What are net-scale-factor and offset of restnet-10 trained with TLT3.0?

Accelerated Computing Intelligent Video Analytics DeepStream SDK
1

• Hardware Platform (Jetson / GPU):Both
• DeepStream Version:5.1
• JetPack Version (valid for Jetson only):4.5
• TensorRT Version:7.1.3

config file:
classifier_1.txt (436 Bytes)

Hi,
I trained resnet-10 classification model on my own dateset, and has 2 classes.

Q1: Doing inference the trained model with tlt evaluation and enable-center-crop=True, I get high accuracy, and I want to know how I can set enable-center-crop=True in the config file? Is it necessary? I test the trained model with tlt evaluation with enable-center-crop=False, I get worse results.

Q2: In the config file, there are net-scale-factor, offset and model-color-format, What are these values?

Q3- The input size of model is 224x224x3, Is it necessary to resize all of images to 244x224x3 before training?

3

No. Currently deepstream does not support it.

4

There are detailed description of the parameters in Gst-nvinfer — DeepStream 6.3 Release documentation

5

Yes.

6

So I have to train the model with enable-center-crop=False?

7

Seems yes.

8

In such way the evaluation accuracy is dropped 2-3%.

9

@Fiona.Chen,
I know offset is the means of each RGB channels, But in practical, because I used fine-tune on my own dataset, The values of offset mode should to be in documentation or config training, I don’t know I have to calculate offset values for my own dataset?

10

offset is used in normalization.

11

@Fiona.Chen, Is it possible to refer to normalization?

12

What do you mean?

13

There is detailed description of “offset” in Gst-nvinfer — DeepStream 6.1.1 Release documentation

14

enable-center-crop=False only impact evaluation, it will not impact training. We do not need to implement center crop with inference processor.

15

enable-center-crop=False only impact evaluation, it will not impact training. We do not need to implement center crop with inference processor.

When we train model with center-crop, It’s better to use center-crop in the inference mode.

If enable-center-crop=False impact in evaluation, So it definitely impact in the inference set If we don’t set center-crop=True.

16

The training is using random crop.

18

This topic was automatically closed 60 days after the last reply. New replies are no longer allowed.