CCS: TDA4x: Input range -1 to 1 for Onnx Model

Vyom Mishra1

Genius 4590 points

Tool/software: Code Composer Studio

Dear Sir,

We are using,

TIDL version: tidl_j7_01_00_00_00

Target Board: TDA4x

We have Onnx based model trained on floating data for which input Image is Normalized by subtracting mean and dividing by Standard Deviation.

In this case, our input range is -1 to 1 float value.

Q1. How to process the same in our model?

Q2. What could be the "inQuantfactor" to be used?

Kindly help us in this.

Thanks and Regards,

Vyom Mishra

over 6 years ago

0 Vyom Mishra1 over 6 years ago

Genius 4590 points

Dear Sir,

Gentle Reminder!!

Thanks and Regards,

Vyom Mishra

0 Anshu Jain over 6 years ago

TI__Guru 56820 points

Hi Vyom,

We already have support to do normalization as part of TIDL processing. You can enable this by setting inDataNorm = 1 in the import config file and you will have to provide the inScale and inMean for all the channels. You can refer the mobileNet import config to refer the same (ti_dl\test\testvecs\config\import\public\caffe\tidl_import_mobilenet_v1.txt )

With these parameter TIDL will add a batch norm layer during the import which will do the scaling and mean shift. If following layer is a convolution layer then this layer can be merged to it.

Regards,

Anshu

0 Vyom Mishra1 over 6 years ago in reply to Anshu Jain

Genius 4590 points

Dear Sir,

Thanks for the suggestion.

We are doing the same on our side.

But we are not sure about the "inQuantfactor" to be used.

Observation:

By following the above method and providing the "inQuantfactor=128"

a) Please find the 84591e31034fac832d29ed8584666c680835268_code-editor-fs">Fullscreen e.ti.com/cfs-filesystemfile/__key/communityserver-discussions-components-files/791/netLog.txt?_=637140908980000000" netLog.txt">netLog.txt e.ti.com/cfs-filesystemfile/__key/communityserver-discussions-components-files/791/netLog.txt?_=637140908980000000" netLog.txt">Download 1e31034fac832d29ed8584666c680835268_code-editor" ;min-width:1230px;margin:0px;" of Layer Detected : 77 ----------------------------------------------------------------------------------------------------------------------------------------------------------------- |Out Data Name |Group |#Ins |#Outs |Inbuf Ids |Outbuf Id |In NCHW |Out NCHW |MACS | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- |input_original | 0| -1| 1| x x x x x x x x | 0 | 0 0 0 0 | 1 3 336 512 | 0 | |input | 0| 1| 1| 0 x x x x x x x | 1 | 1 3 336 512 | 1 3 336 512 | 2064384 | |267 | 0| 1| 1| 1 x x x x x x x | 2 | 1 3 336 512 | 1 27 168 256 | 31352832 | |268 | 0| 1| 1| 1 x x x x x x x | 3 | 1 3 336 512 | 1 3 168 256 | 516096 | |269 | 0| 2| 1| 2 3 x x x x x x | 4 | 1 27 168 256 | 1 30 168 256 | 1290240 | |271 | 0| 1| 1| 4 x x x x x x x | 5 | 1 30 168 256 | 1 30 168 256 | 2580480 | |272 | 0| 1| 1| 5 x x x x x x x | 6 | 1 30 168 256 | 1 30 84 128 | 87091200 | |273 | 0| 1| 1| 5 x x x x x x x | 7 | 1 30 168 256 | 1 30 84 128 | 1290240 | |396 | 0| 1| 1| 5 x x x x x x x | 8 | 1 30 168 256 | 1 6 168 256 | 7741440 | |274 | 0| 2| 1| 6 7 x x x x x x | 9 | 1 30 84 128 | 1 60 84 128 | 645120 | |276 | 0| 1| 1| 9 x x x x x x x | 10 | 1 60 84 128 | 1 60 84 128 | 1290240 | |279 | 0| 1| 1| 10 x x x x x x x | 11 | 1 60 84 128 | 1 30 84 128 | 19998720 | |280 | 0| 1| 1| 11 x x x x x x x | 12 | 1 30 84 128 | 1 30 84 128 | 29030400 | |283 | 0| 1| 1| 12 x x x x x x x | 13 | 1 30 84 128 | 1 30 84 128 | 29675520 | |284 | 0| 1| 1| 13 x x x x x x x | 14 | 1 30 84 128 | 1 30 84 128 | 29030400 | |287 | 0| 1| 1| 14 x x x x x x x | 15 | 1 30 84 128 | 1 30 84 128 | 29675520 | |288 | 0| 2| 1| 15 10 x x x x x x | 16 | 1 30 84 128 | 1 90 84 128 | 967680 | |291 | 0| 1| 1| 16 x x x x x x x | 17 | 1 90 84 128 | 1 30 84 128 | 29675520 | |292 | 0| 1| 1| 17 x x x x x x x | 18 | 1 30 84 128 | 1 30 84 128 | 29030400 | |295 | 0| 1| 1| 18 x x x x x x x | 19 | 1 30 84 128 | 1 30 84 128 | 29675520 | |296 | 0| 1| 1| 19 x x x x x x x | 20 | 1 30 84 128 | 1 30 84 128 | 29030400 | |299 | 0| 1| 1| 20 x x x x x x x | 21 | 1 30 84 128 | 1 30 84 128 | 29675520 | |300 | 0| 2| 1| 21 16 x x x x x x | 22 | 1 30 84 128 | 1 120 84 128 | 1290240 | |303 | 0| 1| 1| 22 x x x x x x x | 23 | 1 120 84 128 | 1 30 84 128 | 39352320 | |304 | 0| 1| 1| 23 x x x x x x x | 24 | 1 30 84 128 | 1 30 84 128 | 29030400 | |307 | 0| 1| 1| 24 x x x x x x x | 25 | 1 30 84 128 | 1 30 84 128 | 29675520 | |308 | 0| 1| 1| 25 x x x x x x x | 26 | 1 30 84 128 | 1 30 84 128 | 29030400 | |311 | 0| 1| 1| 26 x x x x x x x | 27 | 1 30 84 128 | 1 30 84 128 | 29675520 | |312 | 0| 2| 1| 27 22 x x x x x x | 28 | 1 30 84 128 | 1 150 84 128 | 1612800 | |315 | 0| 1| 1| 28 x x x x x x x | 29 | 1 150 84 128 | 1 30 84 128 | 49029120 | |316 | 0| 1| 1| 29 x x x x x x x | 30 | 1 30 84 128 | 1 30 84 128 | 29030400 | |319 | 0| 1| 1| 30 x x x x x x x | 31 | 1 30 84 128 | 1 30 84 128 | 29675520 | |320 | 0| 1| 1| 31 x x x x x x x | 32 | 1 30 84 128 | 1 30 84 128 | 29030400 | |323 | 0| 1| 1| 32 x x x x x x x | 33 | 1 30 84 128 | 1 30 84 128 | 29675520 | |324 | 0| 2| 1| 33 28 x x x x x x | 34 | 1 30 84 128 | 1 180 84 128 | 1935360 | |327 | 0| 1| 1| 34 x x x x x x x | 35 | 1 180 84 128 | 1 30 84 128 | 58705920 | |328 | 0| 1| 1| 35 x x x x x x x | 36 | 1 30 84 128 | 1 30 84 128 | 29030400 | |331 | 0| 1| 1| 36 x x x x x x x | 37 | 1 30 84 128 | 1 30 84 128 | 29675520 | |332 | 0| 1| 1| 37 x x x x x x x | 38 | 1 30 84 128 | 1 30 84 128 | 29030400 | |335 | 0| 1| 1| 38 x x x x x x x | 39 | 1 30 84 128 | 1 30 84 128 | 29675520 | |336 | 0| 2| 1| 39 34 x x x x x x | 40 | 1 30 84 128 | 1 210 84 128 | 2257920 | |339 | 0| 1| 1| 40 x x x x x x x | 41 | 1 210 84 128 | 1 120 42 64 | 122572800 | |392 | 0| 1| 1| 40 x x x x x x x | 42 | 1 210 84 128 | 1 6 84 128 | 13547520 | |342 | 0| 1| 1| 41 x x x x x x x | 43 | 1 120 42 64 | 1 30 42 64 | 9838080 | |343 | 0| 1| 1| 43 x x x x x x x | 44 | 1 30 42 64 | 1 30 42 64 | 7257600 | |346 | 0| 1| 1| 44 x x x x x x x | 45 | 1 30 42 64 | 1 30 42 64 | 7418880 | |347 | 0| 1| 1| 45 x x x x x x x | 46 | 1 30 42 64 | 1 30 42 64 | 7257600 | |350 | 0| 1| 1| 46 x x x x x x x | 47 | 1 30 42 64 | 1 30 42 64 | 7418880 | |351 | 0| 2| 1| 47 41 x x x x x x | 48 | 1 30 42 64 | 1 150 42 64 | 403200 | |354 | 0| 1| 1| 48 x x x x x x x | 49 | 1 150 42 64 | 1 30 42 64 | 12257280 | |355 | 0| 1| 1| 49 x x x x x x x | 50 | 1 30 42 64 | 1 30 42 64 | 7257600 | |358 | 0| 1| 1| 50 x x x x x x x | 51 | 1 30 42 64 | 1 30 42 64 | 7418880 | |359 | 0| 1| 1| 51 x x x x x x x | 52 | 1 30 42 64 | 1 30 42 64 | 7257600 | |362 | 0| 1| 1| 52 x x x x x x x | 53 | 1 30 42 64 | 1 30 42 64 | 7418880 | |363 | 0| 2| 1| 53 48 x x x x x x | 54 | 1 30 42 64 | 1 180 42 64 | 483840 | |366 | 0| 1| 1| 54 x x x x x x x | 55 | 1 180 42 64 | 1 30 42 64 | 14676480 | |367 | 0| 1| 1| 55 x x x x x x x | 56 | 1 30 42 64 | 1 30 42 64 | 7257600 | |370 | 0| 1| 1| 56 x x x x x x x | 57 | 1 30 42 64 | 1 30 42 64 | 7418880 | |371 | 0| 1| 1| 57 x x x x x x x | 58 | 1 30 42 64 | 1 30 42 64 | 7257600 | |374 | 0| 1| 1| 58 x x x x x x x | 59 | 1 30 42 64 | 1 30 42 64 | 7418880 | |375 | 0| 2| 1| 59 54 x x x x x x | 60 | 1 30 42 64 | 1 210 42 64 | 564480 | |378 | 0| 1| 1| 60 x x x x x x x | 61 | 1 210 42 64 | 1 30 42 64 | 17095680 | |379 | 0| 1| 1| 61 x x x x x x x | 62 | 1 30 42 64 | 1 30 42 64 | 7257600 | |382 | 0| 1| 1| 62 x x x x x x x | 63 | 1 30 42 64 | 1 30 42 64 | 7418880 | |383 | 0| 1| 1| 63 x x x x x x x | 64 | 1 30 42 64 | 1 30 42 64 | 7257600 | |386 | 0| 1| 1| 64 x x x x x x x | 65 | 1 30 42 64 | 1 30 42 64 | 7418880 | |387 | 0| 2| 1| 65 60 x x x x x x | 66 | 1 30 42 64 | 1 240 42 64 | 645120 | |390 | 0| 1| 1| 66 x x x x x x x | 67 | 1 240 42 64 | 1 6 42 64 | 3902976 | |391 | 0| 1| 1| 67 x x x x x x x | 68 | 1 6 42 64 | 1 6 84 128 | 6193152 | |393 | 0| 2| 1| 68 42 x x x x x x | 69 | 1 6 84 128 | 1 12 84 128 | 129024 | |394 | 0| 1| 1| 69 x x x x x x x | 70 | 1 12 84 128 | 1 6 84 128 | 774144 | |395 | 0| 1| 1| 70 x x x x x x x | 71 | 1 6 84 128 | 1 6 168 256 | 24772608 | |397 | 0| 2| 1| 71 8 x x x x x x | 72 | 1 6 168 256 | 1 12 168 256 | 516096 | |398 | 0| 1| 1| 72 x x x x x x x | 73 | 1 12 168 256 | 1 6 168 256 | 3096576 | |399 | 0| 1| 1| 73 x x x x x x x | 74 | 1 6 168 256 | 1 6 336 512 | 99090432 | |output | 0| 1| 1| 74 x x x x x x x | 75 | 1 6 336 512 | 1 1 336 512 | 172032 | |output | 0| 1| -1| 75 x x x x x x x | 0 | 1 1 336 512 | 0 0 0 0 | 0 | ----------------------------------------------------------------------------------------------------------------------------------------------------------------- ----------------------------------------------------------------------------------------------------------------------------------------------------------------- = j('#fragment-1a84591e31034fac832d29ed8584666c680835268_code-editor-fs'); = j('#fragment-1a84591e31034fac832d29ed8584666c680835268_code-editor'); function(){ utionCodeEditor('fullscreen')) { utionCodeEditor('fullscreen', false); utionCodeEditor('fullscreen', true);

