timm-info

A utility for exploring and analyzing PyTorch Image Models (timm).

Installation

Using uv (recommended)

uv add timm-info

Using pip

pip install timm-info

Development Setup

Clone the repository and set up your development environment:

git clone https://github.com/belfner/timm-info.git
cd timm-info
uv sync
uv run timminfo --help

Run tests:

uv run pytest

Requirements

• Python 3.10+
• PyTorch 2.0+
• timm 0.9.0+
• click

Usage

timm-info provides the CLI tool timminfo with commands: search and info

Usage: timminfo [OPTIONS] COMMAND [ARGS]...

  Timm model utility.

  Arguments that don't name a command are treated as model names and passed to
  'info', so 'timminfo NAME...' is shorthand for 'timminfo info NAME...' (e.g.
  'timminfo convnext_nano'). Run 'timminfo info -h' for its options.

Options:
  --version   Show the version and exit.
  -h, --help  Show this message and exit.

Commands:
  info    Get information about a particular timm model.
  search  Search for timm models.

info shorthand

Because info is the default command, its model-name arguments can be passed directly to timminfo without naming the command:

timminfo convnext_nano          # equivalent to: timminfo info convnext_nano
timminfo resnet50 efficientnet_b0

Explicit command names (search, info) still take precedence, so a model named like a command must be requested with timminfo info <name>.

search

timminfo search ...

Usage: timminfo search [OPTIONS] [NAME_PATTERN]...

  Search for timm models. Multiple patterns can be passed.

Options:
  -p, --pretrained                Only show pretrained models
  -l, --license                   Also show the licenses across each result's
                                  pretrained weights (comma-separated in
                                  'tsv'/'json' formats)
  -b, --by-license                Group results by license instead of by model
                                  (implies --license)
  -f, --format [pretty|tsv|json]  Output format. 'tsv'/'json' are machine-
                                  parsable.  [default: pretty]
  -h, --help                      Show this message and exit.

Examples

Command

timminfo search 'resnet50d*' 'xception*'

Output

Results for 'resnet50d*':
-------------------------
0. resnet50d

Results for 'xception*':
------------------------
0. xception41
1. xception41p
2. xception65
3. xception65p
4. xception71

Command (-f tsv prints a bare model name per line, for chaining I/O)

timminfo search -f tsv 'resnet50d*' 'xception*'

Output

resnet50d
xception41
xception41p
xception65
xception65p
xception71

Command

timminfo search -p 'resnet50d*' 'xception*'

Results for 'resnet50d*':
-------------------------
0. resnet50d.a1_in1k
1. resnet50d.a2_in1k
2. resnet50d.a3_in1k
3. resnet50d.gluon_in1k
4. resnet50d.ra2_in1k
5. resnet50d.ra4_e3600_r224_in1k

Results for 'xception*':
------------------------
0. xception41.tf_in1k
1. xception41p.ra3_in1k
2. xception65.ra3_in1k
3. xception65.tf_in1k
4. xception65p.ra3_in1k
5. xception71.tf_in1k

Command (-l/--license resolves the licenses across each architecture's pretrained weights. For a bare-architecture search each row becomes a checkbox matrix with one column per distinct license present; a model whose weights carry no resolvable license ticks the unknown column)

timminfo search -l 'convnext_atto*'

Output

Results for 'convnext_atto*':
-----------------------------
+---+-------------------+------------+---------+
| # | Model             | apache-2.0 | unknown |
+---+-------------------+------------+---------+
| 0 | convnext_atto     |     X      |    -    |
| 1 | convnext_atto_ols |     X      |    -    |
| 2 | convnext_atto_rms |     -      |    X    |
+---+-------------------+------------+---------+

Command (with -p each row is a single pretrained weight, so -l collapses to one License column)

timminfo search -p -l 'resnet50d*'

Output

Results for 'resnet50d*':
-------------------------
+---+-------------------------------+------------+
| # | Model                         | License    |
+---+-------------------------------+------------+
| 0 | resnet50d.a1_in1k             | apache-2.0 |
| 1 | resnet50d.a2_in1k             | apache-2.0 |
| 2 | resnet50d.a3_in1k             | apache-2.0 |
| 3 | resnet50d.gluon_in1k          | apache-2.0 |
| 4 | resnet50d.ra2_in1k            | apache-2.0 |
| 5 | resnet50d.ra4_e3600_r224_in1k | apache-2.0 |
+---+-------------------------------+------------+

Command (in tsv format each model's licenses are appended tab-separated, and comma-separated among themselves, for easy parsing)

timminfo search -f tsv -l 'xception*'

Output

xception41	apache-2.0
xception41p	apache-2.0
xception65	apache-2.0
xception65p	apache-2.0
xception71	apache-2.0

Command (-b/--by-license pivots the view: one row per license listing the models that offer it. This implies --license)

timminfo search -b 'convnext_atto*'

Output

Results for 'convnext_atto*' by license:
----------------------------------------
+------------+----------------------------------+
| License    | Models                           |
+------------+----------------------------------+
| apache-2.0 | convnext_atto, convnext_atto_ols |
| unknown    | convnext_atto_rms                |
+------------+----------------------------------+

Command (-f json emits a single-line JSON array over all patterns. The licenses field is null unless -l/-b requested license lookup, in which case it is a list; the -b pivot instead serializes to groups with a null license for the no-license bucket)

timminfo search -f json -l 'xception71'

Output

[{"pattern": "xception71", "pretrained": false, "models": [{"name": "xception71", "licenses": ["apache-2.0"]}]}]

info

timminfo info ...

