TDA2HF: there is difference in middle layers results between import tool tempDir .y file and .bin files generated on eve

chen poca

Part Number: TDA2HF

Hello，

I used import tool tidl_model_import.out to generate the middle layers results .y file in tempDir folder,and using the same input data to generate the middle layers results .bin file on eve, and I found there is difference in some middle layers when compared two corresponding trace file.Here are the results:

trace_dump_1_320x320.y vs trace_dump_1_320x320.bin min: 128 max: 128 avg: 128.0
trace_dump_2_160x160.y vs trace_dump_2_160x160.bin min: 0 max: 0 avg: 0.0
trace_dump_3_160x160.y vs trace_dump_3_160x160.bin min: 0 max: 0 avg: 0.0
trace_dump_4_160x160.y vs trace_dump_4_160x160.bin min: 128 max: 128 avg: 128.0
trace_dump_5_160x160.y vs trace_dump_5_160x160.bin min: 0 max: 0 avg: 0.0
trace_dump_6_80x80.y vs trace_dump_6_80x80.bin min: 0 max: 0 avg: 0.0
trace_dump_7_80x80.y vs trace_dump_7_80x80.bin min: 128 max: 128 avg: 128.0
trace_dump_8_80x80.y vs trace_dump_8_80x80.bin min: 0 max: 0 avg: 0.0
trace_dump_9_80x80.y vs trace_dump_9_80x80.bin min: 0 max: 0 avg: 0.0
trace_dump_10_80x80.y vs trace_dump_10_80x80.bin min: 128 max: 128 avg: 128.0
trace_dump_11_80x80.y vs trace_dump_11_80x80.bin min: 128 max: 128 avg: 128.0
trace_dump_12_80x80.y vs trace_dump_12_80x80.bin min: 0 max: 0 avg: 0.0
trace_dump_13_40x40.y vs trace_dump_13_40x40.bin min: 0 max: 0 avg: 0.0
trace_dump_14_40x40.y vs trace_dump_14_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_15_40x40.y vs trace_dump_15_40x40.bin min: 0 max: 0 avg: 0.0
trace_dump_16_40x40.y vs trace_dump_16_40x40.bin min: 0 max: 0 avg: 0.0
trace_dump_17_40x40.y vs trace_dump_17_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_18_40x40.y vs trace_dump_18_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_19_40x40.y vs trace_dump_19_40x40.bin min: 0 max: 0 avg: 0.0
trace_dump_20_40x40.y vs trace_dump_20_40x40.bin min: 0 max: 0 avg: 0.0
trace_dump_21_40x40.y vs trace_dump_21_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_22_40x40.y vs trace_dump_22_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_23_40x40.y vs trace_dump_23_40x40.bin min: 0 max: 0 avg: 0.0
trace_dump_24_20x20.y vs trace_dump_24_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_25_20x20.y vs trace_dump_25_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_26_20x20.y vs trace_dump_26_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_27_20x20.y vs trace_dump_27_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_28_20x20.y vs trace_dump_28_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_29_20x20.y vs trace_dump_29_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_30_20x20.y vs trace_dump_30_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_31_20x20.y vs trace_dump_31_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_32_20x20.y vs trace_dump_32_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_33_20x20.y vs trace_dump_33_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_34_20x20.y vs trace_dump_34_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_35_20x20.y vs trace_dump_35_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_36_20x20.y vs trace_dump_36_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_37_20x20.y vs trace_dump_37_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_38_20x20.y vs trace_dump_38_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_39_20x20.y vs trace_dump_39_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_40_20x20.y vs trace_dump_40_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_41_20x20.y vs trace_dump_41_20x20.bin min: 128 max: 128 avg: 128.0
trace_dump_42_20x20.y vs trace_dump_42_20x20.bin min: 0 max: 0 avg: 0.0
trace_dump_43_40x40.y vs trace_dump_43_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_44_40x40.y vs trace_dump_44_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_45_40x40.y vs trace_dump_45_40x40.bin min: 128 max: 128 avg: 128.0
trace_dump_46_40x40.y vs trace_dump_46_40x40.bin min: 0 max: 0 avg: 0.0
trace_dump_47_40x40.y vs trace_dump_47_40x40.bin min: 0 max: 220 avg: 55.7348828125
trace_dump_48_80x80.y vs trace_dump_48_80x80.bin min: 84 max: 246 avg: 139.531689453125
trace_dump_49_80x80.y vs trace_dump_49_80x80.bin min: 123 max: 129 avg: 127.6588427734375
trace_dump_50_80x80.y vs trace_dump_50_80x80.bin min: 124 max: 130 avg: 127.7884716796875
trace_dump_51_80x80.y vs trace_dump_51_80x80.bin min: 0 max: 2 avg: 0.000390625
trace_dump_52_80x80.y vs trace_dump_52_80x80.bin min: 0 max: 1 avg: 1.220703125e-05
trace_dump_53_160x160.y vs trace_dump_53_160x160.bin min: 128 max: 129 avg: 128.00000061035155
trace_dump_54_160x160.y vs trace_dump_54_160x160.bin min: 128 max: 128 avg: 128.0
trace_dump_55_160x160.y vs trace_dump_55_160x160.bin min: 128 max: 128 avg: 128.0
trace_dump_56_160x160.y vs trace_dump_56_160x160.bin min: 0 max: 0 avg: 0.0
trace_dump_57_160x160.y vs trace_dump_57_160x160.bin min: 0 max: 0 avg: 0.0
trace_dump_58_320x320.y vs trace_dump_58_320x320.bin min: 128 max: 128 avg: 128.0
trace_dump_59_320x320.y vs trace_dump_59_320x320.bin min: 128 max: 128 avg: 128.0
trace_dump_60_320x320.y vs trace_dump_60_320x320.bin min: 128 max: 128 avg: 128.0
trace_dump_61_320x320.y vs trace_dump_61_320x320.bin min: 0 max: 0 avg: 0.0
trace_dump_62_320x320.y vs trace_dump_62_320x320.bin min: 128 max: 128 avg: 128.0
trace_dump_63_20x20.y vs trace_dump_63_20x20.bin min: 0 max: 214 avg: 78.5892578125
trace_dump_64_10x10.y vs trace_dump_64_10x10.bin min: 0 max: 206 avg: 32.630703125
trace_dump_65_10x10.y vs trace_dump_65_10x10.bin min: 0 max: 254 avg: 44.3418359375
trace_dump_66_10x10.y vs trace_dump_66_10x10.bin min: 18 max: 201 avg: 111.76666666666667

