PyTorch image models, scripts, pretrained weights -- ResNet, ResNeXT, EfficientNet, EfficientNetV2, NFNet, Vision Transformer, MixNet, MobileNet-V3/V2, RegNet, DPN, CSPNet, and more
pytorch
resnet
pretrained-models
mixnet
pretrained-weights
imagenet-classifier
distributed-training
dual-path-networks
cnn-classification
mobilenet-v2
mnasnet
mobile-deep-learning
mobilenetv3
efficientnet
augmix
randaugment
efficientnet-training
nfnets
normalization-free-training
vision-transformer-models
-
Updated
Jun 13, 2022 - Python
Describe the bug
When training a detection model, e.g. SSD Lite MobileNet, the mAP shown in SG (TensorBoard) is much higher than the mAP returned via COCO-API
To Reproduce
Steps to reproduce the behavior:
train_from_recipe.pywithcoco_ssd_lite_mobilenet_v2.yamlorcoco_ssd_mobilenet_v1.yaml