Algorithm 1 Implementation: Few-Shot Model Compression Algorithm.

Input: Base class dataset $D_b$, validation dataset $D_{val}$, and the novel class dataset $D_n$,

the teacher network $\left\{B_{\theta }^t\left(·\right),C_b\left(·\right),C_r\left(·\right)\right\}$, the student network $\left\{B_{\theta }^s\left(·\right),C_s\left(·\right)\right\}$,

temperature parameter τ, hyperparameter ${\alpha }_1$, ${\alpha }_2$, $\beta$.

Output: The predicted value ${\widehat{y}}_q$ of query samples in $D_n$

Stage 1: Teacher network pre-training

While epoch ≤ maximum number of the iteration

A batch of images is randomly selected from $D_b$.

Images are fed into the backbone of the teacher network to extract the feature.

Obtain the base class and rotation class probability values.

Pre-train the teacher network according to Equation (4).

Stage 2: Few-shot model compression

While epoch ≤ maximum number of the iteration

A batch of images is randomly selected from $D_b$.

The image is separately fed into the backbone of the teacher and the student networks to extract features.

Obtain the base class probability values from the teacher network and the student network, respectively.

Calibrate the feature error distribution between the student network and the teacher network according to Equation (16).

Calculate the knowledge distillation loss function for intermediate features according to Equation (17).

Calculate the KL divergence-based loss function between the predicted output values of the student network and the teacher network according to Equation (20).

Calculate the cross-entropy loss function of the student network according to Equation (21).

Train the student network according to Equation (22).

Stage 3: Few-shot model testing

While epoch ≤ maximum number of the iteration

Images from $D_n$ are processed through the feature extractor to obtain the feature.

Train classifier $g_{\varnothing }(·)$ for the novel classes.

Test on the query set from $D_n$.