matlab/matlab-agentic-toolkit / skills-catalog/ai-and-statistics/matlab-import-external-ai-model
matlab-import-external-ai-model
VerifiedImport PyTorch, ONNX, or Keras 3 / TensorFlow 2.16+ deep learning models into MATLAB as dlnetwork objects. Use when importing .pt2 exported programs, traced .pt files, .onnx models, or Keras 3 models via matlabsaver. Covers importNetworkFromPyTorch, importNetworkFromONNX, importNetworkFromKeras, importNetworkFromTensorFlow, torch.export.export, PyTorchInputSizes, InputDataFormats, matlabsaver, tf_keras downgrade, numeric validation against PyTorch or ONNX Runtime, and placeholder/custom layer implementation. Applies when user mentions any of these functions, file formats, or encounters import errors, unsupported operator warnings, 0 learnables, or uninitialized networks.
| File | Kind | Size |
|---|---|---|
| SKILL.md | skill md | 11.1 KB |
| manifest.yaml | config | 626 B |
| references/keras-matlabsaver-workflow.md | markdown | 4.0 KB |
| references/keras-tf-keras-downgrade.md | markdown | 3.8 KB |
| references/onnx-validation-workflow.md | markdown | 6.4 KB |
| references/pytorch-export-guidance.md | markdown | 3.6 KB |
| references/pytorch-import-guidance.md | markdown | 3.9 KB |
| references/pytorch-numeric-validation.md | markdown | 4.3 KB |
| references/pytorch-placeholder-guidance.md | markdown | 4.3 KB |
| scripts/validateImportedNetwork.m | other | 4.3 KB |
references/pytorch-placeholder-guidance.md
View on GitHub# Filling Placeholder Layers After Import
## How Placeholders Work
When `importNetworkFromPyTorch` encounters an unsupported PyTorch operation, it
generates a custom layer `.m` file in the namespace directory. These files contain
a skeleton class with an `error(...)` call in the `predict` method that must be
replaced with a working implementation.
## Identifying Placeholders
1. Check import warnings for lines like:
```
'aten::unbind' operator in 'ModelName_unbind_view_6' layer class
```
2. Find the generated file in the namespace folder:
```
+<namespace>/+ops/pyAtenUnbind.m
```
3. The file contains `error('...')` at the bottom of `predict` — this must be
removed after you implement the operation.
## Implementation Rules
### Rule 1: Dynamic batch sizes
**Never hardcode dimension sizes from the export.** The exported graph may
record specific sizes (e.g., batch=4) but your implementation must work for
any batch size.
```matlab
% WRONG — hardcoded batch size from export
output = reshape(input, [4, -1]);
% CORRECT — dynamic batch
batchSize = size(input, ndims(input)); % batch is last dim in MATLAB
output = reshape(input, [], batchSize);
```
### Rule 2: Preserve dlarray format labels
When your custom op manipulates dimensions, ensure the output retains correct
format labels. Operations that remove or add dimensions need explicit relabeling.
```matlab
% If input has format 'SSCB' and you remove spatial dims:
output = dlarray(extractdata(result), 'CB');
```
Key behaviors of dlarray indexing:
- Indexing with `:` preserves that dimension and its label
- Scalar indexing (e.g., `X(:,:,i,:)`) **removes** that dimension from the format
- Use `squeeze` carefully — it removes all singleton dims and their labels
### Rule 3: Handle arbitrary input ranks
Don't assume 4D input. Build indexing dynamically:
```matlab
ndX = ndims(X);
idx = repmat({':'}, 1, ndX);
idx{targetDim} = i;
slice = X(idx{:});
```
### Rule 4: Ignore template helper references
The auto-generated placeholder may reference helper functions like
`permuteToPyTorchDimensionOrder` or `permutePyTorchToReversePyTorch`.
These **do not exist** in the generated package. Implement the operation
from scratch based on what the PyTorch op actually computes.
### Rule 5: Remove the error call
After implementing, delete the `error(...)` line at the bottom of `predict`.
The import warning will still appear on re-import (it's a static check for the
operator name), but this is cosmetic — the network will execute correctly.
## Example: Implementing aten::unbind
`aten::unbind(tensor, dim)` splits a tensor along `dim` into individual slices,
removing that dimension.
```matlab
function varargout = predict(~, X, dim)
% Convert 0-based PyTorch dim to 1-based MATLAB dim
if isa(dim, 'dlarray')
dim = extractdata(dim);
end
matlabDim = double(dim) + 1;
% Get number of slices dynamically
numSlices = size(X, matlabDim);
% Build dynamic indexing for any rank
ndX = ndims(X);
idx = repmat({':'}, 1, ndX);
% Extract each slice
varargout = cell(1, nargout);
for i = 1:nargout
idx{matlabDim} = i;
slice = X(idx{:});
% Squeeze the indexed dimension (now singleton)
slice = squeeze(slice);
varargout{i} = slice;
end
end
```
## Example: Implementing pyView (reshape with dynamic batch)
```matlab
function Y = predict(~, X, targetShape, ~)
% targetShape comes from the export graph — may have hardcoded batch
if isa(targetShape, 'dlarray')
targetShape = extractdata(targetShape);
end
targetShape = double(targetShape);
% Replace any hardcoded batch dimension with actual batch size
% In MATLAB dlarray 'SSCB' format, batch is the last dimension
actualBatch = size(X, ndims(X));
% Find -1 in target shape (inferred dim) or replace batch
targetShape(targetShape == -1) = [];
newShape = [prod(targetShape), actualBatch];
Y = reshape(X, newShape);
end
```
## Validation After Fixing
After implementing all placeholders:
```matlab
% Test with a sample input
dlIn = dlarray(randn(32, 32, 3, 1, 'single'), 'SSCB');
dlOut = predict(net, dlIn);
% Verify output shape and no errors
fprintf("Output size: [%s]\n", num2str(size(dlOut)));
```
----
Copyright 2026 The MathWorks, Inc.
----
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| slack-gif-creatoranthropics/skillsKnowledge and utilities for creating animated GIFs optimized for Slack. Provides constraints, validation tools, and animation concepts. Use when users request animated GIFs for Slack like "make me a GIF of X doing Y for Slack." | Stars: 198k | Updated | Claude CodeCodex | Verified | Install |
| ponytailDietrichGebert/ponytailLazy senior dev mode for any coding task: the smallest change that fully solves it, and a reply a busy human understands in one read. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailHarvest every shortcut: comment into one debt ledger, so deferrals get tracked instead of forgotten. One-shot report. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-debtDietrichGebert/ponytailList every `shortcut:` comment (and older `ponytail:` ones) as a debt ledger. One-shot report, changes nothing. Use for "ponytail debt", "what did ponytail defer", "list the shortcuts", /ponytail-debt. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |