def evaluate(self,
x=None,
y=None,
batch_per_thread=None,
distributed=False):
if isinstance(x, TFDataset):
x = _standarize_feature_label_dataset(x, self.model)
# todo check arguments
return self._evaluate_distributed(x)
else:
if distributed:
sc = getOrCreateSparkContext()
rdd, types, shapes = _create_rdd_x_y(
x, y, self.model._feed_input_names,
self.model._feed_output_names, sc)
names = self.model._feed_input_names + self.model._feed_output_names
dataset = TFDataset.from_rdd(
rdd,
names=names,
types=types,
shapes=shapes,
batch_per_thread=-1
if batch_per_thread is None else batch_per_thread)
return self._evaluate_distributed(dataset)
else:
return self.model.evaluate(x=x,
y=y,
batch_size=batch_per_thread)
def predict(self, x, batch_per_thread=None, distributed=False):
if isinstance(x, TFDataset):
# todo check arguments
x = _standarize_feature_dataset(x, self.model)
return self._predict_distributed(x)
else:
if distributed:
sc = getOrCreateSparkContext()
rdd, types, shapes = _create_rdd_x(
x, self.model._feed_input_names, sc)
dataset = TFDataset.from_rdd(
rdd,
names=self.model._feed_input_names,
types=types,
shapes=shapes,
batch_per_thread=-1
if batch_per_thread is None else batch_per_thread)
results = self._predict_distributed(dataset).collect()
output_num = len(self.model.outputs)
if output_num == 1:
return np.stack(results)
else:
predictions = []
for i in range(0, output_num):
predictions.append(
np.stack([res[i] for res in results]))
return predictions
else:
return self.model.predict(x=x, batch_size=batch_per_thread)
def evaluate(self,
x=None,
y=None,
batch_per_thread=None,
distributed=False):
if isinstance(x, TFDataset):
if not x.has_batch:
raise ValueError("The batch_per_thread of TFDataset must be " +
"specified when used in KerasModel evaluate.")
x = _standarize_feature_label_dataset(x, self.model)
# todo check arguments
return self._evaluate_distributed(x)
else:
if distributed:
sc = getOrCreateSparkContext()
rdd, types, shapes = _create_rdd_x_y(
x, y, self.model._feed_input_names,
self.model._feed_output_names, sc)
names = self.model._feed_input_names + self.model._feed_output_names
dataset = TFDataset.from_rdd(
rdd,
names=names,
types=types,
shapes=shapes,
batch_per_thread=-1
if batch_per_thread is None else batch_per_thread)
return self._evaluate_distributed(dataset)
else:
return self.model.evaluate(x=x,
y=y,
batch_size=batch_per_thread)
def test_tf_optimizer_with_sparse_gradient(self):
import tensorflow as tf
ids = np.random.randint(0, 10, size=[40])
labels = np.random.randint(0, 5, size=[40])
id_rdd = self.sc.parallelize(ids)
label_rdd = self.sc.parallelize(labels)
training_rdd = id_rdd.zip(label_rdd).map(lambda x: [x[0], x[1]])
with tf.Graph().as_default():
dataset = TFDataset.from_rdd(training_rdd,
names=["ids", "labels"],
shapes=[[], []],
types=[tf.int32, tf.int32],
batch_size=8)
id_tensor, label_tensor = dataset.tensors
embedding_table = tf.get_variable(name="word_embedding",
shape=[10, 5])
embedding = tf.nn.embedding_lookup(embedding_table, id_tensor)
loss = tf.reduce_mean(
tf.losses.sparse_softmax_cross_entropy(logits=embedding,
labels=label_tensor))
optimizer = TFOptimizer(loss, Adam(1e-3))
optimizer.optimize(end_trigger=MaxEpoch(1))
optimizer.sess.close()
def predict(self, x, batch_per_thread=None, distributed=False):
"""
Use a model to do prediction.
:param x: Input data. It could be:
- a TFDataset object
- A Numpy array (or array-like), or a list of arrays
(in case the model has multiple inputs).
- A dict mapping input names to the corresponding array/tensors,
if the model has named inputs.
