ezkl cli gen-settings: [graph] a node is missing required params: padding

[pdibitonto]

I was testing ezkl with a publicly available model YOLOv8s-worldv2 but I'm receiving an error when I fire the gen-settings command, the error is the following: [graph] a node is missing required params: padding

Steps to reproduce the bug

To reproduce the bug:

1. Download the YOLOv8s-worldv2 model here

2. Export the model to onnx with the following python script:

   from ultralytics import YOLO
   model = YOLO("yolov8s-worldv2.pt")
   model.export(format = "onnx")

3. In the same directory, try the gen-settings cli command as follow:

   ezkl gen-settings --model yolov8s-worldv2.onnx --input-visibility hashed --output-visibility public

Device and Operating System

OS: Ubuntu 22.04.4 Version: 22.04.4 LTS Processor: Intel® Core™ i7-10750H Memory: 32 GB

Additional Information

1. Python version: 3.8.16+
2. Required python module for onnx export: ultralytics

[pdibitonto]

Dunno if those details are useful, but I've printed the model (already exported to onnx) graph padding values and they are presents for each node, I've used this script:

import onnx

# Load the ONNX model
model = onnx.load("yolov8s-worldv2.onnx")

# Access the graph
graph = model.graph

for node in graph.node:
    if node.op_type == 'Conv':  # Look for Convolution nodes
        print(f"Node name: {node.name}")
        for attr in node.attribute:
            if attr.name == 'pads':  # 'pads' attribute contains padding info
                print(f"Padding: {attr.ints}")

Let me know if others infos are needed

[lnm98]

Here's a script that creates a tiny onnx file which reproduces the problem:

import tensorflow as tf
import tf2onnx
import onnx

model = tf.keras.Sequential()
model.add(tf.keras.layers.MaxPooling2D(
    pool_size=(4, 4), strides=None, padding="same",
    )
)
model.output_names=['output']
input_signature = [tf.TensorSpec([20, 20, 1, 1], tf.float32, name='x')]
onnx_model, _ = tf2onnx.convert.from_keras(model, input_signature, opset=13)

onnx.save(onnx_model, "samepadding.onnx")

The bug seems to be occurring because the onnx parser doesn't handle pooling ops with size-preserving padding: https://github.com/zkonduit/ezkl/blob/64fbc8a1c9cca0aaee006208ab0484e136c71a73/src/graph/utilities.rs#L99-L106

[alexander-camuto]

@pdibitonto @lnm98 patching this asap :)