Do you have any idea why torchvision vgg16 cannot converge with your training pipeline? about ssd.pytorch HOT 3 CLOSED

The torchvision VGG16 is different because it is not the fully convolutional reduced (atrous) model.
To quote the SSD arxiv paper, "[the model] is pre-trained on the ILSVRC CLS-LOC dataset. Similar to DeepLab-LargeFOV, we convert fc6 and fc7 to convolutional layers, subsample parameters from fc6 and fc7, change pool5 from 2 × 2 − s2 to 3 × 3 − s1, and use the a trous ` algorithm to fill the
”holes”. We remove all the dropout layers and the fc8 layer."

from ssd.pytorch.

Can you post a little more details?

from ssd.pytorch.

@chauhan-utk Basically i just swap the vgg model with the torchvision version, the rest are the same and I couldn't get it to converge.

from ssd.pytorch.