search

Xilinx / Vitis-AI

Vitis AI is Xilinx’s development stack for AI inference on Xilinx hardware platforms, including both edge devices and Alveo cards.
https://www.xilinx.com/ai
Apache License 2.0
1.49k stars 633 forks source link

[UNILOG][FATAL][XCOM_ACGEN_ERROR][instuction generating fail, please contact us.] #726

Open JaviMota opened 2 years ago

JaviMota commented 2 years ago

Hi

Im trying to compile an own model following the 05-Keras_FCN8_UNET_segmentation tutorial using the FCN8 network and I get this error: [UNILOG][FATAL][XCOM_ACGEN_ERROR][instuction generating fail, please contact us.] I only changed the dataset by an own one and the number of classes to 2. The neural network is the same you propose. The weird thing is it has compiled one of several attempts without changing anything. Im using Vitis-ai 2.0 in vitis-ai-tensorflow environment. I have compiled the tutorial with the original classes and it works. Do you know what can be the problem?

Thank you.

wanghong4compiler commented 2 years ago

Hi @JaviMota , which version and which hardware architecture did you use? And, could you share the xmodel used for compiler? So that I can reproduce this issue. Thank you.

JaviMota commented 2 years ago

Hi Vitis-AI Docker version: 2.0.0.1103 Architecture: DPUCZDX8G_ISA0_B4096_MAX_BG2 Environment: conda activate vitis-ai-tensorflow

I share with you the quantized model as Im not able to get the xmodel because of the error. https://drive.google.com/file/d/1wIl82XzjjTUxkMAMmm37vjdLHnLJKclA/view?usp=sharing

Thank you

JaviMota commented 2 years ago

Hi To add a bit more info I have tried to run the 05-Keras_FCN8_UNET_segmentation tutorial using FCN8 with several number of classes and this is what I got:

2,3 classes --> Error in the compilation 4 classes --> After a while, error in the compilation 5 to 12 classes --> It works fine

Could be a bug or Am I doing something wrong?

Thank you.

JaviMota commented 2 years ago

Hi Any news on this topic? have you tried to compile the model? Thank you.

ZiYinHong commented 2 years ago

hi, have you sloved the problem?

wanghong4compiler commented 2 years ago

Hi @ZiYinHong , This issue has been fixed in VAI2.5, could you update the toolchain and try this again?

JaviMota commented 2 years ago

Hi

I have tested it with the latest version of Vitis-AI 2.5 and Im still having the issue. I used to test the https://github.com/Xilinx/Vitis-AI-Tutorials/tree/1.4/Design_Tutorials/05-Keras_FCN8_UNET_segmentation tutorial and I only changed the fcn_config.py to this:

############################################################################### global variables ###############################################################################

Size of images WIDTH = 224 HEIGHT = 224

normalization factor NORM_FACTOR = 127.5

number of classes NUM_CLASSES = 2

names of classes CLASS_NAMES = ("Sky", "Wall")

BATCH_SIZE = 32 EPOCHS = 5

Im also using the DPUCZDX8G_ISA1_B512 architecture. The file Im running is run_fcn8.sh

Regards.