Skip to main content

Table 11 Freezing parameters

From: Pre-Training Without Natural Images

Freezing layer(s) C10 C100 IN100 P30
w/o freezing 93.4 75.7 82.7 75.9
Conv1 92.3 72.2 77.9 74.3
Conv1, Layer1 92.0 72.0 77.5 72.9
Conv1, Layer1–2 89.3 68.0 71.0 68.5
Conv1, Layer1–3 82.7 56.2 55.0 58.3
Conv1, Layer1–4 49.4 24.7 21.2 31.4
  1. We fixed the first convolutional filter (Conv1) and layers (Layer 1–4) in ResNet-50 architecture. The w/o freezing means that the full-layer update which is the same as fine-tuning