b) Output of layer 0( Data Layer)

c) Output of Layer 1(Batch Norm)

d) Import config file

Fullscreen 0842.Import_config.txt Download

#0: Caffe, 1: TensorFlow, 2: ONNX, Default - 0
modelType = 2

# Net definition from Training frames work. Example "deploy.prototxt" from caffe or frozen binary protobuf with parameters from tensorflow
inputNetFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\model.onnx"

# Output TIDL model with Net and Parameters
outputNetFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\net.bin"

# Input and output buffer descriptor file for TIDL ivsion interface
outputParamsFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\param"

# TIDL PC Reference executable file path for range collection. If not specified, uses from default location
tidlStatsTool = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\test\tidl_quant_stats_tool.out.exe"

# Bit depth for model parameters like Kernel, Bias etc; , Default - 8
numParamBits = 8

# Bit depth for Layer activatios; , Default - 8 , Max suported is 16. Any value > 8 is experimentel in the 00.08 release..
numFeatureBits = 8

# 0: Quantization method. Default : Non-power of two scale. 2- Power of 2 scales. can experiment with '3' (power of 2) for stable models
quantizationStyle = 2

# Scale factor of input feature, if the input range used in training is 0 to 1 and the same is passed as tansor range of 0 - 255 to TIDL then this parameters shall be 255
inQuantFactor = 128

# Format for input feature, 0 : 8bit Unsigned, 1 : 8bit Signed Default - 1
inElementType = 1

# Input tensors width
inWidth = 512

# Input tensors Height
inHeight = 336

# Input tensors Number of channels
inNumChannels = 3

# Fold Batch Normalization Layer into TIDL Lib Conv Layer
#foldBnInConv2D = 1

# Combine 2x2 with stride 2 Max Pool layer processing with previous convolution layer if possible
#foldMaxPoolInConv2D = 1

#Input File Format - 0 : Compressed Image (JPEG/PNG), 1: RAW Image, 2: Compressed Image List
inFileFormat = 0

# List of input tensor names in network to be imported. Will be usefull to import portion of network
#inDataNamesList = "input"

# List of output tensor names in network to be imported. Will be usefull to import portion of network
#outDataNamesList = "output"

# Enable / Disable Normalization on input tensor. Scale and mean values are applicable only if this is enabled
inDataNorm = 1

#Mean value needs to be subtracted for each channel of all input tensors
inMean = 94.61594, 94.85598, 93.67584

# List : Scale value needs to be multiplied after means subtract for each channel of all input tensors
inScale = 0.0001394 0.0001347 0.0001308

# List : Input tensor color format. RBG or BGR planar etc. Refer TIDL API guide for supported formats list
inDataFormat = 0

# Image resize type . Refer TIDL API guide for supported formats list
inResizeType = 0

# Image cropping is performed if resizeWidth > imWidth
resizeWidth = 512

# Image cropping is performed if resizeHeight > inHeight
resizeHeight = 336

Post processing on output tensor. 0 : Disable, 1- Classification top 1 and 5 accuracy, 2 – Draw bounding box for OD, 3 - Pixel level color blending
postProcType = 0

#Input tensors File for Reading
inData = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\bgr_2.jpg"

# Number of input tensors to be processed from the input file
numFrames = 1

We feel that our output of Batch Norm is wrong.

