.train_steps, FLAGS.summary_freq)
def construct_graph(self, training, seed):
"""Returns a TensorflowGraph object."""
graph = tf.Graph()
# Lazily created by _get_shared_session().
shared_session = None
# Cache of TensorFlow scopes, to prevent '_1' appended scope names
# when subclass-overridden methods use the same scopes.
name_scopes = {}
# Setup graph
with graph.as_default():
if seed is not None:
tf.set_random_seed(seed)
features, labels, weights = self.add_placeholders(graph, name_scopes)
outputs = self.add_progressive_lattice(graph, name_scopes, training)
if training:
loss = self.add_task_training_costs(graph, name_scopes, outputs, labels,
weights)
else:
loss = None
return TensorflowGraph(
graph=graph,
session=shared_session,
name_scopes=name_scopes,
output=outputs,
labels=labels,
weights=weights,
loss=loss)
def testProbabilitiesCanBeChanged(self):
# Set up graph.
tf.set_random_seed(1234)
lbl1 = 0
lbl2 = 3
# This cond allows the necessary class queues to be populated.
label = tf.cond(tf.greater(0.5, tf.random_uniform([])), lambda: tf.constant(lbl1), lambda: tf.constant(lbl2))
val = [np.array([1, 4]) * label]
probs = tf.placeholder(tf.float32, shape=[5])
batch_size = 2
data_batch, labels = tf.contrib.training.stratified_sample_unknown_dist(val, label, probs, batch_size)
with self.test_session() as sess:
coord = tf.train.Coordinator()
threads = tf.train.start_queue_runners(coord=coord)
for _ in range(5):
[data], lbls = sess.run([data_batch, labels], feed_dict={probs: [1, 0, 0, 0, 0]})
for data_example in data:
self.assertListEqual([0, 0], list(data_example))
self.assertListEqual([0, 0], list(lbls))
# Now change distribution and expect different output.
for _ in range(5):
[data], lbls = sess.run([data_batch, labels], feed_dict={probs: [0, 0, 0, 1, 0]})
for data_example in data:
self.assertListEqual([3, 12], list(data_example))
self.assertListEqual([3, 3], list(lbls))
coord.request_stop()
coord.join(threads)
def _train_model(self, checkpoint_dir, num_steps):
"""Trains a simple classification model.
Note that the data has been configured such that after around 300 steps,
the model has memorized the dataset (e.g. we can expect %100 accuracy).
Args:
checkpoint_dir: The directory where the checkpoint is written to.
num_steps: The number of steps to train for.
"""
with tf.Graph().as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
tf_predictions = logistic_classifier(tf_inputs)
loss = tf.contrib.losses.log_loss(tf_predictions, tf_labels)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
train_op = tf.contrib.training.create_train_op(loss, optimizer)
loss = tf.contrib.training.train(
train_op, checkpoint_dir, hooks=[
tf.train.StopAtStepHook(num_steps)
])
def testCreateOnecloneWithPS(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
model_args = (tf_inputs, tf_labels)
deploy_config = model_deploy.DeploymentConfig(num_clones=1,
num_ps_tasks=1)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, model_args)
self.assertEqual(len(slim.get_variables()), 5)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.assertEqual(len(update_ops), 2)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
total_loss, grads_and_vars = model_deploy.optimize_clones(clones,
optimizer)
self.assertEqual(len(grads_and_vars), len(tf.trainable_variables()))
self.assertEqual(total_loss.op.name, 'total_loss')
for g, v in grads_and_vars:
self.assertDeviceEqual(g.device, '/job:worker/device:GPU:0')
self.assertDeviceEqual(v.device, '/job:ps/task:0/CPU:0')
def testCreateMulticloneWithPS(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
clone_args = (tf_inputs, tf_labels)
deploy_config = model_deploy.DeploymentConfig(num_clones=2,
num_ps_tasks=2)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, clone_args)
self.assertEqual(len(slim.get_variables()), 5)
for i, v in enumerate(slim.get_variables()):
t = i % 2
self.assertDeviceEqual(v.device, '/job:ps/task:%d/device:CPU:0' % t)
self.assertDeviceEqual(v.device, v.value().device)
self.assertEqual(len(clones), 2)
for i, clone in enumerate(clones):
self.assertEqual(
clone.outputs.op.name,
'clone_%d/BatchNormClassifier/fully_connected/Sigmoid' % i)
self.assertEqual(clone.scope, 'clone_%d/' % i)
self.assertDeviceEqual(clone.device, '/job:worker/device:GPU:%d' % i)
def testTrainWithTrace(self):
logdir = os.path.join(tempfile.mkdtemp(prefix=self.get_temp_dir()),
'tmp_logs')
with tf.Graph().as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
tf_predictions = LogisticClassifier(tf_inputs)
slim.losses.log_loss(tf_predictions, tf_labels)
total_loss = slim.losses.get_total_loss()
tf.summary.scalar('total_loss', total_loss)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
train_op = slim.learning.create_train_op(total_loss, optimizer)
loss = slim.learning.train(
train_op,
logdir,
number_of_steps=300,
log_every_n_steps=10,
trace_every_n_steps=100)
self.assertIsNotNone(loss)
for trace_step in [1, 101, 201]:
trace_filename = 'tf_trace-%d.json' % trace_step
self.assertTrue(
os.path.isfile(os.path.join(logdir, trace_filename)))
def testCreateLogisticClassifier(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = LogisticClassifier
clone_args = (tf_inputs, tf_labels)
deploy_config = model_deploy.DeploymentConfig(num_clones=1)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, clone_args)
clone = clones[0]
self.assertEqual(len(slim.get_variables()), 2)
for v in slim.get_variables():
self.assertDeviceEqual(v.device, 'CPU:0')
self.assertDeviceEqual(v.value().device, 'CPU:0')
self.assertEqual(clone.outputs.op.name,
'LogisticClassifier/fully_connected/Sigmoid')
self.assertEqual(clone.scope, '')
self.assertDeviceEqual(clone.device, 'GPU:0')
self.assertEqual(len(slim.losses.get_losses()), 1)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
self.assertEqual(update_ops, [])
def testCreateMulticlone(self):
g = tf.Graph()
with g.as_default():
tf.set_random_seed(0)
tf_inputs = tf.constant(self._inputs, dtype=tf.float32)
tf_labels = tf.constant(self._labels, dtype=tf.float32)
model_fn = BatchNormClassifier
clone_args = (tf_inputs, tf_labels)
num_clones = 4
deploy_config = model_deploy.DeploymentConfig(num_clones=num_clones)
self.assertEqual(slim.get_variables(), [])
clones = model_deploy.create_clones(deploy_config, model_fn, clone_args)
self.assertEqual(len(slim.get_variables()), 5)
for v in slim.get_variables():
self.assertDeviceEqual(v.device, 'CPU:0')
self.assertDeviceEqual(v.value().device, 'CPU:0')
self.assertEqual(len(clones), num_clones)
for i, clone in enumerate(clones):
self.assertEqual(
clone.outputs.op.name,
'clone_%d/BatchNormClassifier/fully_connected/Sigmoid' % i)
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS, clone.scope)
self.assertEqual(len(update_ops), 2)
self.assertEqual(clone.scope, 'clone_%d/' % i)
self.assertDeviceEqual(clone.device, 'GPU:%d' % i)
def simple_test(env_fn, learn_fn, min_reward_fraction, n_trials=N_TRIALS):
def seeded_env_fn():
env = env_fn()
env.seed(0)
return env
np.random.seed(0)
env = DummyVecEnv([seeded_env_fn])
with tf.Graph().as_default(), tf.Session(config=tf.ConfigProto(allow_soft_placement=True)).as_default():
tf.set_random_seed(0)
model = learn_fn(env)
sum_rew = 0
done = True
for i in range(n_trials):
if done:
obs = env.reset()
state = model.initial_state
if state is not None:
a, v, state, _ = model.step(obs, S=state, M=[False])
else:
a, v, _, _ = model.step(obs)
obs, rew, done, _ = env.step(a)
sum_rew += float(rew)
print("Reward in {} trials is {}".format(n_trials, sum_rew))
assert sum_rew > min_reward_fraction * n_trials, \
'sum of rewards {} is less than {} of the total number of trials {}'.format(sum_rew, min_reward_fraction, n_trials)
def main(hps):
# Initialize Horovod.
hvd.init()
# Create tensorflow session
sess = tensorflow_session()
# Download and load dataset.
tf.set_random_seed(hvd.rank() + hvd.size() * hps.seed)
np.random.seed(hvd.rank() + hvd.size() * hps.seed)
# Get data and set train_its and valid_its
train_iterator, test_iterator, data_init = get_data(hps, sess)
hps.train_its, hps.test_its, hps.full_test_its = get_its(hps)
# Create log dir
logdir = os.path.abspath(hps.logdir) + "/"
if not os.path.exists(logdir):
os.mkdir(logdir)
# Create model
import model
model = model.model(sess, hps, train_iterator, test_iterator, data_init)
# Initialize visualization functions
visualise = init_visualizations(hps, model, logdir)
if not hps.inference:
# Perform training
train(sess, model, hps, logdir, visualise)
else:
infer(sess, model, hps, test_iterator)
def __init__(self, env, discount = 0.90, learning_rate = 0.008):
self.env = env
self.observation_space = env.observation_space
self.action_space = env.action_space
self.action_space_n = self.action_space.n
self.n_input = len(self.observation_space.high)
self.n_hidden_1 = 20
#Learning Parameters
self.learning_rate = learning_rate
self.discount = discount
self.num_epochs = 20
self.batch_size = 32
self.graph = tf.Graph()
#Neural network is a Multi-Layered perceptron with one hidden layer containing tanh units
with self.graph.as_default():
tf.set_random_seed(1234)
self.weights = {
'h1': tf.Variable(tf.random_normal([self.n_input, self.n_hidden_1])),
'out': tf.Variable(tf.random_normal([self.n_hidden_1, 1]))
}
self.biases = {
'b1': tf.Variable(tf.random_normal([self.n_hidden_1])),
'out': tf.Variable(tf.random_normal([1]))
}
self.state_input = self.x = tf.placeholder("float", [None, len(self.observation_space.high)])#State input
self.return_input = tf.placeholder("float") #Target return
self.value_pred = self.multilayer_perceptron(self.state_input, self.weights, self.biases)
self.loss = tf.reduce_mean(tf.pow(self.value_pred - self.return_input,2))
self.optim = tf.train.AdamOptimizer(self.learning_rate).minimize(self.loss)
init = tf.initialize_all_variables()
print("Value Graph Constructed")
self.sess = tf.Session(graph = self.graph)
self.sess.run(init)
def init_tf(config_dict=dict()):
if tf.get_default_session() is None:
tf.set_random_seed(np.random.randint(1 << 31))
create_session(config_dict, force_as_default=True)
def __init__(self, params):
#############
## INIT
#############
# Get params, create logger, create TF session
self.params = params
self.logger = Logger(self.params['logdir'])
self.sess = create_tf_session(self.params['use_gpu'],
which_gpu=self.params['which_gpu'])
# Set random seeds
seed = self.params['seed']
tf.set_random_seed(seed)
np.random.seed(seed)
#############
## ENV
#############
# Make the gym environment
self.env = gym.make(self.params['env_name'])
self.env.seed(seed)
