[JIT] Scripting arguments for call are not valid #17106

sto-chastic · 2019-02-14T08:35:07Z

🐛 Bug

I am converting my model to TorchScript in order to be able to save it and load it in C++. After solving various errors that arose along the way, I am left with a RuntimeError: arguments for call are not valid: which I do not know where it originates exactly.

To Reproduce

The latest reference to my code comes from a calling to this function:

def create_yolov3_modules(anchors, anch_mask, n_classes, ignore_thre):
    mlist = nn.ModuleList()
    mlist.append(add_conv(in_ch=3, out_ch=32, ksize=3, stride=1))
    mlist.append(add_conv(in_ch=32, out_ch=64, ksize=3, stride=2))
    mlist.append(resblock(ch=64))
    mlist.append(add_conv(in_ch=64, out_ch=128, ksize=3, stride=2))
    mlist.append(resblock(ch=128, nblocks=2))
    ...

where

def add_conv(in_ch, out_ch, ksize, stride):
    stage = nn.Sequential()
    pad = (ksize - 1) // 2
    stage.add_module('conv', nn.Conv2d(in_channels=in_ch,
                                       out_channels=out_ch, kernel_size=ksize, stride=stride,
                                       padding=pad, bias=False))
    stage.add_module('batch_norm', nn.BatchNorm2d(out_ch))
    stage.add_module('leaky', nn.LeakyReLU(0.1))
    return stage

and

class resblock(ScriptModule):

    __constants__ = ["nblocks","ch","shortcut"]

    def __init__(self, ch, nblocks=1, shortcut=True):

        super().__init__()
        self.shortcut = shortcut
        self.nblocks = nblocks
        self.ch = ch
        self.module_list = nn.ModuleList()
        self.blockt1=add_conv(self.ch, self.ch//2, 1, 1)
        self.blockt2=add_conv(self.ch//2, self.ch, 3, 1)
        for _ in range(nblocks):
            resblock_one = nn.ModuleList()
            self.blockt1
            self.blockt2
            self.module_list.append(resblock_one)

    @script_method
    def forward(self, x):
        for _ in range(self.nblocks):#in_ch, out_ch, ksize, stride
            h = x
            h = self.blockt1(h)
            h = self.blockt2(h)
            x = x + h if self.shortcut else h
        return x

produces this error:

> RuntimeError: 
> arguments for call are not valid:
>   
>   for operator aten::__interpolate(Tensor input, int? size=<default>, float[]? scale_factor=<default>, string mode=<default>, bool? align_corners=<default>) -> Tensor:
>   expected a value of type float[]? for argument 'scale_factor' but found int
>   @weak_script_method
>   def forward(self, input):
>       warnings.warn("nn.{} is deprecated. Use nn.functional.interpolate instead.".format(self.name))
>       return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners)
>                                              ~~~~~~~~~~ <--- HERE
>   
>   for operator aten::__interpolate(Tensor input, int[]? size=<default>, float[]? scale_factor=<default>, string mode=<default>, bool? align_corners=<default>) -> Tensor:
>   expected a value of type float[]? for argument 'scale_factor' but found int
>   @weak_script_method
>   def forward(self, input):
>       warnings.warn("nn.{} is deprecated. Use nn.functional.interpolate instead.".format(self.name))
>       return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners)
>                                              ~~~~~~~~~~ <--- HERE
>   
>   for operator aten::__interpolate(Tensor input, int? size=<default>, float? scale_factor=<default>, string mode=<default>, bool? align_corners=<default>) -> Tensor:
>   expected a value of type float? for argument 'scale_factor' but found int
>   @weak_script_method
>   def forward(self, input):
>       warnings.warn("nn.{} is deprecated. Use nn.functional.interpolate instead.".format(self.name))
>       return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners)
>                                              ~~~~~~~~~~ <--- HERE
>   
>   for operator aten::__interpolate(Tensor input, int[]? size=<default>, float? scale_factor=<default>, string mode=<default>, bool? align_corners=<default>) -> Tensor:
>   expected a value of type float? for argument 'scale_factor' but found int
>   @weak_script_method
>   def forward(self, input):
>       warnings.warn("nn.{} is deprecated. Use nn.functional.interpolate instead.".format(self.name))
>       return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners)
>                                              ~~~~~~~~~~ <--- HERE
> for call at:
> @weak_script_method
> def forward(self, input):
>     warnings.warn("nn.{} is deprecated. Use nn.functional.interpolate instead.".format(self.name))
>     return F.interpolate(input, self.size, self.scale_factor, self.mode, self.align_corners)
>            ~~~~~~~~~~~~~ <--- HERE