Usage: timminfo info [OPTIONS] [NAME]...

  Get information about a particular timm model. Multiple names can be passed.

Options:
  -f, --format [pretty|json]  Output format. 'json' is machine-parsable.
                              [default: pretty]
  -h, --help                  Show this message and exit.

Example

The summary is rendered as a bordered table and includes the per-stage downscaling factors and whether each stage downscales by an exact integer factor. When a model has pretrained weights, a second table lists each weight tag and its license.

Command

timminfo info convnextv2_atto efficientnet_b0

+--------------------------------------+------------------------------+
|                           Model name | convnextv2_atto              |
|                     Number of params | 3,387,400                    |
|                 Estimated model size | 13.550 MB                    |
| Number of in features for classifier | 320                          |
|       Has extractable feature layers | True                         |
|             Number of feature layers | 4                            |
|       Number of channels per feature | [40, 80, 160, 320]           |
|        Downscaling factors per stage | [4x, 2x, 2x, 2x]             |
|           Downscaling type per stage | [exact, exact, exact, exact] |
|                Pretrained Input Size | (3, 224, 224)                |
+--------------------------------------+------------------------------+

+--------------------+--------------+
| Pretrained Weights | License      |
+--------------------+--------------+
|              fcmae | cc-by-nc-4.0 |
|      fcmae_ft_in1k | cc-by-nc-4.0 |
+--------------------+--------------+

+--------------------------------------+-------------------------------------+
|                           Model name | efficientnet_b0                     |
|                     Number of params | 4,007,548                           |
|                 Estimated model size | 16.030 MB                           |
| Number of in features for classifier | 1280                                |
|       Has extractable feature layers | True                                |
|             Number of feature layers | 5                                   |
|       Number of channels per feature | [16, 24, 40, 112, 320]              |
|        Downscaling factors per stage | [2x, 2x, 2x, 2x, 2x]                |
|           Downscaling type per stage | [exact, exact, exact, exact, exact] |
|                Pretrained Input Size | (3, 224, 224)                       |
+--------------------------------------+-------------------------------------+

+---------------------+------------+
|  Pretrained Weights | License    |
+---------------------+------------+
| ra4_e3600_r224_in1k | apache-2.0 |
|             ra_in1k | apache-2.0 |
+---------------------+------------+

Command (-f json emits one single-line JSON array over all requested models, with every ModelInfo field; per-model errors go to stderr so stdout stays valid JSON)

timminfo info -f json convnextv2_atto

[{"name": "convnextv2_atto", "num_params": 3387400, "model_size_mb": 13.5496, "classifier_num_in_features": 320, "has_features": true, "num_feature_layers": 4, "num_channels_per_feature": [40, 80, 160, 320], "downscaling_factors": ["4x", "2x", "2x", "2x"], "downscaling_exact": ["exact", "exact", "exact", "exact"], "pretrained_input_size": [3, 224, 224], "pretrained_weights": [{"tag": "fcmae", "license": "cc-by-nc-4.0"}, {"tag": "fcmae_ft_in1k", "license": "cc-by-nc-4.0"}]}]

Programmatic use

The same functionality is available as an importable API. Processing returns typed dataclasses (ModelInfo, SearchResult), and formatting is separate, so you can consume the data or reuse the CLI's rendering. Importing timm_info is cheap; torch/timm load only when you call a processing function.

from timm_info import get_model_info, search_models

info = get_model_info("resnet50")
info.num_params                # 23508032
info.num_channels_per_feature  # [64, 256, 512, 1024, 2048]
info.downscaling_factors       # ['2x', '2x', '2x', '2x', '2x']
info.pretrained_weights        # [PretrainedWeight(tag='a1_in1k', license='apache-2.0'), ...]

result = search_models("resnet50*", pretrained=True, with_license=True)
[m.name for m in result.models]
result.models[0].licenses      # ['apache-2.0']  (sorted distinct licenses, or [] when none)

Regroup a search result by license with search.pivot_result, which returns a PivotResult of LicenseGroups (the -b/--by-license view):

from timm_info import search

pivot = search.pivot_result(result)
[(g.license, g.models) for g in pivot.groups]   # [('apache-2.0', ['resnet50.a1_in1k', ...]), ...]

Render the same output the CLI prints. format_search_result and format_pivot_result take the format name ("pretty" or "tsv"), and render_json serializes a list of result dataclasses to the CLI's single-line JSON array:

from timm_info import rendering

print(rendering.format_model_info(info))
print(rendering.format_search_result(result, "pretty", show_license=True))
print(rendering.format_pivot_result(pivot, "pretty"))
print(rendering.render_json([result]))   # same JSON as `search -f json`
print(rendering.render_json([info]))      # same JSON as `info -f json`