Hybrid Fine-Tuning Strategy for Few-Shot Classification

Table 3

Comparison results under different pretraining methods on tiered-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	52.100.70	68.990.70	73.210.30	76.820.89	80.381.24	2.66
	HFT-Last1	55.180.72	72.260.66	75.190.62	80.350.63	81.820.62
	FT-all	52.900.78	70.870.69	75.020.28	80.040.72	84.690.96	1.45
	HFT-all	55.180.72	72.260.66	76.570.24	81.500.91	85.240.22
SKD-GEN0	FT-Last1	60.510.75	76.280.80	80.540.71	83.840.67	86.100.61	3.58
	HFT-Last1	64.170.82	79.420.61	83.750.53	87.600.42	90.250.34
	FT-all	61.050.76	76.370.79	83.460.50	87.011.35	90.451.26	1.58
	HFT-all	64.170.82	79.420.61	83.790.89	87.760.98	91.090.93
RFS-simple	FT-Last1	60.450.98	74.090.79	78.860.58	83.360.61	83.511.52	2.86
	HFT-Last1	63.760.88	77.740.57	81.270.53	85.350.50	86.450.57
	FT-all	60.560.97	75.390.80	80.180.43	83.900.47	88.041.35	1.77
	HFT-all	63.760.88	77.740.57	81.420.81	84.980.65	89.010.81
Average gain		3.37	3.37	2.41	2.51	2.13	2.78