Can you please guide us in resolving this issue.

Thanks and Regards,

Vyom Mishra

0 Anshu Jain over 6 years ago in reply to Vyom Mishra1

TI__Guru 56820 points

Hi Vyom,
inQuantFactor should be 1 as your input image will be a fixed point image. The scale to be applied for normalization should be given as part of inScale in the import config file.

Regards,

Anshu

0 Vyom Mishra1 over 6 years ago in reply to Anshu Jain

Genius 4590 points

Dear Sir,

Thanks for the Suggestion for "inQuantfactor".

We will follow and get back to you.

Thanks and Regards,

Vyom Mishra

0 Vyom Mishra1 over 6 years ago in reply to Anshu Jain

Genius 4590 points

Dear Sir,

As per your suggestion, we have used the "inQuantfactor" equals to 1.

The output generated by BatchNorm layer is given below:

And also find the Normalized image obtained in Pytorch below:

As we observed that two images are different and hence our output of the model is also incorrect.

What could be the possible reason for this behaviour.

Kindly help us in this.

Thanks and Regards,

Vyom Mishra

0 Anshu Jain over 6 years ago in reply to Vyom Mishra1

TI__Guru 56820 points

Hi Vyom,

Can you share the import config file which you used?

Regards,

Anshu

0 Vyom Mishra1 over 6 years ago in reply to Anshu Jain

Genius 4590 points

Dear Sir,

Please find the Import config file used for Importing the model.

Fullscreen 5808.Import_config.txt Download

#0: Caffe, 1: TensorFlow, 2: ONNX, Default - 0
modelType = 2

# Net definition from Training frames work. Example "deploy.prototxt" from caffe or frozen binary protobuf with parameters from tensorflow
inputNetFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\model.onnx"

# Output TIDL model with Net and Parameters
outputNetFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\net.bin"

# Input and output buffer descriptor file for TIDL ivsion interface
outputParamsFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\param"

# TIDL PC Reference executable file path for range collection. If not specified, uses from default location
tidlStatsTool = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\test\tidl_quant_stats_tool.out.exe"

# Bit depth for model parameters like Kernel, Bias etc; , Default - 8
numParamBits = 8

# Bit depth for Layer activatios; , Default - 8 , Max suported is 16. Any value > 8 is experimentel in the 00.08 release..
numFeatureBits = 8

# 0: Quantization method. Default : Non-power of two scale. 2- Power of 2 scales. can experiment with '3' (power of 2) for stable models
quantizationStyle = 2

# Scale factor of input feature, if the input range used in training is 0 to 1 and the same is passed as tansor range of 0 - 255 to TIDL then this parameters shall be 255
inQuantFactor = 128

# Format for input feature, 0 : 8bit Unsigned, 1 : 8bit Signed Default - 1
inElementType = 1

# Input tensors width
inWidth = 512

# Input tensors Height
inHeight = 336

# Input tensors Number of channels
inNumChannels = 3

# Fold Batch Normalization Layer into TIDL Lib Conv Layer
#foldBnInConv2D = 1

# Combine 2x2 with stride 2 Max Pool layer processing with previous convolution layer if possible
#foldMaxPoolInConv2D = 1

#Input File Format - 0 : Compressed Image (JPEG/PNG), 1: RAW Image, 2: Compressed Image List
inFileFormat = 0

# List of input tensor names in network to be imported. Will be usefull to import portion of network
#inDataNamesList = "input"

# List of output tensor names in network to be imported. Will be usefull to import portion of network
#outDataNamesList = "output"

# Enable / Disable Normalization on input tensor. Scale and mean values are applicable only if this is enabled
inDataNorm = 1

#Mean value needs to be subtracted for each channel of all input tensors
inMean = 94.61594, 94.85598, 93.67584

# List : Scale value needs to be multiplied after means subtract for each channel of all input tensors
inScale = 0.0001394 0.0001347 0.0001308

# List : Input tensor color format. RBG or BGR planar etc. Refer TIDL API guide for supported formats list
inDataFormat = 0

