Closed ZongHong-Lyu closed 6 years ago
Hi Lyu,
I am testing types of layer related to Cov1D. And there are problems:
GlobalAveragePooling1D error: Wait for command completion failed, issued on layer 1, name=global_average_pooling1d_1: ioctl(DMP_DV_IOC_WAIT) failed for /dev/dv_conv
MaxPooling1D error: expected: [[[-0.71810555] [-0.71810555]]] but get: [ 0. -0.7182617]
Hi Lyu,
I attached the test code, model here. I used PyWrapper:
Hi Tuan,
I've tried but can't reproduce the error for the MaxPooling1D one. I'll report the issue of AveragePooling1D to the HW guy.
EDIT: Alexey also tested it and it seems pooling is not working using non-square configs.
Hi Tuan,
The HW guy says the FPGA can't support 1D average pooling currently, He is working on fix that issue.
Hi Lyu,
Thank you.
But how about MaxPooling1D. Is your output similar to Keras 's output?
Hi Tuan,
It seems work for the one test I did. But as I said above, Alexey made more tests and it seems max pooling is also not working.
Hi Lyu,
Thanks. Hope it will work soon since I have to deploy speech app to board soon.
Hi Tuan,
We have uploaded new FPGA bit files to the DMP FTP. in /Jupiter/JupiterX/20181107 There are two files: boot.bin and boot.bin.SD. boot.bin is for eMMC and boot.bin.SD is for SD card (If you use SD card, rename it to boot.bin)
Can you try that one?
Hi Lyu,
i will test it. Thanks.
Hi Lyu,
It seems the HW was fixed.
Thanks.
Caffe has not explicit 1D layers so only Keras needs to be updated.
This implementation treats those 1D layers as 2D layers with height = 1.
Also I fixed the split_pool_node function so now it can handle asymmetric kernel sizes, since 1D layers are like 2D layers with asymmetric kernels.
And in the fix, I also fixed a bug of pool strides of the split pool node.
This should also closes #16.