:param batch_per_thread:
The default value is 1.
When distributed is True,the total batch size is batch_per_thread * rdd.getNumPartitions.
When distributed is False the total batch size is batch_per_thread * numOfCores.
:param distributed: Boolean. Whether to do prediction in distributed mode or local mode.
Default is True. In local mode, x must be a Numpy array.
"""
if isinstance(x, TFDataset):
# todo check arguments
if not x.has_batch:
raise ValueError(
"The batch_per_thread of TFDataset" +
" must be specified when used in KerasModel predict.")
if isinstance(x, TFNdarrayDataset):
x = _standarize_feature_dataset(x, self.model)
return self._predict_distributed(x)
else:
if distributed:
sc = getOrCreateSparkContext()
rdd, types, shapes = _create_rdd_x(
x, self.model._feed_input_names, sc)
dataset = TFDataset.from_rdd(
rdd,
names=self.model._feed_input_names,
types=types,
shapes=shapes,
batch_per_thread=-1
if batch_per_thread is None else batch_per_thread)
results = self._predict_distributed(dataset).collect()
output_num = len(self.model.outputs)
if output_num == 1:
return np.stack(results)
else:
predictions = []
for i in range(0, output_num):
predictions.append(
np.stack([res[i] for res in results]))
return predictions
else:
return self.model.predict(x=x, batch_size=batch_per_thread)
def fit(self,
x=None,
y=None,
batch_size=None,
epochs=1,
validation_split=0.,
validation_data=None,
distributed=False,
**kwargs):
if isinstance(x, TFDataset):
# todo check arguments
if not x.has_batch:
raise ValueError("The batch_size of TFDataset must be " +
"specified when used in KerasModel fit.")
x = _standarize_feature_label_dataset(x, self.model)
self._fit_distributed(x, validation_split, epochs, **kwargs)
elif distributed:
sc = getOrCreateSparkContext()
train_rdd, types, shapes = _create_rdd_x_y(
x, y, self.model._feed_input_names,
self.model._feed_output_names, sc)
val_rdd = None
if validation_data is not None:
val_rdd, _, _ = _create_rdd_x_y(validation_data[0],
validation_data[1],
self.model._feed_input_names,
self.model._feed_output_names,
sc)
names = self.model._feed_input_names + self.model._feed_output_names
dataset = TFDataset.from_rdd(
train_rdd,
names=names,
shapes=shapes,
types=types,
batch_size=batch_size if batch_size is not None else 32,
val_rdd=val_rdd)
self._fit_distributed(dataset, validation_split, epochs, **kwargs)
else:
self.model.fit(x=x,
y=y,
batch_size=batch_size,
epochs=epochs,
validation_split=validation_split,
validation_data=validation_data,
**kwargs)
def main(max_epoch, data_num):
sc = init_nncontext()
# get data, pre-process and create TFDataset
def get_data_rdd(dataset):
(images_data,
labels_data) = mnist.read_data_sets("/tmp/mnist", dataset)
image_rdd = sc.parallelize(images_data[:data_num])
labels_rdd = sc.parallelize(labels_data[:data_num])
rdd = image_rdd.zip(labels_rdd) \
.map(lambda rec_tuple: [normalizer(rec_tuple[0], mnist.TRAIN_MEAN, mnist.TRAIN_STD),
np.array(rec_tuple[1])])
return rdd
training_rdd = get_data_rdd("train")
testing_rdd = get_data_rdd("test")
dataset = TFDataset.from_rdd(training_rdd,
names=["features", "labels"],
shapes=[[28, 28, 1], []],
types=[tf.float32, tf.int32],
batch_size=280,
val_rdd=testing_rdd)
# construct the model from TFDataset
images, labels = dataset.tensors
with slim.arg_scope(lenet.lenet_arg_scope()):
logits, end_points = lenet.lenet(images,
num_classes=10,
is_training=True)
loss = tf.reduce_mean(
tf.losses.sparse_softmax_cross_entropy(logits=logits, labels=labels))
# create a optimizer
optimizer = TFOptimizer(loss,
Adam(1e-3),
val_outputs=[logits],
val_labels=[labels],
val_method=Top1Accuracy())
optimizer.set_train_summary(TrainSummary("/tmp/az_lenet", "lenet"))
optimizer.set_val_summary(ValidationSummary("/tmp/az_lenet", "lenet"))
# kick off training
optimizer.optimize(end_trigger=MaxEpoch(max_epoch))
saver = tf.train.Saver()
saver.save(optimizer.sess, "/tmp/lenet/")
def main(data_num):
data = Input(shape=[28, 28, 1])
x = Flatten()(data)
x = Dense(64, activation='relu')(x)
x = Dense(64, activation='relu')(x)
predictions = Dense(10, activation='softmax')(x)
model = Model(inputs=data, outputs=predictions)
model.load_weights("/tmp/mnist_keras.h5")
if DISTRIBUTED:
# using RDD api to do distributed evaluation
sc = init_nncontext()
# get data, pre-process and create TFDataset
(images_data, labels_data) = mnist.read_data_sets("/tmp/mnist", "test")
image_rdd = sc.parallelize(images_data[:data_num])
labels_rdd = sc.parallelize(labels_data[:data_num])
rdd = image_rdd.zip(labels_rdd) \
.map(lambda rec_tuple: [normalizer(rec_tuple[0], mnist.TRAIN_MEAN, mnist.TRAIN_STD)])
dataset = TFDataset.from_rdd(rdd,
names=["features"],
shapes=[[28, 28, 1]],
types=[tf.float32],
batch_per_thread=20)
predictor = TFPredictor.from_keras(model, dataset)
accuracy = predictor.predict().zip(labels_rdd).map(
lambda x: np.argmax(x[0]) == x[1]).mean()
print("predict accuracy is %s" % accuracy)
else:
# using keras api for local evaluation
model.compile(optimizer='rmsprop',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
(images_data, labels_data) = mnist.read_data_sets("/tmp/mnist", "test")
images_data = normalizer(images_data, mnist.TRAIN_MEAN,
mnist.TRAIN_STD)
result = model.evaluate(images_data, labels_data)
print(model.metrics_names)
print(result)
def main(max_epoch, data_num):
sc = init_nncontext()
# get data, pre-process and create TFDataset
def get_data_rdd(dataset):
(images_data, labels_data) = mnist.read_data_sets("/tmp/mnist", dataset)
image_rdd = sc.parallelize(images_data[:data_num])
labels_rdd = sc.parallelize(labels_data[:data_num])
rdd = image_rdd.zip(labels_rdd) \
.map(lambda rec_tuple: [normalizer(rec_tuple[0], mnist.TRAIN_MEAN, mnist.TRAIN_STD),
np.array(rec_tuple[1])])
return rdd
training_rdd = get_data_rdd("train")
testing_rdd = get_data_rdd("test")
dataset = TFDataset.from_rdd(training_rdd,
names=["features", "labels"],
shapes=[[28, 28, 1], []],
types=[tf.float32, tf.int32],
batch_size=280,
val_rdd=testing_rdd
)
data = Input(shape=[28, 28, 1])
x = Flatten()(data)
x = Dense(64, activation='relu')(x)
x = Dense(64, activation='relu')(x)
predictions = Dense(10, activation='softmax')(x)
model = Model(input=data, output=predictions)
model.compile(optimizer='rmsprop',
loss='sparse_categorical_crossentropy',
metrics=['accuracy'])
optimizer = TFOptimizer.from_keras(model, dataset)
optimizer.set_train_summary(TrainSummary("/tmp/az_lenet", "lenet"))
optimizer.set_val_summary(ValidationSummary("/tmp/az_lenet", "lenet"))
# kick off training
optimizer.optimize(end_trigger=MaxEpoch(max_epoch))
saver = tf.train.Saver()
saver.save(optimizer.sess, "/tmp/lenet/")
def main():
sc = init_nncontext()
# get data, pre-process and create TFDataset
(images_data, labels_data) = mnist.read_data_sets("/tmp/mnist", "test")
