---

name: pytorch-summary description: "Use and maintain the legacy torchsummary PyTorch model-summary package, including summary calls, device and dtype handling, smoke checks, and focused repo tests." disable-model-invocation: true metadata: disco-role: operating license: MIT

pytorch-summary

Use this repo skill when a task names pytorch-summary or torchsummary, or when a PyTorch model needs the legacy Keras-style summary(...) table for output shapes, parameter counts, and rough size estimates.

This package is small and legacy. The source README recommends the newer torchinfo project for new and more advanced model-inspection work. Use this skill when existing code depends on torchsummary, when a task asks about this repo/package specifically, or when the lightweight legacy API is enough.

Quick install and import check

torchsummary version 1.5.1 imports both PyTorch and NumPy, but its packaging metadata does not declare them as install requirements. Install PyTorch for the intended backend first, then install torchsummary and numpy if needed.

CPU-oriented example:

python -m pip install numpy torchsummary
# install a PyTorch build appropriate for the target platform/backend separately
python - <<'PY'
from torchsummary import summary, summary_string
print(summary, summary_string)
PY

Run the bundled diagnostic when imports or device behavior are uncertain:

python scripts/check_install.py --help
python scripts/check_install.py --device cpu

The diagnostic is safe: it imports torchsummary, torch, and numpy, prints public signatures, reports CUDA visibility, and runs a tiny torch.nn.Linear summary on the requested device.

Route map

• For end-user model inspection with summary(...) or summary_string(...), read sub-skills/model-summary-usage/SKILL.md. This covers single inputs, multiple inputs, dtypes, CPU/CUDA placement, returned counts, table interpretation, and a bundled smoke helper.
• For editing this checkout, fixing tests, changing packaging metadata, or preserving source compatibility, read sub-skills/repo-maintenance/SKILL.md.
• For install/import, backend, dependency, and package-selection failures shared by both routes, read references/troubleshooting.md.
• Before deciding whether this generated skill is stale for a checkout, read references/repo-provenance.md.

Public API facts

The package exports exactly the legacy public functions below:

from torchsummary import summary, summary_string

Verified signatures:

summary(model, input_size, batch_size=-1, device=torch.device("cuda:0"), dtypes=None)
summary_string(model, input_size, batch_size=-1, device=torch.device("cuda:0"), dtypes=None)

Key usage constraints:

• model must be a PyTorch torch.nn.Module and must already be on the same device passed as device.
• input_size excludes batch. Use a tuple for one input and a list of tuples for multiple positional inputs.
• The default device is CUDA; CPU-only use should pass device="cpu" or torch.device("cpu") explicitly.
• dtypes, when provided, is a list aligned with the input_size list.
• summary(...) prints a table and returns (total_params, trainable_params).
• summary_string(...) returns (summary_text, (total_params, trainable_params)).

Use torchinfo instead when

• The task is not tied to legacy torchsummary compatibility.
• The model needs dictionary, keyword-only, or non-tensor inputs.
• You need richer nested-module display, configurable columns/depth, or modern PyTorch behavior.
• Memory estimates must be credible enough for deployment planning.

Verification baseline

This generated skill was built against torchsummary 1.5.1 with CPU PyTorch verification. CUDA is documented as an optional device path, not a required backend for this skill. Use refresh-repo-skill if the package source, packaging metadata, public signatures, tests, or README guidance change.