and the traceback is:

> Traceback (most recent call last):
>   File "train.py", line 204, in <module>
>     main()
>   File "train.py", line 80, in main
>     model = YOLOv3(anchors, anch_mask, n_classes, ignore_thre=ignore_thre)
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 951, in init_then_register
>     original_init(self, *args, **kwargs)
>   File "/home/Documents/yolotorch/training_implementation/models/yolov3.py", line 294, in __init__
>     self.module_list= create_yolov3_modules(anchors, anch_mask, n_classes, ignore_thre)
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 1123, in __setattr__
>     super(ScriptModule, self).__setattr__(attr, _ConstModuleList(value))
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 951, in init_then_register
>     original_init(self, *args, **kwargs)
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 1355, in __init__
>     module = _make_strong(module)
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 1254, in _make_strong
>     proxy = WeakScriptModuleProxy(mod, stubs)
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 951, in init_then_register
>     original_init(self, *args, **kwargs)
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 1191, in __init__
>     _create_methods_from_stubs(self, stubs)
>   File "/home/.conda/envs/python36_ocv_pytorch/lib/python3.6/site-packages/torch/jit/__init__.py", line 913, in _create_methods_from_stubs
>     self._create_methods(defs, rcbs, defaults)

Expected behavior

The code should produce a ScriptModule which I should be able to save.

Environment

PyTorch version: 1.0.1.post2
Is debug build: No
CUDA used to build PyTorch: 10.0.130

OS: Ubuntu 16.04.5 LTS
GCC version: (Ubuntu 5.4.0-6ubuntu1~16.04.11) 5.4.0 20160609
CMake version: version 3.5.1

Python version: 3.6
Is CUDA available: Yes
CUDA runtime version: 9.2.148
GPU models and configuration: GPU 0: GeForce GTX 1070 Ti
Nvidia driver version: 410.48
cuDNN version: Could not collect

Versions of relevant libraries:
[pip] numpy==1.12.1
[pip] torch==1.0.1.post2
[pip] torchvision==0.2.1
[conda] blas 1.0 mkl
[conda] mkl 2018.0.3 1
[conda] pytorch 1.0.1 py3.6_cuda10.0.130_cudnn7.4.2_2 pytorch
[conda] torchvision 0.2.1 py_2 pytorch

The text was updated successfully, but these errors were encountered:

eellison · 2019-02-14T17:41:43Z

We'll look into this, thanks for raising the issue.

driazati · 2019-02-14T23:07:22Z

I can't seem to repro this issue on master, the (slightly modified, look at __constants__) code below runs fine.

import torch
import torch.nn as nn

def add_conv(in_ch, out_ch, ksize, stride):
    stage = nn.Sequential()
    pad = (ksize - 1) // 2
    stage.add_module('conv', nn.Conv2d(in_channels=in_ch,
                                       out_channels=out_ch, kernel_size=ksize, stride=stride,
                                       padding=pad, bias=False))
    stage.add_module('batch_norm', nn.BatchNorm2d(out_ch))
    stage.add_module('leaky', nn.LeakyReLU(0.1))
    return stage

class resblock(torch.jit.ScriptModule):

