Analytics Zoo

A unified analytics + AI platform for distribtued TensoFlow, Keras and BigDL on Apache Spark

What is Analytics Zoo?

Analytics Zoo provides a unified analytics + AI platform that seamlessly unites Spark, TensorFlow, Keras and BigDL programs into an integrated pipeline; the entire pipeline can then transparently scale out to a large Hadoop/Spark cluster for distributed training or inference.

Data wrangling and analysis using PySpark
Deep learning model development using TensorFlow or Keras
Distributed training/inference on Spark and BigDL
All within a single unified pipeline and in a user-transparent fashion!

In addition, Analytics Zoo also provides a rich set of analytics and AI support for the end-to-end pipeline, including:

Easy-to-use abstractions and APIs (e.g., transfer learning support, autograd operations, Spark Dataframe and ML pipeline support, online model serving API, etc.)
Common feature engineering operations (for image, text, 3D image, etc.)
Built-in deep learning models (e.g., object detection, image classification, text classification, recommendation, etc.)
Reference use cases (e.g., anomaly detection, sentiment analysis, fraud detection, image similarity, etc.)

How to use Analytics Zoo?

To get started, please refer to the Python install guide or Scala install guide.
For running distributed TensorFlow on Spark and BigDL, please refer to the quick start here and more examples here.
For more information, You may refer to the Analytis Zoo document website.
For additional questions and discussions, you can join the Google User Group (or subscribe to the Mail List).

Overview

Distributed Tensoflow and Keras on Spark/BigDL
- Data wrangling and analysis using PySpark
- Deep learning model development using TensorFlow or Keras
- Distributed training/inference on Spark and BigDL
- All within a single unified pipeline and in a user-transparent fashion!
High level abstractions and APIs
- Transfer learning: customize pretained model for feature extraction or fine-tuning
- autograd: build custom layer/loss using auto differentiation operations
- nnframes: native deep learning support in Spark DataFrames and ML Pipelines
- Model serving: productionize model serving and inference using POJO APIs
Built-in deep learning models
- Object detection API: high-level API and pretrained models (e.g., SSD and Faster-RCNN) for object detection
- Image classification API: high-level API and pretrained models (e.g., VGG, Inception, ResNet, MobileNet, etc.) for image classification
- Text classification API: high-level API and pre-defined models (using CNN, LSTM, etc.) for text classification
- Recommedation API: high-level API and pre-defined models (e.g., Neural Collaborative Filtering, Wide and Deep Learning, etc.) for recommendation
Reference use cases: a collection of end-to-end reference use cases (e.g., anomaly detection, sentiment analysis, fraud detection, image augmentation, object detection, variational autoencoder, etc.)

Distributed Tensoflow and Keras on Spark/BigDL

To make it easy to build and productionize the deep learning applications for Big Data, Analytics Zoo provides a unified analytics + AI platform that seamlessly unites Spark, TensorFlow, Keras and BigDL programs into an integrated pipeline (as illustrated below), which can then transparently run on a large-scale Hadoop/Spark clusters for distributed training and inference. (Please see more examples here).

Data wrangling and analysis using PySpark

from zoo import init_nncontext
from zoo.pipeline.api.net import TFDataset

sc = init_nncontext()

#Each record in the train_rdd consists of a list of NumPy ndrrays
train_rdd = sc.parallelize(file_list)
  .map(lambda x: read_image_and_label(x))
  .map(lambda image_label: decode_to_ndarrays(image_label))

#TFDataset represents a distributed set of elements,
#in which each element contains one or more Tensorflow Tensor objects. 
dataset = TFDataset.from_rdd(train_rdd,
                             names=["features", "labels"],
                             shapes=[[28, 28, 1], [1]],
                             types=[tf.float32, tf.int32],
                             batch_size=BATCH_SIZE)

Deep learning model development using TensorFlow

import tensorflow as tf

slim = tf.contrib.slim

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))