image_rdd = sc.parallelize(images_data)
labels_rdd = sc.parallelize(labels_data)
rdd = image_rdd.zip(labels_rdd) \
.map(lambda rec_tuple: [normalizer(rec_tuple[0], mnist.TRAIN_MEAN, mnist.TRAIN_STD),
np.array(rec_tuple[1])])
dataset = TFDataset.from_rdd(rdd,
names=["features", "labels"],
shapes=[[28, 28, 1], [1]],
types=[tf.float32, tf.int32],
batch_per_thread=20)
# construct the model from TFDataset
images, labels = dataset.tensors
labels = tf.squeeze(labels)
with slim.arg_scope(lenet.lenet_arg_scope()):
logits, end_points = lenet.lenet(images,
num_classes=10,
is_training=False)
predictions = tf.to_int32(tf.argmax(logits, axis=1))
correct = tf.expand_dims(tf.to_int32(tf.equal(predictions, labels)),
axis=1)
saver = tf.train.Saver()
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
saver.restore(sess, "/tmp/lenet/")
predictor = TFPredictor(sess, [correct])
accuracy = predictor.predict().mean()
print("predict accuracy is %s" % accuracy)
def main():
sc = init_nncontext()
# get data, pre-process and create TFDataset
(images_data, labels_data) = mnist.read_data_sets("/tmp/mnist", "train")
image_rdd = sc.parallelize(images_data)
labels_rdd = sc.parallelize(labels_data)
rdd = image_rdd.zip(labels_rdd) \
.map(lambda rec_tuple: [normalizer(rec_tuple[0], mnist.TRAIN_MEAN, mnist.TRAIN_STD),
np.array(rec_tuple[1])])
dataset = TFDataset.from_rdd(rdd,
names=["features", "labels"],
shapes=[(None, 28, 28, 1), (None, 1)],
types=[tf.float32, tf.int32]
)
# construct the model from TFDataset
images, labels = dataset.inputs
labels = tf.squeeze(labels)
with slim.arg_scope(lenet.lenet_arg_scope()):
logits, end_points = lenet.lenet(images, num_classes=10, is_training=True)
loss = tf.reduce_mean(tf.losses.sparse_softmax_cross_entropy(logits=logits, labels=labels))
# create a optimizer
optimizer = TFOptimizer(loss, Adam(1e-3))
# kick off training
# you may change the MaxIteration to MaxEpoch(5) to make it converge
optimizer.optimize(end_trigger=MaxIteration(20), batch_size=280)
# evaluate
(images_data, labels_data) = mnist.read_data_sets("/tmp/mnist", "test")
images_data = normalizer(images_data, mnist.TRAIN_MEAN, mnist.TRAIN_STD)
predictions = tf.argmax(logits, axis=1)
predictions_data, loss_value = optimizer.sess.run([predictions, loss],
feed_dict={images: images_data,
labels: labels_data})
print(np.mean(np.equal(predictions_data, labels_data)))
print(loss_value)
def main():
sc = init_nncontext()
# get data, pre-process and create TFDataset
(images_data, labels_data) = mnist.read_data_sets("/tmp/mnist", "train")
image_rdd = sc.parallelize(images_data)
labels_rdd = sc.parallelize(labels_data)
rdd = image_rdd.zip(labels_rdd) \
.map(lambda rec_tuple: [normalizer(rec_tuple[0], mnist.TRAIN_MEAN, mnist.TRAIN_STD),
np.array(rec_tuple[1])])
dataset = TFDataset.from_rdd(rdd,
names=["features", "labels"],
shapes=[[28, 28, 1], [1]],
types=[tf.float32, tf.int32],
batch_size=280)
# construct the model from TFDataset
images, labels = dataset.tensors
labels = tf.squeeze(labels)
with slim.arg_scope(lenet.lenet_arg_scope()):
logits, end_points = lenet.lenet(images,
num_classes=10,
is_training=True)
loss = tf.reduce_mean(
tf.losses.sparse_softmax_cross_entropy(logits=logits, labels=labels))
# create a optimizer
optimizer = TFOptimizer(loss, Adam(1e-3))
