Vitis AI Library 1.3 Release Notes - 1.4.1 English - UG1354

Vitis AI Library User Guide (UG1354)
Document ID
UG1354
Release Date
2021-12-11
Version
1.4.1 English

This section contains information regarding the features and updates of the Vitis™ AI Library 1.3 release.

Key Features And Enhancements

This AI Library release includes the following key features and enhancements:

New Board Support
Versal Series: VCK190 is supported in this release.
Xmodel Support for Edge
For Edge, the xmodel is used, which is consistent with the xmodel used on Cloud.
PyTorch Framework Support for Edge
In VAI1.3, PyTorch framework is supported for Edge.
New DPU Support

Enable new DPU target DPUCAHX8L designed for Alveo U50, U50LV, and U280 cards.

New Model Libraries
The following new model libraries are supported.
New Model Support
  • Added 13 new PyTorch models
  • Added 17 new TensorFlow models, including five TensorFlow2 models.
  • Added six new Caffe models

Changes

  • xmodel is used for Edge. The elf model is no longer supported.

Compatibility

The Vitis™ AI Library 1.3 is tested with the following images.

  • xilinx-zcu102-dpu-v2020.2-v1.3.0.img.gz
  • xilinx-zcu104-dpu-v2020.2-v1.3.0.img.gz

Model Support

The following models are supported by this version of the Vitis™ AI Library.

Table 1. Models Supported by the AI Library
No. Neural Network ZCU102/ZCU104 VCK190 U50/U50LV/U280 U50-V3ME/U50LV-V3ME/U280-V3ME Application
1 inception_resnet_v2_tf Y Y Y Y Image Classification
2 inception_v1_tf Y Y Y Y
3 inception_v3_tf Y Y Y Y
4 inception_v4_2016_09_09_tf Y Y Y Y
5 mobilenet_v1_0_25_128_tf Y N/A N/A N/A
6 mobilenet_v1_0_5_160_tf Y N/A N/A Y
7 mobilenet_v1_1_0_224_tf Y N/A N/A Y
8 mobilenet_v2_1_0_224_tf Y N/A N/A Y
9 mobilenet_v2_1_4_224_tf Y N/A N/A Y
10 resnet_v1_101_tf Y Y Y Y
11 resnet_v1_152_tf Y Y Y Y
12 resnet_v1_50_tf Y Y Y Y
13 vgg_16_tf Y Y Y Y
14 vgg_19_tf Y Y Y Y
15 ssd_mobilenet_v1_coco_tf Y N/A N/A Y Object Detection
16 ssd_mobilenet_v2_coco_tf Y N/A N/A Y
17 ssd_resnet_50_fpn_coco_tf Y Y Y N/A
18 yolov3_voc_tf Y Y Y N/A
19 mlperf_ssd_resnet34_tf Y Y Y Y
20 resnet50 Y Y Y Y Image Classification
21 resnet18 Y Y Y Y
22 inception_v1 Y Y Y Y
23 inception_v2 Y Y Y Y
24 inception_v3 Y Y Y Y
25 inception_v4 Y Y Y Y
26 mobilenet_v2 Y N/A N/A Y
27 squeezenet Y Y Y Y
28 ssd_pedestrian_pruned_0_97 Y Y Y Y ADAS Pedestrian Detection
29 ssd_traffic_pruned_0_9 Y Y Y Y Traffic Detection
30 ssd_adas_pruned_0_95 Y Y Y Y ADAS Vehicle Detection
31 ssd_mobilenet_v2 Y N/A N/A Y Object Detection
32 refinedet_pruned_0_8 Y Y Y Y
33 refinedet_pruned_0_92 Y Y Y Y
34 refinedet_pruned_0_96 Y Y Y Y
35 vpgnet_pruned_0_99 Y Y Y Y ADAS Lane Detection
36 fpn Y Y Y Y ADAS Segmentation
37 sp_net Y Y Y Y Pose Estimation
38 openpose_pruned_0_3 Y Y Y Y
39 densebox_320_320 Y Y Y Y Face Detection
40 densebox_640_360 Y Y Y Y
41 face_landmark Y Y Y Y Face Detection and Recognition
42 reid Y Y Y Y Object tracking
43 multi_task Y Y Y Y ADAS
44 yolov3_adas_pruned_0_9 Y Y Y N/A Object Detection
45 yolov3_voc Y Y Y N/A
46 yolov3_bdd Y Y Y N/A
47 yolov2_voc Y Y Y N/A
48 yolov2_voc_pruned_0_66 Y Y Y N/A
49 yolov2_voc_pruned_0_71 Y Y Y N/A
50 yolov2_voc_pruned_0_77 Y Y Y N/A
51 facerec_resnet20 Y Y Y Y Face Recognition
52 facerec_resnet64 Y Y Y Y
53 plate_detection Y Y Y Y Plate Recognition
54 plate_recognition Y Y Y N/A
55 FPN_Res18_Medical_segmentation Y Y Y Y Medical Segmentation
56 refinedet_baseline Y Y Y Y Object Detection
57 resnet50_pt Y Y Y Y Image Classification
58 squeezenet_pt Y Y Y Y
59 inception_v3_pt Y Y Y Y
60

