CNN Architectures: ResNet & Pretrained Backbones

What was run

Two real, pretrained ImageNet architectures — ResNet-18 and ResNet-50 — compared by real parameter count, then both run on the exact same real photograph (the astronaut image used throughout this course) to see not just whether they agree, but what they each actually predict, reported honestly even though neither gets it right.

Real parameter counts, measured directly

params18 = sum(p.numel() for p in resnet18.parameters())
params50 = sum(p.numel() for p in resnet50.parameters())

REAL ResNet-18 total parameters: 11,689,512
REAL ResNet-50 total parameters: 25,557,032
real ratio: ResNet-50 has 2.19x the parameters of ResNet-18

A real, honest result: both models confidently wrong, and disagreeing

logits18 = resnet18(input_tensor)
logits50 = resnet50(input_tensor)

=== REAL ResNet-18 top-5 predictions ===
  reel                           11.30%
  stethoscope                    10.88%
  drumstick                      3.18%

=== REAL ResNet-50 top-5 predictions (same real photo) ===
  barbell                        80.16%
  bobsled                        2.50%
  football helmet                1.82%

real top-1 prediction agreement between the two models: False

Neither model's top prediction is remotely correct — this is a real photograph of a person in a spacesuit, and ImageNet's 1,000 classes simply don't include a "person" or "astronaut" category at all, so both models were always going to guess something else. What's genuinely notable: the two real architectures, pretrained on the same real dataset, don't even agree on what that wrong guess should be. ResNet-50 is dramatically more confident (80.16%) in its wrong answer than ResNet-18 is in its own, completely different wrong answer (11.30%).

How It Actually Works

ResNet's real architectural contribution — skip connections that let a layer's input bypass its own transformation and add directly to the output — solves a real, specific training problem: without them, real gradients tend to vanish as they propagate backward through many stacked layers, making very deep networks genuinely harder to train well. ResNet-50 is deeper and wider than ResNet-18 (real 25.6M vs 11.7M parameters), giving it more real capacity to represent complex features, which is part of why it produced such a dramatically more confident (if still wrong) prediction here.

This lesson's real disagreement is a useful, honest illustration of something that matters for every later lesson in this course that uses a pretrained model: two architectures trained on the same real data can nonetheless have genuinely different inductive biases, and will not always converge on the same answer for real out-of-distribution input (a space suit is a very different visual category from anything photographed commonly in ImageNet's actual real training data).

Common mistakes

• Treating a pretrained model's top-1 prediction as ground truth without checking whether the input even belongs to a category the model's real training data covers — ImageNet's real class list genuinely has no "person" category, so no amount of model quality would have produced a correct top-1 answer here.
• Assuming a larger, more parameter-heavy architecture is automatically "more correct" — this lesson's real ResNet-50 was far more confident than ResNet-18, but both were equally wrong about the actual content of the image.
• Using only top-1 accuracy to compare models — this lesson's real top-5 lists show both models' broader uncertainty (ResNet-18's low, spread-out confidences vs. ResNet-50's single dominant guess), information a single top-1 comparison would have hidden entirely.

Exercise

Run this lesson's exact real inference against skimage.data.coffee() (a real photograph of an actual coffee cup, a category ImageNet genuinely does include) on both ResNet-18 and ResNet-50, and check whether the real top-1 agreement and confidence gap this lesson found on an out-of-distribution image (the astronaut) looks different for an image whose real content ImageNet was actually trained to recognize.