# Maximum length for episodes
self.params['ep_len'] = self.params[
'ep_len'] or self.env.spec.max_episode_steps
# Is this env continuous, or self.discrete?
discrete = isinstance(self.env.action_space, gym.spaces.Discrete)
self.params['agent_params']['discrete'] = discrete
# Observation and action sizes
ob_dim = self.env.observation_space.shape[0]
ac_dim = self.env.action_space.n if discrete else self.env.action_space.shape[
0]
self.params['agent_params']['ac_dim'] = ac_dim
self.params['agent_params']['ob_dim'] = ob_dim
# simulation timestep, will be used for video saving
if 'model' in dir(self.env):
self.fps = 1 / self.env.model.opt.timestep
else:
self.fps = self.env.env.metadata['video.frames_per_second']
#############
## AGENT
#############
agent_class = self.params['agent_class']
self.agent = agent_class(self.sess, self.env,
self.params['agent_params'])
#############
## INIT VARS
#############
## TODO initialize all of the TF variables (that were created by agent, etc.)
## HINT: use global_variables_initializer
self.sess.run(tf.global_variables_initializer())
import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
import matplotlib.pyplot as plt
tf.set_random_seed(1)
np.random.seed(1)
BATCH_SIZE = 50
LR = 0.001 # learning rate
mnist = input_data.read_data_sets('./data', one_hot=True) # they has been normalized to range (0,1)
test_x = mnist.test.images[:2000]
test_y = mnist.test.labels[:2000]
# plot one example
print(mnist.train.images.shape) # (55000, 28 * 28)
print(mnist.train.labels.shape) # (55000, 10)
plt.imshow(mnist.train.images[0].reshape((28, 28)), cmap='gray')
plt.title('%i' % np.argmax(mnist.train.labels[0])); plt.show()
tf_x = tf.placeholder(tf.float32, [None, 28*28]) / 255.
image = tf.reshape(tf_x, [-1, 28, 28, 1]) # (batch, height, width, channel)
tf_y = tf.placeholder(tf.int32, [None, 10]) # input y
# CNN
conv1 = tf.layers.conv2d( # shape (28, 28, 1)
inputs=image,
filters=16,
kernel_size=5,
strides=1,
#multi variable
import tensorflow as tf
tf.set_random_seed(777)
x_data = [[1,2],
[2,3],
[3, 1],
[4,3],
[5,3],
[6,2]]
y_data = [[0],
[0],
[0],
[1],
[1],
[1]
]
x = tf.placeholder(tf.float32, shape=[None,2])
y = tf.placeholder(tf.float32, shape=[None,1])
w = tf.Variable(tf.random_normal([2, 1]), name= 'weight')
#x의 열의 값과 동일해야한다.
b = tf.Variable(tf.random_normal([1]), name= 'bias')
hypothesis = tf.sigmoid(tf.matmul(x, w) + b) # wx+b
# 5, 3 * 3*1 =[5,1]
# cost = tf.reduce_mean(tf.square(hypothesis - y))
from random import randint, uniform
import time
import matplotlib.pyplot as plt
from keras.losses import sparse_categorical_crossentropy
from keras.preprocessing.image import ImageDataGenerator
import tensorflow as t
import random
import cv2
import keras.backend as K
import tensorflow as tf
import pandas as pd
from keras_preprocessing import image
import random as rn
np.random.seed(42)
rn.seed(12345)
tf.set_random_seed(1234)
# -----------------------------------------------------------------------------------------------
# import the essential functions required for computation
# sys.path.insert(0, os.path.expanduser('~//CNN_networks'))
# sys.export PYTHONPATH=/home/yaurehman2/PycharmProjects/face_anti_sp_newidea
print(sys.path)
from cnn_networks.VGG16_A_GAP_dual_inp import cnn_hybrid_color_single
from ess_func import read_pairs, sample_people, prewhiten, store_loss, hog_to_tensor, custom_loss
# -----------------------------------------------------------------------------------------------
def main(args):
# set the image parameters
@author: vp999274
"""
# keras module for building LSTM
from keras.preprocessing.sequence import pad_sequences
from keras.layers import Embedding, LSTM, Dense, Dropout, Bidirectional
from keras.preprocessing.text import Tokenizer
from keras.callbacks import EarlyStopping
from keras.models import Sequential
import keras.utils as ku
import pickle
# set seeds for reproducability
from tensorflow import set_random_seed
from numpy.random import seed
set_random_seed(2)
seed(1)
import pandas as pd
import numpy as np
import string, os
import warnings
warnings.filterwarnings("ignore")
warnings.simplefilter(action='ignore', category=FutureWarning)
curr_dir = 'input_data_headlines/'
all_headlines = []
#reading all files from the directory
for filename in os.listdir(curr_dir):
def _asLoomOpTest(self, max_depth):
with tf.Graph().as_default() as graph:
tf.set_random_seed(8)
np.random.seed(7)
ts = loom.TypeShape(tf.int64, ())
initializer = tf.random_uniform_initializer(dtype=tf.int64,
minval=0,
maxval=1 << 60)
@model_utils.as_loom_op([ts, ts], ts)
def f(x, y):
# Use a variable to make sure variable sharing works correctly
rand = tf.get_variable('rand',
shape=(),
dtype=tf.int64,
initializer=initializer)
return rand - x - y
@model_utils.as_loom_op([ts, ts], [ts, ts])
def g(x, y):
# Test multiple outputs
return x - y, x + y
def make_h():
# Ensure that we can reuse names for different calls to as_loom_op.
# Also test that the name argument to as_loom_op works.
@model_utils.as_loom_op([ts], ts, name='h')
def not_h(x):
v = tf.get_variable('yo',
shape=(),
dtype=tf.int64,
initializer=initializer)
return x + v
return not_h
# Make two h's to ensure they make separate variables
h1 = make_h()
h2 = make_h()
simple_loom = loom.Loom(named_ops={
'f': f,
'g': g,
'h1': h1,
'h2': h2
},
max_depth=max_depth)
self.assertEqual(['f/rand', 'h/yo', 'h_1/yo'],
[v.op.name for v in tf.global_variables()])
# Use f twice and (g,h1,h2) once each
weaver = simple_loom.make_weaver()
x, y, z = np.random.randint(1 << 60, size=3)
wx, wy, wz = weaver(x), weaver(y), weaver(z)
weaver.add_output(weaver.f(wx, weaver.f(wy, wz)))
plus, minus = weaver.g(wx, wy)
weaver.add_output(plus)
weaver.add_output(minus)
weaver.add_output(weaver.h1(wx))
weaver.add_output(weaver.h2(wx))
with self.test_session(graph=graph):
tf.global_variables_initializer().run()
out = simple_loom.output_tensor(ts).eval(
weaver.build_feed_dict())
self.assertEqual(out.shape, (5, ))
# out[0] works only if variables are shared between layers:
# rand - x - (rand - y - z) = y + z - x
self.assertEqual(out[0], y + z - x)
# out[1] and out[2] are simple
self.assertEqual(out[1], x - y)
self.assertEqual(out[2], x + y)
# out[3] and out[4] should use different random variables
self.assertNotEqual(out[3], out[4])
from keras.utils import to_categorical
from keras.callbacks import Callback
from sklearn.metrics import roc_auc_score
from mmoe import MMoE
SEED = 1
# Fix numpy seed for reproducibility
np.random.seed(SEED)
# Fix random seed for reproducibility
random.seed(SEED)
# Fix TensorFlow graph-level seed for reproducibility
tf.set_random_seed(SEED)
tf_session = tf.Session(graph=tf.get_default_graph())
K.set_session(tf_session)
# Simple callback to print out ROC-AUC
class ROCCallback(Callback):
def __init__(self, training_data, validation_data, test_data):
self.train_X = training_data[0]
self.train_Y = training_data[1]
self.validation_X = validation_data[0]
self.validation_Y = validation_data[1]
self.test_X = test_data[0]
self.test_Y = test_data[1]
def on_train_begin(self, logs={}):
import argparse
import sys
import tempfile
import tensorflow as tf
import numpy as np
from imgaug import augmenters as iaa
from utils import *
from taylor_batch_norm_unshared import taylor
np.random.seed(0)
tf.set_random_seed(0)
clip = False
def lr(iteration):
if iteration<3000:
return 1e-3
if iteration<10000:
return 1e-3
if iteration<15000:
return 1e-4
if iteration<20000:
return 1e-5
return 1e-4
def deepnn(x,n_classes):
is_train = tf.placeholder(tf.bool)
keep_prob = tf.placeholder(tf.float32)
h1 = tf.layers.conv2d(x, filters=32, kernel_size=(3, 3), padding='SAME', kernel_regularizer=tf.contrib.layers.l2_regularizer(scale=0.00005))
h = taylor(h1, k=2, is_train=is_train, name="Ac/1")
def main(_):
# Import data
mnist = FMNIST_aug()
mnist.data_augmentation()
n_classes = 10
x = tf.placeholder(tf.float32, [None, 28, 28, 1])
# Define loss and optimizer
y_ = tf.placeholder(tf.float32, [None, n_classes])
print("Gradient Clipping Status: "+str(clip))
learning_rate = tf.placeholder(tf.float32)
# Build the graph for the deep net
y_conv, is_train, keep_prob = deepnn(x, n_classes)
with tf.name_scope('Loss'):
def include_activation(name):
