Disallow shape inference (#128)

The current status quo did shape inference on _some_ layers but not all. This was done through the `computeShapes` method of the ONNXLayer and by examining the implementations of that method, we can see that some calculate the output shape out of the input shapes, but some just blindly copy the tensor's shape.

I propose we disallow shape inference and if we want to properly support this feature we can revert this change and implement it properly or depend on another tool to do it for us (e.g. [ONNX's shape inference module](https://onnx.ai/onnx/api/shape_inference.html)).

Now, I have disallowed shape inference after lowering because some lowerings actually consumed shapeless tensors and produced correct replacements, but some lowerings produced shapeless tensors themeselves, specifically, the MHSA pass in MemPool. I have disabled all tests that used those lowerings which can be re-enabled once the offending lowerings get fixed. I will create an issue following this if this PR gets accepted.

## Added
- Added assertion that all the graph tensors after lowering have a shape annotated

## Changed
- Disabled ICCT_ITA_8 MemPool test because it was using a lowering that created shapeless tensors
- Added missing shape annotation to the testTypeInferenceDifferentTypes

Author: lukamac

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

@@ -69,6 +69,5 @@ jobs:
         ICCT
         ICCT_ITA
         ICCT_8
-        ICCT_ITA_8
         miniMobileNet
         miniMobileNetv2

CHANGELOG.md

@@ -4,6 +4,7 @@ This file contains the changelog for the Deeploy project. The changelog is divid
 ## Unreleased (Planned Release Target: v0.2.1)
 
 ### List of Pull Requests
+- Disallow shape inference [#128](https://github.com/pulp-platform/Deeploy/pull/128)
 - Remove memory-aware node bindings [#123](https://github.com/pulp-platform/Deeploy/pull/123)
 - Fix missing const's layout transformation and refactor NCHWtoNHWC passes [#122](https://github.com/pulp-platform/Deeploy/pull/122)
 - Fix aliasing [#125](https://github.com/pulp-platform/Deeploy/pull/125)
@@ -51,6 +52,7 @@ This file contains the changelog for the Deeploy project. The changelog is divid
 - Per–memory-level usage tracking and worst-case reporting in `NetworkContext`
 - Memory/I/O summaries and input/output logging in deployers
 - RequantHelpers.py for Neureka's TileConstraints
+- Added assertion that all the graph tensors after lowering have a shape annotated
 
 ### Changed
 - Replaced platform-specific tags (`*-amd64`, `*-arm64`) with direct digest references in `Noelware/docker-manifest-action`.
@@ -85,6 +87,8 @@ This file contains the changelog for the Deeploy project. The changelog is divid
 - Removed NodeMemoryLevelChecker, MemoryAwareNodeBinding
 - Removed _parseNode from MemoryNetworkDeployer since we don't need the annotations before typeChecking anymore
 - Removed Wmem variants of bindings and tile constraints from Neureka
+- Disabled ICCT_ITA_8 MemPool test because it was using a lowering that created shapeless tensors
+- Added missing shape annotation to the testTypeInferenceDifferentTypes
 
 ### Fixed
 - Prevent node duplication for graphs generated via GraphSurgeon

Deeploy/DeeployTypes.py

@@ -3372,6 +3372,11 @@ def _removeIdentityNodes(self):
         for node in filter(lambda x: x.op == "Identity", self.graph.nodes):
             self.graph.deleteNode(node)
 
+    def _assertTensorsHaveShape(self) -> None:
+        missingShapes = [name for name, tensor in self.graph.tensors().items() if tensor.shape is None]
+        assert len(missingShapes) == 0, \
+            f"Shape inference is not supported.\nFound tensors with missing shape annotation: {missingShapes}"
+
     def frontEnd(self):
         """API hook to prepare the graph to be deployed and build the initial NetworkContext
 
@@ -3419,6 +3424,9 @@ def frontEnd(self):
         log.info(f"> Export State {_middlewarePostLoweringFilename}[.onnx|.pkl]")
         self.exportDeeployState(self.deeployStateDir, _middlewarePostLoweringFilename)
 
+        log.info(" - Assert all tensors have a shape annotation")
+        self._assertTensorsHaveShape()
+
         log.info("- Perform Graph Parsing")
         try:
             self.parse(self.default_channels_first)  # This reparses the lowered graph