Expand the model zoo (example model set) about singa HOT 39 OPEN

Examples (To be updated):
MobileNetV2
ShuffleNetV2
YOLO
Mask RCNN
Faster RCNN
InceptionNetV3

from singa.

Will Start working with MobileNetV2, InceptionNetV3

from singa.

I will work on Mask RCNN

from singa.

Hi, guys, @Shashankwer , @Alvinnyk , thanks for your help.

Before you start, please check this Jira, https://issues.apache.org/jira/projects/SINGA/issues/SINGA-509?filter=allopenissues

It lists the models which require support and for some of the models, SINGA may lack the necessary operators now.

So I hope you:

1. check the Jira carefully, to see which model we can do now(I guess is: super_resolution, ResNet101_DUC_HDC, ShuffleNet_V2, ShuffleNet_V1, inception_v2, squeezenet).
2. find these models in ONNX model zoo, download them and open them by a tool called Netron, try to understand the model and check each part refer the IR doc of ONNX.
3. try to read the code of SingaRep and SingaBackend in python/singa/soonx.py to understand how we parse and construct the SINGA model.
4. try to read the code examples/onnx, to make out how we load and do the inference for a ONNX model.
5. then you can try to support a model(in first point) by yourself.

from singa.

I will try to work on the ShuffleNet

from singa.

Hi, may I ask how do we get the Amazon AWS link to download the model itself? For example, for ResNet, the link is https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet18v1/resnet18v1.tar.gz. How do we get the link for the others? Because if I copy the link from ONNX's github, the link to download is Github's link, not Amazon AWS link. Thank you!

from singa.

For the models conversion do we also nee to train model on SINGA after conversion from onnx, and have a comparison matrix?

For example super_resolution via subpixel convolution layer was trained on BSD 300. Model is originally present in pytorch and can be converted well into SINGA using ONNX. It gives similar result for an image across pytorch, onnxruntime and SINGA. Do we need to compare the performance of the model across these platforms with the dataset they were trained on or SINGA compatibility is what we are checking at the moment?

Thanks and Regards,
Shashank Nigam

from singa.

There is no need to retrain the model.
But we need to evaluate the model over a test dataset to make sure the conversion is correct.

from singa.

Hi, may I ask how do we get the Amazon AWS link to download the model itself? For example, for ResNet, the link is https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet18v1/resnet18v1.tar.gz. How do we get the link for the others? Because if I copy the link from ONNX's github, the link to download is Github's link, not Amazon AWS link. Thank you!

https://github.com/onnx/models
I think you need to go to the github page for each model and then find the link for the model checkpoint file, which may not all be on aws.
@joddiy any comments?

from singa.

Hi, may I ask how do we get the Amazon AWS link to download the model itself? For example, for ResNet, the link is https://s3.amazonaws.com/onnx-model-zoo/resnet/resnet18v1/resnet18v1.tar.gz. How do we get the link for the others? Because if I copy the link from ONNX's github, the link to download is Github's link, not Amazon AWS link. Thank you!

Hi, @agnesnatasya , it's not necessary to use the AWS download link. As you said, for the image classification models, I found the download link from the backend test cases here. But if for other models, you can use the download link within GitHub repo directly.

from singa.

For the models conversion do we also nee to train model on SINGA after conversion from onnx, and have a comparison matrix?

For example super_resolution via subpixel convolution layer was trained on BSD 300. Model is originally present in pytorch and can be converted well into SINGA using ONNX. It gives similar result for an image across pytorch, onnxruntime and SINGA. Do we need to compare the performance of the model across these platforms with the dataset they were trained on or SINGA compatibility is what we are checking at the moment?

Thanks and Regards,
Shashank Nigam

For this version of SINGA, we don't need to consider the matter of training(or retaining), we just need to test if we can run the model without errors, and the result is correct. In the next version(what we are doing now), the ONNX of SINGA will support retraining.

So for now, we just need to do three things:

1. load the model correctly without errors.
2. verify the model by using the test dataset within the model download package. If you Download (with sample test data), you can see a directory called test_data_set_0, and you can verify the result of the model by using these test data.
3. for demo, we need to do the inference by using a human-readable data(or image), so you need to find a public image and then use the model to do the inference, finally show the result.

from singa.

@joddiy Thank you for the reply, I could find the link to download, and I have completed SqueezeNet. May I ask how do I test if my conversion is already correct? Thank you!

from singa.