return ('activation_coeff' in name)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y_,
logits=y_conv)
cross_entropy = tf.reduce_mean(cross_entropy)
graph = tf.get_default_graph()
temp = [op.values()[0] for op in graph.get_operations() if ((len(op.values()) >= 1) and (include_activation(op.values()[0].name)))]
regl1_loss = 0.01 * tf.add_n([tf.reduce_sum(0.01* tf.abs(tf.cast(v, tf.float32))) for v in temp]),
regl2_loss = 0.01 * tf.add_n([tf.reduce_sum(0.01*tf.nn.l2_loss(tf.cast(v, tf.float32))) for v in temp])
reg_loss = regl1_loss
w_loss = tf.losses.get_regularization_loss()
with tf.name_scope('Adam_optimizer'):
update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
with tf.control_dependencies(update_ops):
if clip == True:
optimizer = tf.train.AdamOptimizer(learning_rate)
gradients, variables = zip(*optimizer.compute_gradients(cross_entropy+reg_loss+w_loss))
gradients, _ = tf.clip_by_global_norm(gradients, 5.0)
# gradients = [
# None if gradient is None else tf.clip_by_norm(gradient, 5.0)
# for gradient in gradients]
train_step = optimizer.apply_gradients(zip(gradients, variables))
else:
train_step = tf.train.AdamOptimizer(learning_rate).minimize(cross_entropy+reg_loss)
# train_step = tf.train.MomentumOptimizer(learning_rate, momentum=0.9).minimize(cross_entropy)
reg_step = tf.train.GradientDescentOptimizer(learning_rate).minimize(w_loss)
train_step = tf.group(train_step, reg_step)
with tf.name_scope('Accuracy'):
correct_prediction = tf.equal(tf.argmax(y_conv, 1), tf.argmax(y_, 1))
correct_prediction = tf.cast(correct_prediction, tf.float32)
accuracy = tf.reduce_mean(correct_prediction)
graph_location = tempfile.mkdtemp()
print('Saving graph to: %s' % graph_location)
train_writer = tf.summary.FileWriter(graph_location)
train_writer.add_graph(tf.get_default_graph())
acc = []
tf.set_random_seed(0)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for i in range(20000):
batch = mnist.next_train_batch(64, augment=False)
if i % 100 == 0:
train_accuracy = accuracy.eval(feed_dict={x: batch[0], y_: batch[1], is_train: True, keep_prob: 0.8})
a = []
for batchx, batchy in iterate_minibatches(mnist.get_test_images(), mnist.get_test_labels(), 1000):
a.append(accuracy.eval(feed_dict={x: batchx, y_: batchy, is_train: False, keep_prob: 1.0}))
print('Step %d, Training accuracy %g, Testing accuracy %g' % (i, train_accuracy, np.mean(a)))
acc.append(a)
train_step.run(feed_dict={x: batch[0], y_: batch[1], is_train: True, learning_rate: lr(i), keep_prob: 0.8})
from tqdm import tqdm
from ccks_all.cut_text import train_text_dic, all_text_dic, kb_all_text_dic, load_cut_text
from keras import backend as K
from ccks_all.modeling.bertmodel.preprocess import PairTokenizer, BertPreProcess, Preprocess, BertNerProcess
from ccks_all.modeling.utils import Metrics
max_q = 30
max_d = 450
seed = 123
random.seed(seed)
np.random.seed(seed)
tf.set_random_seed(seed)
root_dir = r"D:\data\biendata\ccks2019_el\ccks_train_data\{}"
bert_dir = r'D:\data\bert\chinese-bert_chinese_wwm_L-12_H-768_A-12'
config_path = bert_dir + r"\bert_config.json"
checkpoint_path = bert_dir + r"\bert_model.ckpt"
dict_path = bert_dir + r'\vocab.txt'
bert_model = load_trained_model_from_checkpoint(config_path, checkpoint_path)
# train_dir = r"C:\python3workspace\kera_ner_demo\ccks_ner\modeling\pair_model\dt\m3\{}"
model_dir = r"D:\data\biendata\ccks2019_el\entityclf\m21\{}"
log_filepath = model_dir.format(r"log")
model_path = model_dir.format(r"best_model.hdf5")
def train(
self,
training_data: "TrainingData",
cfg: Optional["RasaNLUModelConfig"] = None,
**kwargs: Any,
) -> None:
"""Train the embedding intent classifier on a data set."""
logger.debug("Started training embedding classifier.")
# set numpy random seed
np.random.seed(self.random_seed)
session_data = self.preprocess_train_data(training_data)
possible_to_train = self._check_enough_labels(session_data)
if not possible_to_train:
logger.error("Can not train a classifier. "
"Need at least 2 different classes. "
"Skipping training of classifier.")
return
if self.evaluate_on_num_examples:
session_data, eval_session_data = train_utils.train_val_split(
session_data,
self.evaluate_on_num_examples,
self.random_seed,
label_key="label_ids",
)
else:
eval_session_data = None
self.graph = tf.Graph()
with self.graph.as_default():
# set random seed
tf.set_random_seed(self.random_seed)
# allows increasing batch size
batch_size_in = tf.placeholder(tf.int64)
(
self._iterator,
train_init_op,
eval_init_op,
) = train_utils.create_iterator_init_datasets(
session_data,
eval_session_data,
batch_size_in,
self.batch_in_strategy,
label_key="label_ids",
)
self._is_training = tf.placeholder_with_default(False, shape=())
loss, acc = self._build_tf_train_graph(session_data)
# define which optimizer to use
self._train_op = tf.train.AdamOptimizer().minimize(loss)
# train tensorflow graph
self.session = tf.Session(config=self._tf_config)
train_utils.train_tf_dataset(
train_init_op,
eval_init_op,
batch_size_in,
loss,
acc,
self._train_op,
self.session,
self._is_training,
self.epochs,
self.batch_in_size,
self.evaluate_on_num_examples,
self.evaluate_every_num_epochs,
)
# rebuild the graph for prediction
self.pred_confidence = self._build_tf_pred_graph(session_data)
def main():
args = parser.parse_args()
print(args)
log_dir = args.logdir + '/' + str(args.numX) + '/ep1st' + str(args.epsilon1) + '_ep2nd' + str(args.epsilon2) \
+ '_zeta' + str(args.zeta) + '_seed' + str(args.seed) + '_' + str(datetime.now())
if not os.path.exists(log_dir):
os.makedirs(log_dir)
with open(os.path.join(log_dir, 'args.txt'), 'w') as f:
f.write(str(args))
# set random seed
tf.set_random_seed(args.seed)
np.random.seed(args.seed)
# set gpu
deviceIDs = GPUtil.getAvailable(order='first',
limit=4,
maxLoad=0.1,
maxMemory=0.1,
excludeID=[],
excludeUUID=[])
print('Unloaded gpu:', deviceIDs)
os.environ['CUDA_VISIBLE_DEVICES'] = str(deviceIDs[0])
# dataset 每个像素的数值是在0～1之间的
# Dataset = np.load(os.environ['HOME'] + '/datasets/fashion/fashion.npz')
Dataset = np.load('fashion.npz')
Xul_train = Dataset['xtrain'][0:60000] # 维数为(60000, 784)
Yul_train = Dataset['ytrain'][0:60000]
X_test = Dataset['xtest']
Y_test = Dataset['ytest']
# mask = np.random.choice(60000, 500, False)
mask = np.arange(0, args.numX)
X_train = Xul_train[mask]
Y_train = Yul_train[mask]
for i in range(10):
print('class:', i, np.sum((Y_train[:, i] == 1)))
# build graph
lr = tf.placeholder(tf.float32, [], name='lr')
x = tf.placeholder(tf.float32, [None, 784], name='x')
y = tf.placeholder(tf.float32, [None, 10], name='y')
x_ul = tf.placeholder(tf.float32, [None, 784], name='xul')
vae = cVAE(args.latent_dim)
net = Net()
out = net.classifier(x)
out_ul = net.classifier(x_ul)
mu, logvar = vae.encode(x_ul, out_ul)
z = vae.reparamenterize(mu, logvar)
x_recon = vae.decode(z, out_ul)
x_gen = vae.decode(
tf.random_normal([tf.shape(out_ul)[0], args.latent_dim]), out_ul)
# conditional vae graph
vae_loss = vae.BCE(x_recon, x_ul, mu, logvar) + vae.KLD(mu, logvar)
vae_weight_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES,
scope='cvae')
vae_train_step = tf.train.AdamOptimizer(lr).minimize(
vae_loss, var_list=vae_weight_list)
# TNAR graph
r0 = tf.zeros_like(z, name='zero_holder') # [128 100]
x_recon_r0 = vae.decode(z + r0, out_ul) # [128 784]
diff2 = 0.5 * tf.reduce_sum((x_recon - x_recon_r0)**2, axis=1) # [128,]
# tf.gradients(y, x) 等于 tf.gradients(tf.reduce_sum(y), x)，其shape和x的shape一样
diffJaco = tf.gradients(diff2, r0)[0] # [128 100]
def normalizevector(r):
r /= (1e-12 + tf.reduce_max(tf.abs(r), axis=1, keepdims=True))
return r / tf.sqrt(tf.reduce_sum(r**2, axis=1, keepdims=True) + 1e-6)
# power method
r_adv = normalizevector(tf.random_normal(shape=tf.shape(z)))
for j in range(1):
r_adv = 1e-6 * r_adv
x_r = vae.decode(z + r_adv, out_ul)
out_r = net.classifier(x_r - x_recon + x_ul)
kl = net.kldivergence(out_ul, out_r)
r_adv = tf.stop_gradient(tf.gradients(kl, r_adv)[0]) / 1e-6
r_adv = normalizevector(r_adv)
# begin cg
rk = r_adv + 0
pk = rk + 0
xk = tf.zeros_like(rk)
for k in range(4):
Bpk = tf.stop_gradient(tf.gradients(diffJaco * pk, r0)[0])
pkBpk = tf.reduce_sum(pk * Bpk, axis=1, keepdims=True)
rk2 = tf.reduce_sum(rk * rk, axis=1, keepdims=True)
