masterdevelop

TensorGraph

TensorGraph is a simple, lean, and clean framework on TensorFlow for building any imaginable models.

As deep learning becomes more and more common and the architectures becoming more and more complicated, it seems that we need some easy to use framework to quickly build these models and that's what TensorGraph is designed for. It's a very simple framework that adds a very thin layer above tensorflow. It is for more advanced users who want to have more control and flexibility over his model building and who wants efficiency at the same time.

---

Target Audience

TensorGraph is targeted more at intermediate to advance users who feel keras or other packages is having too much restrictions and too much black box on model building, and someone who don't want to rewrite the standard layers in tensorflow constantly. Also for enterprise users who want to share deep learning models easily between teams.

---

Install

First you need to install tensorflow

To install tensorgraph for bleeding edge version via pip

sudo pip install --upgrade git+https://github.com/hycis/TensorGraph.git@master

or simply clone and add to PYTHONPATH.

git clone https://github.com/hycis/TensorGraph.git
export PYTHONPATH=/path/to/TensorGraph:$PYTHONPATH

in order for the install to persist via export PYTHONPATH. Add PYTHONPATH=/path/to/TensorGraph:$PYTHONPATH to your .bashrc for linux or .bash_profile for mac. While this method works, you will have to ensure that all the dependencies in setup.py are installed.

---

Everything in TensorGraph is about Layers

Everything in TensorGraph is about layers. A model such as VGG or Resnet can be a layer. An identity block from Resnet or a dense block from Densenet can be a layer as well. Building models in TensorGraph is same as building a toy with lego. For example you can create a new model (layer) by subclass the BaseModel layer and use DenseBlock layer inside your ModelA layer.

fromtensorgraph.layersimportDenseBlock, BaseModel, Flatten, Linear, SoftmaximporttensorgraphastgclassModelA(BaseModel):
@BaseModel.init_name_scopedef__init__(self):
layers= []
layers.append(DenseBlock())
layers.append(Flatten())
layers.append(Linear())
layers.append(Softmax())
self.startnode=tg.StartNode(input_vars=[None])
hn=tg.HiddenNode(prev=[self.startnode], layers=layers)
self.endnode=tg.EndNode(prev=[hn])

if someone wants to use your ModelA in his ModelB, he can easily do this

classModelB(BaseModel):
@BaseModel.init_name_scopedef__int__(self):
layers= []
layers.append(ModelA())
layers.append(Linear())
layers.append(Softmax())
self.startnode=tg.StartNode(input_vars=[None])
hn=tg.HiddenNode(prev=[self.startnode], layers=layers)
self.endnode=tg.EndNode(prev=[hn])

creating a layer only created all the Variables. To connect the Variables into a graph, you can do a train_fprop(X) or test_fprop(X) to create the tensorflow graph. By abstracting Variable creation away from linking the Variable nodes into graph prevent the problem of certain tensorflow layers that always reinitialise its weights when it's called, example the tf.nn.batch_normalization layer. Also having a separate channel for training and testing is to cater to layers with different training and testing behaviours such as batchnorm and dropout.

modelb=ModelB()
X_ph=tf.placeholder()
y_train=modelb.train_fprop(X_ph)
y_test=modelb.test_fprop(X_ph)

checkout some well known models in TensorGraph

1. VGG16 code and VGG19 code - Very Deep Convolutional Networks for Large-Scale Image Recognition
2. DenseNet code - Densely Connected Convolutional Networks
3. ResNet code - Deep Residual Learning for Image Recognition
4. Unet code - U-Net: Convolutional Networks for Biomedical Image Segmentation

---

Type of Layers

There are three types of layers, BaseLayer, BaseModel and Merge.

BaseLayer

BaseLayer is a low lying layer that wraps tensorflow codes directly, and define the low level operations that we want the tensorflow to perform within a layer. When implementing BaseLayer we need to implement _train_fprop() and _test_fprop(), by default _test_fprop() calls _train_fprop().

classMyLayer(BaseLayer):
@BaseLayer.init_name_scopedef__init__(self):
''' place all your variables and variables initialization here. '''pass@BaseLayer.init_name_scopedef__init_var__(self, state_below):
'''Define variables which requires input information from state_below, this is called during forward propagation '''passdef_train_fprop(self, state_below):
''' your tensorflow operations for training, do not initialize variables here. '''passdef_test_fprop(self, state_below):
''' your tensorflow operations for testing, do not initialize variables here. Defaults to _train_fprop. '''pass

To use BaseLayer, we can initialize the Variables inside __init__ and/or __init_var__(self, state_below) if our layer requires information from the layer below.

