Skip to content

Latest commit

 

History

History
556 lines (462 loc) · 44.4 KB

uifmodelsetup.md

File metadata and controls

556 lines (462 loc) · 44.4 KB
Raw
-

Unified Inference Frontend (UIF) 1.2 User Guide

Step 2.0: UIF Model Setup

Table of Contents

2.1: UIF Model Zoo Introduction

UIF 1.2 Model Zoo provides 50 models for AMD Instinct™ GPUs (MIGraphX) including 20 new models and 30 models inherited from UIF 1.1. The Vitis™ AI development environment provides 106 reference models for different FPGA adaptive platforms. Also, you could use the 84 models for AMD EPYC™ CPUs (ZenDNN) inherited from UIF1.1.

Model information is located in model-list.

Note: If a model is marked as limited to non-commercial use, you must comply with the AMD license agreement for non-commercial models.

UIF 1.2 Models for MIGraphX

Click here to view details
# Model Original Platform Datatype FP32 Datatype FP16 Pruned Reminder for limited use scope
1 2D-Unet TensorFlow2 √ √ ×
2 2D-Unet pruned0.7 TensorFlow2 √ √ √
3 Albert-base PyTorch √ √ ×
4 Albert-large PyTorch √ √ ×
5 Bert-base TensorFlow2 √ √ ×
6 Bert-large TensorFlow2 √ √ ×
7 DETR PyTorch √ √ ×
8 DistillBert PyTorch √ √ ×
9 DLRM (40M) PyTorch √ √ × Non-Commercial Use Only
10 EfficientDet TensorFlow2 √ √ ×
11 GPT2-large PyTorch √ √ ×
12 GPT2-medium PyTorch √ √ ×
13 GPT2-XL PyTorch √ √ ×
14 MobileBert PyTorch √ √ ×
15 PointPillars PyTorch √ √ × Non-Commercial Use Only
16 Resnet50_v1.5 ofa PyTorch √ √ √ Non-Commercial Use Only
17 RetinaNet PyTorch √ √ ×
18 ViT PyTorch √ √ × Non-Commercial Use Only
19 W&D PyTorch √ √ ×
20 yolov3 TensorFlow2 √ √ ×