alphak = (rk2 / (pkBpk + 1e-8)) * tf.cast((rk2 > 1e-8), tf.float32)
xk += alphak * pk
rk -= alphak * Bpk
betak = tf.reduce_sum(rk * rk, axis=1,
keepdims=True) / (rk2 + 1e-8)
pk = rk + betak * pk
# end cg
r_adv = normalizevector(xk)
x_adv = vae.decode(z + r_adv * args.epsilon1, out_ul)
r_x = x_adv - x_recon
out_adv = net.classifier(x_ul + r_x)
r_x = normalizevector(r_x)
r_adv_orth = normalizevector(tf.random_normal(shape=tf.shape(x_ul)))
for j in range(1):
r_adv_orth1 = 1e-6 * r_adv_orth
out_r = net.classifier(x_ul + r_adv_orth1)
kl = net.kldivergence(out_ul, out_r)
r_adv_orth1 = tf.stop_gradient(tf.gradients(kl, r_adv_orth1)[0]) / 1e-6
r_adv_orth = r_adv_orth1 - args.zeta * (
tf.reduce_sum(r_x * r_adv_orth, axis=1, keepdims=True) *
r_x) + args.zeta * r_adv_orth
r_adv_orth = normalizevector(r_adv_orth)
out_adv_orth = net.classifier(x_ul + r_adv_orth * args.epsilon2) ###
# TNAR loss
vat_loss = net.kldivergence(tf.stop_gradient(out_ul), out_adv)
vat_loss_orth = net.kldivergence(tf.stop_gradient(out_ul), out_adv_orth)
en_loss = net.crossentropy(out_ul, out_ul)
ce_loss = net.crossentropy(y, out)
total_loss = ce_loss + args.coef_vat1 * vat_loss + args.coef_vat2 * vat_loss_orth + args.coef_ent * en_loss
weight_list = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='net')
train_step = tf.train.AdamOptimizer(lr).minimize(total_loss,
var_list=weight_list)
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(out, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
# summary
train_summary_list = [
tf.summary.scalar('loss/total', total_loss),
tf.summary.scalar('loss/crossentropy', ce_loss),
tf.summary.scalar('loss/entropy', en_loss),
tf.summary.scalar('loss/vat', vat_loss),
tf.summary.scalar('acc/train', accuracy),
]
image_summary_list = [
tf.summary.image('x',
tf.reshape(x_ul, [-1, 28, 28, 1]),
max_outputs=32),
tf.summary.image('x_recon',
tf.reshape(x_recon, [-1, 28, 28, 1]),
max_outputs=32),
tf.summary.image('x_gen',
tf.reshape(x_gen, [-1, 28, 28, 1]),
max_outputs=32),
tf.summary.image('x_adv',
tf.reshape(x_adv, [-1, 28, 28, 1]),
max_outputs=32),
tf.summary.image('r_adv',
tf.reshape(x_adv - x_recon, [-1, 28, 28, 1]),
max_outputs=32),
tf.summary.image('r_adv_orth',
tf.reshape(r_adv_orth * args.epsilon2,
[-1, 28, 28, 1]),
max_outputs=32),
tf.summary.image('x_adv_orth',
tf.reshape(x_ul + r_adv_orth * args.epsilon2,
[-1, 28, 28, 1]),
max_outputs=32)
]
_acc = tf.placeholder(tf.float32, name='acc_summary')
test_summary_list = [tf.summary.scalar('acc/test', _acc)]
train_summary_merged = tf.summary.merge(train_summary_list +
image_summary_list)
test_summary_merged = tf.summary.merge(test_summary_list)
saver = tf.train.Saver(max_to_keep=100)
# optimization
gpu_options = tf.GPUOptions(allow_growth=True)
with tf.Session(config=tf.ConfigProto(gpu_options=gpu_options)) as sess:
writer = tf.summary.FileWriter(log_dir, sess.graph)
sess.run(tf.global_variables_initializer(), feed_dict={lr: args.lr})
for ep in range(args.epoch):
if ep < args.lr_decay_epoch:
decayed_lr = args.lr
else:
decayed_lr = args.lr * (
args.epoch - ep) / float(args.epoch - args.lr_decay_epoch)
for i in range(args.iter_per_epoch):
mask = np.random.choice(len(X_train), args.batch_size, False)
mask_ul = np.random.choice(len(Xul_train), args.batch_size_ul,
False)
# optimize cls
_, loss_ce, loss_vat, loss_vat2, train_summary = sess.run(
[
train_step, ce_loss, vat_loss, vat_loss_orth,
train_summary_merged
],
feed_dict={
x: X_train[mask],
y: Y_train[mask],
x_ul: Xul_train[mask_ul],
lr: decayed_lr,
vae.is_train: False
})
# optimize vae
_, loss_vae = sess.run([vae_train_step, vae_loss],
feed_dict={
x_ul: Xul_train[mask_ul],
lr: decayed_lr,
vae.is_train: True
})
acc = 0
for j in range(20):
acc += sess.run(accuracy,
feed_dict={
x: X_test[500 * j:500 * (j + 1)],
y: Y_test[500 * j:500 * (j + 1)]
})
acc /= 20
test_summary = sess.run(test_summary_merged,
feed_dict={
x: X_test[0:128],
y: Y_test[0:128],
_acc: acc
})
writer.add_summary(train_summary, ep)
writer.add_summary(test_summary, ep)
print('epoch', ep, 'ce', loss_ce, 'vat1', loss_vat, 'vat2',
loss_vat2, 'vae', loss_vae)
print('epoch', ep, 'acc', acc)
if ep % 10 == 0:
saver.save(sess, os.path.join(log_dir, 'model'), ep)
print(args)
flags.DEFINE_string("occlude_start_row", 18, "image row to start occlusion")
flags.DEFINE_string("num_generated_images", 9, "number of images to generate")
# Debug
flags.DEFINE_boolean("is_train", True, "training or testing")
flags.DEFINE_string("log_level", "INFO", "log level [DEBUG, INFO, WARNING, ERROR, CRITICAL]")
flags.DEFINE_integer("random_seed", 123, "random seed for python")
conf = flags.FLAGS
# logging
logger = logging.getLogger()
logger.setLevel(conf.log_level)
# random seed
tf.set_random_seed(conf.random_seed)
np.random.seed(conf.random_seed)
def validate_parameters(conf):
if conf.data not in ["mnist", "color-mnist", "cifar"]:
raise ValueError("Configuration parameter 'data' is '{}'. Must be one of [mnist, color-mnist, cifar]"
.format(conf.data))
def preprocess(q_levels):
def preprocess_fcn(images, labels):
# Create the target pixels from the image. Quantize the scalar pixel values into q_level indices.
target_pixels = np.clip(((images * q_levels).astype('int64')), 0, q_levels - 1) # [N,H,W,C]
return (images, target_pixels)
X_test = X_test_flatten/255.
# Convert training and test labels to one hot matrices
Y_train = one_hot_encoding(Y_train_orig, 6)
Y_test = one_hot_encoding(Y_test_orig, 6)
print ("train set size" + str(X_train.shape[1]))
print ("test set size" + str(X_test.shape[1]))
print ("X_train shape " + str(X_train.shape))
print ("Y_train shape " + str(Y_train.shape))
print ("X_test shape " + str(X_test.shape))
print ("Y_test shape " + str(Y_test.shape))
print("START TRAINING")
ops.reset_default_graph() # Clears the default graph stack and resets the global default graph. To rerun the model without overwriting tf variables
tf.set_random_seed(1) # 1 to keep consistent results
(n_x, m) = X_train.shape # (n_x: input size, m : number of examples in the train set)
n_y = Y_train.shape[0] # n_y : output size
costs = [] # To keep track of the cost
epoch_n = []
learning_rate = 0.0001
num_epochs = 1500
minibatch_size = 32
seed=3
### CREATE PLACEHOLDERS
X = tf.placeholder(dtype = tf.float32, shape = [n_x, None], name = "X")
Y = tf.placeholder(dtype = tf.float32, shape = [n_y, None], name = "Y")
###INITIALIZE PARAMETERS
# Lab 11 MNIST and Deep learning CNN
import tensorflow as tf
import random
from tensorflow.examples.tutorials.mnist import input_data
tf.set_random_seed(777) # reproducibility
mnist = input_data.read_data_sets("MNIST_data", one_hot=True)
# hyper parameters
learning_rate = 0.001
trainig_epochs = 15
batch_size = 100
# dropout (keep_prob) rate 0.7~0.5 on training, but should be 1 for testing
keep_prob = tf.placeholder(tf.float32)
# input place holders
X = tf.placeholder(tf.float32, [None, 784])
Y = tf.placeholder(tf.float32, [None, 10])
# img 28x28x1 (black/white)
# input 사이즈는 모르기 때문에 -1로 셋팅
X_img = tf.reshape(X, [-1, 28, 28, 1])
# L1 ImageIn shape = (?, 28, 28, 1)
W1 = tf.Variable(tf.random_normal([3, 3, 1, 32], stddev=0.01))
# Conv -> (?, 28, 28, 32)
# Pool -> (?, 14, 14, 32)
L1 = tf.nn.conv2d(X_img, W1, strides=[1, 1, 1, 1], padding='SAME')
L1 = tf.nn.relu(L1)
# print(action)
# observation, reward, done = env.step(action)
def one_hot(num, maximum):
one_hot = np.zeros(maximum)
one_hot[num] = 1
return one_hot
# def main(_):
with tf.Session() as sess:
train_step = 0
env = env2D(ENVIRONMENT)
np.random.seed(RANDOM_SEED)
tf.set_random_seed(RANDOM_SEED)
env.seed = random.seed(RANDOM_SEED)
state_dim = env.observation_flat().shape[0]
action_dim = env.action_space.shape[0]
# create network
net = Network(sess, state_dim, action_dim, LEARNING_RATE, TAU)
# train(sess, env, net)
sess.run(tf.global_variables_initializer())
saver = tf.train.Saver()
writer = tf.summary.FileWriter(SUMMARY_DIR, sess.graph)
from model.model_fn import model_fn
from model.training import train_and_evaluate
parser = argparse.ArgumentParser()
parser.add_argument('--model_dir', default='experiments/base_model',
help="Experiment directory containing params.json")
parser.add_argument('--data_dir', default='data/processed_data',
help="Directory containing the dataset")
