Import all initialized tensors as dense constants

[tungld]

This patch is to:

• Fix value attribute format for ConstantOp by using dense tensor,
• Import all initialized tensors as dense constants

By this patch, I successfully ran MNIST model (which was obtained from: https://github.com/onnx/models/tree/master/vision/classification/mnist)

Python script (Implementation of DummyBackend is not shown here):

import onnx
from onnx import numpy_helper
import numpy as np
test_data_dir = '/home/tungld/dl/dlc-misc/mnist/'
model = onnx.load(test_data_dir + 'model.onnx')

inputs = []
input_file = test_data_dir + 'test_data_set_0/input_0.pb'
tensor = onnx.TensorProto()
with open(input_file, 'rb') as f:
    tensor.ParseFromString(f.read())
inputs.append(numpy_helper.to_array(tensor))

# Run inference with onnx-mlir
outputs = DummyBackend.run_model(model, inputs)
print("Predicted output: \n")
print(outputs)

true_outputs = []
output_file = test_data_dir + 'test_data_set_0/output_0.pb'
output_tensor = onnx.TensorProto()
with open(output_file, 'rb') as f:
    output_tensor.ParseFromString(f.read())
true_outputs.append(numpy_helper.to_array(output_tensor))
print("True output: \n")
print(true_outputs)

Here is output (ignored unrelated parts):

[  1%] Built target cruntime
[  3%] Built target pyruntime
[  5%] Built target gen_onnx_decompose
[  7%] Built target gen_shape_inference
[ 10%] Built target gen_onnx_combine
[ 15%] Built target gen_onnx_rewrite
[ 21%] Built target gen_onnx
[ 28%] Built target gen_krnl_ops
[ 35%] Built target compiler
[ 35%] Built target gen_onnx_proto
[ 40%] Built target onnx_proto
[ 75%] Built target onnx
[ 78%] Built target builder
[ 80%] Built target onnf_shape_inference
[ 94%] Built target onnf_lower_frontend
[ 96%] Built target onnf_onnx_decompose
[ 98%] Built target onnf_transform
[100%] Built target onnf
Predicted output:

[array([[  975.6703  ,  -618.7241  ,  6574.5684  ,   668.02795 ,
         -917.2716  , -1671.636   , -1952.7598  ,   -61.549263,
         -777.1762  , -1439.5306  ]], dtype=float32)]
True output:

[array([[  975.6701  ,  -618.72394 ,  6574.5684  ,   668.02893 ,
         -917.27094 , -1671.6359  , -1952.7599  ,   -61.549873,
         -777.17664 , -1439.5316  ]], dtype=float32)]

[tjingrant]

Looks great, congrats on getting mnist to work!

[doru1004]

I think it's great that we now have confirmation that the code we have been generating so far is correct and can correctly run MNIST.

I have some reservations in terms of the solution of emitting large constants in the source code. We can avoid that and we should try to do so.

I am now working on a more robust way of dealing with constant inputs.

[doru1004]

LG