UIF 1.2 Models for Vitis AI

Click here to view details
# Model Platform Datatype FP32 Datatype INT8 Pruned Reminder for limited use scope
1 inceptionv1 TensorFlow √ √ × Non-Commercial Use Only
2 inceptionv1 pruned0.09 TensorFlow √ √ √ Non-Commercial Use Only
3 inceptionv1 pruned0.16 TensorFlow √ √ √ Non-Commercial Use Only
4 inceptionv3 TensorFlow √ √ × Non-Commercial Use Only
5 inceptionv3 pruned0.2 TensorFlow √ √ √ Non-Commercial Use Only
6 inceptionv3 pruned0.4 TensorFlow √ √ √ Non-Commercial Use Only
7 inceptionv4 TensorFlow √ √ × Non-Commercial Use Only
8 inceptionv4 pruned0.2 TensorFlow √ √ √ Non-Commercial Use Only
9 inceptionv4 pruned0.4 TensorFlow √ √ √ Non-Commercial Use Only
10 mobilenetv1_0.25 TensorFlow √ √ × Non-Commercial Use Only
11 mobilenetv1_1.0 TensorFlow √ √ × Non-Commercial Use Only
12 mobilenetv1_1.0 pruned0.11 TensorFlow √ √ √ Non-Commercial Use Only
13 mobilenetv1_1.0 pruned0.12 TensorFlow √ √ √ Non-Commercial Use Only
14 mobilenetv2_1.0 TensorFlow √ √ × Non-Commercial Use Only
15 mobilenetv2_1.4 TensorFlow √ √ × Non-Commercial Use Only
16 resnetv1_50 TensorFlow √ √ × Non-Commercial Use Only
17 resnetv1_50 pruned0.38 TensorFlow √ √ √ Non-Commercial Use Only
18 resnetv1_50 pruned0.65 TensorFlow √ √ √ Non-Commercial Use Only
19 resnetv1_101 TensorFlow √ √ × Non-Commercial Use Only
20 resnetv1_101 pruned0.35 TensorFlow √ √ √ Non-Commercial Use Only
21 resnetv1_101 pruned0.57 TensorFlow √ √ √ Non-Commercial Use Only
22 resnetv1_152 TensorFlow √ √ × Non-Commercial Use Only
23 resnetv1_152 pruned0.51 TensorFlow √ √ √ Non-Commercial Use Only
24 resnetv1_152pruned0.60 TensorFlow √ √ √ Non-Commercial Use Only
25 vgg16 TensorFlow √ √ × Non-Commercial Use Only
26 vgg16 pruned0.43 TensorFlow √ √ √ Non-Commercial Use Only
27 vgg16 pruned0.50 TensorFlow √ √ √ Non-Commercial Use Only
28 vgg19 TensorFlow √ √ × Non-Commercial Use Only
29 vgg19 pruned0.24 TensorFlow √ √ √ Non-Commercial Use Only
30 vgg19 pruned0.39 TensorFlow √ √ √ Non-Commercial Use Only
31 resnetv2_50 TensorFlow √ √ × Non-Commercial Use Only
32 resnetv2_101 TensorFlow √ √ × Non-Commercial Use Only
33 resnetv2_152 TensorFlow √ √ × Non-Commercial Use Only
34 efficientnet-edgetpu-S TensorFlow √ √ × Non-Commercial Use Only
35 efficientnet-edgetpu-M TensorFlow √ √ × Non-Commercial Use Only
36 efficientnet-edgetpu-L TensorFlow √ √ × Non-Commercial Use Only
37 mlperf_resnet50 TensorFlow √ √ × Non-Commercial Use Only
38 resnet50 TensorFlow √ √ × Non-Commercial Use Only
39 mobilenetv1 TensorFlow √ √ × Non-Commercial Use Only
40 inceptionv3 TensorFlow √ √ × Non-Commercial Use Only
41 efficientnet-b0 TensorFlow √ √ × Non-Commercial Use Only
42 mobilenetv3 TensorFlow √ √ × Non-Commercial Use Only
