Mobilenet-V1 training

[forflafor]

Hello everyone.
I am trying to train the Mobilenet-V1 network (available in vision.nn.mobilenet.py) with the script made available in the following path: jetson-inference → python → training → classification → train.py.
Unfortunately I was unable to load the model, not even used the --resume flag.
Can anyone help me out? Thank you.

[MtHiker]

Would you share the command to train?
If you follow the instruction, cat and dog, and put the dataset into the proper position,
it would seem like this;
(supposed to 5 epochs)

$ python3 train.py --model-dir=models/cat_dog data/cat_dog --epochs=5

[AastaLLL]

Hi,

Do you meet any error when loading the model?
To give more suggestion, would you mind sharing the detailed output log with us first?

Thanks.

[forflafor]

Thanks for the replies @MtHiker @AastaLLL .

The command I run is the following:

python3 train.py --resume=jetson_nano/KD/model/mobilenet.pth --batch-size=4 --workers=2 --epochs=1 jetson_nano/KD/data/

I don’t have any kind of error but the script starts downloading the pre-trained model “ResNet18”. This makes me understand that he didn’t accept my pre-trained model (Mobilenet).

How can I solve? Thank you.

[MtHiker]

IF you do not define --arch flag, the script will use Resnet-18 by default.

And try to use --resume and --start-epoch flags together to resume your custom model.