Anton Lokhmotov

"ck compile program:caffe-time-opencl" fails

9

``` g++-5 -c -I../ -DANDROID_USE_OPENMP=ON -DBLAS=Open -DCK_HOST_OS_NAME2_LINUX=1 -DCK_HOST_OS_NAME_LINUX=1 -DCK_TARGET_OS_NAME2_LINUX=1 -DCK_TARGET_OS_NAME_LINUX=1 -DUSE_GREENTEA=ON -DUSE_LMDB=OFF -DUSE_OPENCV=ON -DXOPENME=ON -I/home/anton/CK_TOOLS/lib-caffe-bvlc-opencl-clblast-trunk-gcc-5.4.0-linux-64/install/include -I/home/anton/CK_TOOLS/lib-gflags-2.2.0-gcc-5.4.0-linux-64/install/include -I/home/anton/CK_TOOLS/lib-glog-development-gcc-5.4.0-linux-64/install/include -I/home/anton/CK_TOOLS/lib-openblas-0.2.19-gcc-5.4.0-linux-64/install/include -I/home/anton/CK_TOOLS/lib-protobuf-host-3.1.0-linux-64/install/include -I/usr/include -I/usr/include -I/home/anton/usr-local/cuda-8.0.61/include -I/usr/include -I/home/anton/CK_TOOLS/lib-viennacl-src-dvdt-master-linux-64/src -I/home/anton/CK_TOOLS/lib-rtl-xopenme-0.3-gcc-5.4.0-linux-64/include -I/home/anton/CK_TOOLS/lib-caffe-bvlc-opencl-clblast-trunk-gcc-5.4.0-linux-64/install/.build_release/src ../caffe.cpp -o caffe.o In...

Implement imagenet-val command and/or caffe-test program

1

The [CK-TensorRT](https://github.com/dividiti/ck-tensorrt) project has a program called [`tensorrt-time`](https://github.com/dividiti/ck-tensorrt/tree/master/program/tensorrt-time). Effectively, it's a classifier that for a given image outputs a vector of probabilities for it to belong to each of the...

Add support for VGG

3

The [Visual Geometry Group](http://www.robots.ox.ac.uk/~vgg/research/very_deep/) at the University of Oxford designed several [highly accurate DNNs](https://arxiv.org/abs/1409.1556). Add CK-Caffe support for: - the VGG16 model ([weights](http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_16_layers.caffemodel), [topology](https://gist.githubusercontent.com/ksimonyan/211839e770f7b538e2d8/raw/0067c9b32f60362c74f4c445a080beed06b07eb3/VGG_ILSVRC_16_layers_deploy.prototxt)) - the VGG19 model ([weights](http://www.robots.ox.ac.uk/~vgg/software/very_deep/caffe/VGG_ILSVRC_19_layers.caffemodel), [topology](https://gist.githubusercontent.com/ksimonyan/3785162f95cd2d5fee77/raw/f02f8769e64494bcd3d7e97d5d747ac275825721/VGG_ILSVRC_19_layers_deploy.prototxt))...

caffe-classification-opencl mispredicts and segfaults on 32-bit Android

11

When running on the T-Firefly board (Android 5.1.1, 32-bit, Mali-T760 w/ driver v6.0): ``` I0306 23:35:31.354465 6560 caffe.cpp:678] Start with command: device_query I0306 23:35:31.356835 6560 common.cpp:379] Total devices: 1 I0306...

"ck compile program:caffe-time --target_os=android21-arm" fails with Android NDK GCC compiler

4

``` $ ck compile program:caffe-time --target_os=android21-arm In file included from /home/anton/CK_TOOLS/lib-opencv-development-android-ndk-4.9.x-android21-arm-32/src/3rdparty/libwebp/cpu-features/cpu-features.c:61:0:android-ndk-r13b/platforms/android-21/arch-arm/usr/include/machine/cpu-features.h:52:6: error: #error Unknown or unsupported ARM architecture # error Unknown or unsupported ARM architecture ^ make[2]: *** [3rdparty/libwebp/CMakeFiles/libwebp.dir/cpu-features/cpu-features.c.o] Error...

"ck compile program:caffe-time" fails with Android NDK LLVM compiler

2

``` $ ck compile program:caffe-time --target_os=android21-arm64 *************************************************************************************** Current directory: /home/anton/CK_REPOS/ck-caffe/program/caffe-time/tmp *************************************************************************************** ----------------------------------- Resolving software dependencies ... *** Dependency 1 = compiler (C++ compiler): More than one environment found for "C++...

[Quantization] MatMul result precision

4

When inspecting a quantized graph for a simple FC network, I was a bit surprised that the result of MatMul nodes was coerced into i8, possibly resulting in a huge...

Odroid N2+ @ 2.2 GHz slower than Raspberry Pi 4 @ 1.5 GHz

5

I've noticed a curious thing when [benchmarking ResNet50](https://github.com/krai/ck-mlperf/blob/master/program/image-classification-armnn-tflite-loadgen/README.singlestream.md#all-in-one) via ArmNN v21.11 with the Neon backend on Odroid N2+ @ 2208 MHz and Raspberry Pi 4 @ 1500 MHz. Despite the...

[URGENT] RetinaNet mAP is 37.50% on main, 37.55% on r2.1 branch

10

According to the [README in the main branch](https://github.com/mlcommons/inference/tree/master/vision/classification_and_detection), the reference fp32 accuracy of RetinaNet is `37.5%` or `37.50%`. With this refence value, the 99% threshold for valid submissions is `37.50%...