BaseModel

BaseModel is a higher level layer that can be made up of BaseLayers and BaseModels. For BaseModel, a default implementation of _train_fprop and _test_fprop has been done for a single StartNode and single EndNode Graph, to use this default implementation, we have to define self.startnode and self.endnode inside BaseModel's __init__.

For Graph defined inside BaseModel, BaseModel will automatically call the _train_fprop and _test_fprop within each layer inside its model.

classMyLayer(BaseModel):
def__init__(self):
''' place all your layers inside here and define self.startnode and self.endnode example: layers = [] layers.append(DenseBlock()) layers.append(Flatten()) layers.append(Linear()) layers.append(Softmax()) self.startnode = tg.StartNode(input_vars=[None]) hn = tg.HiddenNode(prev=[self.startnode], layers=layers) self.endnode = tg.EndNode(prev=[hn]) '''pass

It is possible for BaseModel to return multiple outputs, example

classMyLayerFork(BaseModel):
@BaseModel.init_name_scopedef__init__(self):
# a Y shape model, where we have one input and two outputsself.startnode=tg.StartNode(input_vars=[None])
# first fork outputlayers= []
layers.append(Linear())
layers.append(Softmax())
hn=tg.HiddenNode(prev=[self.startnode], layers=layers)
# second fork outputlayers2= []
layers2.append(Linear())
layers2.append(Softmax())
hn2=tg.HiddenNode(prev=[self.startnode], layers=layers2)
# two forks outputsself.endnode=tg.EndNode(prev=[hn, h2])

In this case, a call to train_fprop will return two outputs

mylayer=MylayerFork()
y1, y2=mylayer.train_fprop(X_ph)

Customize inputs and outputs for BaseModel

Another way to customize your own inputs and outputs is to redefine _train_fprop and _test_fprop within BaseModel.

The default _train_fprop and _test_fprop in BaseModel looks like this

classBaseModel(Template):
@staticmethoddefcheck_y(y):
iflen(y) ==1:
returny[0]
eliflen(y) >1:
returnyelse:
raiseException('{} is empty or not a list'.format(y))
def_train_fprop(self, *state_belows):
self.startnode.input_vars=state_belowsgraph=Graph(start=[self.startnode], end=[self.endnode])
y=graph.train_fprop()
returnBaseModel.check_y(y)
def_test_fprop(self, *state_belows):
self.startnode.input_vars=state_belowsgraph=Graph(start=[self.startnode], end=[self.endnode])
y=graph.test_fprop()
returnBaseModel.check_y(y)

for the MyLayerFork Model, for two inputs and two outputs, we can redefine it with multiple StartNodes and EndNodes within _train_fprop and _test_fprop.

classMyLayerFork(BaseModel):
@BaseModel.init_name_scopedef__init__(self):
# multiple inputs and multiple outputsself.startnode1=tg.StartNode(input_vars=[None])
self.startnode2=tg.StartNode(input_vars=[None])
layers1= []
layers1.append(Linear())
layers1.append(Softmax())
hn1=tg.HiddenNode(prev=[self.startnode1], layers=layers)
layers2= []
layers2.append(Linear())
layers2.append(Softmax())
hn2=tg.HiddenNode(prev=[self.startnode2], layers=layers2)
# two forks outputsself.endnode1=tg.EndNode(prev=[hn1])
self.endnode2=tg.EndNode(prev=[hn2])
def_train_fprop(self, input1, input2):
self.startnode1.input_vars= [input1]
self.startnode2.input_vars= [input2]
graph=Graph(start=[self.startnode1, self.startnode2], end=[self.endnode1, self.endnode2])
y=graph.train_fprop()
returnBaseModel.check_y(y)
def_test_fprop(self, input1, input2):
self.startnode1.input_vars= [input1]
self.startnode2.input_vars= [input2]
graph=Graph(start=[self.startnode1, self.startnode2], end=[self.endnode1, self.endnode2])
y=graph.test_fprop()
returnBaseModel.check_y(y)
if__name__=='__main__':
model=MyLayerFork()
y1, y2=model.train_fprop(X1, X2)

Merge

When we have more than one outputs from previous layer and we want to merge them, we can use the Merge layer in tensorgraph.layers.merge.Merge to merge multiple inputs into one.

classConcat(Merge):
@Merge.init_name_scopedef__init__(self, axis=1):
''' Concat which is a Merge layer is used to concat the list of states from layer below into one state '''self.axis=axisdef_train_fprop(self, state_list):
returntf.concat(axis=self.axis, values=state_list)

We can use Merge layer in conjunction with BaseModel layer with multiple outputs, example

classMyLayerMergeFork(BaseModel):
def__init__(self):
layers= []
# fork layer from above examplelayers.append(MyLayerFork())
# merge layerlayers.append(Concat())
self.startnode=tg.StartNode(input_vars=[None])
hn=tg.HiddenNode(prev=[self.startnode], input_merge_mode=NoChange(), layers=layers)
self.endnode=tg.EndNode(prev=[hn])

---

How TensorGraph Works?