Distributed training on Spark and BigDL

from zoo.pipeline.api.net import TFOptimizer
from bigdl.optim.optimizer import MaxIteration, Adam, MaxEpoch, TrainSummary

optimizer = TFOptimizer(loss, Adam(1e-3))
optimizer.set_train_summary(TrainSummary("/tmp/az_lenet", "lenet"))
optimizer.optimize(end_trigger=MaxEpoch(5))

Alternatively, using Keras APIs for model development and distribtued training

from zoo.pipeline.api.keras.models import *
from zoo.pipeline.api.keras.layers import *

model = Sequential()
model.add(Reshape((1, 28, 28), input_shape=(28, 28, 1)))
model.add(Convolution2D(6, 5, 5, activation="tanh", name="conv1_5x5"))
model.add(MaxPooling2D())
model.add(Convolution2D(12, 5, 5, activation="tanh", name="conv2_5x5"))
model.add(MaxPooling2D())
model.add(Flatten())
model.add(Dense(100, activation="tanh", name="fc1"))
model.add(Dense(class_num, activation="softmax", name="fc2"))

model.compile(loss='sparse_categorical_crossentropy',
              optimizer='adam')
model.fit(train_rdd, batch_size=BATCH_SIZE, nb_epoch=5)

High level abstractions and APIs

Analytics Zoo provides a set of easy-to-use, high level abstractions and APIs that natively transfer learning, autograd and custom layer/loss, Spark DataFrames and ML Pipelines, online model serving, etc. etc.

Transfer learning

Using the high level transfer learning APIs, you can easily customize pretrained models for feature extraction or fine-tuning. (See more details here)

Load an existing model (pretrained in Caffe)

from zoo.pipeline.api.net import *
full_model = Net.load_caffe(def_path, model_path)

Remove the last few layers

# create a new model by removing layers after pool5/drop_7x7_s1
model = full_model.new_graph(["pool5/drop_7x7_s1"])

Freeze the first few layers

# freeze layers from input to pool4/3x3_s2 inclusive
model.freeze_up_to(["pool4/3x3_s2"])

Add a few new layers

from zoo.pipeline.api.keras.layers import *
from zoo.pipeline.api.keras.models import *
inputs = Input(name="input", shape=(3, 224, 224))
inception = model.to_keras()(inputs)
flatten = Flatten()(inception)
logits = Dense(2)(flatten)
newModel = Model(inputs, logits)

`autograd`

autograd provides automatic differentiation for math operations, so that you can easily build your own custom loss and layer (in both Python and Scala), as illustracted below. (See more details here)

Define model using Keras-style API and autograd

import zoo.pipeline.api.autograd as A
from zoo.pipeline.api.keras.layers import *
from zoo.pipeline.api.keras.models import *

input = Input(shape=[2, 20])
features = TimeDistributed(layer=Dense(30))(input)
f1 = features.index_select(1, 0)
f2 = features.index_select(1, 1)
diff = A.abs(f1 - f2)
model = Model(input, diff)

Optionally define custom loss function using autograd

def mean_absolute_error(y_true, y_pred):
    return mean(abs(y_true - y_pred), axis=1)

Train model with custom loss function

model.compile(optimizer=SGD(), loss=mean_absolute_error)
model.fit(x=..., y=...)

`nnframes`

nnframes provides native deep learning support in Spark DataFrames and ML Pipelines, so that you can easily build complex deep learning pipelines in just a few lines, as illustrated below. (See more details here)

Initialize NNContext and load images into DataFrames using NNImageReader

from zoo.common.nncontext import *
from zoo.pipeline.nnframes import *
from zoo.feature.image import *
sc = init_nncontext()
imageDF = NNImageReader.readImages(image_path, sc)

Process loaded data using DataFrames transformations

getName = udf(lambda row: ...)
getLabel = udf(lambda name: ...)
df = imageDF.withColumn("name", getName(col("image"))).withColumn("label", getLabel(col('name')))

Processing image using built-in feature engineering operations

transformer = RowToImageFeature() -> ImageResize(64, 64) -> ImageChannelNormalize(123.0, 117.0, 104.0) \
              -> ImageMatToTensor() -> ImageFeatureToTensor())