parser.add_argument('--restore_from', default=None,
help="Optional, directory or file containing weights to reload before training")
if __name__ == '__main__':
# Set the random seed for the whole graph for reproducible experiments
tf.set_random_seed(230)
# Load the parameters from json file
args = parser.parse_args()
json_path = os.path.join(args.model_dir, 'params.json')
assert os.path.isfile(json_path), "No json configuration file found at {}".format(json_path)
params = Params(json_path)
# Check that we are not overwriting some previous experiment
# Comment these lines if you are developing your model and don't care about overwritting
# model_dir_has_best_weights = os.path.isdir(os.path.join(args.model_dir, "best_weights"))
# overwritting = model_dir_has_best_weights and args.restore_from is None
# assert not overwritting, "Weights found in model_dir, aborting to avoid overwrite"
# Set the logger
set_logger(os.path.join(args.model_dir, 'train.log'))
scaling = args.scaling
### initialize random seed generator of numpy
np.random.seed(random_seed)
#--------------------------------------------------------------------
# import keras and its backend (e.g., tensorflow)
#--------------------------------------------------------------------
os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" # see issue #152
if gpu_id >= 0:
os.environ["CUDA_VISIBLE_DEVICES"] = str(gpu_id)
else:
os.environ["CUDA_VISIBLE_DEVICES"] = ''
os.environ['TF_CPP_MIN_LOG_LEVEL']='2' # supress warning messages
import tensorflow as tf
tf.set_random_seed(random_seed) # initialize random seed generator of tensorflow
from keras.layers import Dense, Dropout
from keras.models import Sequential, load_model
train_df = pd.read_csv(path_train, header=0) # pass header=0 to be able to replace existing names
train_AP_features = scale(np.asarray(train_df.iloc[:,0:520]).astype(float), axis=1) # convert integer to float and scale jointly (axis=1)
train_df['REFPOINT'] = train_df.apply(lambda row: str(int(row['SPACEID'])) + str(int(row['RELATIVEPOSITION'])), axis=1) # add a new column
# map reference points to sequential IDs per building-floor before building labels
blds = np.unique(train_df[['BUILDINGID']])
flrs = np.unique(train_df[['FLOOR']])
for bld in blds:
for flr in flrs:
cond = (train_df['BUILDINGID']==bld) & (train_df['FLOOR']==flr)
_, idx = np.unique(train_df.loc[cond, 'REFPOINT'], return_inverse=True) # refer to numpy.unique manual
train_df.loc[cond, 'REFPOINT'] = idx
def main():
np.random.seed(1234)
tf.set_random_seed(1237)
M, N, train_data, valid_data, test_data, user_movie, user_movie_score, \
movie_user, movie_user_score = dataset.load_movielens1m_mapped(
os.path.join(conf.data_dir, 'ml-1m.zip'))
# set configurations and hyper parameters
D = 30
batch_size = 100000
K = 8
n_epochs = 500
eval_freq = 10
# paralleled
chunk_size = 50
N = (N + chunk_size - 1) // chunk_size
N *= chunk_size
M = (M + chunk_size - 1) // chunk_size
M *= chunk_size
# Selection
neighbor_u = tf.placeholder(tf.int32, shape=[None], name="neighbor_u")
neighbor_v = tf.placeholder(tf.int32, shape=[None], name="neighbor_v")
select_u = tf.placeholder(tf.int32, shape=[None], name="select_u")
select_v = tf.placeholder(tf.int32, shape=[None], name="select_v")
alpha_u = 1.0
alpha_v = 1.0
alpha_pred = 0.2 / 4.0
# Define samples as variables
Us = []
Vs = []
for i in range(N // chunk_size):
ui = tf.get_variable('u_chunk_%d' % i,
shape=[K, chunk_size, D],
initializer=tf.random_normal_initializer(0, 0.1),
trainable=False)
Us.append(ui)
for i in range(M // chunk_size):
vi = tf.get_variable('v_chunk_%d' % i,
shape=[K, chunk_size, D],
initializer=tf.random_normal_initializer(0, 0.1),
trainable=False)
Vs.append(vi)
U = tf.concat(Us, axis=1)
V = tf.concat(Vs, axis=1)
n = tf.placeholder(tf.int32, shape=[], name='n')
m = tf.placeholder(tf.int32, shape=[], name='m')
model = pmf(n, m, D, K, select_u, select_v, alpha_u, alpha_v, alpha_pred)
# prediction
true_rating = tf.placeholder(tf.float32, shape=[None], name='true_rating')
normalized_rating = (true_rating - 1.0) / 4.0
pred_rating = model.observe(u=U, v=V)["r"]
pred_rating = tf.reduce_mean(pred_rating, axis=0)
rmse = tf.sqrt(tf.reduce_mean(
tf.square(pred_rating - normalized_rating))) * 4
hmc_u = zs.HMC(step_size=1e-3,
n_leapfrogs=10,
adapt_step_size=None,
target_acceptance_rate=0.9)
hmc_v = zs.HMC(step_size=1e-3,
n_leapfrogs=10,
adapt_step_size=None,
target_acceptance_rate=0.9)
target_u = tf.gather(U, neighbor_u, axis=1)
target_v = tf.gather(V, neighbor_v, axis=1)
candidate_sample_u = tf.get_variable(
'cand_sample_chunk_u',
shape=[K, chunk_size, D],
initializer=tf.random_normal_initializer(0, 0.1),
trainable=True)
candidate_sample_v = tf.get_variable(
'cand_sample_chunk_v',
shape=[K, chunk_size, D],
initializer=tf.random_normal_initializer(0, 0.1),
trainable=True)
def log_joint(bn):
log_pu, log_pv = bn.cond_log_prob(['u', 'v']) # [K, N], [K, M]
log_pr = bn.cond_log_prob('r') # [K, batch]
log_pu = tf.reduce_sum(log_pu, axis=-1)
log_pv = tf.reduce_sum(log_pv, axis=-1)
log_pr = tf.reduce_sum(log_pr, axis=-1)
return log_pu + log_pv + log_pr
model.log_joint = log_joint
sample_u_op, sample_u_info = hmc_u.sample(model, {
"r": normalized_rating,
"v": target_v
}, {"u": candidate_sample_u})
sample_v_op, sample_v_info = hmc_v.sample(model, {
"r": normalized_rating,
"u": target_u
}, {"v": candidate_sample_v})
candidate_idx_u = tf.placeholder(tf.int32,
shape=[chunk_size],
name='cand_u_chunk')
candidate_idx_v = tf.placeholder(tf.int32,
shape=[chunk_size],
name='cand_v_chunk')
candidate_u = tf.gather(U, candidate_idx_u, axis=1) # [K, chunk_size, D]
candidate_v = tf.gather(V, candidate_idx_v, axis=1) # [K, chunk_size, D]
trans_cand_U = tf.assign(candidate_sample_u, candidate_u)
trans_cand_V = tf.assign(candidate_sample_v, candidate_v)
trans_us_cand = []
for i in range(N // chunk_size):
trans_us_cand.append(tf.assign(Us[i], candidate_sample_u))
trans_vs_cand = []
for i in range(M // chunk_size):
trans_vs_cand.append(tf.assign(Vs[i], candidate_sample_v))
# Run the inference
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(1, n_epochs + 1):
epoch_time = -time.time()
for i in range(N // chunk_size):
nv, sv, tr, ssu, ssv = select_from_corpus(
i * chunk_size, (i + 1) * chunk_size, user_movie,
user_movie_score)
_ = sess.run(trans_cand_U,
feed_dict={
candidate_idx_u:
list(
range(i * chunk_size,
(i + 1) * chunk_size))
})
_ = sess.run(sample_u_op,
feed_dict={
neighbor_v: sv,
true_rating: tr,
select_u: ssu,
select_v: ssv,
n: chunk_size,
m: nv
})
_ = sess.run(trans_us_cand[i])
for i in range(M // chunk_size):
nu, su, tr, ssv, ssu = select_from_corpus(
i * chunk_size, (i + 1) * chunk_size, movie_user,
movie_user_score)
_ = sess.run(trans_cand_V,
feed_dict={
candidate_idx_v:
list(
range(i * chunk_size,
(i + 1) * chunk_size))
})
_ = sess.run(sample_v_op,
feed_dict={
neighbor_u: su,
true_rating: tr,
select_u: ssu,
select_v: ssv,
n: nu,
m: chunk_size
})
_ = sess.run(trans_vs_cand[i])
epoch_time += time.time()
print("Epoch {}: {:.1f}s".format(epoch, epoch_time))
def _eval(phase, data, batch_size):
rmses = []
sizes = []
time_eval = -time.time()
n_iters = (data.shape[0] + batch_size - 1) // batch_size
for t in range(n_iters):
su = data[t * batch_size:(t + 1) * batch_size, 0]
sv = data[t * batch_size:(t + 1) * batch_size, 1]
tr = data[t * batch_size:(t + 1) * batch_size, 2]
re = sess.run(rmse,
feed_dict={
select_u: su,
select_v: sv,
n: N,
m: M,
true_rating: tr
})
rmses.append(re)
sizes.append(tr.shape[0])
time_eval += time.time()
print('>>> {} ({:.1f}s): rmse = {}'.format(
phase, time_eval, average_rmse_over_batches(rmses, sizes)))
_eval("Train", train_data, batch_size)
if epoch % eval_freq == 0:
_eval("Validation", valid_data, batch_size)
_eval("Test", test_data, batch_size)
def reset_graph(seed = 42):
#to make the results reproducible across runs
tf.reset_default_graph()
tf.set_random_seed(seed)
np.random.seed(seed)
def run(data_dir, data_fname, split_seed, ntrain_div_classes,
attr_normalization, alpha, eps, topk, ppr_normalization, hidden_size,
nlayers, weight_decay, dropout, lr, max_epochs, batch_size,
batch_mult_val, eval_step, run_val, early_stop, patience,
nprop_inference, inf_fraction):
'''
Run training and inference.