In TensorGraph models, layers are put into nodes and nodes are connected together into graph. When we create nodes and layers, we also initializes all the tensorflow Variables, then we connect the nodes together to form a computational graph. The initialization of Variables and the linking of Variables into a computational graph are two separate steps. By splitting them into two separate steps, we ensure the flexibility of building our computational graph without the worry of accidental reinitialization of the Variables. We defined three types of nodes

1. StartNode : for inputs to the graph
2. HiddenNode : for putting sequential layers inside
3. EndNode : for getting outputs from the model

We put all the sequential layers into a HiddenNode, HiddenNode can be connected to another HiddenNode or StartNode, the nodes are connected together to form an architecture. The graph always starts with StartNode and ends with EndNode. Once we have defined an architecture, we can use the Graph object to connect the path we want in the architecture, there can be multiple StartNodes (s1, s2, etc) and multiple EndNodes (e1, e2, etc), we can define which path we want in the entire architecture, example to link from s2 to e1. The StartNode is where you place your starting point, it can be a placeholder, a symbolic output from another graph, or data output from tfrecords. EndNode is where you want to get an output from the graph, where the output can be used to calculate loss or simply just a peek at the outputs at that particular layer. Below shows an example of building a tensor graph.

---

Graph Example

First define the StartNode for putting the input placeholder

y1_dim=50y2_dim=100batchsize=32learning_rate=0.01y1=tf.placeholder('float32', [None, y1_dim])
y2=tf.placeholder('float32', [None, y2_dim])
s1=StartNode(input_vars=[y1])
s2=StartNode(input_vars=[y2])

Then define the HiddenNode for putting the sequential layers in each HiddenNode

h1=HiddenNode(prev=[s1, s2],
input_merge_mode=Concat(),
layers=[Linear(y2_dim), RELU()])
h2=HiddenNode(prev=[s2],
layers=[Linear(y2_dim), RELU()])
h3=HiddenNode(prev=[h1, h2],
input_merge_mode=Sum(),
layers=[Linear(y1_dim), RELU()])
layers=[Linear(y1_dim+y2_dim, y2_dim), RELU()])
h2=HiddenNode(prev=[s2],
layers=[Linear(y2_dim, y2_dim), RELU()])
h3=HiddenNode(prev=[h1, h2],
input_merge_mode=Sum(),
layers=[Linear(y2_dim, y1_dim), RELU()])

Then define the EndNode. EndNode is used to back-trace the graph to connect the nodes together.

e1=EndNode(prev=[h3])
e2=EndNode(prev=[h2])

Finally build the graph by putting StartNodes and EndNodes into Graph, we can choose to use the entire architecture by using all the StartNodes and EndNodes and run the forward propagation to get symbolic output from train mode. The number of outputs from graph.train_fprop is the same as the number of EndNodes put into Graph

graph=Graph(start=[s1, s2], end=[e1, e2])
o1, o2=graph.train_fprop()

or we can choose which node to start and which node to end, example

graph=Graph(start=[s2], end=[e1])
o1, =graph.train_fprop()

Finally build an optimizer to optimize the objective function

o1_mse=tf.reduce_mean((y1-o1)**2)
o2_mse=tf.reduce_mean((y2-o2)**2)
mse=o1_mse+o2_mseoptimizer=tf.train.AdamOptimizer(learning_rate).minimize(mse)

---

TensorGraph on Multiple GPUS

To use tensorgraph on multiple gpus, you can easily integrate it with horovod.

