add_graph exception on autoencoder network

[Nikronic]

This is my network architecture:

# %% Import libraries
import torch
import torch.nn as nn
import torch.nn.functional as F

# %% Submodules
class CL(nn.Module):
    def __init__(self, input_channel, output_channel):
        """
        It consists of the 4x4 convolutions with stride=2, padding=1, each followed by
        a leaky rectified linear unit (Leaky ReLU)

        :param input_channel: input channel size
        :param output_channel: output channel size
        """

        assert (input_channel > 0 and output_channel > 0)

        super(CL, self).__init__()
        layers = [nn.Conv2d(input_channel, output_channel, kernel_size=4, stride=2, padding=1), nn.LeakyReLU(0.2)]
        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        return self.layers(x)

# %%
class CBL(nn.Module):
    def __init__(self, input_channel, output_channel):
        """
        It consists of the 4x4 convolutions with stride=2, padding=1, and a batch normalization, followed by
        a leaky rectified linear unit (ReLU)

        :param input_channel: input channel size
        :param output_channel: output channel size
        """
        assert (input_channel > 0 and output_channel > 0)

        super(CBL, self).__init__()
        layers = [nn.Conv2d(input_channel, output_channel, kernel_size=4, stride=2, padding=1),
                  nn.BatchNorm2d(num_features=output_channel), nn.LeakyReLU(0.2)]
        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        return self.layers(x)

# %%
class CE(nn.Module):
    def __init__(self, input_channel, output_channel, ks=4, s=2):
        """
        It consists of the 4x4 convolutions with stride=2, padding=1, each followed by
        a exponential linear unit (ELU)

        :param input_channel: input channel size
        :param output_channel: output channel size
        :param ks: kernel size
        :param s: stride size
        """
        assert (input_channel > 0 and output_channel > 0)

        super(CE, self).__init__()
        layers = [nn.ConvTranspose2d(input_channel, output_channel, kernel_size=ks, stride=s, padding=1),
                  nn.ELU(alpha=1)]
        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        return self.layers(x)

# %%
class Contract(nn.Module):
    def __init__(self, input_channel, output_channel, module='cbl'):
        """
        It consists of a CL or CBL followed by a 2x2 MaxPooling operation with stride 2 for down sampling.

        :param input_channel: input channel size
        :param output_channel: output channel size
        :param module: using Convolution->ReLU (CL class) or Convolution->BathNorm->ReLU (CBL class)
                Convolution->ELU (CE class) for first layer of Expand (decoder) path
        """

        assert (input_channel > 0 and output_channel > 0)

        super(Contract, self).__init__()

        layers = []
        if module == 'cl':
            layers.append(CL(input_channel, output_channel))
        elif module == 'ce':
            layers.append(CE(input_channel, output_channel))
        else:
            layers.append(CBL(input_channel, output_channel))

        self.layers = nn.Sequential(*layers)

    def forward(self, x):
        return self.layers(x)

# %%
class Expand(nn.Module):
    def __init__(self, input_channel, output_channel, ks=4, s=2):
        """
        This path consists of an up sampling of the feature map followed by a
        4x4 convolution ("up-convolution" or Transformed Convolution) that halves the number of
        feature channels, a concatenation with the correspondingly cropped feature map from Contract phase

        :param input_channel: input channel size
        :param output_channel: output channel size
        """
        super(Expand, self).__init__()
        self.layers = CE(input_channel * 2, output_channel, ks, s)

    def forward(self, x1, x2):
        delta_x = x1.size()[2] - x2.size()[2]
        delta_y = x1.size()[3] - x2.size()[3]
        x2 = F.pad(x2, pad=(delta_x // 2, delta_y // 2, delta_x // 2, delta_y // 2), mode='constant', value=0)
        x = torch.cat((x2, x1), dim=1)
        x = self.layers(x)
        return x

# %%
class C(nn.Module):
    def __init__(self, input_channel, output_channel):
        """
        At the final layer, a 3x3 convolution is used to map each 64-component feature vector to the desired
        number of classes.