# Image resize type . Refer TIDL API guide for supported formats list
inResizeType = 0

# Image cropping is performed if resizeWidth > imWidth
resizeWidth = 512

# Image cropping is performed if resizeHeight > inHeight
resizeHeight = 336

Post processing on output tensor. 0 : Disable, 1- Classification top 1 and 5 accuracy, 2 – Draw bounding box for OD, 3 - Pixel level color blending
postProcType = 0

#Input tensors File for Reading
inData = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\bgr_2.jpg"

# Number of input tensors to be processed from the input file
numFrames = 1

Thanks and Regards,

Vyom Mishra

0 Anshu Jain over 6 years ago in reply to Vyom Mishra1

TI__Guru 56820 points

Hi Vyom,

I see that you have separated inMean values by comma. Use space to separate values for different channels :

inMean = 94.61594, 94.85598, 93.67584

should be

inMean = 94.61594 94.85598 93.67584

Regards,

Anshu

0 Vyom Mishra1 over 6 years ago in reply to Anshu Jain

Genius 4590 points

Dear Sir,

We removed the commas from the parameter "inMean"

but still facing a similar issue,

please find the image below

Please find the Import config used

Fullscreen 56804.config.txt Download

#0: Caffe, 1: TensorFlow, 2: ONNX, Default - 0
modelType = 2

# Net definition from Training frames work. Example "deploy.prototxt" from caffe or frozen binary protobuf with parameters from tensorflow
inputNetFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\model_ir_3_opset_8_h_w_4x4_deconv_fix.onnx"

# Output TIDL model with Net and Parameters
outputNetFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\eda_side_deconv_4x4_padding\eda_net.bin"

# Input and output buffer descriptor file for TIDL ivsion interface
outputParamsFile = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\eda_side_deconv_4x4_padding\eda_param"

# TIDL PC Reference executable file path for range collection. If not specified, uses from default location
tidlStatsTool = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\test\tidl_quant_stats_tool.out.exe"

# Bit depth for model parameters like Kernel, Bias etc; , Default - 8
numParamBits = 8

# Bit depth for Layer activatios; , Default - 8 , Max suported is 16. Any value > 8 is experimentel in the 00.08 release..
numFeatureBits = 8

# 0: Quantization method. Default : Non-power of two scale. 2- Power of 2 scales. can experiment with '3' (power of 2) for stable models
quantizationStyle = 2

# Scale factor of input feature, if the input range used in training is 0 to 1 and the same is passed as tansor range of 0 - 255 to TIDL then this parameters shall be 255
inQuantFactor = 1

# Format for input feature, 0 : 8bit Unsigned, 1 : 8bit Signed Default - 1
inElementType = 1

# Input tensors width
inWidth = 512

# Input tensors Height
inHeight = 336 

# Input tensors Number of channels
inNumChannels = 3

# Fold Batch Normalization Layer into TIDL Lib Conv Layer
#foldBnInConv2D = 1

# Combine 2x2 with stride 2 Max Pool layer processing with previous convolution layer if possible
#foldMaxPoolInConv2D = 1

#Input File Format - 0 : Compressed Image (JPEG/PNG), 1: RAW Image, 2: Compressed Image List
inFileFormat = 0

# List of input tensor names in network to be imported. Will be usefull to import portion of network
#inDataNamesList = "input"

# List of output tensor names in network to be imported. Will be usefull to import portion of network
#outDataNamesList = "output"

# Enable / Disable Normalization on input tensor. Scale and mean values are applicable only if this is enabled
inDataNorm = 1

#Mean value needs to be subtracted for each channel of all input tensors
inMean = 94.61594 94.85598 93.67584

# List : Scale value needs to be multiplied after means subtract for each channel of all input tensors
inScale = 0.0001394 0.0001347 0.0001308

# List : Input tensor color format. RBG or BGR planar etc. Refer TIDL API guide for supported formats list
inDataFormat = 0

# Image resize type . Refer TIDL API guide for supported formats list
inResizeType = 0

# Image cropping is performed if resizeWidth > imWidth
resizeWidth = 512

# Image cropping is performed if resizeHeight > inHeight
resizeHeight = 336

Post processing on output tensor. 0 : Disable, 1- Classification top 1 and 5 accuracy, 2 – Draw bounding box for OD, 3 - Pixel level color blending
postProcType = 0

#Input tensors File for Reading
inData = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\bgr_2.jpg"

# Number of input tensors to be processed from the input file
numFrames = 1

Thanks and Regards,

Vyom Mishra

0 kumar.desappan over 6 years ago in reply to Vyom Mishra1

TI__Mastermind 22145 points

Since you are enabling inDataNorm, comment below two lines in your import config file.