Define model using Keras-style APIs

from zoo.pipeline.api.keras.layers import *
from zoo.pipeline.api.keras.models import *
model = Sequential().add(Convolution2D(32, 3, 3, activation='relu', input_shape=(1, 28, 28))) \
                .add(MaxPooling2D(pool_size=(2, 2))).add(Flatten()).add(Dense(10, activation='softmax')))

Train model using Spark ML Pipelines

classifier = NNClassifier(model, CrossEntropyCriterion(),transformer).setLearningRate(0.003) \
                .setBatchSize(40).setMaxEpoch(1).setFeaturesCol("image").setCachingSample(False)
nnModel = classifier.fit(df)

Model Serving

Using the POJO model serving API, you can productionize model serving and infernece in any Java based frameworks (e.g., Spring Framework, Apache Storm, Kafka or Flink, etc.), as illustrated below:

import com.intel.analytics.zoo.pipeline.inference.AbstractInferenceModel;
import com.intel.analytics.zoo.pipeline.inference.JTensor;

public class TextClassificationModel extends AbstractInferenceModel {
    public TextClassificationModel() {
        super();
    }
}

TextClassificationModel model = new TextClassificationModel();
model.load(modelPath, weightPath);

List<JTensor> inputs = preprocess(...);
List<List<JTensor>> result = model.predict(inputs);
...

Built-in deep learning models

Analytics Zoo provides several built-in deep learning models that you can use for a variety of problem types, such as object detection, image classification, text classification, recommendation, etc.

Object detection API

Using Analytics Zoo Object Detection API (including a set of pretrained detection models such as SSD and Faster-RCNN), you can easily build your object detection applications (e.g., localizing and identifying multiple objects in images and videos), as illustrated below. (See more details here)

Download object detection models in Analytics Zoo

You can download a collection of detection models (pretrained on the PSCAL VOC dataset and COCO dataset) from detection model zoo.

Use Object Detection API for off-the-shell inference

from zoo.models.image.objectdetection import *
model = ObjectDetector.load_model(model_path)
image_set = ImageSet.read(img_path, sc)
output = model.predict_image_set(image_set)

Image classification API

Using Analytics Zoo Image Classification API (including a set of pretrained detection models such as VGG, Inception, ResNet, MobileNet, etc.), you can easily build your image classification applications, as illustrated below. (See more details here)

Download image classification models in Analytics Zoo

You can download a collection of image classification models (pretrained on the ImageNet dataset) from image classification model zoo.

Use Image classification API for off-the-shell inference

from zoo.models.image.imageclassification import *
model = ImageClassifier.load_model(model_path)
image_set = ImageSet.read(img_path, sc)
output = model.predict_image_set(image_set)

Text classification API

Analytics Zoo Text Classification API provides a set of pre-defined models (using CNN, LSTM, etc.) for text classifications. (See more details here)

Recommendation API

Analytics Zoo Recommendation API provides a set of pre-defined models (such as Neural Collaborative Filtering, Wide and Deep Learning, etc.) for recommendations. (See more details here)

Reference use cases

Analytics Zoo provides a collection of end-to-end reference use cases, including time series anomaly detection, sentiment analysis, fraud detection, image similarity, etc. (See more details here)

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

analytics-zoo.md

analytics-zoo.md

Analytics Zoo

What is Analytics Zoo?

How to use Analytics Zoo?

Overview

Distributed Tensoflow and Keras on Spark/BigDL

High level abstractions and APIs

Transfer learning

`autograd`

`nnframes`

Model Serving

Built-in deep learning models

Object detection API

Image classification API

Text classification API

Recommendation API

Reference use cases

Files

analytics-zoo.md

Latest commit

History

analytics-zoo.md

File metadata and controls

Analytics Zoo

What is Analytics Zoo?

How to use Analytics Zoo?

Overview

Distributed Tensoflow and Keras on Spark/BigDL

High level abstractions and APIs

Transfer learning

autograd

nnframes

Model Serving

Built-in deep learning models

Object detection API

Image classification API

Text classification API

Recommendation API

Reference use cases

`autograd`

`nnframes`