Please refer to the complete report while interpreting the results, as it is divided into three sections, each providing different insights:
- AutoDL Leaderboard: This section displays the best model and its corresponding backbone based on the overall performance. Only refer to this section while choosing the best model and backbone, as here the models are ranked based on the accuracy they achieved.
- Network-wise Study Details: This section presents network-specific details from the Optuna study. It shows the best parameter combination for each network, which might or might not correspond to the best overall model. In your case, when AutoDL tried to perform parameter tuning in advanced mode, it identified ResNet-152 as the best backbone with the other stated optimal parameter combination (such as dice_loss_avarage:micro, for example). If you had chosen other networks while training, this section would show the best parameter combination for each of those models.
- Best Performing Model Report: This section shows the results from the model that performed best based on evaluation metrics after training. These results correspond to the best network from the AutoDL leaderboard.
Summary:
In your case, when AutoDL tried to perform parameter tuning in advanced mode, it identified ResNet-152 as the best backbone with the other stated optimal parameter combination. However, it's important to note that this doesn't necessarily mean ResNet-152 is the best model overall; it's just the best when paired with the parameters chosen by AutoDL during the tuning process.
The overall results indicate that ResNet-50 performed better without any parameter tuning. Therefore, based on the evaluation metrics, I recommend selecting ResNet-50 for the best performance, as it outperforms ResNet-152 in terms of actual results.