and we would find there is difference in some layers

trace_dump_47_40x40.y vs trace_dump_47_40x40.bin min: 0 max: 220 avg: 55.7348828125
trace_dump_48_80x80.y vs trace_dump_48_80x80.bin min: 84 max: 246 avg: 139.531689453125
trace_dump_49_80x80.y vs trace_dump_49_80x80.bin min: 123 max: 129 avg: 127.6588427734375
trace_dump_50_80x80.y vs trace_dump_50_80x80.bin min: 124 max: 130 avg: 127.7884716796875
trace_dump_51_80x80.y vs trace_dump_51_80x80.bin min: 0 max: 2 avg: 0.000390625
trace_dump_52_80x80.y vs trace_dump_52_80x80.bin min: 0 max: 1 avg: 1.220703125e-05
trace_dump_53_160x160.y vs trace_dump_53_160x160.bin min: 128 max: 129 avg: 128.00000061035155

trace_dump_63_20x20.y vs trace_dump_63_20x20.bin min: 0 max: 214 avg: 78.5892578125
trace_dump_64_10x10.y vs trace_dump_64_10x10.bin min: 0 max: 206 avg: 32.630703125
trace_dump_65_10x10.y vs trace_dump_65_10x10.bin min: 0 max: 254 avg: 44.3418359375
trace_dump_66_10x10.y vs trace_dump_66_10x10.bin min: 18 max: 201 avg: 111.76666666666667

and from the import tool debug information,we konw that:

47, TIDL_ConvolutionLayer , conv35
48, TIDL_Deconv2DLayer , conv_transpose2
49, TIDL_BatchNormLayer , batch_norm37
50, TIDL_BatchNormLayer , bn_scale37
51, TIDL_BatchNormLayer , relu24
52, TIDL_ConvolutionLayer , conv36
53, TIDL_Deconv2DLayer , conv_transpose3

54, TIDL_BatchNormLayer , batch_norm39
55, TIDL_BatchNormLayer , bn_scale39
56, TIDL_BatchNormLayer , relu26
57, TIDL_ConvolutionLayer , conv37
58, TIDL_Deconv2DLayer , conv_transpose_seg
59, TIDL_BatchNormLayer , batch_norm_seg
60, TIDL_BatchNormLayer , bn_scale_seg
61, TIDL_BatchNormLayer , relu_seg
62, TIDL_ConvolutionLayer , conv_seg

63, TIDL_ConvolutionLayer , conv40
64, TIDL_ConvolutionLayer , conv41
65, TIDL_ConvolutionLayer , conv42
66, TIDL_ConvolutionLayer , conv_pairs

and below is my part model that happen this:

So Why does this happen?

