Transfer learning

What

The torchvision wrapper supports a replaced classifier head, three stem adaptations, freeze modes, BatchNorm policy, and regex parameter groups. The transfer benchmark uses three 32×32 ResNet-18 configs: scratch, linear probe, and fine-tune.

Why

Transfer learning is only a fair comparison if the architecture, data, and evaluation are identical and the pretraining is the only changed variable. The toolkit makes that structural: all three configs use the same split, augmentation family, epoch budget, early stopping, and test evaluation. The stem is adapted the same way in all three.

Config keys

Knob	Config key	Notes
Weights	model.transfer.weights	null or a torchvision weights name
Stem	model.params.stem_adaptation	none, keep_7x7_stride1, adapt_3x3
Freeze	model.transfer.freeze	none, backbone, all_but_head
BN policy	model.transfer.train_bn	train BN parameters while frozen
Differential LR	model.transfer.param_groups	pattern, multipliers
Head dropout	model.params.dropout	applied in the replacement head

flowchart LR
    subgraph Scratch [Scratch]
        SB["Backbone: trainable"] --> SH["Head: trainable"]
    end
    subgraph Probe [Linear probe]
        PB["Backbone: frozen"] --> PH["Head: trainable"]
    end
    subgraph Fine [Fine-tune]
        FB["Backbone: trainable, lr 1e-4"] --> FH["Head: trainable, lr 1e-3"]
    end

Implementation

models.torchvision builds the backbone and replaces the classifier. models.transfer.apply_stem_adaptation changes the stem; apply_transfer freezes parameters and applies the BN policy; build_param_groups builds disjoint groups and rejects patterns that match nothing. A frozen adapt_3x3 stem is rejected because the random stem would block the pretrained features.

Evidence

• The linear-probe config has exactly 5,130 trainable parameters before training (the 512→10 head and bias).
• All three configs use keep_7x7_stride1, ImageNet normalization, the same 45k/5k random split, 30 epochs, and validation-based early stopping.
• Measured test accuracy: fine-tune 94.98%, scratch 90.88%, linear probe 62.66%. All three share the same test fingerprint and have per-run bootstrap intervals in the comparison.
• The reference scratch baseline uses CIFAR mean/std; the transfer anchor uses ImageNet mean/std because the pretrained model expects it. The comparison labels that difference rather than hiding it.
• BatchNorm running statistics update even while probe parameters are frozen, so the linear probe is a frozen-parameter probe rather than a strict frozen-feature probe.

Caveat

The 32×32 protocol is one compute budget, not a claim that linear probes are always best. This site covers the 32×32 benchmark; a 224×224 study is outside its scope. The reference baseline and the transfer anchor are separate experiments; they should not be conflated.