@joddiy Thank you for the reply, I could find the link to download, and I have completed SqueezeNet. May I ask how do I test if my conversion is already correct? Thank you!

you can use this code to verify the model by using its test data set:

    # verifty the test dataset
    from utils import load_dataset
    inputs, ref_outputs = load_dataset(
        os.path.join('/tmp', 'resnet100', 'test_data_set_0'))
    x_batch = tensor.Tensor(device=dev, data=inputs[0])
    outputs = model.forward(x_batch)
    for ref_o, o in zip(ref_outputs, outputs):
        np.testing.assert_almost_equal(ref_o, tensor.to_numpy(o), 4)

from singa.

Hi,

Can anyone verify if for SuperResolutionNet, attached notebook would be sufficient?

Thanks and Regards,
Shashank Nigam
SuperResolution.zip

from singa.

Hi,

Shufflenetv2 for average pooling layer the conversion from onnx fails on below error:
ValueError: Not implemented yet for count_include_pad or ceil_mode

with below stack trace

ValueError Traceback (most recent call last)
in ()
----> 1 sg_ir = sonnx.prepare(onnx_model, device=dev)

3 frames
/usr/local/lib/python3.7/site-packages/singa/sonnx.py in prepare(cls, model, device, **kwargs)
2033 opset_version = 1
2034 weights, singa_ops = cls._onnx_model_to_singa_net(
-> 2035 model, init_inputs, device, opset_version)
2036 return SingaRep(model, weights, singa_ops,
2037 cls.keep_initializers_as_inputs)

/usr/local/lib/python3.7/site-packages/singa/sonnx.py in _onnx_model_to_singa_net(cls, model, init_inputs, device, opset_version)
1971 ]
1972 handle, forward = cls._onnx_node_to_singa_op(
-> 1973 node, inputs, opset_version)
1974 # if it is Constant, we hanlde it as a weight
1975 # otherwise, we run it and add its output into map for being used by later operators

/usr/local/lib/python3.7/site-packages/singa/sonnx.py in _onnx_node_to_singa_op(cls, onnx_node, inputs, opset_version)
1816 else:
1817 translator = cls._common_onnx_node_to_singa_op
-> 1818 return translator(onnx_node, inputs, opset_version)
1819
1820 classmethod

/usr/local/lib/python3.7/site-packages/singa/sonnx.py in _create_max_avg_pool(cls, onnx_node, inputs, opset_version)
1635 if "count_include_pad" in onnx_node.attrs or "ceil_mode" in onnx_node.attrs:
1636 raise ValueError(
-> 1637 "Not implemented yet for count_include_pad or ceil_mode")
1638
1639 # only support 2d

ValueError: Not implemented yet for count_include_pad or ceil_mode.

Kindly let us know the further steps to be taken

Thanks and Regards,
Shashank Nigam

from singa.

@Shashankwer I got the same error too, there is a "not implemented error" too when I try to run ShufflenetV1

I got a floating point exception: 8 when loading GPT2, does anyone know what causes this?

from singa.

@joddiy Hi, can I ask, for the other models that is not in ONNX, how do we find the reliable reference for it? I look at the papers, and they usually only discuss about the general idea and comparison between models, but not the details on operations to be done on every layer.

from singa.

@Shashankwer I got the same error too, there is a "not implemented error" too when I try to run ShufflenetV1

I got a floating point exception: 8 when loading GPT2, does anyone know what causes this?

Hi @agnesnatasya ,

For ShufflenetV1 issue is it requires cuda gpu for implementation. It makes use of grouped 1x1 convolution which does not seem to be supported by cpu module and is written only for cuda. It refers get_default_device as '-1' and gpu devices starting from 0. For me it failed on the check condition
device.id()==-1 in sonnx.py.
Running the same on colab did resolve the issue.

Thanks,
Shashank

from singa.

@Shashankwer Oh, I see, thank you for that!

from singa.

@joddiy Hi, can I ask, for the other models that is not in ONNX, how do we find the reliable reference for it? I look at the papers, and they usually only discuss about the general idea and comparison between models, but not the details on operations to be done on every layer. If I look at https://modelzoo.co/, there are a lot of different implementations. How do we choose the one that we could reference from? Thank you!

from singa.

@joddiy Hi, can I ask, for the other models that is not in ONNX, how do we find the reliable reference for it? I look at the papers, and they usually only discuss about the general idea and comparison between models, but not the details on operations to be done on every layer. If I look at https://modelzoo.co/, there are a lot of different implementations. How do we choose the one that we could reference from? Thank you!

Typically, when an author published his paper, he usually had to release his code at the same time. So, I guess if you cannot find the code on the GitHub, you can try to send an email to the author to ask for the code.

from singa.

