[feature request] Support output size parameter in Upsample ONNX operator

[aachigorin]

Right now Upsample in onnx supports fixed (calculated for specific input size) scale factors. Which are calculated during graph tracing based on current input and output sizes.

At the same time F.upsample allows to specify exact output size, which is very convenient if we want to evaluate models with backward connections (like Feature Pyramid Network) on inputs with varying sizes.

Could we extend onnx Upsample operator to support desired output size (exactly like in F.upsample), taken as additional input to the operator?

Simple example to check (this should export the graph with upsample taking its output size dynamically as h1.size()[-2:]):

class SimpleNet(nn.Module):
    def __init__(self):
        super(SimpleNet, self).__init__()
        self.conv1 = nn.Conv2d(3, 8, kernel_size=3, stride=1, padding=1)
        self.conv2 = nn.Conv2d(3, 8, kernel_size=3, stride=2, padding=1)
        
    def forward(self, x):
        h1 = self.conv1(x)
        h2 = F.upsample(x, size=h1.size()[-2:], mode='bilinear', align_corners=False)        
        h = torch.cat([h1, h2], dim=1)
        return h


model = SimpleNet()
save_path = '/Users/aachigorin/temp/test_bilinear_onnx_export.txt'
x = Variable(torch.randn(1, 3, 10, 10)).cpu()
torch_out = torch.onnx._export(model, x, save_path, export_params=True, verbose=True)
print('finished onnx export')

model = onnx.load(save_path)
print('check = {}'.format(onnx.checker.check_model(model)))
print(onnx.helper.printable_graph(model.graph))

cc @BowenBao @neginraoof

[ahirner]

Dynamic upsampling is a PR on onnx side: onnx/onnx#1467

[wadimkehl]

Any update on this, seeing that the dependent PR has been merged? The code from above still produces an error with the latest nightly conda package 1.0.0.dev20181204-py3.6_cuda9.0.176_cudnn7.4.1_0

raise RuntimeError("ONNX symbolic expected a constant value in the trace")

Running the code with 0.41 gives me
ONNX export failed on upsample_bilinear2d because align_corners == True not supported
although align_corners=False. This is probably another bug.

Lastly, if I run the code with 0.41 and mode='nearest', I get
TypeError: 'torch._C.Value' object does not support indexing (occurred when translating upsample_nearest2d)

So it seems that this is still a persistent problem.

[glenn-jocher]

I'm experiencing the same problem when exporting from PyTorch 1.0.0 to ONNX to CoreML. Any known solution?

[Kay-Tian]

I'm experiencing the same issue, and by the way, the nightly version of '1.2.0.dev20190725+cpu' raise another fault：
/anaconda3/envs/torch-nightly/lib/python3.6/site-packages/torch/nn/modules/conv.py(342)conv2d_forward() 341 return F.conv2d(input, weight, self.bias, self.stride, --> 342 self.padding, self.dilation, self.groups) 343 Illegal instruction (core dumped)

[soumith]

@Kay-Tian this will be fixed in tomorrow's nightlies

[garymm]

F.upsample has been deprecated in favor of torch.nn.functional.interpolate, and the Upsample ONNX op has been deprecated since ONNX op set 10. The exporter now emits Resize instead.
Does this code produce the output you are asking for with the latest nightly pytorch? If not, could you please explain what you expect to see vs what you see?

import torch
from torch import nn

class SimpleNet(nn.Module):                                                                                                                    
    def __init__(self):                                                                                                                        
        super(SimpleNet, self).__init__()                     
        self.conv1 = nn.Conv2d(3, 8, kernel_size=3, stride=1, padding=1)                                                                       
                                                                                                                                                                                                                                                                                                        
    def forward(self, x):                                      
        h1 = self.conv1(x)                                                                                                                     
        h2 = nn.functional.interpolate(x, size=h1.size()[-2:], mode='bilinear', align_corners=False)                                              
        h = torch.cat([h1, h2], dim=1)                                                                                                         
        return h                                                
                                                                       
                                                                                                                                                                                                                                                                                                           
model = SimpleNet()                                                                                                                                                                                                                                                                                    
x = torch.randn(1, 3, 10, 10).cpu()                                                                                                            
torch.onnx.export(model, x, "/dev/null", verbose=True, opset_version=13)

If it's not correct, what if you replace the last line with?

torch.onnx.export(torch.jit.script(model), x, "/dev/null", example_outputs=model(x), verbose=True, opset_version=13)

[garymm]

Closing due to lack of replies. If this is still an issue please open a new issue.