bohaohuang / mrs Goto Github PK

Implement DeepLabV3

• VGG, ResNet, Inception familiy backbones compatibility
• Inria building + deepglobe building benchmark score

Add SqueezeNet

• UNet, PSPNet, LinkNet compatability
• SqueezeNet + Unet on Inria toyset
• SqueezeNet + Unet on Inria buildings
• Dilation

Hello, I think there are error in this line the second Adam optimizer for discriminator not for generate

Implement DLinkNet

• VGG & ResNet family Backbones compatibility
• ResNet50 backbone on Inria toyset
• Deepglobe road benchmark score

Implement Inception

• UNet, PSPNet, LinkNet compatability
• InceptionV3 + UNet on Inria toyset
• InceptionV3 + UNet on Inria building

Object-based scoring method

• Confmap parsing function
• Object link function
• PR scoring
• mAP versus link radius

Also tried "24" and '24'. None of them worked.

• "resume_epoch" is used when the pre-defined number of epochs of training isn't completed. A model's state dict and the optimizer will be loaded from a snapshot. config["trainer"]["resume_epoch"] has to be no greater than config["trainer"]["epochs"].
• Fine-tuning is for a new dataset and the optimizer will be re-initialized.
• "further train" is for situations where the finished training is deemed insufficient and we want to train for some more epochs from a snapshot model without re-initializing the optimizer. Basically this is equivalent to fine-tuning but with network_utils.load_epoch() in stead of network_utils.load()

@bohaohuang let me know if this makes sense to you.

Mean and std should be precomputed for each dataset, see https://forums.fast.ai/t/image-normalization-in-pytorch/7534/7
We need a helper function for computing mean and std for any dataset or list of img files. Also, we need to store this mean and std stats so it can be easily read when defining dataloader. Which means we might not able to have it defined in .config files anymore...