    __constants__ = ["nblocks","ch","shortcut", "blockt1", "blockt2"]

    def __init__(self, ch, nblocks=1, shortcut=True):
        super(resblock, self).__init__()
        self.shortcut = shortcut
        self.nblocks = nblocks
        self.ch = ch
        self.module_list = nn.ModuleList()
        self.blockt1=add_conv(self.ch, self.ch//2, 1, 1)
        self.blockt2=add_conv(self.ch//2, self.ch, 3, 1)
        for _ in range(nblocks):
            resblock_one = nn.ModuleList()
            self.blockt1
            self.blockt2
            self.module_list.append(resblock_one)

    @torch.jit.script_method
    def forward(self, x):
        for _ in range(self.nblocks):#in_ch, out_ch, ksize, stride
            h = x
            h = self.blockt1(h)
            h = self.blockt2(h)
            x = x + h if self.shortcut else h
        return x

mlist = nn.ModuleList()
mlist.append(add_conv(in_ch=3, out_ch=32, ksize=3, stride=1))
mlist.append(add_conv(in_ch=32, out_ch=64, ksize=3, stride=2))
mlist.append(resblock(ch=64))
mlist.append(add_conv(in_ch=64, out_ch=128, ksize=3, stride=2))
mlist.append(resblock(ch=128, nblocks=2))
print(mlist)
r = resblock(2)
print(r.graph)

driazati · 2019-02-14T23:08:08Z

The error is related to our lack of int to float implicit conversion, e.g.

@torch.jit.script
def inner(x):
    # type: (float) -> float
    return x

@torch.jit.script
def use_float(x):
    # type: (int) -> float
    return inner(x)

So leaving this issue open to track that

sto-chastic · 2019-02-15T00:03:34Z

Thank you, I'll look into that

luisa-uber · 2019-03-05T17:53:33Z

Hi, is there an update on this issue? I also ran into this error. Here's my code snippet

import torch
from torch import nn
class CrossScale(torch.jit.ScriptModule):
    __constants__ = ['xc']

    def __init__(self):
        super(CrossScale, self).__init__()
        self.xc = nn.ModuleList((nn.Conv2d(2, 2, 2, bias=True), nn.Upsample(scale_factor=2, mode='bilinear')))

    @torch.jit.script_method
    def forward(self, x):
        cols = []
        i = 0
        for xc in self.xc:
            out = xc(x[i])
            i += 1
            cols.append(out)
        return cols


if __name__ == "__main__":
    cs = CrossScale()

Changing scale_factor=2 to scale_factor=float(2) fixes the issue.

pytorchbot added the oncall: jit Add this issue/PR to JIT oncall triage queue label Feb 14, 2019

eellison added the jit-triaged label Feb 14, 2019

suo assigned driazati Feb 14, 2019

driazati mentioned this issue Mar 6, 2019

Cast nn.Upsample.scale_factor to a float #17732

Closed

facebook-github-bot closed this as completed in 0955592 Mar 8, 2019

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[JIT] Scripting arguments for call are not valid #17106

[JIT] Scripting arguments for call are not valid #17106

sto-chastic commented Feb 14, 2019 •

edited by driazati

eellison commented Feb 14, 2019

driazati commented Feb 14, 2019

driazati commented Feb 14, 2019

sto-chastic commented Feb 15, 2019

luisa-uber commented Mar 5, 2019

[JIT] Scripting arguments for call are not valid #17106

[JIT] Scripting arguments for call are not valid #17106

Comments

sto-chastic commented Feb 14, 2019 • edited by driazati

🐛 Bug

To Reproduce

Expected behavior

Environment

eellison commented Feb 14, 2019

driazati commented Feb 14, 2019

driazati commented Feb 14, 2019

sto-chastic commented Feb 15, 2019

luisa-uber commented Mar 5, 2019

sto-chastic commented Feb 14, 2019 •

edited by driazati