personreid-res50_pt

 Y Y Y Y Object Tracking
61

facereid-large_pt

 Y Y Y N/A
62

facereid-small_pt

 Y Y Y Y
63

SemanticFPN_cityscapes_pt

 Y Y Y Y Segmentation
64

facerec-resnet20_mixed_pt

 Y Y Y Y Face Recognition
65 face-quality_pt Y Y Y Y
66 MT-resnet18_mixed_pt Y N/A N/A N/A ADAS
67 salsanext_pt Y Y Y Y Point Cloud
68 pointpillars_kitti_12000_0_pt

pointpillars_kitti_12000_1_pt

 Y N/A N/A N/A
69 unet_chaos-CT_pt Y Y Y N/A CT Segmentation
70 FPN-resnet18_covid19-seg_pt Y Y Y Y Covid-19 Segmentation
71 ENet_cityscapes_pt Y Y Y Y Segmentation
72 personreid-res18_pt Y Y Y N/A Object Tracking
73 yolov4_leaky_spp_m Y Y Y N/A Object Detection
74 hourglass-pe_mpii Y N/A N/A N/A Pose Estimation
75 retinaface Y N/A N/A N/A Face Detection
76 FPN-resnet18_Endov Y N/A N/A N/A Robot Instrument Segmentation
77 tiny_yolov3_vmss Y Y Y N/A Object Detection
78 face-quality Y Y Y Y Face Recognition
79 ssdlite_mobilenet_v2_coco_tf Y N/A N/A Y Object Detection
80 ssd_inception_v2_coco_tf Y N/A N/A N/A
81 MLPerf_resnet50_v1.5_tf Y Y Y Y Image Classification
82 mobilenet_edge_1_0_tf Y N/A N/A N/A
83 mobilenet_edge_0_75_tf Y N/A N/A N/A
84 refinedet_VOC_tf Y Y Y Y Object Detection
85 RefineDet-Medical_EDD_tf Y Y Y Y Medical Detection
86 resnet_v2_50_tf Y N/A N/A N/A Image Classification
87 resnet_v2_101_tf Y N/A N/A N/A
88 resnet_v2_152_tf Y N/A N/A N/A
89 mobilenet_v2_cityscapes_tf Y N/A N/A N/A Segmentation
90 inception_v2_tf Y N/A N/A Y Image Classification
91 resnet50_tf2 Y Y Y Y
92 mobilenet_1_0_224_tf2 Y N/A N/A Y
93 inception_v3_tf2 Y Y Y Y
94 medical_seg_cell_tf2 Y Y Y Y Medical Segmentation
95 semantic_seg_citys_tf2 Y Y Y Y Segmentation
  1. Networks with the suffix "_tf" or "_tf2" were trained on TensorFlow.
  2. Networks with the suffix "_pt" were trained on PyTorch.
  3. Networks with no suffix were trained on Caffe.
  4. The column of U50-V3ME/U50LV-V3ME/U280-V3ME is for DPUCAHX8L.

Device Support

The following platforms and evaluation boards (EVB) are supported by the Vitis™ AI Library 1.3.

Table 2. Edge Device Support
Platform EVB Version
Zynq UltraScale+ MPSoC ZU9EG Xilinx ZCU102 V1.1
Zynq® UltraScale+™ MPSoC ZU7EV Xilinx ZCU104 V1.0
Versal AI Core series VC1902 Xilinx VCK190 V1.0
Table 3. Cloud Board Support
Accelerator Cards
Xilinx Alveo U50
Xilinx Alveo U50LV
Xilinx Alveo U200
Xilinx Alveo U250
Xilinx Alveo U280

Limitations

  • Some neural networks with mobilenet as the backbone are not supported on Versal VCK190 board and on the Alveo U50, U50LV, and U280 cards.
  • Due to limitations of the Docker environment, Multi-task demos cannot run in the DRM mode on Cloud boards.