optimizer.set_train_summary(TrainSummary("/tmp/az_lenet", "lenet"))
# kick off training
for i in range(5):
optimizer.optimize(end_trigger=MaxEpoch(i + 1))
saver = tf.train.Saver()
saver.save(optimizer.sess, "/tmp/lenet/")
def test_tf_optimizer_with_sparse_gradient_using_keras(self):
import tensorflow as tf
ids = np.random.randint(0, 10, size=[40])
labels = np.random.randint(0, 5, size=[40])
id_rdd = self.sc.parallelize(ids)
label_rdd = self.sc.parallelize(labels)
training_rdd = id_rdd.zip(label_rdd).map(lambda x: [x[0], x[1]])
with tf.Graph().as_default():
dataset = TFDataset.from_rdd(training_rdd,
names=["ids", "labels"],
shapes=[[], []],
types=[tf.int32, tf.int32],
batch_size=8)
from tensorflow.python.ops import variable_scope
def variable_creator(**kwargs):
kwargs["use_resource"] = False
return variable_scope.default_variable_creator(None, **kwargs)
getter = lambda next_creator, **kwargs: variable_creator(**kwargs)
with variable_scope.variable_creator_scope(getter):
words_input = tf.keras.layers.Input(shape=(),
name='words_input')
embedding_layer = tf.keras.layers.Embedding(
input_dim=10, output_dim=5, name='word_embedding')
word_embeddings = embedding_layer(words_input)
embedding = tf.keras.layers.Flatten()(word_embeddings)
output = tf.keras.layers.Dense(5,
activation="softmax")(embedding)
model = tf.keras.models.Model(inputs=[words_input],
outputs=[output])
model.compile(optimizer="sgd",
loss="sparse_categorical_crossentropy")
optimizer = TFOptimizer.from_keras(model, dataset)
optimizer.optimize(end_trigger=MaxEpoch(1))
optimizer.sess.close()
# In[10]:
# Let's have a look at one element of trainRDD
trainRDD.take(1)
# We can see that `features` is now composed by the list of 801 particles with 19 features each (`shape=[801 19]`) plus the HLF (`shape=[14]`) and the encoded label (`shape=[3]`).
# In[11]:
from zoo.pipeline.api.net import TFDataset
from zoo.tfpark.model import KerasModel
# create TFDataset for TF training
dataset = TFDataset.from_rdd(trainRDD,
features=[(tf.float32, [801, 19]),
(tf.float32, [14])],
labels=(tf.float32, [3]),
batch_size=256,
val_rdd=testRDD)
# ## Optimizer setup and training
# In[12]:
# Set of hyperparameters
numEpochs = 8
# The batch used by BDL must be a multiple of numExecutors * executorCores
# Because data will be equally distibuted inside each executor
workerBatch = 64
# numExecutors = int(sc._conf.get('spark.executor.instances'))
output_dim=FC_LINEAR_DIMENSION, # 尺寸: 32 -> 64.
activation="sigmoid"))
# BigDL 不支持 parameter sharing, 不得已而为之.
both_feature = TimeDistributed(layer=convolve_net,
input_shape=input_shape)(both_input)
encode_left = both_feature.index_select(1, 0)
encode_right = both_feature.index_select(1, 1)
distance = autograd.abs(encode_left - encode_right)
predict = Dense(output_dim=NUM_CLASS_LABEL, activation="sigmoid")(distance)
siamese_net = Model(input=both_input, output=predict)
siamese_net.compile(optimizer="adam",
loss='sparse_categorical_crossentropy',
metrics=["accuracy"])
# 构造分布式的数据集对象.
data_set = TFDataset.from_rdd(train_rdd,
shapes=[input_shape, [1]],
batch_size=args.batch_size,
val_rdd=test_rdd)
optimizer = TFOptimizer.from_keras(siamese_net, data_set)
app_name = "Siamese Network"
optimizer.set_train_summary(TrainSummary("tmp", app_name))
optimizer.set_val_summary(ValidationSummary("tmp", app_name))
optimizer.optimize(end_trigger=MaxEpoch(args.num_epoch))
