Search Space Zoo¶

DartsCell¶

DartsCell is extracted from CNN model designed here. A DartsCell is a directed acyclic graph containing an ordered sequence of N nodes and each node stands for a latent representation (e.g. feature map in a convolutional network). Directed edges from Node 1 to Node 2 are associated with some operations that transform Node 1 and the result is stored on Node 2. The operations between nodes is predefined and unchangeable. One edge represents an operation that chosen from the predefined ones to be applied to the starting node of the edge. One cell contains two input nodes, a single output node, and other n_node nodes. The input nodes are defined as the cell outputs in the previous two layers. The output of the cell is obtained by applying a reduction operation (e.g. concatenation) to all the intermediate nodes. To make the search space continuous, the categorical choice of a particular operation is relaxed to a softmax over all possible operations. By adjusting the weight of softmax on every node, the operation with the highest probability is chosen to be part of the final structure. A CNN model can be formed by stacking several cells together, which builds a search space. Note that, in DARTS paper all cells in the model share the same structure.

One structure in the Darts search space is shown below. Note that, NNI merges the last one of the four intermediate nodes and the output node.

The predefined operations are shown in references.

class nni.nas.pytorch.search_space_zoo.DartsCell(n_nodes, channels_pp, channels_p, channels, reduction_p, reduction)[source]¶

Builtin Darts Cell structure. There are n_nodes nodes in one cell, in which the first two nodes’ values are fixed to the results of previous previous cell and previous cell respectively. One node will connect all the nodes after with predefined operations in a mutable way. The last node accepts five inputs from nodes before and it concats all inputs in channels as the output of the current cell, and the number of output channels is n_nodes times channels.

Parameters:

n_nodes (int) – the number of nodes contained in this cell
channels_pp (int) – the number of previous previous cell’s output channels
channels_p (int) – the number of previous cell’s output channels
channels (int) – the number of output channels for each node
reduction_p (bool) – Is previous cell a reduction cell
reduction (bool) – is current cell a reduction cell

forward(pprev, prev)[source]¶

Parameters:	pprev (torch.Tensor) – the output of the previous previous layer prev (torch.Tensor) – the output of the previous layer

Example code¶

example code

git clone https://github.com/Microsoft/nni.git
cd nni/examples/nas/search_space_zoo
# search the best structure
python3 darts_example.py

References¶

All supported operations for Darts are listed below.

MaxPool / AvgPool

MaxPool: Call torch.nn.MaxPool2d. This operation applies a 2D max pooling over all input channels. Its parameters kernel_size=3 and padding=1 are fixed. The pooling result will pass through a BatchNorm2d then return as the result.

AvgPool: Call torch.nn.AvgPool2d. This operation applies a 2D average pooling over all input channels. Its parameters kernel_size=3 and padding=1 are fixed. The pooling result will pass through a BatchNorm2d then return as the result.

MaxPool / AvgPool with kernel_size=3 and padding=1 followed by BatchNorm2d

class nni.nas.pytorch.search_space_zoo.darts_ops.PoolBN(pool_type, C, kernel_size, stride, padding, affine=True)[source]¶

AvgPool or MaxPool with BN. pool_type must be max or avg.

Parameters:	pool_type (str) – choose operation C (int) – number of channels kernal_size (int) – size of the convolving kernel stride (int) – stride of the convolution padding (int) – zero-padding added to both sides of the input affine (bool) – is using affine in BatchNorm

SkipConnect

There is no operation between two nodes. Call torch.nn.Identity to forward what it gets to the output.
Zero operation

There is no connection between two nodes.

DilConv3x3 / DilConv5x5

DilConv3x3: (Dilated) depthwise separable Conv. It’s a 3x3 depthwise convolution with C_in groups, followed by a 1x1 pointwise convolution. It reduces the amount of parameters. Input is first passed through relu, then DilConv and finally batchNorm2d. Note that the operation is not Dilated Convolution, but we follow the convention in NAS papers to name it DilConv. 3x3 DilConv has parameters kernel_size=3, padding=1 and 5x5 DilConv has parameters kernel_size=5, padding=4.

class nni.nas.pytorch.search_space_zoo.darts_ops.DilConv(C_in, C_out, kernel_size, stride, padding, dilation, affine=True)[source]¶

(Dilated) depthwise separable conv. ReLU - (Dilated) depthwise separable - Pointwise - BN. If dilation == 2, 3x3 conv => 5x5 receptive field, 5x5 conv => 9x9 receptive field.

