Error compiling custom yolov5m

General
, ,
1

I tried compiling yolov5 model, the optimization process went through but giving error at compilation.
This is the error

[error] Mapping Failed (allocation time: 19m 22s)
[critical] Node proposal_generator0 mapping is not feasible, with resources: Subclusters: 1, APUs: 1, IAs: 1, Parameters memories per subclusters: 0, ROW_PER_CUT: 0, Total I/O output memory: 10, Total input memory: 30

[error] Failed to produce compiled graph

The model script i used is

alls_lines_compile = [

    'compilation_param(proposal_generator0, resources_allocation_strategy=manual_scs_selection, number_of_subclusters=8, use_16x4_sc=disabled)\n',

    'compilation_param(proposal_generator0_concat, resources_allocation_strategy=manual_scs_selection, number_of_subclusters=1, use_16x4_sc=disabled)\n',
]

I had to add this compilation param, since a previous error failed because of using 16x4

2

Hey @nmitikiri,

That error message you’re seeing for the proposal_generator0 node is a classic resource allocation snag. It looks like the Hailo Dataflow Compiler is hitting a wall trying to fit that part of your model with the specific configuration you’ve given it:

[critical] Node proposal_generator0 mapping is not feasible, with resources: Subclusters: 1, APUs: 1, IAs: 1, Parameters memories per subclusters: 0, ROW_PER_CUT: 0, Total I/O output memory: 10, Total input memory: 30

The Short Version of the Problem:

Basically, those manual parameters you set—like turning off 16x4 mode and fixing the number of subclusters—are too restrictive for a node that naturally needs a lot of resources. Proposal generators (common in detection models) are usually resource hogs. By telling the compiler exactly how to map it, you’re not giving it enough flexibility to find a working solution. The fact that it failed with and without 16x4 mode suggests those manual overrides are the main culprit.

What I’d Suggest Trying:

  1. Drop the Manual Settings: The easiest fix is usually to let the compiler do its job. Try compiling the model without the compilation_param commands for proposal_generator0 and proposal_generator0_concat. Let it use the automatic subcluster selection—it’s designed to figure out the best global resource strategy for the whole model.
  2. Double-Check Your Model: Just to be safe, make sure your model architecture and every single layer operation are fully supported by your current Hailo compiler version. Sometimes an unsupported operation can cause cascading resource issues.
  3. If You Must Manually Tweak: If you have a critical reason for manual control, you’ll need to carefully explore different number_of_subclusters values and re-evaluate that 16x4 mode setting. You’ll definitely want to consult the Dataflow Compiler documentation tables to see what’s actually feasible for those specific layer types.

Hope this helps!

3

Hi @omria, i did try without any model script. but got the below error

[error] Mapping Failed (allocation time: 1m 34s)

[error] Failed to produce compiled graph
No successful assignments: proposal_generator0_concat errors:
	Agent infeasible


BackendAllocatorException: Compilation failed: No successful assignments: proposal_generator0_concat errors:
	Agent infeasible