Thanks,

chen poca

over 2 years ago

0 chen poca over 2 years ago

Prodigy 30 points

this is the caffe prototxt that happen this:

layer {
name: "relu21"
type: "ReLU"
bottom: "batch_norm_blob34"
top: "relu_blob21"
}
layer {
name: "conv_transpose1"
type: "Deconvolution"
bottom: "relu_blob21"
top: "conv_transpose_blob1"
convolution_param {
num_output: 64
bias_term: false
pad: 1
kernel_size: 4
stride: 2
group: 64
weight_filler {
type: "xavier"
}
dilation: 1
engine: CAFFE
}
}
layer {
name: "batch_norm35"
type: "BatchNorm"
bottom: "conv_transpose_blob1"
top: "batch_norm_blob35"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale35"
type: "Scale"
bottom: "batch_norm_blob35"
top: "batch_norm_blob35"
scale_param {
bias_term: true
}
}
layer {
name: "relu22"
type: "ReLU"
bottom: "batch_norm_blob35"
top: "relu_blob22"
}
layer {
name: "conv35"
type: "Convolution"
bottom: "relu_blob22"
top: "conv_blob35"
convolution_param {
num_output: 64
bias_term: false
pad: 2
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "xavier"
}
dilation: 2
}
}
layer {
name: "batch_norm36"
type: "BatchNorm"
bottom: "conv_blob35"
top: "batch_norm_blob36"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale36"
type: "Scale"
bottom: "batch_norm_blob36"
top: "batch_norm_blob36"
scale_param {
bias_term: true
}
}
layer {
name: "relu23"
type: "ReLU"
bottom: "batch_norm_blob36"
top: "relu_blob23"
}
layer {
name: "conv_transpose2"
type: "Deconvolution"
bottom: "relu_blob23"
top: "conv_transpose_blob2"
convolution_param {
num_output: 64
bias_term: false
pad: 1
kernel_size: 4
stride: 2
group: 64
weight_filler {
type: "xavier"
}
dilation: 1
engine: CAFFE
}
}
layer {
name: "batch_norm37"
type: "BatchNorm"
bottom: "conv_transpose_blob2"
top: "batch_norm_blob37"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale37"
type: "Scale"
bottom: "batch_norm_blob37"
top: "batch_norm_blob37"
scale_param {
bias_term: true
}
}
layer {
name: "relu24"
type: "ReLU"
bottom: "batch_norm_blob37"
top: "relu_blob24"
}
layer {
name: "conv36"
type: "Convolution"
bottom: "relu_blob24"
top: "conv_blob36"
convolution_param {
num_output: 64
bias_term: false
pad: 4
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "xavier"
}
dilation: 4
}
}
layer {
name: "batch_norm38"
type: "BatchNorm"
bottom: "conv_blob36"
top: "batch_norm_blob38"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale38"
type: "Scale"
bottom: "batch_norm_blob38"
top: "batch_norm_blob38"
scale_param {
bias_term: true
}
}
layer {
name: "relu25"
type: "ReLU"
bottom: "batch_norm_blob38"
top: "relu_blob25"
}
layer {
name: "conv_transpose3"
type: "Deconvolution"
bottom: "relu_blob25"
top: "conv_transpose_blob3"
convolution_param {
num_output: 64
bias_term: false
pad: 1
kernel_size: 4
stride: 2
group: 64
weight_filler {
type: "xavier"
}
dilation: 1
engine: CAFFE
}
}
layer {
name: "batch_norm39"
type: "BatchNorm"
bottom: "conv_transpose_blob3"
top: "batch_norm_blob39"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale39"
type: "Scale"
bottom: "batch_norm_blob39"
top: "batch_norm_blob39"
scale_param {
bias_term: true
}
}
layer {
name: "relu26"
type: "ReLU"
bottom: "batch_norm_blob39"
top: "relu_blob26"
}
layer {
name: "conv37"
type: "Convolution"
bottom: "relu_blob26"
top: "conv_blob37"
convolution_param {
num_output: 32
bias_term: false
pad: 4
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "xavier"
}
dilation: 4
}
}
layer {
name: "batch_norm40"
type: "BatchNorm"
bottom: "conv_blob37"
top: "batch_norm_blob40"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale40"
type: "Scale"
bottom: "batch_norm_blob40"
top: "batch_norm_blob40"
scale_param {
bias_term: true
}
}
layer {
name: "relu27"
type: "ReLU"
bottom: "batch_norm_blob40"
top: "relu_blob27"
}