43 efficientnet-lite TensorFlow √ √ × Non-Commercial Use Only
44 ViT TensorFlow √ √ × Non-Commercial Use Only
45 ssdmobilenetv1 TensorFlow √ √ ×
46 ssdmobilenetv2 TensorFlow √ √ ×
47 yolov3 TensorFlow √ √ ×
48 mlperf_resnet34 TensorFlow √ √ ×
49 efficientdet-d2 TensorFlow √ √ ×
50 yolov3 TensorFlow √ √ ×
51 yolov4_416 TensorFlow √ √ ×
52 yolov4_512 TensorFlow √ √ ×
53 RefineDet-Medical TensorFlow √ √ ×
54 RefineDet-Medical pruned0.50 TensorFlow √ √ √
55 RefineDet-Medical pruned0.75 TensorFlow √ √ √
56 RefineDet-Medical pruned0.85 TensorFlow √ √ √
57 RefineDet-Medical pruned0.88 TensorFlow √ √ √
58 bert-base TensorFlow √ √ ×
59 superpoint TensorFlow √ √ ×
60 HFNet TensorFlow √ √ ×
61 rcan TensorFlow √ √ ×
62 inceptionv3 PyTorch √ √ × Non-Commercial Use Only
63 inceptionv3 pruned0.3 PyTorch √ √ √ Non-Commercial Use Only
64 inceptionv3 pruned0.4 PyTorch √ √ √ Non-Commercial Use Only
65 inceptionv3 pruned0.5 PyTorch √ √ √ Non-Commercial Use Only
66 inceptionv3 pruned0.6 PyTorch √ √ √ Non-Commercial Use Only
67 squeezenet PyTorch √ √ × Non-Commercial Use Only
68 resnet50_v1.5 PyTorch √ √ × Non-Commercial Use Only
69 resnet50_v1.5 pruned0.3 PyTorch √ √ √ Non-Commercial Use Only
70 resnet50_v1.5 pruned0.4 PyTorch √ √ √ Non-Commercial Use Only
71 resnet50_v1.5 pruned0.5 PyTorch √ √ √ Non-Commercial Use Only
72 resnet50_v1.5 pruned0.6 PyTorch √ √ √ Non-Commercial Use Only
73 resnet50_v1.5 pruned0.7 PyTorch √ √ √ Non-Commercial Use Only
74 OFA-resnet50 PyTorch √ √ × Non-Commercial Use Only
75 OFA-resnet50 pruned0.45 PyTorch √ √ √ Non-Commercial Use Only
76 OFA-resnet50 pruned0.60 PyTorch √ √ √ Non-Commercial Use Only
77 OFA-resnet50 pruned0.74 PyTorch √ √ √ Non-Commercial Use Only
78 OFA-resnet50 pruned0.88 PyTorch √ √ √ Non-Commercial Use Only
79 OFA-depthwise-resnet50 PyTorch √ √ × Non-Commercial Use Only
80 vehicle type classification PyTorch √ √ ×
81 vehicle make classification PyTorch √ √ ×
82 vehicle color classification PyTorch √ √ ×
83 OFA-yolo PyTorch √ √ ×
84 OFA-yolo pruned0.3 PyTorch √ √ √
85 OFA-yolo pruned0.6 PyTorch √ √ √
86 yolox-nano PyTorch √ √ ×
87 yolov4csp PyTorch √ √ ×
88 yolov6m PyTorch √ √ ×
89 pointpillars PyTorch √ √ ×
90 HRNet PyTorch √ √ ×
91 3D-UNET PyTorch √ √ ×
92 bert-base PyTorch √ √ ×
93 bert-large PyTorch √ √ ×
94 bert-tiny PyTorch √ √ ×
95 face-mask-detection PyTorch √ √ ×
96 movenet PyTorch √ √ ×
97 fadnet PyTorch √ √ ×
98 fadnet pruned0.65 PyTorch √ √ √
99 fadnetv2 PyTorch √ √ ×
100 fadnetv2 pruned0.51 PyTorch √ √ √
101 psmnet pruned0.68 PyTorch √ √ √
102 SESR-S PyTorch √ √ ×
103 OFA-rcan PyTorch √ √ ×
104 xilinxSR PyTorch √ √ ×
105 yolov7 PyTorch √ √ ×
106 2D-UNET PyTorch √ √ ×

