Export and Import pre-trained model

1

I would like to tweak a pre-trained model, export it and import in another system.
I am having issues as soon as I change the classes (see code below).
Ideally I would like to export architecture and weight and then use gluon.nn.SymbolBlock.imports

from gluoncv import model_zoo, data, utils
from matplotlib import pyplot as plt
net = model_zoo.get_model('yolo3_darknet53_voc', pretrained=True)
my_classes = ['car']
net.reset_class(classes=my_classes, reuse_weights=my_classes)
net.save_parameters('mymodel.params')
new_net = net = model_zoo.get_model('yolo3_darknet53_voc', pretrained=False)
new_net.load_parameters('mymodel.params')

I’ve got error shape incompatible.
Could someone explain how to do that using gluon.nn.SymbolBlock.imports
Thanks

2

I found the right way

from gluoncv import model_zoo, data, utils
from matplotlib import pyplot as plt
net = model_zoo.get_model('yolo3_darknet53_voc', pretrained=True)
my_classes = ['car']
net.reset_class(classes=my_classes, reuse_weights=my_classes)
x = nd.random.uniform(shape=(1, 3, 512, 512))
_ = net(x)
net.export("sample_model")

Then, the model can be imported with gluon.nn.SymbolBlock.imports.
Note that the classes parameters won’t be available and so you can run an assignment as

imported_model.class = my_classes