Add script for accuracy validation

[kuke]

Resolve #5

To run this validation script, need to install Caffe2 to use its ONNX backend for the backend in the onnx/onnx repo has a bug when executing run_model(). And at the same time, need to install the latest ONNX from source, otherwise some other problems will happen.

[kuke]

E.g. to verify the fit_a_line model conversion, run

python validate.py --fluid_model ./extras/fit_a_line.inference.model/ --onnx_model ./fit_a_line.onnx

and get the output

-----------  Configuration Arguments -----------
a: 0.0
b: 1.0
batch_size: 10
expected_decimal: 6
fluid_model: ./extras/fit_a_line.inference.model/
onnx_model: ./fit_a_line.onnx
------------------------------------------------
Inference results for fluid model:
[array([[11.572977],
       [15.133337],
       [13.47306 ],
       [13.345606],
       [15.656847],
       [13.662115],
       [16.10838 ],
       [12.218645],
       [14.332058],
       [11.419915]], dtype=float32)]


Inference results for ONNX model:
Outputs(_0=array([[11.572977],
       [15.133337],
       [13.47306 ],
       [13.345606],
       [15.656848],
       [13.662115],
       [16.10838 ],
       [12.218645],
       [14.332058],
       [11.419915]], dtype=float32))


The exported model achieves 5-decimal precision.

[varunarora]

I think this is really well done.

[varunarora]

I know one is input and the other is output, but in #9, this arg is named --onnx_model_dir. Let's just make it consistent

[kuke]

OK. I will modify the argument in #9

[varunarora]

Should we add the requirement for this in our requirements.txt?

[kuke]

Caffe2 cannot be installed by pip and its installation is not that direct without Anaconda environment.

[varunarora]

Got it, just add some notes to the README then

[kuke]

OK

[varunarora]

Does this use paddle_onnx_shape or is this completely different

[kuke]

paddle_onnx_shape only changes the unknown dim from -1 to 0 (required by ONNX). And this line is used to set the batch size.

[pkuyym]

Excellent Job. Some questions:

1. FC --> matmul ? why not FC --> FC ?
2. Current weight FC.W and FC.b using a 'Constant' type, I think we should consider this field
   https://github.com/onnx/onnx/blob/master/onnx/onnx-ml.proto3#L242 which should be a general solution.

[kuke]

@pkuyym

1. No FC op in Fluid

2. I cannot find the proper way to initialize tensor, and using Constant op to store parameters also makes sense.

[varunarora]

Hey @kuke, isnt there an FC in fluid? http://paddlepaddle.org/docs/develop/api/fluid/en/layers.html#permalink-36-fc. Is this actually not an op?

[kuke]

@varunarora No. The fc layer consists of mul op and bias add op in Fluid

https://github.com/PaddlePaddle/Paddle/blob/a64edbf14e9c2f499a0dbb34417c50b8ded67c58/python/paddle/fluid/layers/nn.py#L192

[pkuyym]

LGTM