How to read individual layers' weight & bias values from ONNX model?

Question

How to get weight/bias matrix values from ONNX model, I can currently get the inputs, Kernel size, stride and pad values from model.onnx. I load the model and then read the graph nodes to get the same:

import onnx
m = onnx.load('model.onnx')
print(m.graph.node)

Nitin Ashutosh · Answer 1 · 2018-09-29T07:57:11.303

6

from onnx import numpy_helper
MODEL_PATH = "....../resnet50"
_model = onnx.load(MODEL_PATH + "/model.onnx")
INTIALIZERS=_model.graph.initializer
Weight=[]
for initializer in INTIALIZERS:
    W= numpy_helper.to_array(initializer)
    Weight.append(W)

edited Sep 29 '18 at 07:57

answered Sep 20 '18 at 11:29

Nitin Ashutosh

111
5

Please add some explanation to your answer. – not2qubit Sep 27 '18 at 13:30
_model.graph.initializer will give list of all initializer of model likewise weight and bias. – Nitin Ashutosh Sep 29 '18 at 08:06
How can we achieve this in C++? – user9875189 Oct 27 '22 at 15:05
in c++, https://stackoverflow.com/a/67317076/11998382 `for (auto& tensor : graph.initializer()) auto& bytes_as_string = tensor.raw_data()` – Tom Huntington Nov 14 '22 at 02:09

score 2 · Answer 2 · answered Jul 29 '20 at 14:59

2

The following code helps you to create a state dictionary from onnx model.

import onnx
from onnx import numpy_helper
onnx_model   = onnx.load("model.onnx")
INTIALIZERS  = onnx_model.graph.initializer
onnx_weights = {}
for initializer in INTIALIZERS:
    W = numpy_helper.to_array(initializer)
    onnx_weights[initializer.name] = W

answered Jul 29 '20 at 14:59

viswanatha reddy

21
2

`onnx_weights = {i.name: numpy_helper.to_array(i) for i in onnx_model.graph.initalizer}` – xaviersjs Aug 16 '23 at 00:52

25b3nk · Accepted Answer · 2018-09-20T11:26:50.063

1

After posting an issue on official git repo, I got the answer to the question. We can access the weight values from initializers in m.graph.

weights = m.graph.initializer

To get the weight matrix, you need to use numpy_helper from onnx.

from onnx import numpy_helper
w1 = numpy_helper.to_array(weights[0])

edited Sep 20 '18 at 11:26

answered Sep 20 '18 at 06:13

25b3nk

164
2
16

I am still figuring out on how to convert the raw data into weight matrix. – 25b3nk Sep 20 '18 at 06:14

How to read individual layers' weight & bias values from ONNX model?

3 Answers3