UIF 1.1 Models for ZenDNN

Click here to view details
# Model Original Platform Converted Format Datatype FP32 Datatype BF16 Datatype INT8 Pruned Reminder for limited use scope
1 RefineDet TensorFlow .PB √ × √ ×
2 RefineDet pruned0.5 TensorFlow .PB √ × √ √
3 RefineDet pruned0.75 TensorFlow .PB √ × √ √
4 RefineDet pruned0.85 TensorFlow .PB √ × √ √
5 RefineDet pruned0.88 TensorFlow .PB √ × √ √
6 EfficientNet-L TensorFlow .PB √ × √ × Non-Commercial Use Only
7 EfficientNet-M TensorFlow .PB √ × √ × Non-Commercial Use Only
8 EfficientNet-S TensorFlow .PB √ × √ × Non-Commercial Use Only
9 Resnetv1_50 TensorFlow .PB √ √ √ × Non-Commercial Use Only
10 Resnetv1_50 pruned0.38 TensorFlow .PB √ √ √ √ Non-Commercial Use Only
11 Resnetv1_50 pruned0.65 TensorFlow .PB √ √ √ √ Non-Commercial Use Only
12 Resnetv1_101 TensorFlow .PB √ × √ × Non-Commercial Use Only
13 Resnetv1_101 pruned0.34 TensorFlow .PB √ × √ √ Non-Commercial Use Only
14 Resnetv1_101 pruned0.57 TensorFlow .PB √ × √ √ Non-Commercial Use Only
15 Resnetv1_152 TensorFlow .PB √ × √ × Non-Commercial Use Only
16 Resnetv1_152 pruned0.51 TensorFlow .PB √ × √ √ Non-Commercial Use Only
17 Resnetv1_152 pruned0.60 TensorFlow .PB √ × √ √ Non-Commercial Use Only
18 Inceptionv1 TensorFlow .PB √ × √ × Non-Commercial Use Only
19 Inceptionv1 pruned0.09 TensorFlow .PB √ × √ √ Non-Commercial Use Only
20 Inceptionv1 pruned0.16 TensorFlow .PB √ × √ √ Non-Commercial Use Only
21 Inceptionv3 TensorFlow .PB √ × √ × Non-Commercial Use Only
22 Inceptionv3 pruned0.2 TensorFlow .PB √ × √ √ Non-Commercial Use Only
23 Inceptionv3 pruned0.4 TensorFlow .PB √ × √ √ Non-Commercial Use Only
24 Inceptionv4 TensorFlow .PB √ × √ × Non-Commercial Use Only
25 Inceptionv4 pruned0.2 TensorFlow .PB √ × √ √ Non-Commercial Use Only
26 Inceptionv4 pruned0.4 TensorFlow .PB √ × √ √ Non-Commercial Use Only
27 Mobilenetv1 TensorFlow .PB √ × √ × Non-Commercial Use Only
28 Mobilenetv1 pruned0.11 TensorFlow .PB √ × √ √ Non-Commercial Use Only
29 Mobilenetv1 pruned0.12 TensorFlow .PB √ × √ √ Non-Commercial Use Only
30 VGG16 TensorFlow .PB √ √ √ × Non-Commercial Use Only
31 VGG16 pruned0.43 TensorFlow .PB √ √ √ √ Non-Commercial Use Only
32 VGG16 pruned0.50 TensorFlow .PB √ √ √ √ Non-Commercial Use Only
33 VGG19 TensorFlow .PB √ × √ × Non-Commercial Use Only
34 VGG19 pruned0.24 TensorFlow .PB √ × √ √ Non-Commercial Use Only
35 VGG19 pruned0.39 TensorFlow .PB √ × √ √ Non-Commercial Use Only
36 Bert-base TensorFlow .PB √ × × ×
37 Albert-base TensorFlow .PB √ × × ×
38 Albert-large TensorFlow .PB √ × × ×
39 Mobilebert TensorFlow .PB √ × × ×
40 Resnet50_v1.5 PyTorch .PT √ √ √ × Non-Commercial Use Only
41 Resnet50_v1.5 pruned0.3 PyTorch .PT √ √ √ √ Non-Commercial Use Only
42 Resnet50_v1.5 pruned0.4 PyTorch .PT √ √ √ √ Non-Commercial Use Only
43 Resnet50_v1.5 pruned0.5 PyTorch .PT √ √ √ √ Non-Commercial Use Only
44 Resnet50_v1.5 pruned0.6 PyTorch .PT √ √ √ √ Non-Commercial Use Only
45 Resnet50_v1.5 pruned0.7 PyTorch .PT √ √ √ √ Non-Commercial Use Only
46 Inceptionv3 PyTorch .PT √ × √ × Non-Commercial Use Only
47 Inceptionv3 pruned0.3 PyTorch .PT √ × √ √ Non-Commercial Use Only
48 Inceptionv3 pruned0.4 PyTorch .PT √ × √ √ Non-Commercial Use Only
49 Inceptionv3 pruned0.5 PyTorch .PT √ × √ √ Non-Commercial Use Only
50 Inceptionv3 pruned0.6 PyTorch .PT √ × √ √ Non-Commercial Use Only
51 Reid_resnet50 PyTorch .PT √ × √ × Non-Commercial Use Only
52 Reid_resnet50 pruned0.4 PyTorch .PT √ × √ √ Non-Commercial Use Only
53 Reid_resnet50 pruned0.5 PyTorch .PT √ × √ √ Non-Commercial Use Only
54 Reid_resnet50 pruned0.6 PyTorch .PT √ × √ √ Non-Commercial Use Only
55 Reid_resnet50 pruned0.7 PyTorch .PT √ × √ √ Non-Commercial Use Only
56 OFA_resnet50 PyTorch .PT √ × √ × Non-Commercial Use Only
57 OFA_resnet50 pruned0.45 PyTorch .PT √ × √ √ Non-Commercial Use Only
58 OFA_resnet50 pruned0.60 PyTorch .PT √ × √ √ Non-Commercial Use Only
59 OFA_resnet50 pruned0.74 PyTorch .PT √ × √ √ Non-Commercial Use Only
60 OFA_resnet50 pruned0.88 PyTorch .PT √ × √ √ Non-Commercial Use Only
61 3D-Unet PyTorch .PT √ × × ×
62 3D-Unet pruned0.1 PyTorch .PT √ × × √
63 3D-Unet pruned0.2 PyTorch .PT √ × × √
64 3D-Unet pruned0.3 PyTorch .PT √ × × √
65 3D-Unet pruned0.4 PyTorch .PT √ × × √
66 3D-Unet pruned0.5 PyTorch .PT √ × × √
67 DLRM PyTorch .PT √ × √ × Non-Commercial Use Only
68 GPT-2 small PyTorch .PT √ × × ×
69 DistilGPT-2 PyTorch .PT √ × × ×
70 Albert-base PyTorch .PT √ × × ×
71 Albert-large PyTorch .PT √ × × ×
72 Mobilebert PyTorch .PT √ × × ×
73 Resnetv1_50 ONNXRT .ONNX √ × √ × Non-Commercial Use Only
74 Resnetv1_50 pruned0.38 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only
75 Resnetv1_50 pruned0.65 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only
76 Inceptionv3 ONNXRT .ONNX √ × √ × Non-Commercial Use Only
77 Inceptionv3 pruned0.2 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only
78 Inceptionv3 pruned0.4 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only
79 Mobilenetv1 ONNXRT .ONNX √ × √ × Non-Commercial Use Only
80 Mobilenetv1 pruned0.11 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only
81 Mobilenetv1 pruned0.12 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only
82 VGG16 ONNXRT .ONNX √ × √ × Non-Commercial Use Only
83 VGG16 pruned0.43 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only
84 VGG16 pruned0.50 ONNXRT .ONNX √ × √ √ Non-Commercial Use Only

