Vitis AI Compiler for Caffe model error #6

[JerrySciangula]

Hi everyone,

I'm a student at Scuola Superiore Sant'Anna in Pisa.

I'm trying to accelerate a network model on a ZCU102 ultrascale+ using Vitis-AI tools v1.2. I can successfully quantize the model, but when launching the vai_c_caffe command, it reports:

_[VAI_C][Warning] Only 'channel' axis supported on DPU for Concat, current layer is [Concat1]. [VAI_C][Warning] Layer [deconv_out] is not compressed. [VAI_C][Warning] layer [Slice1] (type: Slice) is not supported in DPU, deploy it in CPU instead. [VAI_C][Warning] layer [deconv_out] is not compressed, deploy it in CPU instead. [VAI_C][Warning] layer [softmax] (type: Softmax) is not supported in DPU, deploy it in CPU instead. [VAI_C][Fatal] Check failed for condition [layersmap.find(layer->name()) == layersmap.end()] in [/home/xbuild/conda-bld/dnnc_1592904456005/work/dnncimpl/compile/kernel.cc:149] :Duplicate names exist in current layer graph, name is [Slice1].

The last is the fatal error that blocks the compilation process.

Could anyone explain me the error and how to fix it, if it is possible?

[Mookel]

Hi, @JerrySciangula could you share the deploy.prototxt file, thus I can reproduce the issue? thanks very much.

[JerrySciangula]

Hi @Mookel.

I have to say that I'm trying to resolve with a trick the previous issue at #180. I decided to add a slice point since in #180 you reported me that there was a bug in the compiler. I was thinking that editing the quantized deploy.prototxt and the corresponding quantized deploy.caffemodel adding a slice point would have resolved the problem. However the compiler reported the fatal error mentioned in this issue.

So, I think it is better to share with you the modified quantized deploy.prototxt and the modified deploy.caffemodel, to reproduce this issue directly using the VAI_C compiler.

Since the caffemodel file is too big, I share the link to my Drive with the zipped files. Here they are: link

[Mookel]

Hi, @JerrySciangula Well noted, I will give feedbacks ASAP.

[Mookel]

Hi, @JerrySciangula

I successfully reproduced the issue and found a bug about the partition module and I will fix it, thanks so much:-).

But I found another two issues:

ISSUE-1: For the 'fc1' layer, the backend will output error: "[VAI_C-BACKEND][Check Failed: kernel_param * input_channel_group <= img_buf_depth]". The 'kernel_param', 'input_channel_group' and 'img_buf_depth' are all parameters used internally by the backend and they are tightly related to the DPU type you are using(see https://www.xilinx.com/support/documentation/ip_documentation/dpu/v3_2/pg338-dpu.pdf page20 more details about the parameter and constraints of our DPU.), in our case:

• kernel_param = kernel_h kernel_w = 20 15 = 300

• input_channel_group = ceil(input_channel / channel_parallel) = 256 / 16 = 16 (taken ZCU104 for example)

• img_buf_depth = bank_depth = 1024

(_Note: bank depth and channel_parallel are DPU related, while kernel_h/kernel_w, inputchannel are model related.) In our case, img_buf_depth < 300 * 16 = 4800, which is too large for the DPU.

There are usually several workarouds to solve this issue:

• Use larger core DPU, such as B2304 or B4096, which has larger bank memory size; <--- didn't work in our case.

• Modify the model, such as using smaller kernel size, and smaller input channel;

• Prune the model, reduce the input or output of channel;

ISSUE-2: For the 'Concat1' layer, the backend will fail to generate DPU instructions , the root cause is: For the DPU, only concatenation in channel axis is supported(In our case, the axis is 2 for 'Concat1'- concatenation in H axis); Although In some cases, some optimizations can be leveraged by the middle end to get rid of such constraint, while in our case the constraint still exists after passes of middle end, this is why the backend output an error.

Thanks very much.

[JerrySciangula]

Hi @Mookel, thank you for your exhaustive reply. I will try to resolve according your advices. Thank you again