CCS: TDA4x: convolution initialization error in TF based model

Hi Vyom,
     We currently don't support depth wise separable convolution with kernel size of 1x1. We have not come across a netowork where depth wise separable convolution is used with 1x1 coefficient dimensions. Typically 1x1 kernels are used in normal CNN style convolution (non depthwise separable convolution), which is supported in TIDL. 

Regards,

Anshu

Hi Anshu,

The model ran in PC emulation. If the depthwise separable convolutions with 1x1 kernel were not supported in TIDL, shouldn't it have failed in PC emulation too.

Best Regards,

Sankalp 

Hi Sankalp,

    PC emulation mode run reference code which is a generic code for all convolution, on the other hand  target code has specific optimization for depth wise separable layers and currently doesn't support 1x1 convolution.

Regards,

Anshu

Hi Anshu,

We have replaced the kernel size from 1*1 to 3*3

After successfully importing , we have verified the results from trace dumps. 

When we are running the model on the board, we are facing the same issue with layer number 1. Please find error for your reference.

Alg Init for Layer # - 0
Alg Init for Layer # - 1
MMALIB_CNN_convolve_col_smallNo_ixX_ixX_oxX_init 1 :FAILED
Algorithm Init failed with error number: -1
Error at line: 479 : in file src/tidl_tb.c, of function : tidlMultiInstanceTest
Invalid Error Type!

I am sharing you the information of layer number 1 as below:

1.       Layer name in tensorflow graph - DASNet/depthwise_conv/separable_conv2d/separable_conv2d/depthwise
2.       Type - Depthwise convolution.
3.       Mapped to TIDL layer type - TIDL_ConvolutionLayer
4.       Operation name - DepthwiseConv2dNative
5.       Input tensor dimensions (in NCHW format) – (1, 3, 608, 608)
6.       Output tensor dimensions (in NCHW format) - (1, 3, 602, 602)
7.       Convolution kernel size – (7, 7)
8.       Number of convolution kernels kernel – 3
9.       Strides – (1, 1)
10.   Padding type used – VALID
11.   Dilation rate – (1, 1)

Kindly suggest your views on this issue.

Regards,

Anshuman

Hi Anshuman,
    Are you running this layer in 8-bit mode or 16-bit mode?

Regards,

Anshu

Hi Anshu,

Can you please specify what do you mean by running any layer in 8/16 bit mode?

As of my understanding 

we have used 

numparambits = 8

numfeaturebit = 8

while importing the model.

Regards,

Anshuman

Hi Anshuman,

   From the parameters you are running in 8 bit mode. I don't see any specific reason for init to fail. Can you share a small dummy network with just this layer so that we can see the behavior at our end.

Regards,

Anshu

Dear Sir,

Please find the model attached.single_layer_model.zip

Thanks and Regards,

Vyom Mishra

Dear Sir,

Gentle Reminder!

Any updates on this.

Thanks and Regards,

Vyom Mishra

Hi Vyom,

   We need a bit of time for this. Can you also share the import and infer config files which you have used?

Regards,

Anshu

Dear Sir,

Please find the import and inference configuration file for your reference.

modelType = 1
inputNetFile = "C:\Users\vipin\Desktop\single_layer_model\protobuf_files\optimized_single_layer_model.pb"
inputParamsFile = "NA"
outputNetFile = "C:\Users\vipin\Desktop\single_layer_model\TIDL_files\single_layer_net.bin"
outputParamsFile   = "C:\Users\vipin\Desktop\single_layer_model\TIDL_files\single_layer_param_"
tidlStatsTool = "C:\Users\vipin\Desktop\TI-TF_import\TDA4x_raw\tidl_j7_01_00_00_00\ti_dl\test\tidl_quant_stats_tool.out.exe"
numParamBits = 8
numFeatureBits = 8
quantizationStyle = 2
inQuantFactor = 255.0
foldBnInConv2D = 1
metaArchType = -1
inFileFormat = 1
numFrames = 1
inElementType = 0
inWidth  = 608
inHeight = 608
inNumChannels = 3
postProcType = 0
inDataFormat = 1
preProcType  = 0
inData = "C:\Users\vipin\Desktop\TI-TF_import\Import_Tools\TIDL_import_tool_v1.4\model_data\test_images_tidl_raw\BGT-07_20190225121648_G_360_resized_raw.raw"
perfSimConfig = "C:\Users\vipin\Desktop\TI-TF_import\TDA4x_raw\tidl_j7_01_00_00_00\ti_dl\test\testvecs\config\import\perfsim_base.cfg"
perfSimTool = "C:\Users\vipin\Desktop\TI-TF_import\TDA4x_raw\tidl_j7_01_00_00_00\ti_dl\utils\perfsim\ti_cnnperfsim.out.exe"

inFileFormat    = 1
postProcType = 2
numFrames   = 1
#padInBuffInTB = 0
netBinFile      = "/home/vyom/psdk_rtos_auto_j7_06_00_00_00/tidl_j7_00_09_00_00/ti_dl/test/testvecs/config/tidl_models/tensorflow/DASNet/tidl_net_DASNet_608x608.bin"
ioConfigFile   = "/home/vyom/psdk_rtos_auto_j7_06_00_00_00/tidl_j7_00_09_00_00/ti_dl/test/testvecs/config/tidl_models/tensorflow/DASNet/tidl_io_DASNet_608x608_1.bin"
inData  =   "/home/vyom/psdk_rtos_auto_j7_06_00_00_00/tidl_j7_00_09_00_00/ti_dl/test/testvecs/input/Image_DASNet.raw"
outData =   testvecs/output/DASNet_out_PC.bin
debugTraceLevel = 1
writeTraceLevel = 1
writeOutput = 1

Thanks and Regards,

Vyom Mishra

Dear Sir,

Gentle Reminder!

Any updates on this.

Thanks and Regards,

Vyom Mishra

Dear Sir,

Thanks for the information!

We are trying on our side to modify the network as per your suggestion.

Is there any update from your side on this.

Thanks and Regards,

Vyom Mishra

Dear Anshu,

Could you let us know when you would be releasing the next release so that we can sync up our development activities accordingly.

Best Regards,

Sankalp