MobileNetV2

The core structure of MobileNetV2 includes depthwise separable convolution and the inverted residual block.

The inverted residual structure of MobileNetV2 first uses a 1×1 convolution to expand the number of channels (usually by a factor of 6), then performs a depthwise convolution, and finally uses a 1×1 convolution to compress back to low dimension, with the residual connection made at the low-dimensional point.

The first convolution layer of MobileNetV2 has only 32 kernels, and the number of channels in the shallow layers is very small, so freezing the shallow layers saves limited computation but costs no small amount of adaptability.

In MobileNetV2 fine-tuning, the closer the data volume is to the original pretraining distribution and the fewer the samples, the higher the proportion that can be frozen; conversely, more layers should be released to participate in training.

Training only the classification head (freezing the entire backbone) suits scenarios with a sample size in the hundreds to thousands and with classes semantically close to ImageNet, but the model's ability to adapt to domain shift is almost zero.

Freezing the shallow layers and fine-tuning the deep layers (keeping the earlier feature extraction layers and releasing the last few inverted residual blocks and the classification head) is the compromise that appears most frequently in MobileNetV2 fine-tuning discussions.

Progressive unfreezing (first freezing and training for a few rounds to let the classification head converge, then gradually unfreezing the deep layers) is considered in discussions to positively help final accuracy, but it lengthens the training cycle.

MobileNetV2 is sensitive to input resolution, and after changing the resolution the applicability of the shallow features decreases.

On Reddit, developers discussed MobileNetV2 fine-tuning and layer freezing strategies, and that discussion did not give a one-size-fits-all configuration.