Parameters
----------
data_dir:
Directory containing .npz data files.
data_fname:
Name of .npz data file.
split_seed:
Seed for splitting the dataset into train/val/test.
ntrain_div_classes:
Number of training nodes divided by number of classes.
attr_normalization:
Attribute normalization. Not used in the paper.
alpha:
PPR teleport probability.
eps:
Stopping threshold for ACL's ApproximatePR.
topk:
Number of PPR neighbors for each node.
ppr_normalization:
Adjacency matrix normalization for weighting neighbors.
hidden_size:
Size of the MLP's hidden layer.
nlayers:
Number of MLP layers.
weight_decay:
Weight decay used for training the MLP.
dropout:
Dropout used for training.
lr:
Learning rate.
max_epochs:
Maximum number of epochs (exact number if no early stopping).
batch_size:
Batch size for training.
batch_mult_val:
Multiplier for validation batch size.
eval_step:
Accuracy is evaluated after every this number of steps.
run_val:
Evaluate accuracy on validation set during training.
early_stop:
Use early stopping.
patience:
Patience for early stopping.
nprop_inference:
Number of propagation steps during inference
inf_fraction:
Fraction of nodes for which local predictions are computed during inference.
'''
start = time.time()
(adj_matrix, attr_matrix, labels, train_idx, val_idx,
test_idx) = utils.get_data(f"{data_dir}/{data_fname}",
seed=split_seed,
ntrain_div_classes=ntrain_div_classes,
normalize_attr=attr_normalization)
try:
d = attr_matrix.n_columns
except AttributeError:
d = attr_matrix.shape[1]
nc = labels.max() + 1
time_loading = time.time() - start
logging.info('Loading done.')
# compute the ppr vectors for train/val nodes using ACL's ApproximatePR
start = time.time()
topk_train = ppr.topk_ppr_matrix(adj_matrix,
alpha,
eps,
train_idx,
topk,
normalization=ppr_normalization)
if run_val:
topk_val = ppr.topk_ppr_matrix(adj_matrix,
alpha,
eps,
val_idx,
topk,
normalization=ppr_normalization)
else:
topk_val = None
time_preprocessing = time.time() - start
logging.info('Preprocessing done.')
start = time.time()
tf.reset_default_graph()
tf.set_random_seed(0)
model = pprgo.PPRGo(d,
nc,
hidden_size,
nlayers,
lr,
weight_decay,
dropout,
sparse_features=type(attr_matrix) is not np.ndarray)
sess = tf.compat.v1.Session()
with sess.as_default():
tf.compat.v1.global_variables_initializer().run()
nepochs, loss_hist, acc_hist, f1_hist = pprgo.train(
sess=sess,
model=model,
attr_matrix=attr_matrix,
train_idx=train_idx,
val_idx=val_idx,
topk_train=topk_train,
topk_val=topk_val,
labels=labels,
max_epochs=max_epochs,
batch_size=batch_size,
batch_mult_val=batch_mult_val,
eval_step=eval_step,
early_stop=early_stop,
patience=patience,
ex=ex)
time_training = time.time() - start
logging.info('Training done.')
start = time.time()
predictions, time_logits, time_propagation = model.predict(
sess=sess,
adj_matrix=adj_matrix,
attr_matrix=attr_matrix,
alpha=alpha,
nprop=nprop_inference,
inf_fraction=inf_fraction,
ppr_normalization=ppr_normalization)
time_inference = time.time() - start
logging.info('Inference done.')
results = {
'accuracy_train':
100 * accuracy_score(labels[train_idx], predictions[train_idx]),
'accuracy_val':
100 * accuracy_score(labels[val_idx], predictions[val_idx]),
'accuracy_test':
100 * accuracy_score(labels[test_idx], predictions[test_idx]),
'f1_train':
f1_score(labels[train_idx], predictions[train_idx], average='macro'),
'f1_val':
f1_score(labels[val_idx], predictions[val_idx], average='macro'),
'f1_test':
f1_score(labels[test_idx], predictions[test_idx], average='macro'),
}
gpu_max_bytes = tf.contrib.memory_stats.MaxBytesInUse()
results.update({
'time_loading': time_loading,
'time_preprocessing': time_preprocessing,
'time_training': time_training,
'time_inference': time_inference,
'time_logits': time_logits,
'time_propagation': time_propagation,
'gpu_memory': sess.run(gpu_max_bytes),
'memory': utils.get_max_memory_bytes(),
'nepochs': nepochs,
})
return results
def mnist_model(features, labels, mode, params):
# TODO: Using features["images"], compute `logits` using:
# - convolutional layer with 8 channels, kernel size 3 and ReLu activation
# - max pooling layer with pool size 2 and stride 2
# - convolutional layer with 16 channels, kernel size 3 and ReLu activation
# - max pooling layer with pool size 2 and stride 2
# - flattening layer
# - dense layer with 256 neurons and ReLU activation
# - dense layer with 10 neurons and no activation
predictions = tf.argmax(logits, axis=1)
if mode == tf.estimator.ModeKeys.PREDICT:
# TODO: Return EstimatorSpec with `mode` and `predictions` parameters
# TODO: Compute loss using `tf.losses.sparse_softmax_cross_entropy`.
if mode == tf.estimator.ModeKeys.TRAIN:
# TODO: Get optimizer class, using `params.get("optimizer", None)`.
# TODO: Create optimizer, using `params.get("learning_rate", None)` parameter.
# TODO: Define `train_op` as `optimizer.minimize`, with `tf.train.get_global_step` as `global_step`.
# TODO: Return EstimatorSpec with `mode`, `loss`, `train_op` and `eval_metric_ops` arguments,
# the latter being a dictionary with "accuracy" key and `tf.metrics.accuracy` value.
if mode == tf.estimator.ModeKeys.EVAL:
# TODO: Return EstimatorSpec with `mode`, `loss` and `eval_metric_ops` arguments,
# the latter being a dictionary with "accuracy" key and `tf.metrics.accuracy` value.
if __name__ == "__main__":
import argparse
import datetime
import os
import re
# Fix random seed
np.random.seed(42)
tf.set_random_seed(42)
# Parse arguments
parser = argparse.ArgumentParser()
parser.add_argument("--batch_size", default=50, type=int, help="Batch size.")
parser.add_argument("--epochs", default=3, type=int, help="Number of epochs.")
parser.add_argument("--threads", default=1, type=int, help="Maximum number of threads to use.")
args = parser.parse_args()
# Create logdir name
args.logdir = "logs/{}-{}-{}".format(
os.path.basename(__file__),
datetime.datetime.now().strftime("%Y-%m-%d_%H%M%S"),
",".join(("{}={}".format(re.sub("(.)[^_]*_?", r"\1", key), value) for key, value in sorted(vars(args).items())))
)
if not os.path.exists("logs"): os.mkdir("logs") # TF 1.6 will do this by itself
# Construct the model
model = tf.estimator.Estimator(
model_fn=mnist_model,
model_dir=args.logdir,
config=tf.estimator.RunConfig(tf_random_seed=42,
session_config=tf.ConfigProto(inter_op_parallelism_threads=args.threads,
intra_op_parallelism_threads=args.threads)),
params={
"optimizer": tf.train.AdamOptimizer,
"learning_rate": 0.001,
})
# Load the data
from tensorflow.examples.tutorials import mnist
mnist = mnist.input_data.read_data_sets(".", reshape=False, seed=42)
# Train
for i in range(args.epochs):
def setRandomSeed(seed):
# eliminate random factors
random.seed(seed)
tf.set_random_seed(seed)
import datetime as dt
import statsmodels.api as sm
import itertools
plt.style.use('seaborn-white')
from statsmodels.tsa.arima_model import ARIMA
from tensorflow.keras.layers import Dense
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras import Sequential
print(tf.__version__)
print(keras.__version__)
np.random.seed(1337)
PYTHONHASHSEED = 0
tf.random.set_random_seed(1337)
tf.set_random_seed(1337)
featSelect = 0
#%% load data
finalDataDf = pd.read_csv('data/MyData12.csv')
finalData = finalDataDf.values
X = finalData[:, :-1]
y = finalData[:, -1]
#%% feature normalization and preprocess
testRate = 0.2
numTrain = int(len(X) * (1 - testRate))
numTest = 1
m = X[:numTrain].mean(axis=0)
def model(X_train,
Y_train,
X_test,
Y_test,
learning_rate=0.009,
num_epochs=100,
minibatch_size=64,
print_cost=True,
isPlot=True):
"""
使用TensorFlow实现三层的卷积神经网络
CONV2D -> RELU -> MAXPOOL -> CONV2D -> RELU -> MAXPOOL -> FLATTEN -> FULLYCONNECTED
参数：
X_train - 训练数据，维度为(None, 64, 64, 3)
Y_train - 训练数据对应的标签，维度为(None, n_y = 6)
X_test - 测试数据，维度为(None, 64, 64, 3)
Y_test - 训练数据对应的标签，维度为(None, n_y = 6)
learning_rate - 学习率
num_epochs - 遍历整个数据集的次数
minibatch_size - 每个小批量数据块的大小
print_cost - 是否打印成本值，每遍历100次整个数据集打印一次
isPlot - 是否绘制图谱
返回：
train_accuracy - 实数，训练集的准确度
test_accuracy - 实数，测试集的准确度
parameters - 学习后的参数
"""
ops.reset_default_graph() #能够重新运行模型而不覆盖tf变量
tf.set_random_seed(1) #确保你的数据和我一样
seed = 3 #指定numpy的随机种子
(m, n_H0, n_W0, n_C0) = X_train.shape
n_y = Y_train.shape[1]
costs = []
#为当前维度创建占位符
X, Y = create_placeholders(n_H0, n_W0, n_C0, n_y)
#初始化参数
parameters = initialize_parameters()
#前向传播
Z3 = forward_propagation(X, parameters)
#计算成本
cost = compute_cost(Z3, Y)
#反向传播，由于框架已经实现了反向传播，我们只需要选择一个优化器就行了
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(cost)
#全局初始化所有变量
init = tf.global_variables_initializer()
#开始运行
config = tf.ConfigProto(allow_soft_placement=True)
