blueoil
blueoil copied to clipboard
blue-oil
Bring Deep Learning to small devices
Currently, we supports only `DivideBy255` preprocessor. However, internally, we've prepared `PerImageStandardization`. @yasumura-lm san mentioned that sometimes this preprocessing method is is very effective. Hence, we want to support `PerImageStandardization`.
This is a new feature to enable FPGA with 64bit ARM processor. Now FPGA available only with 32bit ARM processor and we want to add 64 bit as well.
I converted a model from Blueoil to run on de10 nano. I want to measure the memory usage so I edit the code to print some value and trying tome...
When running different inferences at the same time, FPGA will be stuck and inferences are stopped. And after that, I cannot run any inference which needs FPGA, these will be...
lm_fpga.elf doesn't work well with LMBiSeNetQuantize. The program gets stucked (freezed) when we run it on FPGA. The error point is: ``` Conv2D_struct.input_height = 1; Conv2D_struct.input_width = 1; Conv2D_struct.kernel_height =...
According to the performance measurement of LmResnet for ImageNet on FPGA, there is some low hanging fruits, `Add` operator is now consuming over 20ms.
There is a resize process in pre_process, but the default interpolation method differs depending on the Pillow version, so it seems to affect learning. https://github.com/blue-oil/blueoil/blob/master/lmnet/lmnet/pre_processor.py#L38 Pillow v7.0.0 https://github.com/python-pillow/Pillow/blob/7.0.0/src/PIL/Image.py#L1786 Pillow v6.2.2...
Now this repository has many not active issues and PRs. They will blind us to find out which issue should be solved first. Those with no progress long time like...
MappedMem class (blueoil/converter/templates/include/memdriver.h) and DMA_Buffer class (blueoil/converter/templates/include/dma_buffer.h) has very similar features to handle the physical memory addresses. So it's better to refactor to use same code for common parts between...
A network of LmSegnetV1Quantize has 12 output channels from first space_to_depth operation. The number of channels is not good to make it faster by optimization. Multiple of 32 output channels...