Failed internal assertion with torch::pickle_load in LibTorch 1.5.0

[ankane]

🐛 Bug

Hi, I'm running into a failed assertion with torch::pickle_load on PyTorch models.

libc++abi.dylib: terminating with uncaught exception of type c10::Error: isGenericDict() INTERNAL ASSERT FAILED at ../aten/src/ATen/core/ivalue_inl.h:715, please report a bug to PyTorch. Expected GenericDict but got None (toGenericDict at ../aten/src/ATen/core/ivalue_inl.h:715)

To Reproduce

Steps to reproduce the behavior:

1. Save a tensor and a model in PyTorch

import torch
from torch import nn

torch.save(torch.tensor([1, 2, 3]), "tensor.pt", _use_new_zipfile_serialization=True)

class TheModelClass(nn.Module):
    def __init__(self):
        super(TheModelClass, self).__init__()
        self.conv1 = nn.Conv2d(3, 6, 5)
        self.pool = nn.MaxPool2d(2, 2)
        self.conv2 = nn.Conv2d(6, 16, 5)
        self.fc1 = nn.Linear(16 * 5 * 5, 120)
        self.fc2 = nn.Linear(120, 84)
        self.fc3 = nn.Linear(84, 10)

    def forward(self, x):
        x = self.pool(F.relu(self.conv1(x)))
        x = self.pool(F.relu(self.conv2(x)))
        x = x.view(-1, 16 * 5 * 5)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

model = TheModelClass()

torch.save(model.state_dict(), "model.pt", _use_new_zipfile_serialization=True)

2. Create and compile a C++ application

#include <torch/torch.h>
#include <iostream>
#include <iterator>

int main() {
  // tensor
  std::ifstream file("../tensor.pt", std::ios::binary);
  std::vector<char> data((std::istreambuf_iterator<char>(file)), std::istreambuf_iterator<char>());
  torch::IValue ivalue = torch::pickle_load(data);
  torch::Tensor tensor = ivalue.toTensor();
  std::cout << tensor << std::endl;

  // model
  std::ifstream file2("../model.pt", std::ios::binary);
  std::vector<char> data2((std::istreambuf_iterator<char>(file2)), std::istreambuf_iterator<char>());
  torch::IValue ivalue2 = torch::pickle_load(data2);
}

3. Run it

The tensor loads as expected, but the model throws an exception.

Stack Trace

libc++abi.dylib: terminating with uncaught exception of type c10::Error: isGenericDict() INTERNAL ASSERT FAILED at ../aten/src/ATen/core/ivalue_inl.h:715, please report a bug to PyTorch. Expected GenericDict but got None (toGenericDict at ../aten/src/ATen/core/ivalue_inl.h:715)
frame #0: c10::Error::Error(c10::SourceLocation, std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > const&) + 135 (0x104aaaf47 in libc10.dylib)
frame #1: c10::IValue::toGenericDict() const & + 280 (0x104b547b8 in libtorch_cpu.dylib)
frame #2: torch::jit::Unpickler::readInstruction() + 5580 (0x10754cc1c in libtorch_cpu.dylib)
frame #3: torch::jit::Unpickler::run() + 472 (0x10754b2d8 in libtorch_cpu.dylib)
frame #4: torch::jit::Unpickler::parse_ivalue() + 34 (0x10754ae32 in libtorch_cpu.dylib)
frame #5: torch::jit::readArchiveAndTensors(std::__1::basic_string<char, std::__1::char_traits<char>, std::__1::allocator<char> > const&, c10::optional<std::__1::function<c10::StrongTypePtr (c10::QualifiedName const&)> >, c10::optional<std::__1::function<c10::intrusive_ptr<c10::ivalue::Object, c10::detail::intrusive_target_default_null_type<c10::ivalue::Object> > (c10::StrongTypePtr, c10::IValue)> >, c10::optional<c10::Device>, caffe2::serialize::PyTorchStreamReader&) + 838 (0x107580496 in libtorch_cpu.dylib)
frame #6: torch::jit::pickle_load(std::__1::vector<char, std::__1::allocator<char> > const&) + 347 (0x10758539b in libtorch_cpu.dylib)
frame #7: torch::pickle_load(std::__1::vector<char, std::__1::allocator<char> > const&) + 14 (0x1079944ee in libtorch_cpu.dylib)
frame #8: main + 341 (0x104a70745 in example-app)
frame #9: start + 1 (0x7fff79ac53d5 in libdyld.dylib)
frame #10: 0x0 + 1 (0x1 in ???)

Expected behavior

No exception.

Environment

• PyTorch Version (e.g., 1.0): 1.5.0
• OS (e.g., Linux): macOS 10.14
• How you installed PyTorch (conda, pip, source): pip
• Build command you used (if compiling from source):
• Python version: 3.7
• CUDA/cuDNN version: n/a
• GPU models and configuration: n/a

cc @ezyang @gchanan @zou3519 @suo

[dfalbel]

It seems that if you transform all nn.Parameters to true Tensors everything works as expected.

[ankane]

Cleaning up stale issues. As @dfalbel mentioned, converting Parameters to Tensors is a way around this.