2.1.1: Standard Naming Rules

Model name: F_M_(P)_V_Z

  • F specifies the training framework: tf is TensorFlow 1.x, tf2 is TensorFlow 2.x, pt is PyTorch, onnx is ONNXRT.
  • M specifies the model.
  • P specifies the pruning ratio, meaning how much computation is reduced. It is optional depending on whether the model is pruned or not.
  • V specifies the version of UIF.
  • Z specifies the version of ZenDNN or MIGraphX.

For example, pt_resnet50v1.5_pruned0.6_1.2_M2.6 is a resnet50 v1.5 model trained with PyTorch, 60% pruned, the UIF release version is 1.2, and the MIGraphX version is 2.6.

2.1.2: Model List

Visit model-list The models are named according to standard naming rules.

You can get a download link and MD5 checksum of all released models running on different hardware platforms from here. You can download it manually or use the the following automatic download script to search for models by keyword.

2.1.3: Once-For-All (OFA): Efficient Model Customization for Various Platforms

Once-For-All (OFA) is an efficient neural architecture search (NAS) method that enables the customization of sub-networks for various hardware platforms. It decouples training and search, enabling quick derivation of specialized models optimized for specific platforms with efficient inference performance. OFA offers the following benefits:

  1. OFA requires only one training process and can specialize in diverse hardware platforms (for example, DPU, GPU, CPU, IPU) without incurring the heavy computation costs associated with manual design or conventional RL-based NAS methods.

  2. By decoupling supernet training and subnetwork searching, OFA optimizes networks with actual latency constraints of a specific hardware platform, avoiding discrepancies between estimated and actual latencies.

  3. OFA offers strong flexibility and scalability regarding the search space, supporting various network architectures (for example, CNN and Transformer) with different granularities (for example, operation-wise, layer-wise, block-wise) to cater to different tasks (for example, CV, NLP).

  4. Extensive experiments with OFA demonstrate consistent performance acceleration while maintaining similar accuracy levels compared to baselines on different devices. For instance, OFA-ResNet achieves a speedup ratio of 69% on MI100, 78% on MI210, and 97% on Navi2 compared to ResNet50 baselines with a batch size of 1 and prune ratio of 78%.

  5. OFA has the potential to search for optimized large language models (LLM) by stitching pretrained models for accuracy-efficiency trade-offs. A proof of concept for Vision Transformer (ViT) shows that OFA can derive an optimized ViT with a 20% to 40% speedup ratio compared to ViT-base on MI100.

