Coming from timm features_only¶

This page is for people who build timm backbones with features_only=True to get multi-scale feature maps and want the same maps, or others from the same model, from TorchLens. The examples were written against the timm range TorchLens's test environment installs (timm~=1.0, in the hf and test extras). Checked 2026-11-03.

The same task in both¶

Read the stage outputs of a ResNet-18 backbone at out_indices=(1, 2). Without timm installed, the test that runs this page skips both examples.

In timm¶

# migration-test: tool=timm expected=[[1, 64, 8, 8], [1, 128, 4, 4]]
import timm
import torch

torch.manual_seed(0)
model = timm.create_model("resnet18", pretrained=False, features_only=True, out_indices=(1, 2))
model.eval()
with torch.no_grad():
    features = model(torch.randn(1, 3, 32, 32))
RESULT = [list(f.shape) for f in features]

In TorchLens¶

# migration-test: tool=timm_torchlens expected=([[1, 64, 8, 8], [1, 128, 4, 4]], True)
import timm
import torch
import torchlens as tl

torch.manual_seed(0)
model = timm.create_model("resnet18", pretrained=False, features_only=True, out_indices=(1, 2))
model.eval()
x = torch.randn(1, 3, 32, 32)
names = model.feature_info.module_name()
features = tl.extract(model, x, names)

with torch.no_grad():
    native = model(x)
RESULT = (
    [list(features[name].shape) for name in names],
    all(torch.equal(features[name], out) for name, out in zip(names, native)),
)

Translation table¶

What TorchLens adds¶

• Lookups are not limited to the stages the model author exposed: any module address or op label works (tl.trace(model, x).modules and .layers).
• The plain timm.create_model(name) classifier works as it is; no second, rebuilt model is needed to read an intermediate stage.
• tl.extract_dataset writes the same lookups over a dataloader to disk; see Multi-GPU extraction.
• A trace records every executed op between stages, so a feature map and the arithmetic that produced it come from one call.

What timm features_only does that TorchLens does not¶

• It returns a model whose forward is the list of feature maps, ready to plug into a detector or segmentation head as a backbone you can train. tl.extract returns tensors for one call.
• It carries feature_info, the author-curated stage names, channel counts and reduction factors. TorchLens reports shapes of what ran but not the author's choice of which stages count as the pyramid.
• It handles architectures timm rebuilds for feature output (for example by dropping the classifier). TorchLens leaves the model whole.

See where to go instead for tools outside TorchLens's scope.

Using both¶

timm feeds TorchLens¶

A timm model, with or without features_only, is an nn.Module: pass it to tl.extract or tl.trace (the example above does). Use model.feature_info.module_name() to get the stage lookups. There is no timm-specific bridge.

TorchLens feeds timm¶

There is no bridge. Module addresses read from trace.modules are the same strings timm uses for stages when they are modules, so you can find a name with TorchLens and check it against feature_info.

Semantic differences¶

• Output shape. timm returns a list indexed by out_indices; tl.extract returns a dict keyed by the lookups you passed.
• Detached by default. tl.extract returns detached copies; pass detach_saved_activations=False to keep gradients, as a timm backbone's outputs have.
• Same tensors. On a features_only model the TorchLens lookups equal the model's own outputs, as the second example checks.