importhorovod.tensorflowashvdfromtensorflow.python.frameworkimportopsimporttensorflowastfhvd.init()
# tensorgraph model derived previouslymodelb=ModelB()
X_ph=tf.placeholder()
y_ph=tf.placeholder()
y_train=modelb.train_fprop(X_ph)
y_test=modelb.test_fprop(X_ph)
train_cost=mse(y_train, y_ph)
test_cost=mse(y_test, y_ph)
opt=tf.train.RMSPropOptimizer(0.001)
opt=hvd.DistributedOptimizer(opt)
# required for BatchNormalization layerupdate_ops=ops.get_collection(ops.GraphKeys.UPDATE_OPS)
withops.control_dependencies(update_ops):
train_op=opt.minimize(train_cost)
init_op=tf.group(tf.global_variables_initializer(),
tf.local_variables_initializer())
bcast=hvd.broadcast_global_variables(0)
# Pin GPU to be used to process local rank (one GPU per process)config=tf.ConfigProto()
config.gpu_options.allow_growth=Trueconfig.gpu_options.visible_device_list=str(hvd.local_rank())
withtf.Session(graph=graph, config=config) assess:
sess.run(init_op)
bcast.run()
# training modelforepochinrange(100):
forX,yintrain_data:
_, loss_train=sess.run([train_op, train_cost], feed_dict={X_ph:X, y_ph:y})

for a full example on tensorgraph on horovod

---

Hierachical Softmax Example

Below is another example for building a more powerful hierachical softmax whereby the lower hierachical softmax layer can be conditioned on all the upper hierachical softmax layers.

## paramsx_dim=50component_dim=100batchsize=32learning_rate=0.01x_ph=tf.placeholder('float32', [None, x_dim])
# the three hierachical levely1_ph=tf.placeholder('float32', [None, component_dim])
y2_ph=tf.placeholder('float32', [None, component_dim])
y3_ph=tf.placeholder('float32', [None, component_dim])
# define the graph model structurestart=StartNode(input_vars=[x_ph])
h1=HiddenNode(prev=[start], layers=[Linear(component_dim), Softmax()])
h2=HiddenNode(prev=[h1], layers=[Linear(component_dim), Softmax()])
h3=HiddenNode(prev=[h2], layers=[Linear(component_dim), Softmax()])
h1=HiddenNode(prev=[start], layers=[Linear(x_dim, component_dim), Softmax()])
h2=HiddenNode(prev=[h1], layers=[Linear(component_dim, component_dim), Softmax()])
h3=HiddenNode(prev=[h2], layers=[Linear(component_dim, component_dim), Softmax()])
e1=EndNode(prev=[h1], input_merge_mode=Sum())
e2=EndNode(prev=[h1, h2], input_merge_mode=Sum())
e3=EndNode(prev=[h1, h2, h3], input_merge_mode=Sum())
graph=Graph(start=[start], end=[e1, e2, e3])
o1, o2, o3=graph.train_fprop()
o1_mse=tf.reduce_mean((y1_ph-o1)**2)
o2_mse=tf.reduce_mean((y2_ph-o2)**2)
o3_mse=tf.reduce_mean((y3_ph-o3)**2)
mse=o1_mse+o2_mse+o3_mseoptimizer=tf.train.AdamOptimizer(learning_rate).minimize(mse)

---

Transfer Learning Example

Below is an example on transfer learning with bi-modality inputs and merge at the middle layer with shared representation, in fact, TensorGraph can be used to build any number of modalities for transfer learning.

## paramsx1_dim=50x2_dim=100shared_dim=200y_dim=100batchsize=32learning_rate=0.01x1_ph=tf.placeholder('float32', [None, x1_dim])
x2_ph=tf.placeholder('float32', [None, x2_dim])
y_ph=tf.placeholder('float32', [None, y_dim])
# define the graph model structures1=StartNode(input_vars=[x1_ph])
s2=StartNode(input_vars=[x2_ph])
h1=HiddenNode(prev=[s1], layers=[Linear(shared_dim), RELU()])
h2=HiddenNode(prev=[s2], layers=[Linear(shared_dim), RELU()])
h3=HiddenNode(prev=[h1,h2], input_merge_mode=Sum(),
layers=[Linear(y_dim), Softmax()])
h1=HiddenNode(prev=[s1], layers=[Linear(x1_dim, shared_dim), RELU()])
h2=HiddenNode(prev=[s2], layers=[Linear(x2_dim, shared_dim), RELU()])
h3=HiddenNode(prev=[h1,h2], input_merge_mode=Sum(),
layers=[Linear(shared_dim, y_dim), Softmax()])
e1=EndNode(prev=[h3])
graph=Graph(start=[s1, s2], end=[e1])
o1, =graph.train_fprop()
mse=tf.reduce_mean((y_ph-o1)**2)
optimizer=tf.train.AdamOptimizer(learning_rate).minimize(mse)