config.gpu_options.allow_growth = True
with tf.Session(config=config) as sess:
#初始化参数
sess.run(init)
#开始遍历数据集
minibatch_cost = 1
epoch = 0
while minibatch_cost > 0.4 and epoch < num_epochs:
minibatch_cost = 0
num_minibatches = int(m / minibatch_size) #获取数据块的数量
seed = seed + 1
minibatches = cnn_utils.random_mini_batches(
X_train, Y_train, minibatch_size, seed)
#对每个数据块进行处理
for minibatch in minibatches:
#选择一个数据块
(minibatch_X, minibatch_Y) = minibatch
#最小化这个数据块的成本
_, temp_cost = sess.run([optimizer, cost],
feed_dict={
X: minibatch_X,
Y: minibatch_Y
})
#累加数据块的成本值
minibatch_cost += temp_cost / num_minibatches
#是否打印成本
if print_cost:
#每5代打印一次
if epoch % 5 == 0:
print("当前是第 " + str(epoch) + " 代，成本值为：" +
str(minibatch_cost))
#记录成本
if epoch % 1 == 0:
costs.append(minibatch_cost)
epoch += 1
#数据处理完毕，绘制成本曲线
if isPlot:
plt.plot(np.squeeze(costs))
plt.ylabel('cost')
plt.xlabel('iterations (per tens)')
plt.title("Learning rate =" + str(learning_rate))
plt.show()
#开始预测数据
## 计算当前的预测情况
predict_op = tf.arg_max(Z3, 1)
corrent_prediction = tf.equal(predict_op, tf.arg_max(Y, 1))
##计算准确度
accuracy = tf.reduce_mean(tf.cast(corrent_prediction, "float"))
print("corrent_prediction accuracy= " + str(accuracy))
train_accuracy = accuracy.eval({X: X_train, Y: Y_train})
test_accuary = accuracy.eval({X: X_test, Y: Y_test})
print("训练集准确度：" + str(train_accuracy))
print("测试集准确度：" + str(test_accuary))
return (train_accuracy, test_accuary, parameters)
def __init__(self, layer1_dim, layer2_dim, layer3_dim, layer4_dim, x_dim1, x_dim2,
y_dim, learning_rate, data_num):
# in order to generate same random sequences
tf.set_random_seed(1)
"""
input parameter
"""
# 否则传不到MLP() 里面，要不然MLP() function 得写成MLP(layer1_dim) 的形式
self.layer1_dim = layer1_dim
self.layer2_dim = layer2_dim
self.layer3_dim = layer3_dim
self.layer4_dim = layer4_dim
self.x_dim1 = x_dim1
self.x_dim2 = x_dim2
self.y_dim = y_dim
self.learning_rate = learning_rate
self.data_num = data_num
"""
input data
"""
# training data: record and label
self.dropout_keep = tf.placeholder(dtype=tf.float32, name='dropout_keep')
self.xa = tf.placeholder(tf.float32, shape=(None, self.x_dim1), name='xa-input')
self.xb = tf.placeholder(tf.float32, shape=(None, self.x_dim2), name='xb-input')
self.y = tf.placeholder(tf.float32, shape=(None, self.y_dim), name='y-input')
"""
batch norm
"""
# if self.is_batch_norm:
# update_ops = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
# with tf.control_dependencies(update_ops):
# self.loss = -tf.reduce_mean(self.y * tf.log(tf.clip_by_value(self.y_pred, 1e-10, 1.0)))
# self.loss = self.loss + tf.add_n(tf.get_collection('loss')) #L2 regularization
# else:
# self.loss = -tf.reduce_mean(self.y * tf.log(tf.clip_by_value(self.y_pred, 1e-10, 1.0)))
# self.loss = self.loss + tf.add_n(tf.get_collection('loss')) #L2 regularization
"""
graph structure
"""
# predict data: label
self.y_pred = self.MLP()
self.y_pred_softmax = tf.nn.softmax(self.y_pred)
# print(self.y_pred_softmax)
# acc
self.acc = tf.equal(tf.argmax(self.y_pred_softmax, 1), tf.argmax(self.y, 1))
self.acc = tf.reduce_mean(tf.cast(self.acc, tf.float32))
"""
model training
"""
self.loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=self.y_pred, labels=self.y))
self.loss_metric = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=self.y_pred, labels=self.y))
# self.loss = -tf.reduce_mean(self.y * tf.log(tf.clip_by_value(self.y_pred_softmax, 1e-10, 1.0)))
# self.loss = tf.losses.mean_squared_error(self.y, self.y_pred_softmax)
# self.loss = tf.reduce_mean(tf.square(self.y - self.y_pred_softmax))
# loss_less = 10
# loss_more = 0.1
# self.loss = tf.reduce_sum(tf.where(tf.greater(self.y_pred_softmax, self.y),
# (self.y_pred_softmax-self.y) * loss_more, (self.y-self.y_pred_softmax) * loss_less))
# optimizer
self.optimizer = tf.train.AdamOptimizer(learning_rate=self.learning_rate, name='optimizer')
# self.optimizer = tf.train.RMSPropOptimizer(learning_rate=self.learning_rate,
# decay=0.9, momentum=0.0, epsilon=1e-5, name='optimizer')
# self.optimizer = tf.train.AdadeltaOptimizer(learning_rate=self.learning_rate, name='optimizer')
# self.optimizer = tf.train.AdagradOptimizer(learning_rate=self.learning_rate, name='optimizer')
# self.optimizer = tf.train.FtrlOptimizer(learning_rate=self.learning_rate, name='optimizer')
# self.optimizer = tf.train.GradientDescentOptimizer(learning_rate=self.learning_rate, name='optimizer')
# self.optimizer = tf.train.MomentumOptimizer(learning_rate=self.learning_rate, momentum=0.5, name='optimizer')
# self.optimizer = tf.train.ProximalAdagradOptimizer(learning_rate=self.learning_rate, name='optimizer')
# self.optimizer = tf.train.ProximalGradientDescentOptimizer(learning_rate=self.learning_rate, name='optimizer')
self.train_op = self.optimizer.minimize(self.loss, name='train_op')
def main():
"""Create the model and start the training."""
start = time.time()
args = get_arguments()
h, w = map(int, args.input_size.split(','))
input_size = (h, w)
tf.set_random_seed(args.random_seed)
# Create queue coordinator.
coord = tf.train.Coordinator()
# Load reader.
with tf.name_scope("create_inputs"):
reader = ImageReader(
args.data_dir,
args.data_list,
input_size,
args.random_scale,
args.random_mirror,
args.ignore_label,
coord)
image_batch, label_batch = reader.dequeue(args.batch_size)
# Create network.
net = DeepLabResNetModel({'data': image_batch}, is_training=args.is_training, num_classes=args.num_classes)
# Predictions.
raw_output = net.layers['fc1_voc12']
restore_var = [v for v in tf.global_variables() if 'fc' not in v.name or not args.not_restore_last]
all_trainable = [v for v in tf.trainable_variables() if 'beta' not in v.name and 'gamma' not in v.name]
fc_trainable = [v for v in all_trainable if 'fc' in v.name]
conv_trainable = [v for v in all_trainable if 'fc' not in v.name] # lr * 1.0
fc_w_trainable = [v for v in fc_trainable if 'weights' in v.name] # lr * 10.0
fc_b_trainable = [v for v in fc_trainable if 'biases' in v.name] # lr * 20.0
assert (len(all_trainable) == len(fc_trainable) + len(conv_trainable))
assert (len(fc_trainable) == len(fc_w_trainable) + len(fc_b_trainable))
# Predictions: ignoring all predictions with labels greater or equal than n_classes
raw_prediction = tf.reshape(raw_output, [-1, args.num_classes])
label_proc = prepare_label(label_batch, tf.stack(raw_output.get_shape()[1:3]), num_classes=args.num_classes,
one_hot=False) # [batch_size, h, w]
raw_gt = tf.reshape(label_proc, [-1, ])
indices = tf.squeeze(tf.where(tf.less_equal(raw_gt, args.num_classes - 1)), 1)
gt = tf.cast(tf.gather(raw_gt, indices), tf.int32)
prediction = tf.gather(raw_prediction, indices)
# Pixel-wise softmax loss.
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=prediction, labels=gt)
l2_losses = [args.weight_decay * tf.nn.l2_loss(v) for v in tf.trainable_variables() if 'weights' in v.name]
reduced_loss = tf.reduce_mean(loss) + tf.add_n(l2_losses)
# Processed predictions: for visualisation.
raw_output_up = tf.image.resize_bilinear(raw_output, tf.shape(image_batch)[1:3, ])
raw_output_up = tf.argmax(raw_output_up, axis=3)
pred = tf.expand_dims(raw_output_up, dim=3)
# Define loss and optimisation parameters.
base_lr = tf.constant(args.learning_rate)
step_ph = tf.placeholder(dtype=tf.float32, shape=())
learning_rate = tf.scalar_mul(base_lr, tf.pow((1 - step_ph / args.num_steps), args.power))
opt_conv = tf.train.MomentumOptimizer(learning_rate, args.momentum)
opt_fc_w = tf.train.MomentumOptimizer(learning_rate * 10.0, args.momentum)
opt_fc_b = tf.train.MomentumOptimizer(learning_rate * 20.0, args.momentum)
grads = tf.gradients(reduced_loss, conv_trainable + fc_w_trainable + fc_b_trainable)
grads_conv = grads[:len(conv_trainable)]
grads_fc_w = grads[len(conv_trainable): (len(conv_trainable) + len(fc_w_trainable))]
grads_fc_b = grads[(len(conv_trainable) + len(fc_w_trainable)):]
train_op_conv = opt_conv.apply_gradients(zip(grads_conv, conv_trainable))
train_op_fc_w = opt_fc_w.apply_gradients(zip(grads_fc_w, fc_w_trainable))
train_op_fc_b = opt_fc_b.apply_gradients(zip(grads_fc_b, fc_b_trainable))
train_op = tf.group(train_op_conv, train_op_fc_w, train_op_fc_b)
# Set up tf session and initialize variables.
config = tf.ConfigProto()
config.gpu_options.allow_growth = True
sess = tf.Session(config=config)
init = tf.global_variables_initializer()
sess.run(init)
# Saver for storing checkpoints of the model.
saver = tf.train.Saver(var_list=tf.global_variables(), max_to_keep=10)
# Load variables if the checkpoint is provided.
if args.restore_from is not None:
loader = tf.train.Saver(var_list=restore_var)
load(loader, sess, args.restore_from)
