sponge.metrics.BalancedMSE

View Source On Gitee
class sponge.metrics.BalancedMSE(first_break, last_break, num_bins, beta=0.99, reducer_flag=False)[source]

Balanced MSE error Compute Balanced MSE error between the prediction and the ground truth to solve unbalanced labels in regression task.

Refer to Ren, Jiawei, et al. 'Balanced MSE for Imbalanced Visual Regression'.

\[L =-\log \mathcal{N}\left(\boldsymbol{y} ; \boldsymbol{y}_{\text {pred }}, \sigma_{\text {noise }}^{2} \mathrm{I}\right) +\log \sum_{i=1}^{N} p_{\text {train }}\left(\boldsymbol{y}_{(i)}\right) \cdot \mathcal{N}\left(\boldsymbol{y}_{(i)} ; \boldsymbol{y}_{\text {pred }}, \sigma_{\text {noise }}^{2} \mathrm{I}\right)\]
Parameters

  • first_break (float) – The begin value of bin.

  • last_break (float) – The end value of bin.

  • num_bins (int) – The bin numbers.

  • beta (float, optional) – The moving average coefficient, default: 0.99.

  • reducer_flag (bool, optional) – Whether to aggregate the label values of multiple devices, default: False.

Inputs:

  • prediction (Tensor) - Predict values, shape is \((batch\_size, ndim)\).

  • target (Tensor) - Label values, shape is \((batch\_size, ndim)\).

Outputs:

Tensor, shape is \((batch\_size, ndim)\).

Supported Platforms:

Ascend GPU

Examples

>>> import numpy as np
>>> from sponge.metrics import BalancedMSE
>>> from mindspore import Tensor
>>> net = BalancedMSE(0, 1, 20)
>>> prediction = Tensor(np.random.randn(32, 10).astype(np.float32))
>>> target = Tensor(np.random.randn(32, 10).astype(np.float32))
>>> out = net(prediction, target)
>>> print(out.shape)
(32, 10)