@Shashankwer I got the same error too, there is a "not implemented error" too when I try to run ShufflenetV1

I got a floating point exception: 8 when loading GPT2, does anyone know what causes this?

@Shashankwer @agnesnatasya , sorry for the late, I saw this reply just now. Yes, as @agnesnatasya has said, some features in ONNX have not been implemented in SINGA. For these errors, please skip this model. I guess we will implement these features soon.

from singa.

Hi @joddiy,

For Shufflenetv2 the model conversion is failing due to use of ceil_mode. Although the value is set to false in onnx model (0) its still failing at a condition.

I have opened a new issue for the same. If we can split the condition checking

if "count_include_pad" in onnx_node.attrs or "ceil_mode" in onnx_node.attrs

to

if "ceil_mode" in onnx_node.attrs and onnx_node.attrs["ceil_mode"]:
  RaiseError
if "count_include_pad" in onnx_node.attrs:
  RaiseError

Shufflenetv2 can be successfully converted to singa.

from singa.

Hi @joddiy,

For Shufflenetv2 the model conversion is failing due to use of ceil_mode. Although the value is set to false in onnx model (0) its still failing at a condition.

I have opened a new issue for the same. If we can split the condition checking

if "count_include_pad" in onnx_node.attrs or "ceil_mode" in onnx_node.attrs

to

if "ceil_mode" in onnx_node.attrs and onnx_node.attrs["ceil_mode"]:
  RaiseError
if "count_include_pad" in onnx_node.attrs:
  RaiseError

Shufflenetv2 can be successfully converted to singa.

Thanks, @Shashankwer , you can open a PR update like this:

ceil_mode = onnx_node.getattr("ceil_mode", 0)
count_include_pad = onnx_node.getattr("count_include_pad", 0)
if ceil_mode != 0 or count_include_pad != 0:
...

from singa.

@joddiy Hi, can I ask, for the other models that is not in ONNX, how do we find the reliable reference for it? I look at the papers, and they usually only discuss about the general idea and comparison between models, but not the details on operations to be done on every layer. If I look at https://modelzoo.co/, there are a lot of different implementations. How do we choose the one that we could reference from? Thank you!

Typically, when an author published his paper, he usually had to release his code at the same time. So, I guess if you cannot find the code on the GitHub, you can try to send an email to the author to ask for the code.

Hi @joddiy ,

Do we need to prepare the model architecture or train it as well? For training what dataset should we include? (some models are trained on imagenet where dataset can be huge).

from singa.

@joddiy Hi, can I ask, for the other models that is not in ONNX, how do we find the reliable reference for it? I look at the papers, and they usually only discuss about the general idea and comparison between models, but not the details on operations to be done on every layer. If I look at https://modelzoo.co/, there are a lot of different implementations. How do we choose the one that we could reference from? Thank you!

Typically, when an author published his paper, he usually had to release his code at the same time. So, I guess if you cannot find the code on the GitHub, you can try to send an email to the author to ask for the code.

Hi @joddiy ,

Do we need to prepare the model architecture or train it as well? For training what dataset should we include? (some models are trained on imagenet where dataset can be huge).

Sure, all these works have been done in this PR #703 , if you are interested at it, you can check the new SONNX code.

from singa.

Hi @joddiy I tried to implement InceptionV1, but it involves LRN thus it is not yet supported. I tried to implement InceptionV2 but I received this error

F0606 13:54:23.837922 336954816 tensor.cc:1177] Check failed: size == t.Size() / t.shape(axis) (196 vs. 169) The size of all axis should  be the same except the concatenated axis
*** Check failure stack trace: ***
Abort trap: 6

May I ask if there is any other model (outside ONNX) that I could implement? Thank you!

from singa.

Thanks @agnesnatasya , I'll test the InceptionV2 again to check the error.

And thanks you all for your contribution! @Shashankwer @agnesnatasya @Alvinnyk

For the next plan, I list some necessary operators we have to implement:

Type 1

Abs 1h
Floor 1h

Type 2

Exp 1h
Round 6h
Pad 6h
Expand 6h
Upsample 6h

Type 3

SpaceToDepth 12h
ScatterElements 12h
RoiAlign 12h
NonMaxSuppression 12h
Resize 12h

Type 4

TopK c++
ReduceMin c++
ReduceMax c++

Type 1 means the operators have been implemented in autograd.py, we only need to add it into the sonnx.py.
Type 2 and Type 3 mean the operators don't exist in autograd.py, we need to implement them first. Type 2 is easier but Type 3 is harder.
Type 4 need to modify the c++ code, we'll consider it later.
Xh means the workload hours.

