jiri-kula / efficientdet

My implementation of EfficientDet.

Python

Use v3.8 wget https://www.python.org/ftp/python/3.9.11/Python-3.9.11.tar.xz xz -d Python-3.9.11.tar.xz tar -xvf Python-3.9.11.tar cd Python-3.9.11 sudo apt install
libffi-dev
libgl1
libsqlite3-dev
libssl-dev
-y ./configure make -j12 sudo make altinstall

EfficientDet

Repository with my implementation of EfficientDet.

CUDNN_PATH=$(dirname $(python -c "import nvidia.cudnn;print(nvidia.cudnn.file)")) CUPTI_PATH=/usr/local/cuda-12.2/targets/x86_64-linux/lib/ export LD_LIBRARY_PATH=$CUDNN_PATH/lib:$CUPTI_PATH:$LD_LIBRARY_PATH python3 -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"

Ubuntu

sudo sh cuda_12.2.2_535.104.05_linux.run

Driver: Not Selected Toolkit: Installed in /usr/local/cuda-12.2/

Please make sure that

• PATH includes /usr/local/cuda-12.2/bin
• LD_LIBRARY_PATH includes /usr/local/cuda-12.2/lib64, or, add /usr/local/cuda-12.2/lib64 to /etc/ld.so.conf and run ldconfig as root

To uninstall the CUDA Toolkit, run cuda-uninstaller in /usr/local/cuda-12.2/bin ***WARNING: Incomplete installation! This installation did not install the CUDA Driver. A driver of version at least 535.00 is required for CUDA 12.2 functionality to work. To install the driver using this installer, run the following command, replacing with the name of this run file: sudo .run --silent --driver

Training checklist

1. Check your data annotation datacheck.py
2. Measure how to set anchor boxes anchor_histogram.py, modify model/anchors.py at aspects and areas.

CustomOp

Object detection z tflite model maker: síť má na konci vrstvu TFLite_Detection_PostProcess, která je build-in v tflite. Podívej se na tensorflow/tensorflow/lite/kernels/register.cc.

TFLite_Detection_PostProcess je zřejmně kompatibilní s EdgeTPU compiler, ale neobsahuje extrakci sin/cosine, kterou potřebuji pro rotaci blade v Edwards.

Sít je typu SSD. Na výstupu má surové [x1, x2, y1, y2, sin, cos, class logits], kde x, y, jsou offset a scale Anchorů a class logits je také třeba dále zpravovat na třídu a score.

Existují dvě strategie:

1. Add post processing layer like TFLite_Detection_PostProcess, but it lacks angle. Adding custom-op is not supported in edgetpu_compiler resulting to error

Edge TPU Compiler version 16.0.384591198
Started a compilation timeout timer of 180 seconds.
ERROR: Encountered unresolved custom op: Atan.
ERROR: Node number 1 (Atan) failed to prepare.

Compilation failed: Model failed in Tflite interpreter. Please ensure model can be loaded/run in Tflite interpreter.
Compilation child process completed within timeout period.
Compilation failed! 

2. Do post processing in custom c++ code as we did that for palm and landmark detection. Anchors need to be generated in c++ and kept in sync with python training code manually.

Reference

Better rotation representations for accurate pose estimation

Inspiration

By tflite model maker /home/jiri/custom_model_maker/examples/tensorflow_examples/lite/model_maker/pip_package/src/tensorflow_examples/lite/model_maker/third_party/efficientdet/keras/postprocess.py

See comments in def postprocess_tflite(params, cls_outputs, box_outputs):

https://keras.io/examples/vision/retinanet/

Rotation

/home/jiri/EfficientDet/model/rotation.py in line 7 40 # %% 41 # see 42 # [1] https://www.tensorflow.org/api_docs/python/tf/linalg/matmul 43 # [2] https://math.stackexchange.com/questions/744736/rotation-matrix-to-axis-angle 45 a = 10.0 / 180.0 * 3.14 ----> 46 R = Rx(0.0) @ Ry(0.0) @ Rz(0.0) # if label rotation 47 Q = Rx(0.0) @ Ry(0.0) @ Rz(a) # and predicted rotation differ by angle a 49 QT = tf.transpose(Q)

Mount windows partition into Ubuntu

mkdir -p ~/winpart

jiri@jiri-81Y6:~/EfficientDet$ lsblk
NAME        MAJ:MIN RM   SIZE RO TYPE MOUNTPOINTS
loop0         7:0    0     4K  1 loop /snap/bare/5
loop1         7:1    0  55.7M  1 loop /snap/core18/2785
loop2         7:2    0  55.7M  1 loop /snap/core18/2790
loop3         7:3    0  63.4M  1 loop /snap/core20/1974
loop4         7:4    0  63.5M  1 loop /snap/core20/2015
loop5         7:5    0  73.9M  1 loop /snap/core22/858
loop6         7:6    0  73.9M  1 loop /snap/core22/864
loop7         7:7    0 240.5M  1 loop /snap/firefox/3206
loop8         7:8    0 238.8M  1 loop /snap/firefox/3252
loop9         7:9    0 218.4M  1 loop /snap/gnome-3-34-1804/93
loop10        7:10   0 485.5M  1 loop /snap/gnome-42-2204/126
loop11        7:11   0   497M  1 loop /snap/gnome-42-2204/141
loop12        7:12   0  91.7M  1 loop /snap/gtk-common-themes/1535
loop13        7:13   0 115.7M  1 loop /snap/slack/105
loop14        7:14   0 113.3M  1 loop /snap/slack/89
loop15        7:15   0  12.3M  1 loop /snap/snap-store/959
loop16        7:16   0  40.8M  1 loop /snap/snapd/20092
loop17        7:17   0  40.9M  1 loop /snap/snapd/20290
loop18        7:18   0   452K  1 loop /snap/snapd-desktop-integration/83
loop19        7:19   0 320.4M  1 loop /snap/vlc/3078
nvme1n1     259:0    0   1.8T  0 disk 
├─nvme1n1p1 259:1    0   100M  0 part 
├─nvme1n1p2 259:2    0    16M  0 part 
├─nvme1n1p3 259:3    0   1.8T  0 part <<=== for example this big drive on Legion5 of (jiri)
└─nvme1n1p4 259:4    0   530M  0 part 
nvme0n1     259:5    0 465.8G  0 disk 
├─nvme0n1p1 259:6    0 209.8G  0 part 
├─nvme0n1p2 259:7    0     1G  0 part /boot/efi
└─nvme0n1p3 259:8    0   255G  0 part /var/snap/firefox/common/host-hunspell

1. read-only access

sudo mount -t ntfs -o ro /dev/nvme1n1p3 ~/winpart

2. read-write access

sudo mount -t ntfs /dev/nvme1n1p3 ~/winpart

Known issues

• Access is denied because the NTFS volume is already exclusively opened use lsblk to check if the partition is already mounted, if so use

sudo umount /media/jiri/D6667DDE667DBFB3 (or whatever after /jiri/...)