Debug error for vision maskrcnn

[anijain2305]

Torchscript error, repro is below

import torch
from torch.nn import *

class FxModule(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.register_buffer('_tensor_constant0', torch.zeros([], dtype=torch.int64))
        self.register_buffer('_tensor_constant1', torch.zeros([], dtype=torch.int64))
        self.register_buffer('_tensor_constant2', torch.zeros([], dtype=torch.int64))
        self.register_buffer('_tensor_constant3', torch.zeros([], dtype=torch.int64))
        self.register_buffer('_tensor_constant4', torch.zeros([], dtype=torch.int64))
        self.register_buffer('_tensor_constant5', torch.zeros([4, 1], dtype=torch.int64))
        self.register_buffer('_tensor_constant6', torch.zeros([4, 359613], dtype=torch.int64))



    def forward(self, primals_1, primals_2):
        view = torch.ops.aten.view(primals_2, [4, 359613]);  primals_2 = None
        split_with_sizes = torch.ops.aten.split_with_sizes(view, [269952, 67488, 16872, 4218, 1083], 1)
        getitem = split_with_sizes[0]
        getitem_1 = split_with_sizes[1]
        getitem_2 = split_with_sizes[2]
        getitem_3 = split_with_sizes[3]
        getitem_4 = split_with_sizes[4];  split_with_sizes = None
        topk = torch.ops.aten.topk(getitem, 1000, 1);  getitem = None
        getitem_6 = topk[1];  topk = None
        _tensor_constant0 = self._tensor_constant0
        add = torch.ops.aten.add(getitem_6, _tensor_constant0);  getitem_6 = _tensor_constant0 = None
        topk_1 = torch.ops.aten.topk(getitem_1, 1000, 1);  getitem_1 = None
        getitem_8 = topk_1[1];  topk_1 = None
        _tensor_constant1 = self._tensor_constant1
        add_1 = torch.ops.aten.add(getitem_8, _tensor_constant1);  getitem_8 = _tensor_constant1 = None
        topk_2 = torch.ops.aten.topk(getitem_2, 1000, 1);  getitem_2 = None
        getitem_10 = topk_2[1];  topk_2 = None
        _tensor_constant2 = self._tensor_constant2
        add_2 = torch.ops.aten.add(getitem_10, _tensor_constant2);  getitem_10 = _tensor_constant2 = None
        topk_3 = torch.ops.aten.topk(getitem_3, 1000, 1);  getitem_3 = None
        getitem_12 = topk_3[1];  topk_3 = None
        _tensor_constant3 = self._tensor_constant3
        add_3 = torch.ops.aten.add(getitem_12, _tensor_constant3);  getitem_12 = _tensor_constant3 = None
        topk_4 = torch.ops.aten.topk(getitem_4, 1000, 1);  getitem_4 = None
        getitem_14 = topk_4[1];  topk_4 = None
        _tensor_constant4 = self._tensor_constant4
        add_4 = torch.ops.aten.add(getitem_14, _tensor_constant4);  getitem_14 = _tensor_constant4 = None
        cat = torch.ops.aten.cat([add, add_1, add_2, add_3, add_4], 1);  add = add_1 = add_2 = add_3 = add_4 = None
        _tensor_constant5 = self._tensor_constant5
        index = torch.ops.aten.index(view, [_tensor_constant5, cat]);  view = _tensor_constant5 = None
        _tensor_constant6 = self._tensor_constant6
        _tensor_constant5_1 = self._tensor_constant5
        index_1 = torch.ops.aten.index(_tensor_constant6, [_tensor_constant5_1, cat]);  _tensor_constant6 = _tensor_constant5_1 = None
        _tensor_constant5_2 = self._tensor_constant5
        index_2 = torch.ops.aten.index(primals_1, [_tensor_constant5_2, cat]);  primals_1 = _tensor_constant5_2 = cat = None
        sigmoid = torch.ops.aten.sigmoid(index);  index = None
        select_1 = torch.ops.aten.select(index_2, 0, 1)
        select_5 = torch.ops.aten.select(sigmoid, 0, 1)
        return [select_1, select_5, ]


sizes =  [torch.Size([4, 359613, 4]), torch.Size([1438452, 1])]
dtypes =  [torch.float32, torch.float32]
inps = [torch.empty(size, dtype=dtype, device="cuda") for (size, dtype) in zip(sizes, dtypes)]


mod = FxModule().cuda()
ref = mod(*inps)

ts_mod = torch.jit.script(mod)
res = ts_mod(*inps)

[anijain2305]

cc @eellison Can you please take a look?

[eellison]

@anijain2305 this didn't repro for me as of 3326fa60ccfb7c717ecfacbb2a216544561f08d2 on master. In general, do you mind posting a dump of the error when you file an issue ?

[anijain2305]

I see. I will build pytorch and retry.

Yeah, I will start adding the error log as well from next time.

[anijain2305]

Issue is resolved at PyTorch master. Closing.