        :param input_channel: input channel size
        :param output_channel: output channel size
        """
        super(C, self).__init__()
        layers = [nn.Conv2d(input_channel, output_channel, kernel_size=3, padding=1, stride=1), nn.Sigmoid()]
        self.layer = nn.Sequential(*layers)

    def forward(self, x):
        return self.layer(x)

#%% Main CLass
class CoarseNet(nn.Module):
    def __init__(self, input_channels=3, output_channels=3):
        """
        Implementation of CoarseNet, a modified version of UNet.
        (https://arxiv.org/abs/1505.04597 - Convolutional Networks for Biomedical Image Segmentation (Ronneberger et al., 2015))

        :param input_channels: number of input channels of input images to network.
        :param output_channels: number of output channels of output images of network.
        """

        super(CoarseNet, self).__init__()
        self.input_channels = input_channels
        self.output_channels = output_channels

        # Encoder
        self.cl0 = Contract(input_channels, 64, module='cl')
        self.cbl0 = Contract(64, 128)
        self.cbl1 = Contract(128, 256)
        self.cbl2 = Contract(256, 512)
        self.cl1 = Contract(512, 512, module='cl')

        # Decoder
        self.ce0 = Contract(512, 512, module='ce')
        self.ce1 = Expand(512, 256)
        self.ce2 = Expand(256, 128)
        self.ce3 = Expand(128, 64)
        self.ce4 = Expand(64, 64)
        self.ce5 = CE(64, 64, ks=3, s=1)

        # final
        self.final = C(64, self.output_channels)

    def forward(self, x):
        out = self.cl0(x)  # 3>64
        out2 = self.cbl0(out)  # 64>128
        out3 = self.cbl1(out2)  # 128>256
        out4 = self.cbl2(out3)  # 256>512
        out5 = self.cl1(out4)  # 512>512
        in0 = self.ce0(out5)

        in1 = self.ce1(out4, in0)  # 512>512
        in2 = self.ce2(out3, in1)  # 512>256
        in3 = self.ce3(out2, in2)  # 256>128
        in4 = self.ce4(out, in3)  # 128>64
        f = self.ce5(in4)
        f = self.final(f)
        return f

And I am trying to show graph in tensorboard using this code:

from tensorboardX import SummaryWriter
dummy_input = (torch.rand(1, 3, 256, 256),)

with SummaryWriter() as w:
    w.add_graph(CoarseNet(input_channels=3, output_channels=1), dummy_input, True)

And the result:

RuntimeError                              Traceback (most recent call last)
<ipython-input-8-cc778a7f1d82> in <module>()
     11 
     12 with SummaryWriter(comment='LinearInLinear') as w:
---> 13     w.add_graph(CoarseNet(input_channels=3, output_channels=1), dummy_input, True)

/usr/local/lib/python3.6/dist-packages/tensorboardX/writer.py in add_graph(self, model, input_to_model, verbose, **kwargs)
    564                     print('add_graph() only supports PyTorch v0.2.')
    565                     return
--> 566             self.file_writer.add_graph(graph(model, input_to_model, verbose))
    567         else:
    568             # Caffe2 models do not have the 'forward' method

/usr/local/lib/python3.6/dist-packages/tensorboardX/pytorch_graph.py in graph(model, args, verbose, omit_useless_nodes)
    233             return GraphDef(versions=VersionDef(producer=22))
    234 
--> 235     _optimize_trace(trace, torch.onnx.utils.OperatorExportTypes.ONNX)
    236 
    237     graph = trace.graph()

/usr/local/lib/python3.6/dist-packages/tensorboardX/pytorch_graph.py in _optimize_trace(trace, operator_export_type)
    173 
    174     def _optimize_trace(trace, operator_export_type):
--> 175         trace.set_graph(_optimize_graph(trace.graph(), operator_export_type))
    176 
    177     def _optimize_graph(graph, operator_export_type):

