Squeezenet1.0 models give wrong prediction results

[mingmingtasd]

Describe the issue

Move the onnx/models#659 here:

All squeezenet 1.0 models from https://github.com/onnx/models/tree/main/validated/vision/classification/squeezenet can't provide correct prediction results.

To reproduce

Select any one squeezenet 1.0 model to try:

import onnx
import onnxruntime
import numpy as np
from PIL import Image

# Load SqueezeNet ONNX
model_path = 'squeezenet1.0-12-fp32.onnx'
model = onnx.load(model_path)

# Create ONNX session
session = onnxruntime.InferenceSession(model_path)

# Load image
image_path = 'dog.jpg'
image = Image.open(image_path)
image = image.resize((224, 224)) 

# Image preprocessing
image = np.array(image).astype(np.float32)
image /= 255.0 

# Normalize image
mean = [0.485, 0.456, 0.406]
std = [0.229, 0.224, 0.225]
image = (image - mean) / std

image = np.transpose(image, (2, 0, 1))  # Adjust channel order of the image
image = np.expand_dims(image, axis=0)  # Add batch dimension
# convert the input tensor to float type
image = image.astype(np.float32)

# Predict
input_name = session.get_inputs()[0].name
output_name = session.get_outputs()[0].name
input_feed = {input_name: image}
output = session.run([output_name], input_feed)

# Load the labels file
labels_path = 'synset.txt'
with open(labels_path, 'r') as f:
    labels = f.read().splitlines()

# Get results
predicted_idx = np.argmax(output[0])
predicted_label = labels[predicted_idx]

print("Predicted label:", predicted_label)
...

The prediction result is always:

Predicted label: n03788365 mosquito net

You can also use onnxruntime web demo to reproduce: use squeezenet 1_0 rather than 1_1 in https://github.com/microsoft/onnxruntime-web-demo/blob/main/src/components/models/Squeezenet.vue

Platform

Windows

OS Version

11

[fdwr]

@mingmingtasd This is marked as "EP == DirectML", but shouldn't this be the WebNN EP? I'm confused by the issue because it includes Python code, points to an onnxruntime web demo link for the repro, but also says DML EP (which isn't used in ORTW).

[mingmingtasd]

@mingmingtasd This is marked as "EP == DirectML", but shouldn't this be the WebNN EP? I'm confused by the issue because it includes Python code, points to an onnxruntime web demo link for the repro, but also says DML EP (which isn't used in ORTW).

I firstly opened an issue : onnx/models#659 , but someone told me to open the issue to onnxruntime. Sorry for confusion, I give the onnxruntime web demo and python code mainly to provide two ways to reproduce.

[mingmingtasd]

The label should be webnn ep, but it seems I can't change the label...

[fdwr]

The label should be webnn ep, but it seems I can't change the label...

I guess you have to be part of microsoft/onnxruntime to change labels - I changed it 👍.

Did you already compare against the CPU EP and DML EP, and did those return different results than the WebNN EP? If so, then @Honry will want to be aware.

[github-actions]

This issue has been automatically marked as stale due to inactivity and will be closed in 30 days if no further activity occurs. If further support is needed, please provide an update and/or more details.