Parameters:	C_in (int) – the number of input channels C_out (int) – the number of output channels kernal_size – size of the convolving kernel padding – zero-padding added to both sides of the input dilation (int) – spacing between kernel elements. affine (bool) – is using affine in BatchNorm

SepConv3x3 / SepConv5x5

Composed of two DilConvs with fixed kernel_size=3, padding=1 or kernel_size=5, padding=2 sequentially.

class nni.nas.pytorch.search_space_zoo.darts_ops.SepConv(C_in, C_out, kernel_size, stride, padding, affine=True)[source]¶

Depthwise separable conv. DilConv(dilation=1) * 2.

Parameters:	C_in (int) – the number of input channels C_out (int) – the number of output channels kernal_size – size of the convolving kernel padding – zero-padding added to both sides of the input dilation (int) – spacing between kernel elements. affine (bool) – is using affine in BatchNorm

ENASMicroLayer¶

This layer is extracted from the model designed here. A model contains several blocks that share the same architecture. A block is made up of some normal layers and reduction layers, ENASMicroLayer is a unified implementation of the two types of layers. The only difference between the two layers is that reduction layers apply all operations with stride=2.

ENAS Micro employs a DAG with N nodes in one cell, where the nodes represent local computations, and the edges represent the flow of information between the N nodes. One cell contains two input nodes and a single output node. The following nodes choose two previous nodes as input and apply two operations from predefined ones then add them as the output of this node. For example, Node 4 chooses Node 1 and Node 3 as inputs then applies MaxPool and AvgPool on the inputs respectively, then adds and sums them as the output of Node 4. Nodes that are not served as input for any other node are viewed as the output of the layer. If there are multiple output nodes, the model will calculate the average of these nodes as the layer output.

One structure in the ENAS micro search space is shown below.

The predefined operations can be seen here.

class nni.nas.pytorch.search_space_zoo.ENASMicroLayer(num_nodes, in_channels_pp, in_channels_p, out_channels, reduction)[source]¶

Builtin EnasMicroLayer. Micro search designs only one building block whose architecture is repeated throughout the final architecture. A cell has num_nodes nodes and searches the topology and operations among them in RL way. The first two nodes in a layer stand for the outputs from previous previous layer and previous layer respectively. For the following nodes, the controller chooses two previous nodes and applies two operations respectively for each node. Nodes that are not served as input for any other node are viewed as the output of the layer. If there are multiple output nodes, the model will calculate the average of these nodes as the layer output. Every node’s output has out_channels channels so the result of the layer has the same number of channels as each node.

Parameters:	num_nodes (int) – the number of nodes contained in this layer in_channles_pp (int) – the number of previous previous layer’s output channels in_channels_p (int) – the number of previous layer’s output channels out_channels (int) – output channels of this layer reduction (bool) – is reduction operation empolyed before this layer

forward(pprev, prev)[source]¶

Parameters:	pprev (torch.Tensor) – the output of the previous previous layer prev (torch.Tensor) – the output of the previous layer

The Reduction Layer is made up of two Conv operations followed by BatchNorm, each of them will output C_out//2 channels and concat them in channels as the output. The Convolution has kernel_size=1 and stride=2, and they perform alternate sampling on the input to reduce the resolution without loss of information. This layer is wrapped in ENASMicroLayer.

Example code¶

example code

git clone https://github.com/Microsoft/nni.git
cd nni/examples/nas/search_space_zoo
# search the best cell structure
python3 enas_micro_example.py

References¶

All supported operations for ENAS micro search are listed below.

MaxPool / AvgPool

MaxPool: Call torch.nn.MaxPool2d. This operation applies a 2D max pooling over all input channels followed by BatchNorm2d. Its parameters are fixed to kernel_size=3, stride=1 and padding=1.
AvgPool: Call torch.nn.AvgPool2d. This operation applies a 2D average pooling over all input channels followed by BatchNorm2d. Its parameters are fixed to kernel_size=3, stride=1 and padding=1.

class nni.nas.pytorch.search_space_zoo.enas_ops.Pool(pool_type, kernel_size, stride, padding)[source]¶

Pooling structure

Parameters:	pool_type (str) – only accept `max` for MaxPool and `avg` for AvgPool kernal_size (int) – size of the convolving kernel stride (int) – stride of the convolution padding (int) – zero-padding added to both sides of the input

SepConv

SepConvBN3x3: ReLU followed by a DilConv and BatchNorm. Convolution parameters are kernel_size=3, stride=1 and padding=1.
SepConvBN5x5: Do the same operation as the previous one but it has different kernel sizes and paddings, which is set to 5 and 2 respectively.

class nni.nas.pytorch.search_space_zoo.enas_ops.SepConvBN(C_in, C_out, kernel_size, padding)[source]¶

Implement SepConv followed by BatchNorm. The structure is ReLU ==> SepConv ==> BN.

