Algorithm 1.

Proposed L2TFL algorithm.

Step 1 Read one raw PTM network $M_0$,

Step 2 Remove the last NLR l-learnable layers from $M_0$ and get $M_1$, $M_1=F_{rl}\left(M_0,L\right)$,

Step 3 Add two new fully connected layers, $M_2=F_{afcl}\left(M_1,2\right)$,

Step 4 Freeze early layers, $\overrightarrow{l_r}\left[M_2\left(1:L_0-L\right)\right]\leftarrow 0$,

Step 5 Let last two layers retrainable, $\overrightarrow{l_r}\left[M_2\left(L_2-1:L_2\right)\right]\leftarrow 1$,

Step 6 Retrain the whole network, and obtain the new network $M_3=F_{rt}\left(M_2,X\right)$,

Step 7 Output learnt features $f_N=F_{ac}\left(M_3,L_2-1\right)$.