OrdinalEntropy

The official code of "Improving Deep Regression with Ordinal Entropy" in ICLR 2023. [PDF].

We currently provide a detailed code for experiments on the synthetic dataset, with a new visualization experiments for easy reproduction.

Experiments on the synthetic dataset

Obtain experiments results on the synthetic dataset

• run main.py

Visualization experiment on the synthetic dataset

We add a new visualization experiment with the synthetic dataset for easy reproduction, as the visualization experiments in our paper is on depth estimation task, which may take some effort to reproduce.

• run vis_tsne.py to obtain the features
• run vis_sphere.py to visualize the obtained features on a sphere

Dataset

For the Linear task:

• train.npy : the traning set
• test.npy: the test set, please download it here.

For the non-linear task:

• train_sde.npy : the traning set
• test_sde.npy: the test set

The dataset above is generated with this code: DeepONet.

Experiments on the Depth Estimation and Crowd Counting

The code for the Depth Baseline can be found here:

• NeW-CRFs.

The code for the Crowd Counting Baseline can be found here:

• CSRNet.

The ordinal entropy code for the two tasks can be found here:

• ./DepthEstimation&CrowdCounting/OrdinalEntropy.py

The ordinal entropy can be added into the New-CRFs and CSRNet baselines by:

• change the output of models from

        returen x

to

        if self.training:
            return x, encoding
        else:
            return x

• add the ordinal entropy into the loss: change

outputs = model(inputs, targets, epoch)

to

outputs, features = model(inputs, targets, epoch)
oe_loss = ordinalentropy(features, targets)
loss = loss + oe_loss

Visualization results on depth-estimation with NYU-v2

The visualization results can be obtained by:

• run vis_sphere.py to visualize the obtained features on a sphere

Experiments on the Age Estimation

The code for the Baseline can be found here:

• Imbalanced Regression.

The ordinal entropy code for Age Estimation can be found here:

• ./AgeEstimation/OrdinalEntropy.py

The ordinal entropy can be added into the Age Estimation baselines in a similar way shown above.

Reference

S. Zhang, L. Yang, M. Bi Mi, X. Zheng, A. Yao, "Improving Deep Regression with Ordinal Entropy," in ICLR, 2023. [PDF].