#inQuantFactor = 1

#inElementType = 1

0 Vyom Mishra1 over 6 years ago in reply to kumar.desappan

Genius 4590 points

Dear Sir,

Thanks for the Suggestion.

We have experimented the same on our side and got the below output from BatchNorm, which is still not matching with Pytorch output.

we suspect there is swapping of channels (Blue and Red).

Background: Pytorch model is trained with BGR input, So we provided BGR input to the import tool and "inDataFormat = 0"i.e., BGR_Planar in the config file.

Please find the updated Import_config file for your reference

Fullscreen tidl_import_msi.txt Download

modelType          = 2
numParamBits       = 8
inputNetFile      = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\model.onnx"
outputNetFile      = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\net.bin"
outputParamsFile   = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\param.bin"
inDataNorm  = 1
inMean = 94.61594 94.85598 93.67584
inScale = 0.0001394 0.0001347 0.0001308
resizeWidth = 512
resizeHeight = 336
inWidth  = 512
inHeight = 336 
inNumChannels = 3
inData = E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\tidlModelImport\out\bgr_2.jpg
postProcType = 0
tidlStatsTool = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\test\tidl_quant_stats_tool.out.exe"
perfSimConfig = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\test\testvecs\config\import\perfsim_base.cfg"
perfSimTool = "E:\TIDL_workshop\TIDL_import_tool_v14_updated\tidl_j7_01_00_00_00\ti_dl\utils\perfsim\ti_cnnperfsim.out.exe"
numFrames = 1
inDataFormat = 0
inFileFormat = 0

Kindly help us to resolve the issue.

Thanks and Regards,

Vyom Mishra

0 Vyom Mishra1 over 6 years ago in reply to Vyom Mishra1

Genius 4590 points

Dear Sir,

Gentle Reminder!!

Any suggestions.

Thanks and Regards,

Vyom Mishra

0 kumar.desappan over 6 years ago in reply to Vyom Mishra1

TI__Mastermind 22145 points

Hi Vyom Mishra,

How did you generate the image that you have shared here?

Can you share the simail image for input layer as well

0 Vyom Mishra1 over 6 years ago in reply to kumar.desappan

Genius 4590 points

Dear Sir,

a) we have read the float bin file for that layer using python, reshaped it to the respective dimension.

Used "cv2normalize" method to normalize the image from 0-255 and then wrote an image using "cv2.imwrite"

b) Please find the output of the data layer

Thanks and Regards,

Vyom Mishra

0 Vyom Mishra1 over 6 years ago in reply to Vyom Mishra1

Genius 4590 points

Dear Sir,

Gentle Reminder!!

Thanks and Regards,

Vyom Mishra

0 kumar.desappan over 6 years ago in reply to Vyom Mishra1

TI__Mastermind 22145 points

Vyom Mishra,

This image shall match with the image that is passed to the model during your training step. Can you compare this with your input in PyTorch.

0 Vyom Mishra1 over 6 years ago in reply to kumar.desappan

Genius 4590 points

Dear Sir,

We have checked on our side and we have shared the images in the above replies.

We suspect that Swapping is happening in DataLayer, Kindly confirm this or please provide us with the reason for this mismatch.

Thanks and Regards,

Vyom Mishra

0 kumar.desappan over 6 years ago in reply to Vyom Mishra1

TI__Mastermind 22145 points

Is this issue related to the below thread? If yes. I will close this read and we continue in the below

https://e2e.ti.com/support/j721e/f/1026/t/864182

Processors

Processors forum

CCS: TDA4x: Input range -1 to 1 for Onnx Model