2.1.3.1: OFA-ResNet for AMD GPUs with MIGraphX

  • Task Description: This case aims to search for ResNet-like models optimized for AMD GPUs with MIGraphX. The latency on MI100 is used as constraints for optimization. Channel numbers are aligned with GPU capabilities.
  • Search Space Design: The search space has the following parameters:
    • Stage output ratio (The ratio of output channel for each stage of the model): [0.65, 0.8, 1.0]
    • Depth (The number of blocks for each stage): [0, 1, 2]
    • Expand ratio (The ratio of output channel for each block in each stage): [0.2, 0.25, 0.35]
    • Resolution (The input size of the model): [128, 160, 192, 224]
  • Results: The OFA-ResNet model achieves a speedup ratio of 69% on MI100, 78% on MI210, 97% on Navi2 compared to ResNet50 baseline (with batch size of 1 and pruned ratio of 78%) while maintaining similar accuracy.

2.1.3.2: Performance Comparison

Model Float Accuracy (ImageNet 1K) FLOPs (G) Pruned Ratio Speedup ratio on MI100 Speedup ratio on MI210 Speedup ratio on Navi2
ResNet50 76.1% 8.2 0% - - -
OFA-ResNet 75.8% 1.77 78% 1.69x 1.78x 1.97x

2.2: Get MIGraphX Models from UIF Model Zoo

Perform the following steps to install UIF 1.2 models:

  1. Set up and run the model downloader tool.

    git clone https://github.com/AMD/uif.git
cd uif/docs/2_model_setup
python3 downloader.py

    It has the following provision to specify the frameworks:

    Tip:
You need to input framework and model name. Use space divide such as tf vgg16
tf:tensorflow1.x  tf2:tensorflow2.x  onnx:onnxruntime  dk:darknet  pt:pytorch  all: list all model
input:

  2. Download UIF 1.2 GPU models. Provide pt as input to get the list of models.

    input:pt
chose model
0 : all
1 : pt_inceptionv3_1.2_M2.6
2 : pt_bert_base_1.2_M2.6
3 : pt_bert_large_1.2_M2.6


...
input num:

The models with 1.2 as suffix are UIF 1.2 models. MI100 means the model has been tuned for MI-100 GPU, and MI210 indicates the model has been tuned for MI-210 GPU. Without either of these suffixes, the model is a model version that should be used for training, including the GPU example later in this section.

```
input num:1
chose model type
0: all
1: GPU
2: MI100
3: MI200
...
...
input num:

```

Provide 1 as input to download the GPU model.

```
input num:1
pt_inceptionv3_1.2_M2.6.zip
                                              100.0%|100%
done
```

2.3: Set Up MIGraphX YModel

YModel is designed to provide significant inference performance through MIGraphX. Prior to the introduction of YModel, performance tuning was conditioned by having tuned kernel configs stored in a /home local User DB. If users move their model to a different server or allow a different user to use it, they would have to run through the MIOpen tuning process again to populate the next User DB with the best kernel configs and corresponding solvers. Tuning is time consuming, and if the users have not performed tuning, they would see discrepancies between expected or claimed inference performance and the actual inference performance. This leads to repetitive and time-consuming tuning tasks for each user.

MIGraphX introduces a feature known as YModel that stores the kernel config parameters found during tuning in the same file as the model itself. This ensures the same level of expected performance even when a model is copied to a different user or system.

Note: The YModel feature is available starting with the ROCm™ v5.4.1 and UIF v1.1 releases. For more information on ROCm and the YModel feature, see https://rocm.docs.amd.com/en/latest/examples/machine_learning/migraphx_optimization.html#ymodel.

Note: YModel does not support MIOpen fusions and must be disabled while generating YModel.

To set up the YModel for GPUs:

  1. Tune the kernels for your architecture with MIOPEN_FIND_ENFORCE=3 migraphx-driver run <file.onnx>.

  2. Build a *.mxr file with a compile option:

	migraphx-driver compile file.onnx --enable-offload-copy --gpu --binary -o file.mxr
  1. Run the (Python) program with a command line:
	model=migraphx.load("file.mxr")

Note: If the Model Zoo contains a *.mxr file for your architecture, you can skip steps 1 and 2.