/usr/local/lib/python3.6/dist-packages/tensorboardX/pytorch_graph.py in _optimize_graph(graph, operator_export_type)
    204 
    205         if operator_export_type != OperatorExportTypes.RAW:
--> 206             graph = torch._C._jit_pass_onnx(graph, operator_export_type)
    207             torch._C._jit_pass_lint(graph)
    208             # torch._C._jit_pass_onnx_peephole(graph)

/usr/local/lib/python3.6/dist-packages/torch/onnx/__init__.py in _run_symbolic_function(*args, **kwargs)
     50 def _run_symbolic_function(*args, **kwargs):
     51     from torch.onnx import utils
---> 52     return utils._run_symbolic_function(*args, **kwargs)
     53 
     54 

/usr/local/lib/python3.6/dist-packages/torch/onnx/utils.py in _run_symbolic_function(g, n, inputs, env, operator_export_type)
    502                     return None
    503                 fn = getattr(torch.onnx.symbolic, op_name)
--> 504                 return fn(g, *inputs, **attrs)
    505 
    506         elif ns == "prim":

/usr/local/lib/python3.6/dist-packages/torch/onnx/symbolic.py in wrapper(g, *args)
     86         def wrapper(g, *args):
     87             assert len(arg_descriptors) == len(args)
---> 88             args = [_parse_arg(arg, arg_desc) for arg, arg_desc in zip(args, arg_descriptors)]
     89             return fn(g, *args)
     90         # In Python 2 functools.wraps chokes on partially applied functions, so we need this as a workaround

/usr/local/lib/python3.6/dist-packages/torch/onnx/symbolic.py in <listcomp>(.0)
     86         def wrapper(g, *args):
     87             assert len(arg_descriptors) == len(args)
---> 88             args = [_parse_arg(arg, arg_desc) for arg, arg_desc in zip(args, arg_descriptors)]
     89             return fn(g, *args)
     90         # In Python 2 functools.wraps chokes on partially applied functions, so we need this as a workaround

/usr/local/lib/python3.6/dist-packages/torch/onnx/symbolic.py in _parse_arg(value, desc)
     43         return value
     44     if value.node().kind() != 'onnx::Constant':
---> 45         raise RuntimeError("ONNX symbolic expected a constant value in the trace")
     46     tval = value.node()['value']
     47     if desc == 'i':

RuntimeError: ONNX symbolic expected a constant value in the trace

Environment:

1. OS: Windows 10 enterprise build 10.0.17134.648 version 1803 X64
2. Python: 3.6.2
3. PyTorch: 1.0.1
4. TensorboardX 1.6
5. Tensorflow-tensorboad 1.5.1
6. Torchvision 0.2.2.post3
7. Tensorflow 2.0.0a0 (I also tested on 1.13.1)

[lanpa]

Which version are you using? Please follow the issue template. Thanks.

[Nikronic]

Sorry, the previous problem solved and was my fault. But I tried this one on Google colab with latest versions of all packages, and my local system with packages as I mentioned. Thanks for help

[lanpa]

So does the problem solved?

[Nikronic]

Not actually, I edited the whole post!!!

[xwjBupt]

@Nikronic did you solve the problem?i have met the same one

[Nikronic]

@xwjBupt No. I still have the same issue. I do not know why they do not answer anything about this issue, but I tried to use Visdom by Facebookresearch but they do not provide any API for graph creation, so I am trying to Impelement my model in Tensorflow and use Tensorboard to handle my problems then convert final model to PyTorch again.

If you find anything helpful, I would be appreciated if you comment. Thank you

[lanpa]

Hi, please try: w.add_graph(CoarseNet(input_channels=3, output_channels=1), dummy_input, True, operator_export_type="RAW") with the master branch.

[Nikronic]

Thank you so much.

[Nikronic]

@xwjBupt problem solved. Look at the guide provided by @lanpa