# Start queue threads.
threads = tf.train.start_queue_runners(coord=coord, sess=sess)
losses = []
# Iterate over training steps.
for step in range(args.num_steps):
start_time = time.time()
feed_dict = {step_ph: step}
if step % args.save_pred_every == 0:
loss_value, images, labels, preds, _ = sess.run([reduced_loss, image_batch, label_batch, pred, train_op],
feed_dict=feed_dict)
save(saver, sess, args.snapshot_dir, step)
else:
loss_value, _ = sess.run([reduced_loss, train_op], feed_dict=feed_dict)
duration = time.time() - start_time
print('step {:d} \t loss = {:.3f}, ({:.3f} sec/step)'.format(step, loss_value, duration))
losses.append(loss_value)
plotLoss(losses)
coord.request_stop()
coord.join(threads)
end = time.time()
print("Duration taken to complete the training process is {} seconds.".format(end - start))
savingParams = []
savingParams.append(loss_value)
savingParams.append(end - start)
# write saved parameters from training
with open(savingOutputDir, 'w') as f:
print(savingParams, file=f)
def __init__(self, num_vocabulary, batch_size, emb_dim, hidden_dim, sequence_length, start_token, ):
self.num_vocabulary = num_vocabulary
self.batch_size = batch_size
self.emb_dim = emb_dim
self.hidden_dim = hidden_dim
self.sequence_length = sequence_length
self.start_token = tf.constant([start_token] * self.batch_size, dtype=tf.int32)
self.g_params = []
self.temperature = 1.0
with tf.variable_scope('generator'):
tf.set_random_seed(1234)
self.g_embeddings = tf.Variable(
tf.random_normal([self.num_vocabulary, self.emb_dim], 0.0, 1.0, seed=123314154))
self.g_params.append(self.g_embeddings)
self.g_recurrent_unit = self.create_recurrent_unit(self.g_params) # maps h_tm1 to h_t for generator
self.g_output_unit = self.create_output_unit(self.g_params) # maps h_t to o_t (output token logits)
# placeholder definition
self.x = tf.placeholder(tf.int32, shape=[self.batch_size,
self.sequence_length]) # sequence of tokens generated by generator
# processed for batch
with tf.device("/cpu:0"):
tf.set_random_seed(1234)
self.processed_x = tf.transpose(tf.nn.embedding_lookup(self.g_embeddings, self.x),
perm=[1, 0, 2]) # seq_length x batch_size x emb_dim
# initial states
self.h0 = tf.zeros([self.batch_size, self.hidden_dim])
self.h0 = tf.stack(self.h0)
# generator on initial randomness
gen_o = tensor_array_ops.TensorArray(dtype=tf.float32, size=self.sequence_length,
dynamic_size=False, infer_shape=True)
gen_x = tensor_array_ops.TensorArray(dtype=tf.int32, size=self.sequence_length,
dynamic_size=False, infer_shape=True)
def _g_recurrence(i, x_t, h_tm1, gen_o, gen_x):
h_t = self.g_recurrent_unit(x_t, h_tm1) # hidden_memory_tuple
o_t = self.g_output_unit(h_t) # batch x vocab , logits not prob
log_prob = tf.log(tf.nn.softmax(o_t))
next_token = tf.cast(tf.reshape(tf.multinomial(log_prob, 1), [self.batch_size]), tf.int32)
x_tp1 = tf.nn.embedding_lookup(self.g_embeddings, next_token) # batch x emb_dim
gen_o = gen_o.write(i, tf.reduce_sum(
tf.multiply(tf.one_hot(next_token, self.num_vocabulary, 1.0, 0.0), tf.nn.softmax(o_t)),
1)) # [batch_size] , prob
gen_x = gen_x.write(i, next_token) # indices, batch_size
return i + 1, x_tp1, h_t, gen_o, gen_x
_, _, _, self.gen_o, self.gen_x = control_flow_ops.while_loop(
cond=lambda i, _1, _2, _3, _4: i < self.sequence_length,
body=_g_recurrence,
loop_vars=(tf.constant(0, dtype=tf.int32),
tf.nn.embedding_lookup(self.g_embeddings, self.start_token), self.h0, gen_o, gen_x)
)
self.gen_x = self.gen_x.stack() # seq_length x batch_size
self.gen_x = tf.transpose(self.gen_x, perm=[1, 0]) # batch_size x seq_length
# supervised pretraining for generator
g_predictions = tensor_array_ops.TensorArray(
dtype=tf.float32, size=self.sequence_length,
dynamic_size=False, infer_shape=True)
ta_emb_x = tensor_array_ops.TensorArray(
dtype=tf.float32, size=self.sequence_length)
ta_emb_x = ta_emb_x.unstack(self.processed_x)
def _pretrain_recurrence(i, x_t, h_tm1, g_predictions):
h_t = self.g_recurrent_unit(x_t, h_tm1)
o_t = self.g_output_unit(h_t)
g_predictions = g_predictions.write(i, tf.nn.softmax(o_t)) # batch x vocab_size
x_tp1 = ta_emb_x.read(i)
return i + 1, x_tp1, h_t, g_predictions
_, _, _, self.g_predictions = control_flow_ops.while_loop(
cond=lambda i, _1, _2, _3: i < self.sequence_length,
body=_pretrain_recurrence,
loop_vars=(tf.constant(0, dtype=tf.int32),
tf.nn.embedding_lookup(self.g_embeddings, self.start_token),
self.h0, g_predictions))
self.g_predictions = tf.transpose(
self.g_predictions.stack(), perm=[1, 0, 2]) # batch_size x seq_length x vocab_size
# pretraining loss
self.pretrain_loss = -tf.reduce_sum(
tf.one_hot(tf.to_int32(tf.reshape(self.x, [-1])), self.num_vocabulary, 1.0, 0.0) * tf.log(
tf.reshape(self.g_predictions, [-1, self.num_vocabulary]))) / (self.sequence_length * self.batch_size)
self.out_loss = tf.reduce_sum(
tf.reshape(
-tf.reduce_sum(
tf.one_hot(tf.to_int32(tf.reshape(self.x, [-1])), self.num_vocabulary, 1.0, 0.0) * tf.log(
tf.reshape(self.g_predictions, [-1, self.num_vocabulary])), 1
), [-1, self.sequence_length]
), 1
) # batch_size
def model(X_train,
Y_train,
X_test,
Y_test,
learning_rate=0.0001,
num_epochs=1500,
minibatch_size=32,
print_cost=True):
"""
Implements a three-layer tensorflow neural network: LINEAR->RELU->LINEAR->RELU->LINEAR->SOFTMAX.
Arguments:
X_train -- training set, of shape (input size = 12288, number of training examples = 1080)
Y_train -- test set, of shape (output size = 6, number of training examples = 1080)
X_test -- training set, of shape (input size = 12288, number of training examples = 120)
Y_test -- test set, of shape (output size = 6, number of test examples = 120)
learning_rate -- learning rate of the optimization
num_epochs -- number of epochs of the optimization loop
minibatch_size -- size of a minibatch
print_cost -- True to print the cost every 100 epochs
Returns:
parameters -- parameters learnt by the model. They can then be used to predict.
"""
ops.reset_default_graph(
) # to be able to rerun the model without overwriting tf variables
tf.set_random_seed(1) # to keep consistent results
seed = 3 # to keep consistent results
(
n_x, m
) = X_train.shape # (n_x: input size, m : number of examples in the train set)
n_y = Y_train.shape[0] # n_y : output size
costs = [] # To keep track of the cost
# Create Placeholders of shape (n_x, n_y)
### START CODE HERE ### (1 line)
X, Y = create_placeholders(n_x, n_y)
### END CODE HERE ###
# Initialize parameters
### START CODE HERE ### (1 line)
parameters = initialize_parameters()
### END CODE HERE ###
# Forward propagation: Build the forward propagation in the tensorflow graph
### START CODE HERE ### (1 line)
Z3 = forward_propagation(X, parameters)
### END CODE HERE ###
# Cost function: Add cost function to tensorflow graph
### START CODE HERE ### (1 line)
cost = compute_cost(Z3, Y)
### END CODE HERE ###
# Backpropagation: Define the tensorflow optimizer. Use an AdamOptimizer.
### START CODE HERE ### (1 line)
optimizer = tf.train.AdamOptimizer(
learning_rate=learning_rate).minimize(cost)
### END CODE HERE ###
# Initialize all the variables
init = tf.global_variables_initializer()
# Start the session to compute the tensorflow graph
with tf.Session() as sess:
# Run the initialization
sess.run(init)
# Do the training loop
for epoch in range(num_epochs):
epoch_cost = 0. # Defines a cost related to an epoch
num_minibatches = int(
m / minibatch_size
) # number of minibatches of size minibatch_size in the train set
seed = seed + 1
minibatches = random_mini_batches(X_train, Y_train, minibatch_size,
seed)
for minibatch in minibatches:
# Select a minibatch
(minibatch_X, minibatch_Y) = minibatch
# IMPORTANT: The line that runs the graph on a minibatch.
# Run the session to execute the "optimizer" and the "cost", the feedict should contain a minibatch for (X,Y).
### START CODE HERE ### (1 line)
_, minibatch_cost = sess.run([optimizer, cost],
feed_dict={
X: minibatch_X,
Y: minibatch_Y
})
### END CODE HERE ###
epoch_cost += minibatch_cost / num_minibatches
# Print the cost every epoch
if print_cost == True and epoch % 100 == 0:
print("Cost after epoch %i: %f" % (epoch, epoch_cost))
if print_cost == True and epoch % 5 == 0:
costs.append(epoch_cost)
# plot the cost
plt.plot(np.squeeze(costs))
plt.ylabel('cost')
plt.xlabel('iterations (per tens)')
plt.title("Learning rate =" + str(learning_rate))
plt.show()
# lets save the parameters in a variable
parameters = sess.run(parameters)
print("Parameters have been trained!")
# Calculate the correct predictions
correct_prediction = tf.equal(tf.argmax(Z3), tf.argmax(Y))
# Calculate accuracy on the test set
accuracy = tf.reduce_mean(tf.cast(correct_prediction, "float"))
print("Train Accuracy:", accuracy.eval({X: X_train, Y: Y_train}))
print("Test Accuracy:", accuracy.eval({X: X_test, Y: Y_test}))
return parameters