So the plan is:

1. You should read the code of soonx.py under dev branch. Because we have updated the soonx to support the re-training, so it's a little different from the old version.
2. You try to add an operator at the Type 1, you may add a key in the SingaBackend._rename_operators, and if you need a function to modify the operator, add a key in the SingaBackend._special_operators to indicate such a function.
3. Try to add the test cases to test your new operator, the test cases list at here, and you should check the code in test_onnx_backend.py. There are two patterns, the include pattern means the operators which do not in it will be excluded, and the exclude patterns will exclude the pattern in it.
4. When you finish the Type 1, I guess you must have been familiar with the sonnx, so you can try the Type 2 and Type 3. These need you to add operators in autograd.py, you can read the code in it and add test cases in test_operation.py. In the autograd.py, you need to add a class with forward and backward functions. and you need to add a function to call the class's forward function. Don't forget to add test cases in test_operation.py. After you add the operator in the autograd, you can add it to the sonnx following the point 2 and 3.

I know it's a little difficult at first. So I hope you don't be urgent to do, just read the code carefully first and if you have any questions, feel free to ask me.

By the way, please add your code to dev branch. And please skip the frontend code in sonnx.py, only read the backend one. I'm working on the frontend to upgrade it.

from singa.

Hi, you can follow my PR as an example.

#734

and

#736

from singa.

Hi @joddiy,

For The Type I operator floor is not present in autograd.py can we implement the same in autograd.py on taking the implementation of Ceil as the reference. Also Equal is implemented in the existing code.

Thanks,
Shashank

from singa.

from singa.

Hi @joddiy,

For The Type I operator floor is not present in autograd.py can we implement the same in autograd.py on taking the implementation of Ceil as the reference. Also Equal is implemented in the existing code.

Thanks,
Shashank

Yes, @Shashankwer , you can create a new operator in autograd.py by using this func singa.Floor

from singa.

Hi,

Exp operator is already implemented in autograd.py

Thanks and Regards,
Shashank Nigam

from singa.

Hi,

Exp operator is already implemented in autograd.py

Thanks and Regards,
Shashank Nigam

Hi, @Shashankwer , the Type 1 means the operators have been implemented in autograd.py, we only need to add it into the sonnx.py.

You can follow my PR to see how to add an operator to sonnx.py. By the way, you should test the test_onnx_backend.py.

@Shashankwer @Alvinnyk @agnesnatasya
There are two points I should explain a little more:

onnx_node.set_attr_inputs(onnx_node.inputs[1], 'depth')
onnx_node.set_weight_inputs(onnx_node.inputs[2], 'scale')

set_attr_inputs means in ONNX Operator Schemas, they regard an element as input, but we regard it as an attribute, so we mark the second input onnx_node.inputs[1] as attribute depth in SINGA.

set_weight_inputs means in ONNX Operator Schemas, they regard an element as input and they store its value, but we regard it as a weight in layer.py. so we mark the third input onnx_node.inputs[2] as weight scale in SINGA.

from singa.

Hi,

Just one question, does apache singa CTensor support advanced indexing as what numpy supports?

Thanks,
Shashank

from singa.

Hi,

Just one question, does apache singa CTensor support advanced indexing as what numpy supports?

Thanks,
Shashank

cannot yet. I guess for this operator, you can use tensor.to_numpy(tensor.from_raw_tensor(CTensor)) to convert it to numpy array to do it for now. We will consider to implement it at c++ end later.

from singa.

Hi,

Will work on RoiAlign. Also can GatherElements be implemented similar to ScatterElements?

Thanks and Regards,
Shashank Nigam

from singa.

Hi,

Will work on RoiAlign. Also can GatherElements be implemented similar to ScatterElements?

Thanks and Regards,
Shashank Nigam

Thanks @Shashankwer . I just saw your PR of ScatterElements.

I guess you can implement the NonMaxSuppression firstly, it's easier than RoiAlign. There are some operators we cannot support for RoiAlign, for example, the crop. But the NonMaxSuppression is straightforward now.

You can follow this one:

https://github.com/pytorch/vision/blob/master/torchvision/ops/boxes.py

And for RoiAlign, we have found an implement on GitHub, we're discussing how to implement on python or c++.

https://github.com/longcw/RoIAlign.pytorch

from singa.

Thanks for the reference will try implementing NonMaxSuppression first

from singa.