MODELZOO.md

CHaiDNN-v2
Analysis and Eval
	Supported Layers		Performance/Resource Utilization
	Performance Eval
Design and Development
	API Reference	Quantization User Guide for CHaiDNN	Model Zoo	Running Inference on new Network
	Creating SDx GUI Project	Configurable Parameters	Custom Platform Generation	Software Layer Plugin
	SDSoC Environment User Guide		Hardware-Software Partitioning for Performance

ModelZoo

CHaiDNN ModelZoo contains a collection of pre-trained reduced precision Convolutional Neural Networks. These models were carefully analyzed to find the best precision details per layer.

CHai ModelZoo consists of 3 sets of example-models:

1. Xilinx's Quantizer Models: Models generated using the technique described here
2. Dynamic Fixed Models: Models generated using the technique described here
3. User Defined Software Layer Model: CHai supports software layer plug-in. For more details, see Software Layer Plugin.

Xilinx's Quantizer Model Details

Models for 8-bit and 6-bit bit-width.

DNN model	Single Precision Top-1 (Top-5)	S/W*** 8Bit Top-1 (Top-5)	S/W*** 6Bit Top-1 (Top-5)
AlexNet_NoLRN	55.54 (78.75)	55.18 (78.60)	55.24 (78.44) *
AlexNet_FCN	39.74 (MeanIOU)	39.14 (MeanIOU)	35.84 (MeanIOU) **
VGG-16	68.3 (88.4)	68.18 (88.3)	70.99 (90.23) *
VGG-SSD300	78.21 (mAP)	78.06 (mAP)	77.28 (mAP)
GoogleNet_V1_NoLRN	67.35 (88.18)	67.09 (87.97)	67.27 (87.99) *
ResNet-50-BNfused	72.86 (91.11)	73.4 (91.47) *	68.8 (89.04)**

^{* indicates accuracies for the shared models is obtained by re-trained/finetuned to re-gain lost accuracy due to quantization. ** indicates accuracies for the shared models can be improved with retraining. *** refers to the Caffe + Xilinx quantization.}

Dynamic Fixed Model Details

Bitwidth specifies the weights/ activations precisions.

DNN model	Bitwidth 8-bit or 6-bit
GoogleNet_V1_NoLRN	Yes

Custom Layer Model Details

DNN model	Bitwidth 8-bit or 6-bit	Custom Layers
GoogleNet_V1_NoLRN	Yes	Fully-connected/ Innerproduct layer

---

^{Copyright© 2018 Xilinx}