Structured Pruning for Efficient Transfer Learning on Edge Devices

Sammanfattning

Modern pretrained CNN backbones have become attractive as feature extractors for transfer learning, since frozen representations can often be reused across visual domains by training only a lightweight classifier on top. However, deploying such backbones on resource-constrained edge devices remains challenging, where limited memory, computational budget, and latency requirements impose strict constraints on the model architecture. This thesis explores structured pruning as a post-training compression approach for frozen pretrained backbones, without backbone fine-tuning or retraining after pruning. The proposed pipeline combines a frozen MobileNetV3-Large or MobileNetV3-Small backbone pretrained on ImageNet-1K with a downstream linear classifier trained on pooled channel-wise feature descriptors from multiple layers. Channel importance is estimated directly from the classifier weights, providing a task-specific and data-driven pruning signal without gradient flow through the backbone. The study evaluates how extraction scope, Batch Normalization recalibration, pruning schedule, pruning ratio, and initial backbone capacity affect the resulting compression--performance trade-off. Experiments are conducted on 18 datasets, consisting of three multi-class classification datasets and 15 supervised binary classification tasks derived from MVTec-AD categories. The results indicate that intermediate convolutional features provide a stronger transfer representation than the penultimate layer alone, motivating the use of depthwise features with combined average and max pooling in the pruning experiments. The main compression analysis is performed per dataset using the Forward Feature Selection frontier, which represents the best observed performance--resource trade-offs among the evaluated configurations. The attainable compression varies substantially across datasets, backbones, and whether FLOPs or parameter count is used as the resource metric. Thus, the results do not indicate a single generally valid pruning ratio. The architecture comparison further suggests that pruning a larger backbone can, in some cases, yield a more favourable transfer-performance trade-off than directly deploying a smaller unpruned backbone at a comparable resource budget. Overall, the thesis presents classifier-weight-based structured pruning as a low-cost approach for reducing the footprint of frozen-backbone transfer learning models.