Hybrid Fine-Tuning Strategy for Few-Shot Classification

Table 2

Comparison results under different pretraining methods on mini-ImageNet. “Pre-tra” and “Lay-fro” are short for the pretraining method and the layer frozen policy, respectively. We report the mean accuracy of 600 episodes and the 95% confidence intervals.

Pre-tra	Lay-fro	1-Shot	5-Shot	10-Shot	20-Shot	30-Shot	Average gain
R2D2	FT-Last1	50.580.74	66.150.36	71.070.71	75.630.87	76.560.96	3.83
	HFT-Last1	53.470.61	70.130.50	74.720.45	79.670.45	81.160.47
	FT-all	51.390.81	68.630.40	73.380.61	79.430.66	80.840.99	1.90
	HFT-all	53.470.61	70.130.50	75.490.59	81.410.70	82.660.38
SKD-GEN0	FT-Last1	57.830.53	73.910.53	78.191.03	85.030.76	87.010.45	2.36
	HFT-Last1	60.740.68	77.450.49	81.300.38	86.310.42	87.960.39
	FT-all	59.940.77	74.340.56	81.691.09	86.960.32	87.430.74	1.19
	HFT-all	60.740.68	77.450.49	82.341.01	87.150.38	88.630.52
RFS-simple	FT-Last1	56.990.60	72.430.32	76.270.29	82.971.29	84.020.94	1.38
	HFT-Last1	58.410.71	73.660.51	78.850.45	83.580.56	85.100.55
	FT-all	57.100.21	72.790.59	79.310.28	83.010.51	85.230.91	0.86
	HFT-all	58.410.71	73.660.51	79.690.75	83.810.73	86.160.42
Average gain		2.29	2.85	2.50	1.78	2.12	2.30