layer {
name: "conv_transpose_seg"
type: "Deconvolution"
bottom: "relu_blob27"
top: "conv_transpose_seg"
convolution_param {
num_output: 32
bias_term: false
pad: 1
kernel_size: 4
stride: 2
group: 32
weight_filler {
type: "xavier"
}
dilation: 1
engine: CAFFE
}
}
layer {
name: "batch_norm_seg"
type: "BatchNorm"
bottom: "conv_transpose_seg"
top: "batch_norm_seg"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: false
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale_seg"
type: "Scale"
bottom: "batch_norm_seg"
top: "batch_norm_seg"
scale_param {
bias_term: true
}
}
layer {
name: "relu_seg"
type: "ReLU"
bottom: "batch_norm_seg"
top: "relu_blob_seg"
}

layer {
name: "conv_seg"
type: "Convolution"
bottom: "relu_blob_seg"
top: "conv_seg"
convolution_param {
num_output: 2
bias_term: true
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "constant"
value: 0
}
bias_filler {
type: "constant"
value: -1
}
dilation: 1
}
}
layer {
name: "conv40"
type: "Convolution"
bottom: "relu_blob21"
top: "conv_blob40"
convolution_param {
num_output: 64
bias_term: false
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "xavier"
}
dilation: 1
}
}
layer {
name: "batch_norm43"
type: "BatchNorm"
bottom: "conv_blob40"
top: "batch_norm_blob43"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale43"
type: "Scale"
bottom: "batch_norm_blob43"
top: "batch_norm_blob43"
scale_param {
bias_term: true
}
}
layer {
name: "relu30"
type: "ReLU"
bottom: "batch_norm_blob43"
top: "relu_blob30"
}
layer {
name: "conv41"
type: "Convolution"
bottom: "relu_blob30"
top: "conv_blob41"
convolution_param {
num_output: 256
bias_term: false
pad: 1
kernel_size: 3
group: 1
stride: 2
weight_filler {
type: "xavier"
}
dilation: 1
}
}
layer {
name: "batch_norm44"
type: "BatchNorm"
bottom: "conv_blob41"
top: "batch_norm_blob44"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale44"
type: "Scale"
bottom: "batch_norm_blob44"
top: "batch_norm_blob44"
scale_param {
bias_term: true
}
}
layer {
name: "relu31"
type: "ReLU"
bottom: "batch_norm_blob44"
top: "relu_blob31"
}
layer {
name: "conv42"
type: "Convolution"
bottom: "relu_blob31"
top: "conv_blob42"
convolution_param {
num_output: 256
bias_term: false
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "xavier"
}
dilation: 1
}
}
layer {
name: "batch_norm45"
type: "BatchNorm"
bottom: "conv_blob42"
top: "batch_norm_blob45"
param {
lr_mult: 0
}
param {
lr_mult: 0
}
param {
lr_mult: 0
}
batch_norm_param {
use_global_stats: true
eps: 9.9999997e-06
}
}
layer {
name: "bn_scale45"
type: "Scale"
bottom: "batch_norm_blob45"
top: "batch_norm_blob45"
scale_param {
bias_term: true
}
}
layer {
name: "relu32"
type: "ReLU"
bottom: "batch_norm_blob45"
top: "relu_blob32"
}

layer {
name: "conv_pairs"
type: "Convolution"
bottom: "relu_blob32"
top: "conv_pairs"
convolution_param {
num_output: 6
bias_term: true
pad: 1
kernel_size: 3
group: 1
stride: 1
weight_filler {
type: "xavier"
}
bias_filler {
type: "constant"
}
dilation: 1
}
}

0 Praveen Eppa1 over 2 years ago in reply to chen poca

TI__Genius 17580 points

Hi,

How did you generate .bin files on eve ? Kindly generate .y files in the eve also and compare .

Thanks,

Praveen

0 chen poca over 2 years ago in reply to Praveen Eppa1

Prodigy 30 points

Hello，

Thank you for your replying . I have found the problem ,because TIDL do not support scale layer.

Thanks.

Poca

0 Praveen Eppa1 over 2 years ago in reply to chen poca

TI__Genius 17580 points

Thanks for the update.

Processors

Processors forum

TDA2HF: there is difference in middle layers results between import tool tempDir .y file and .bin files generated on eve