eepdemo_arm is the FREE-TPU's demo application.
./eepdemo_arm -h
Usage:
./eepdemo_arm [-options]
where options include:
--help(-h) # print this help message
--bin <path> # eep tpu bin file path.
--image <path> # image jpg path or folder path.
--label <label> # label or label file path.
--crop <mode> # crop mode; 0: just resize; 1: 1-crop
(1-crop mode: for classify neural networks)
--prefetch <flag> # images read mode. if set 1, will prefetch all images before test.
--net <net name> # if no mean/norm, set 'default'.
Free-TPU currently supports: 'lenet', 'mobilenet'
'squeezenet', 'resnet50', 'inception', 'myolo3', 'icnet'
--version(-v) # show version.
Crop mode: for classify neural networks. If set to 1, will do crop. The crop method is: <1> set a default crop scale(<=1.0); <2> find the short edge value of image size; <3> short edge value multiple the crop scale, as the new short edge value; <4> crop the center part of image by short_edge * short_edge square; <5> proportionally resize the square crop image to the input size. Suppose the input height is equal to width. The default crop scale is: if input width < 256, crop_scale = width / 256.0; else crop_scale = 1.0.
Root privileges: Need.
The '--net ' option currently supports the following networks: (more will update)
lenet: Lenet-5
mobilenet: Mobilenet_v1 / Mobilenet_v2
squeezenet: Squeezenet_v1.1
resnet50: Resnet-50
inception: Inception_v3
myolo3: Mobilenet_yolo_v3
icnet: ICNET
Free TPU supports 2 types of bin files:
-
freetpu_c8a1: For hardware boards which have 512MB PS memory.
-
freetpu_c8a3: For hardward boards which have >=1GB PS memory.
There are only a few Free-TPU bin files (lenet / mobilenet_v1) on github for test.
You can download ALL Free-TPU bin files from BaiDu Cloud (Extraction code: w462 ; Updated on March 1, 2019).
Suppose the demo have 3 folders: " images, tpubin, eepdemo ", and we are in eepdemo folder in command line. Take "freetpu_c8a1" for example, you could change the bin file according to your board.
- Test 1 image:
sudo ./eepdemo_arm --bin ../tpubin/freetpu_c8a1/mobilenet1/eeptpu.bin --image ../images/classify/ILSVRC2012_val_00000013.JPEG --label 370 --net mobilenet --crop 1
- Test all images in folder:
sudo ./eepdemo_arm --bin ../tpubin/freetpu_c8a1/mobilenet1/eeptpu.bin --image ../images/classify/ --label ../images/classify/val.txt --net mobilenet --crop 1
sudo ./eepdemo_arm --prefetch 1 --bin ../tpubin/freetpu_c8a1/mobilenet1/eeptpu.bin --image ../images/classify/ --label ../images/classify/val.txt --net mobilenet --crop 1
Note: The demo application can pre-read all images in the folder by using '--prefetch 1', but do not place too many images in the folder if using prefetch.
Free-TPU demo application runs in Linux(ARM-v7 architecture), compatible with raspberry pi file system.
| Caffe ( Top1 / Top5 ) | Free-TPU ( Top1 / Top5 ) | Free-TPU Vs. Caffe ( Top1 / Top5 ) | |
|---|---|---|---|
| Lenet5 | 99.06% / 100% | 99.01% / 100% | -0.05% / 0.00% |
| Mobilenet_v1 | 70.74% / 89.84% | 70.76% / 89.81% | +0.02% / -0.03% |
| Mobilenet_v2 | 71.82% / 90.45% | 71.84% / 90.41% | +0.02% / -0.04% |
| Squeezenet_v1.1 | 56.50% / 79.67% | 56.41% / 79.63% | -0.09% / -0.04% |
| ResNet-50 | 75.15% / 92.19% | 75.16% / 92.15% | +0.01% / -0.04% |
| Inception_v3 | 78.38% / 94.11% | 78.38% / 94.11% | 0.00% / 0.00% |