TDA4VM: tidl problem

user6406170

Part Number: TDA4VM

Dear TI，

I have a private pytorch model，I converted it to caffe model, and there is a conv with kernel size = 3X3 , stride = 2, padding = 1 in this model, after import ,the imported model can be running in tda4 target, I modified the vision_apps/app_tidl to do my work. But the target's conv output is different from the pytorch output. when I modified the padding to 0, the output is same.

I'm confused, please help me.

over 5 years ago

0 user6406170 over 5 years ago

Intellectual 375 points

can someone help me?thanks

0 user6406170 over 5 years ago in reply to user6406170

Intellectual 375 points

In tidl_import_txt, I set quantizationStyle = 3 (power of two scale)

1、nn.Conv2d(3, 32, kernel_size = （3， 3）, stride=2,bias=False, padding = 1)

pytorch top 10 output is [ 69.40277 88.37494 88.46049 88.92659 88.65936 88.1016 86.986145 86.76906 87.08494 87.4024]

tda4 top 10 output is [3667 5485 5482 5516 5499 5468 5395 5374 5398 5423]

the scale = 64, （tda4 top 10 output） / 64 should is same as （pytorch top 10 output ）,but it has big diff. it‘s Wrong.

2、nn.Conv2d(3, 32, kernel_size = （3， 3）, stride=2,bias=False, padding = 0)

pytorch top 10 output is [106.67845 106.14711 106.34976 106.440445 105.71505 104.268654 103.866455 103.97082 104.25949 105.15342]

tda4 top 10 output is [6827 6793 6806 6812 6766 6673 6647 6654 6673 6730]

the scale = 64, （tda4 top 10 output） / 64 should is Approximately equal to （pytorch top 10 output ）,it has little diff. it‘s right.

0 kumar.desappan over 5 years ago in reply to user6406170

TI__Mastermind 22145 points

Please use onnx model format for Pytorch models instead caffe.

Not very sure, about the top 10 outputs here.

Follow the steps mentioned in the below to debug this issue

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_02_00_09/ti_dl/docs/user_guide_html/md_tidl_fsg_steps_to_debug_mismatch.html

0 user6406170 over 5 years ago in reply to kumar.desappan

Intellectual 375 points

my output code is :

output_sizes[0] = ioBufDesc->outWidth[id] + ioBufDesc->outPadL[id] + ioBufDesc->outPadR[id];

output_sizes[1] = ioBufDesc->outHeight[id] + ioBufDesc->outPadT[id] + ioBufDesc->outPadB[id];

output_sizes[2] = ioBufDesc->outNumChannels[id];

...

status = tivxMapTensorPatch(output_tensors[id], 3, start, output_sizes, &map_id_output, output_strides, &output_buffer, VX_READ_ONLY, VX_MEMORY_TYPE_HOST);

if (VX_SUCCESS == status)

{

if(data_type == VX_TYPE_INT16)

{

vx_int16 *pOut;

pOut = (vx_int16 *)output_buffer + (ioBufDesc->outPadT[id] * output_sizes[0]) + ioBufDesc->outPadL[id];

vx_int16 *p = pOut;

for (size_t i = 0; i < 10; i++)

{

int16_t fdata;

memcpy((void *)&fdata, p, sizeof(vx_int16))

printf("%d\n", fdata);

p++;

}

Can i use it like this？

0 user6406170 over 5 years ago in reply to kumar.desappan

Intellectual 375 points

I convert pytorch to ONNX, then convert it to tidl, but the tda4 top 10 outputs is [3667 5485 5482 5516 5499 5468 5395 5374 5398 5423], it's same as caffe.it’s wrong.

this is my import txt.

modelType = 2
numFrames = 1
numParamBits = 16
numFeatureBits = 16
quantizationStyle = 3
inputNetFile = "../../test/testvecs/models/public/onnx/efficientdet/efficientdet.onnx"
outputNetFile = "../../test/testvecs/config/tidl_models/onnx/efficientdet.bin"
outputParamsFile = "../../test/testvecs/config/tidl_models/onnx/efficientdet_"
inWidth = 1024
inHeight = 512
inNumChannels = 3
inDataFormat = 0
inData = ../../test/testvecs/config/imageNet_sample_val.txt
postProcType = 0

0 user6406170 over 5 years ago in reply to user6406170

Intellectual 375 points

and the caffe prototxt is:

name: "efficientdet"
input: "blob1"
input_dim: 1
input_dim: 3
input_dim: 512
input_dim: 1024
layer {
name: "conv1"
type: "Convolution"
bottom: "blob1"
top: "conv_blob1"
convolution_param {
num_output: 32
bias_term: false
pad: 1
kernel_size: 3
group: 1
stride: 2
dilation: 1
}
}

it only has a input and a conv

0 user6406170 over 5 years ago in reply to kumar.desappan

Intellectual 375 points

I found that when padding = 1, the four outermost edges of the feature map are not equal to the output of pytorch. The values of the feature map except for these four edges all correspond to the value of the feature map one-to-one.

0 user6406170 over 5 years ago in reply to kumar.desappan

Intellectual 375 points

I did a test, when padding = 1, I set input featue map as 0， but after a conv, the output [-1612 -539 -769 -308 -457 -309 -442 -635 -155 -43 ...]. They should be zero. Why？I‘m confused. please help me. thanks.

0 kumar.desappan over 5 years ago in reply to user6406170

TI__Mastermind 22145 points

Please follow the steps in below and run the inference in float mode first in PC as suggested in the below

https://software-dl.ti.com/jacinto7/esd/processor-sdk-rtos-jacinto7/latest/exports/docs/tidl_j7_01_02_00_09/ti_dl/docs/user_guide_html/md_tidl_fsg_steps_to_debug_mismatch.html

Processors

Processors forum

TDA4VM: tidl problem