H
i noted that only efficientnet-b0 backbone is available for tlt and noted that for pytoch efficient-b4 is available in ngc. Can i used tlt for efficientnet-b4? thanks
Could you please share the link for pytorch efficient-b4 in ngc?
Hi @Morganh
Can you suggest some reference for training efficientnet-b0 (available in NGC) on custom data using tlt?
@skim1
See Overview — TAO Toolkit 3.22.05 documentation and Transfer Learning Toolkit — Transfer Learning Toolkit 3.0 documentation , you can download faster_rcnn jupyter notebook and try run with efficientnet backbone.
1 Like
hi
i am looking at this:
That one is not compatible with TLT. In TLT, see Transfer Learning Toolkit — Transfer Learning Toolkit 3.0 documentation , then https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b0_relu or https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b0_swish , currently, only b0 version is available.
thank you. Any chance that other than b0, b4 and above will be made available for transfer learning toolkit?
They are in the roadmap.
1 Like
thank you. could i check on the roadmap what estimated time frame for it? thank you
Yesterday, the b1 versions are available at https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b1_swish and
https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b1_relu
Currently, EfficientNet B2-B7 are not supported in TLT.
thank you very much. Would hope if other such as B4 can be supported as well since it is already in the ngc.
Can i check what is the differences between those in the TLT and those offered in the NGC(non-TLT)? thanks
In your above-mentioned link efficientnet-b4 pretrained weights (PyTorch, AMP, ImageNet) | NVIDIA NGC , it is for classification. https://github.com/NVIDIA/DeepLearningExamples/tree/master/PyTorch/Classification/ConvNets
It is not compatible with TLT.
In TLT pretrained model link what I mentioned above, they are the pretrained models for TLT classification network or object detection network(FasterRCNN, SSD, DSSD,RetinaNet ) . They are compatible with TLT. Below are the links.
Efficinetnet b0/b1 pretrained models for TLT Classification network: https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_classification/files?version=efficientnet_b1_swish, https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_classification/files?version=efficientnet_b1_relu , https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_classification/files?version=efficientnet_b0_relu,
https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_classification/files?version=efficientnet_b0_swish
Efficinetnet b0/b1 pretrained models for TLT object detection networks: https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b1_swish,
https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b1_relu,
https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b0_relu,
https://ngc.nvidia.com/catalog/models/nvidia:tlt_pretrained_object_detection/files?version=efficientnet_b0_swish