Fix UnsqueezeParser for ONNX Opset 13+ (#119)

This PR updates the `Squeeze` and `Unsqueeze` operators to align with the ONNX Opset 13+ standard, where axes is now provided as an input instead of a node attribute. The change is backward compatible, as the old logic remains for single-input cases. A new test case, `CCT/CCT_2_32_32_128_Opset20`, has been added using models exported from PyTorch with ONNX Opset 20 to verify compatibility with recent versions.

## Added
- Added support for ONNX Opset 13 and higher.

## Changed
- UnsqueezeParser in Generic NodeParser
   - Check for the presence of `axes` in node attributes and use the old workflow otherwise check for exactly 2 inputs (data and axes).
   - Node context was changes accordingly; 1 single input follows the old workflow, 2 inputs uses the new 2 input Op. format.

## Fixed
- Breaking compilation with ONNX Opset 13 and higher when using `Squeeze` Op.

Author: MatteoFasulo

.github/workflows/ci-platform-generic.yml

@@ -94,4 +94,5 @@ jobs:
         miniMobileNet
         miniMobileNetv2
         CCT/CCT_1_16_16_8
+        CCT/CCT_2_32_32_128_Opset20
         testFloatDemoTinyViT

.github/workflows/ci-platform-siracusa.yml

@@ -85,5 +85,6 @@ jobs:
         MLPerf/ImageClassification
         MLPerf/AnomalyDetection
         CCT/CCT_1_16_16_8
+        CCT/CCT_2_32_32_128_Opset20
         testTrainCCT/CCT1_Classifier_Training/CCT_1_16_16_8
       num-cores: 8

CHANGELOG.md

@@ -14,6 +14,7 @@ This file contains the changelog for the Deeploy project. The changelog is divid
 - Node Mangling to avoid duplication [#93](https://github.com/pulp-platform/Deeploy/pull/93)
 - Prepare Post v0.2.0 Release [#104](https://github.com/pulp-platform/Deeploy/pull/104)
 - Use Docker digests instead of arch-specific tags [#106](https://github.com/pulp-platform/Deeploy/pull/106)
+- Fix `Unsqueeze` Op. when using ONNX opset 13 or higher (from attribute to input) [#119](https://github.com/pulp-platform/Deeploy/pull/119)
 
 ### Added
 - Add manual type inference feature (CLI: `--input-type-map`/`--input-offset-map`) to resolve ambiguities when test inputs are not representative enough
@@ -81,6 +82,7 @@ This file contains the changelog for the Deeploy project. The changelog is divid
 - Fixed multiple typos in variable and method names, such as changing `includeGobalReferences` to `includeGlobalReferences` and `dicardedMappers` to `discardedMappers`
 - Corrected method usage in `importDeeployState` to call `NetworkContext.importNetworkContext` instead of the incorrect method name
 - Correctly return `signProp` from `setupDeployer` instead of hardcoding the value to `False` in `testMVP.py`
+- Fixed `Unsqueeze` Op. when using ONNX opset 13 or higher (from attribute to input)
 
 ### Removed
 - Delete outdated and unused `.gitlab-ci.yml` file

Deeploy/Targets/Generic/Parsers.py

@@ -940,10 +940,19 @@ def __init__(self):
 
     def parseNode(self, node: gs.Node) -> (bool):
 
-        ret = all(['axes' in node.attrs, len(node.inputs) == 1, len(node.outputs) == 1])
+        # ONNX v11: 'axes' is a node attribute
+        if 'axes' in node.attrs:
+            ret = all(['axes' in node.attrs, len(node.inputs) == 1, len(node.outputs) == 1])
+        # ONNX v13+: 'axes' becomes an input with the data
+        # Source: https://onnx.ai/onnx/operators/onnx__Unsqueeze.html
+        else:
+            ret = all([len(node.inputs) == 2, len(node.outputs) == 1])
 
-        if ret:
-            self.operatorRepresentation['axes'] = node.attrs['axes']
+        if ret and 'axes' in node.attrs:
+            axes_attr = node.attrs['axes']
+            self.operatorRepresentation['axes'] = [int(axes_attr)] if isinstance(axes_attr, int) \
+                else [int(a) for a in axes_attr]
+        # For opset 13+, axes will be extracted from the second input in parseNodeCtxt
 
         return ret
 
@@ -952,13 +961,26 @@ def parseNodeCtxt(self,
                       node: gs.Node,
                       channels_first: bool = True) -> Tuple[NetworkContext, bool]:
 
-        inputs = ['data_in']
         outputs = ['data_out']
-
-        for idx, inputNode in enumerate(node.inputs):
-            self.operatorRepresentation[inputs[idx]] = ctxt.lookup(inputNode.name).name
-        for idx, outputNode in enumerate(node.outputs):
-            self.operatorRepresentation[outputs[idx]] = ctxt.lookup(outputNode.name).name
+        if len(node.inputs) == 1:
+            inputs = ['data_in']
+            for idx, inputNode in enumerate(node.inputs):
+                self.operatorRepresentation[inputs[idx]] = ctxt.lookup(inputNode.name).name
+            for idx, outputNode in enumerate(node.outputs):
+                self.operatorRepresentation[outputs[idx]] = ctxt.lookup(outputNode.name).name
+        else:
+            data_in = ctxt.lookup(node.inputs[0].name)
+            data_out = ctxt.lookup(node.outputs[0].name)
+            self.operatorRepresentation['data_in'] = data_in.name
+            self.operatorRepresentation['data_out'] = data_out.name
+            # axes must be a constant; extract values
+            axes_buf = ctxt.lookup(node.inputs[1].name)
+            assert hasattr(axes_buf, 'values'), "Unsqueeze: expected constant 'axes' input for opset 13+"
+            axes_vals = np.array(axes_buf.values).astype(int).flatten().tolist()
+            self.operatorRepresentation['axes'] = axes_vals
+            # Do not deploy the axes tensor
+            axes_buf._live = False
+            axes_buf._deploy = False
 
         return ctxt, True