Akida models zoo

Overview

Brainchip Akida Models package is a model zoo that offers a set of pre-built akida compatible models (e.g MobileNet, AkidaNet, ViT), pretrained weights for those models and training scripts. See the model zoo API reference for a complete list of the available models. The performance of all models from the zoo are reported for both Akida 1.0 and Akida 2.0 in the model zoo performance page. Akida Models also contains a set of layer blocks that are used to define the above models.

Command-line interface for model creation

In addition to the programming API, the Akida Models toolkit provides a command-line interface to instantiate and save models from the zoo.

Instantiating models using the CLI makes use of the model definitions from the programming interface with default values. To quantize a given model, the QuantizeML quantize CLI should be used.

Examples

Instantiate a DS-CNN network for KWS (keyword spotting):

akida_models create ds_cnn_kws

The model is automatically saved to ds_cnn_kws.h5.

Some models come with additional parameters that allow a deeper configuration. Examples are given below.

To build an AkidaNet model with a 64x64 input size, alpha parameter (model width) equal to 0.35 and 250 classes:

akida_models create akidanet_imagenet -i 64 -a 0.35 -c 250

To create a YOLO model with 20 classes, 5 anchors and a model width of 0.5:

akida_models create yolo_base -c 20 -na 5 -a 0.5

The full parameter list with description can be obtained using the -h or --help argument for each model:

akida_models create akidanet_imagenet -h

usage: akida_models create akidanet_imagenet [-h]
                                             [-i {32,64,96,128,160,192,224}]
                                             [-a ALPHA] [-c CLASSES]

optional arguments:
    -h, --help            show this help message and exit
     -c CLASSES, --classes CLASSES
                          The number of classes, by default 1000.
     -i {32,64,96,128,160,192,224}, --image_size {32,64,96,128,160,192,224}
                          The square input image size, by default 224.
     -a ALPHA, --alpha ALPHA
                          The width of the model, by default 1.0.

Current available models for creation are:

  • vgg_utk_face

  • convtiny_dvs_handy

  • convtiny_dvs_gesture

  • ds_cnn_kws

  • pointnet_plus_modelnet40

  • mobilenet_imagenet

  • akidanet_imagenet

  • akidanet_imagenet_edge

  • akidanet18_imagenet

  • yolo_base

  • centernet

  • gxnor_mnist

  • akida_unet_portrait128

  • vit_ti16

  • bc_vit_ti16

  • deit_ti16

  • bc_deit_ti16

Command-line interface for model training

The package also comes with a CLI to train models from the zoo.

Training models first requires that a model is created and saved using the CLI described above. Once a model is ready, training will use dedicated scripts to load and preprocess a dataset and perform training.

As shown in the examples below, the training CLI should be used along with akida_models create and quantizeml quantize.

If the quantized model offers acceptable performance, it can be converted into an Akida model, ready to be loaded on the Akida NSoC using the CNN2SNN convert CLI.

KWS training

KWS training pipeline uses the ds_cnn_kws model and the QuantizeML quantize CLI. Dataset loading and preprocessing is done within the training script called by the kws_train CLI.

Example

Create a DS-CNN model for KWS, train it over 16 epochs, then quantize it to 4-bit weights and activations (using a set of samples for calibration only), perform a 16 epochs QAT to recover accuracy and evaluate.

akida_models create -s ds_cnn_kws.h5 ds_cnn_kws
kws_train train -m ds_cnn_kws.h5 -s ds_cnn_kws.h5 -e 16

wget https://data.brainchip.com/dataset-mirror/samples/kws/kws_batch1024.npz
quantizeml quantize -m ds_cnn_kws.h5 -w 4 -a 4 -e 2 -bs 100 -sa kws_batch1024.npz
kws_train train -m ds_cnn_kws_i8_w4_a4.h5 -e 16 -s ds_cnn_kws_i8_w4_a4.h5
kws_train eval -m ds_cnn_kws_i8_w4_a4.h5

AkidaNet training

AkidaNet training pipeline uses the akidanet_imagenet model and the QuantizeML quantize CLI. Dataset loading and preprocessing is done within the training script called by the imagenet_train CLI. Note that ImageNet data must be downloaded from https://www.image-net.org/ first.

Example

Create an AkidaNet 0.5 with resolution 160, train it for 90 epochs then quantize to 4-bit weights and activations, perform a 10 epochs QAT to recover accuracy, upscale to resolution 224 and evaluate.

akida_models create -s akidanet_imagenet_160_alpha_0.5.h5 akidanet_imagenet -a 0.5 -i 160
imagenet_train train -d path/to/imagenet/ -e 90 -m akidanet_imagenet_160_alpha_0.5.h5 \
                     -s akidanet_imagenet_160_alpha_0.5.h5

wget https://data.brainchip.com/dataset-mirror/samples/imagenet/imagenet_batch1024_160.npz
quantizeml quantize -m akidanet_imagenet_160_alpha_0.5.h5 -w 4 -a 4 -e 2 -bs 100 \
                     -sa imagenet_batch1024_160.npz
imagenet_train tune -d path/to/imagenet/ -e 10 -m akidanet_imagenet_160_alpha_0.5_i8_w4_a4.h5 \
                    -s akidanet_imagenet_160_alpha_50_i8_w4_a4.h5
imagenet_train rescale -i 224 -m akidanet_imagenet_160_alpha_0.5_i8_w4_a4.h5 \
                       -s akidanet_imagenet_224_alpha_0.5_i8_w4_a4.h5
imagenet_train eval -d path/to/imagenet/ -m akidanet_imagenet_224_alpha_0.5_i8_w4_a4.h5

Current training pipelines available are:

  • utk_face_train

  • kws_train

  • modelnet40_train

  • yolo_train

  • dvs_train

  • mnist_train

  • imagenet_train

  • portrait128_train

  • centernet_train

Command-line interface for model evaluation

The CLI also comes with an eval action that allows to evaluate model performance, the -ak or --akida option allows to convert to Akida then evaluate the model.

kws_train eval -m ds_cnn_kws_i8_w4_a4.h5

kws_train eval -m ds_cnn_kws_i8_w4_a4.h5 -ak

Command-line interface to evaluate model MACS

CLI comes with a macs action that allows to compute the number of multiply and accumulate (MACS) in a model.

akida_models macs -m akidanet_imagenet_224_alpha_0.5.h5 -v

Layer Blocks

In Keras, it is very common for activations or other functions to be defined along with the processing layer, e.g.:

x = Dense(64)(x)
x = BatchNormalization()(x)
x = Activation('relu')(x)

In order to ease the design of a Keras model compatible for conversion into an Akida model, a higher-level interface is proposed with the use of layer blocks. These blocks are available in the package through:

import akida_models.layer_blocks

For instance, the following code snippet sets up the same trio of layers as those above:

x = dense_block(x, 64, add_batchnorm=True, relu_activation='ReLU')

The dense_block function will produce a group of layers that we call a “block”.

Note

  • To avoid adding the activation layer, add the parameter relu_activation = False to the block.

  • The ReLU activation max_value can be set in the parameter using a string expression, that is relu_activation='ReLU6' will create a ReLU activation with max_value set to 6.

  • The ReLu activation can also be defined as unbounded, that is relu_activation='ReLU' (only supported for models targeting Akida 2.0)

Separable layers can be defined as fused (Akida 1.0) or unfused (Akida 2.0):

x = separable_conv_block(x, 64, 3, add_batchnorm=True, relu_activation='ReLU6', fused=False)

Placement of the GlobalAveragePooling (GAP) operation is also configurable in layer blocks so that it comes before the activation (post_relu_gap=False for Akida 1.0) or after (post_relu_gap=True for Akida 2.0):

x = conv_block(x, 64, 3, relu_activation='ReLU', post_relu_gap=True)

The option of including pooling, BatchNormalization layers or activation is directly built into the provided block modules.

The layer block functions provided are:

Most of the parameters for these blocks are identical to those passed to the corresponding inner processing layers, such as strides and bias. The detailed API is given in the dedicated section.

Handling Akida 1.0 and Akida 2.0 specificities

Akida 1.0 and 2.0 specific model architecture requirements are embedded in the returned models (pretrained or not). By default, the returned models and pretrained model target Akida 2.0. It is however possible to build and instantiate Akida 1.0 models.

Using the programming interface:

from akida_models import ds_cnn_kws, ds_cnn_kws_pretrained
from cnn2snn import set_akida_version, AkidaVersion

with set_akida_version(AkidaVersion.v1):
   model = ds_cnn_kws()
   pretrained = ds_cnn_kws_pretrained()

Using the CLI interface:

CNN2SNN_TARGET_AKIDA_VERSION=v1 akida_models create ds_cnn_kws