2.4: Get Vitis AI Models from UIF Model Zoo

Follow the instructions in the Vitis AI Model Zoo page.

2.5: Get ZenDNN Models from UIF Model Zoo

Perform the following steps to install UIF 1.1 models:

  1. Set up and run the model downloader tool.

    git clone https://github.com/AMD/uif.git
cd uif/docs/2_model_setup
python3 downloader.py

    It has the following provision to specify the frameworks:

    Tip:
You need to input framework and model name, use space divide such as tf vgg16
tf:tensorflow1.x  tf2:tensorflow2.x  onnx:onnxruntime  dk:darknet  pt:pytorch  all: list all model
input:

  2. Download UIF 1.1 TensorFlow+ZenDNN models. Provide tf as input to get the list of models.

    input:tf
chose model
0 : all
1 : tf_refinedet_edd_320_320_81.28G_1.1_Z4.0
2 : tf_refinedet_edd_320_320_0.5_41.42G_1.1_Z4.0
3 : tf_refinedet_edd_320_320_0.75_20.54G_1.1_Z4.0
4 : tf_refinedet_edd_320_320_0.85_12.32G_1.1_Z4.0
...
...
input num:

    The models with 1.1 as suffix are UIF 1.1 models. Z4.0 means ZenDNN version is 4.0. Provide the model number corresponding to the required model as input. For example, select 1 for a RefineDet baseline model.

    input num:1
chose model type
0: all
1 : CPU

...
...
input num:

    Provide 1 as input to download the ZenDNN CPU model.

    input num:1
tf_refinedet_edd_320_320_81.28G_1.1_Z4.0.zip
                                              100.0%|100%
done

  3. Download UIF 1.1 PyTorch+ZenDNN models. Provide pt as input to get the list of models.

    input:pt
chose model
0 : all
1 : pt_inceptionv3_imagenet_299_299_11.4G_1.1_M2.4
2 : pt_3dunet_kits19_128_128_128_0.3_763.8G_1.1_Z4.0
3 : pt_bert_base_SQuADv1.1_384_70.66G_1.1_M2.4
4 : pt_bert_large_SQuADv1.1_384_246.42G_1.1_M2.4
5 : pt_personreid-res50_market1501_256_128_0.4_3.3G_1.1_Z4.0
...
...
input num:

    Models with 1.1 as suffix are UIF 1.1 models. Provide the model number corresponding to the required model as input. For example, select 3 for resnet50 pruned 40% model.

    input num:2
chose model type
0: all
1 : CPU

...
...
input num:

    Provide 1 as input to download the UIF 1.1 ZenDNN CPU model.

    input num:1
pt_3dunet_kits19_128_128_128_0.3_763.8G_1.1_Z4.0.zip
                                              100.0%|100%
done

  4. Download UIF 1.1 ONNXRT+ZenDNN models. Provide onnx as input to get the list of models.

    input:onnx
chose model
0 : all
1 : onnx_vgg16_imagenet_224_224_30.96G_1.1_Z4.0
2 : onnx_inceptionv3_imagenet_299_299_0.4_6.9G_1.1_Z4.0
3 : onnx_resnetv1_50_imagenet_224_224_0.38_4.3G_1.1_Z4.0
4 : onnx_mobilenetv1_1.0_imagenet_224_224_1.14G_1.1_Z4.0
...
...
input num:

    Models with 1.1 as suffix are UIF 1.1 models. Provide the model number corresponding to the required model as input. For example, select 3 for resnet50 pruned 40% model.

    input num:3
chose model type
0: all
1 : CPU

...
...
input num:

    Provide 1 as input to download the UIF 1.1 ZenDNN CPU model.

    input num:1
onnx_resnetv1_50_imagenet_224_224_0.38_4.3G_1.1_Z4.0.zip
                                              100.0%|100%
done

< Previous | Next >

License

UIF is licensed under Apache License Version 2.0. Refer to the LICENSE file for the full license text and copyright notice.

Technical Support

Contact [email protected] for questions, issues, and feedback on UIF.

Submit your questions, feature requests, and bug reports on the GitHub issues page.