How to show Metrics in Custom Object Detection Model?

Robotics & Edge Computing Jetson & Embedded Systems Jetson Nano
,
1

Hi! I’ve trained a custom model following the Re-training SSD-Mobilenet tutorial. It works fine, but now I’d like to show metrics like F1 score, Accuracy and Loss.

For example, in others projects I use this code:

acc = historial.history['accuracy']
val_acc = historial.history['val_accuracy']
historial = modelo.fit(
data_gen_entrenamiento, epochs=50, batch_size=32,
validation_data=data_gen_pruebas)

But, in the Re-training SSD-Mobilenet tutorial, for training use a diference code:

!python3 train_ssd.py --dataset-type=voc --data=./ --model-dir=models/Orange --batch-size=4 --workers=2 --epochs=50

So, I don’t know how to call ‘accuracy’ and ‘loss’ in order to show them into a graphic. By the way, I used Colab for training 'cause I have only a 2GB Jetson Nano.

2

Hi @juanpalaciosvargas, you can use the --validation-mean-ap flag to train_ssd.py to calculate the mean average precision (mAP) of each class during the validation phase of each training epoch.

4

Can you give me an example please? I’ve used this flag without any success. By the way, what I want to do is to show the results of training (“loss” and “accuracy”) into a graphic (“loss” vs" epochs" / “accuracy” vs “epochs”) and create a" confusion matrix".

5

Do you mean that it fails to run with that flag? If so, perhaps you are running an older version of train_ssd.py - in that case, try running this:

cd jetson-inference/python/training/detection/ssd
git pull

There is support for TensorBoard in train_ssd.py, it will output to your model’s tensorboard/ directory. You can then run TensorBoard from another terminal like tensorboard --log-dir=models/your_model/tensorboard and navigate your browser to it to see the Loss vs Epochs graphs and MeanAP vs Epochs graphs.

1 Like
7

This is me again. Regarding the flag, yes it fails when I run the code.
First, I used this code:

!python3 train_ssd.py --dataset-type=voc --data=./ --model-dir=models/Orange --batch-size=4 --workers=2 --epochs=50

But, I tried adding the flag:

!python3 train_ssd.py --dataset-type=voc --data=./ --model-dir=models/Orange --batch-size=4 --workers=2 --epochs=50 --validation-mean-ap

So, I tried adding the indicated lines code:

cd jetson-inference/python/training/detection/ssd
git pull

But I got this error:
You are not currently on a branch.
Please apecify which branch you want to merge with.
See git-pull(1) for details.

git pull <remote> <branch>

Any idea? By the way, I’m currently training the model using Google COLAB.

8

Hi @juanpalaciosvargas, can you try running git checkout master and git pull origin master instead?

If you are on Google COLAB, you can just clone pytorch-ssd repo directly (and it will be the latest master) - git clone https://github.com/dusty-nv/pytorch-ssd

9

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.