Parameters:	C_in (int) – the number of imput channels C_out (int) – the number of output channels kernal_size (int) – size of the convolving kernel padding (int) – zero-padding added to both sides of the input

SkipConnect

Call torch.nn.Identity to connect directly to the next cell.

ENASMacroLayer¶

In Macro search, the controller makes two decisions for each layer: i) the operation to perform on the result of the previous layer, ii) which the previous layer to connect to for SkipConnects. ENAS uses a controller to design the whole model architecture instead of one of its components. The output of operations is going to concat with the tensor of the chosen layer for SkipConnect. NNI provides predefined operations for macro search, which are listed in references.

Part of one structure in the ENAS macro search space is shown below.

class nni.nas.pytorch.search_space_zoo.ENASMacroLayer(key, prev_labels, in_filters, out_filters)[source]¶

Builtin ENAS Marco Layer. With search space changing to layer level, the controller decides what operation is employed and the previous layer to connect to for skip connections. The model is made up of the same layers but the choice of each layer may be different.

Parameters:	key (str) – the name of this layer prev_labels (str) – names of all previous layers in_filters (int) – the number of input channels out_filters – the number of output channels

forward(prev_list)[source]¶

Parameters:	prev_list (list) – The cell selects the last element of the list as input and applies an operation on it. The cell chooses none/one/multiple tensor(s) as SkipConnect(s) from the list excluding the last element.

To describe the whole search space, NNI provides a model, which is built by stacking the layers.

class nni.nas.pytorch.search_space_zoo.ENASMacroGeneralModel(num_layers=12, out_filters=24, in_channels=3, num_classes=10, dropout_rate=0.0)[source]¶

The network is made up by stacking ENASMacroLayer. The Macro search space contains these layers. Each layer chooses an operation from predefined ones and SkipConnect then forms a network.

Parameters:	num_layers (int) – The number of layers contained in the network. out_filters (int) – The number of each layer’s output channels. in_channel (int) – The number of input’s channels. num_classes (int) – The number of classes for classification. dropout_rate (float) – Dropout layer’s dropout rate before the final dense layer.

forward(x)[source]¶

Parameters:	x (torch.Tensor) – the input of the network

Example code¶

example code

git clone https://github.com/Microsoft/nni.git
cd nni/examples/nas/search_space_zoo
# search the best cell structure
python3 enas_macro_example.py

References¶

All supported operations for ENAS macro search are listed below.

ConvBranch

All input first passes into a StdConv, which is made up of a 1x1Conv followed by BatchNorm2d and ReLU. Then the intermediate result goes through one of the operations listed below. The final result is calculated through a BatchNorm2d and ReLU as post-procedure.

Separable Conv3x3: If separable=True, the cell will use SepConv instead of normal Conv operation. SepConv’s kernel_size=3, stride=1 and padding=1.
Separable Conv5x5: SepConv’s kernel_size=5, stride=1 and padding=2.
Normal Conv3x3: If separable=False, the cell will use a normal Conv operations with kernel_size=3, stride=1 and padding=1.
Normal Conv5x5: Conv’s kernel_size=5, stride=1 and padding=2.

class nni.nas.pytorch.search_space_zoo.enas_ops.ConvBranch(C_in, C_out, kernel_size, stride, padding, separable)[source]¶

Conv structure for Macro search. First pass through a 1x1 Conv, then Conv operation with kernal_size equals 3 or 5 followed by BatchNorm and ReLU.

Parameters:	C_in (int) – the number of input channels C_out (int) – the number of output channels kernal_size (int) – size of the convolving kernel stride (int) – stride of the convolution padding (int) – zero-padding added to both sides of the input separable (True) – is separable Conv is used

PoolBranch

All input first passes into a StdConv, which is made up of a 1x1Conv followed by BatchNorm2d and ReLU. Then the intermediate goes through pooling operation followed by BatchNorm.

AvgPool: Call torch.nn.AvgPool2d. This operation applies a 2D average pooling over all input channels. Its parameters are fixed to kernel_size=3, stride=1 and padding=1.
MaxPool: Call torch.nn.MaxPool2d. This operation applies a 2D max pooling over all input channels. Its parameters are fixed to kernel_size=3, stride=1 and padding=1.

class nni.nas.pytorch.search_space_zoo.enas_ops.PoolBranch(pool_type, C_in, C_out, kernel_size, stride, padding, affine=False)[source]¶

Pooling structure for Macro search. First pass through a 1x1 Conv, then pooling operation followed by BatchNorm2d.

Parameters:	pool_type (str) – only accept `max` for MaxPool and `avg` for AvgPool C_in (int) – the number of input channels C_out (int) – the number of output channels kernal_size (int) – size of the convolving kernel stride (int) – stride of the convolution padding (int) – zero-padding added to both sides of the input