🚀 FlowSentry-Wake | Stage 2 Update

EdgeFlowNet Optical Flow. Successfully Deployed On Board

1️⃣ EdgeFlowNet Is Now Running on the Board

We’ve successfully compiled and deployed EdgeFlowNet,

a modern encoder–decoder optical flow model,

directly on the Orange Pi 5 Plus + Metis M.2 platform.

2️⃣ Calibration for Optical Flow Is Not “Plug-and-Play”

Unlike single-image models, optical flow operates on consecutive frame pairs.

To make this work inside the Voyager SDK calibration flow,

we implemented a custom DataAdapter dedicated to optical flow:

• Takes two consecutive RGB frames
• Concatenates them into a 6-channel tensor
• Ensures layout, normalization, and ordering match model expectations

This adapter is used only for calibration,

but it is critical for correct quantization and deployment.

3️⃣ A Mysterious Shape Mismatch — and How We Found It

During deployment, we encountered a confusing Relay type error:

Error: The Relaytype checker is unable to show the following types match:
Tensor[(192,260,4,64),int8]
Tensor[(192,256,1,64),int8]

This dimensional mismatch could not be traced back to the model itself despite multiple confirmations

In the end, we found the right way to debug this in compiler_configs_full.md.

To locate the issue, we enabled two debug options in the model’s YAML file.

compilation_config:
	save_error_artifact:true
	trace_tvm_passes:true

This exposed the real culprit:

👉 ICR was enabled by default.

The fix was explicit and clean:

compilation_config:
    enable_icr:false
    enable_swicr:false

Once disabled, the mismatch disappeared.

4️⃣ Non-Symmetric Padding in Transposed Convolutions

EdgeFlowNet’s decoder contains several stride-2 transposed convolutions with odd kernels.

This leads to non-symmetric padding, for example:

• Padding: [0, 0, 1, 1]

However, AIPU requires symmetric padding.

Retraining the model was not an option —

we wanted to preserve the original weights.

So we applied a targeted transformation strategy:

• Detect non-symmetric padding
• Convert it into:
  • Symmetric padding
  • Explicit output padding compensation

Example:

Before the change:

Padding was set to [0, 0, 1, 1]. Output padding was [0, 0]. AutoPad was set to SAME_UPPER.

After the change:

Padding was adjusted to [1, 1, 1, 1]. Output padding became [1, 1]. AutoPad was changed to NOTSET.

This satisfies hardware constraints

without changing the final output behavior or performance.

A pure deployment-side fix.

5️⃣ Where We Are Now

At this point, we have confirmed:

• EdgeFlowNet compiles successfully
• Runs fully on-board
• Uses a custom calibration pipeline
• Passes hardware constraints without retraining

This is a solid foundation.

🔜 Next Stage: Making It Useful

Coming up next:

• Optical flow visualization
• Runtime cost analysis
• Integration into the FlowSentry-Wake multi-stage perception framework
• Coordinated operation of:
  • Object detection
  • Motion estimation
  • Power-aware stage switching

This is where the system-level design really starts to shine.

Stay tuned.

!-->