ResNet18_v2 under the directory of /mxnet/example/gluon #15962
-
I tried to train ResNet18_v2 with CIFAR-10, the initial learning rate is 0.1, and decrease to 0.01 and 0.001 at the 80 epoch and 160 epoch (The image_classification.py doest not implement the lr_scheduler function and I add that myself). Even training on a single node with 4 GPUs, the final test accuracy is around 81%, why is that? |
Beta Was this translation helpful? Give feedback.
Replies: 5 comments
-
Hey, this is the MXNet Label Bot. |
Beta Was this translation helpful? Give feedback.
-
The following is a sample of the training log:
|
Beta Was this translation helpful? Give feedback.
-
@mxnet-label-bot add [training, question] |
Beta Was this translation helpful? Give feedback.
-
Your training accuracy is going to 1.0 so your model is probably overfitting the dataset (https://www.d2l.ai/chapter_multilayer-perceptrons/underfit-overfit.html) |
Beta Was this translation helpful? Give feedback.
-
The other reason is that no data augmentation is applied in the training process, therefore, the final test accuracy is around 81% because of the limited training dataset. When the data augmentation is applied, the test accuracy is improved. |
Beta Was this translation helpful? Give feedback.
The other reason is that no data augmentation is applied in the training process, therefore, the final test accuracy is around 81% because of the limited training dataset. When the data augmentation is applied, the test accuracy is improved.