我遇到以下代码错误。我想做的事情相当简单：我想使用 TensorBoard 可视化 PyTorch 模型。我完全不知道为什么会发生这个错误。
import torch
from torch import nn
class MyModel(nn.Module):
def __init__(self):
super().__init__()
self.mul = nn.MultiheadAttention(128, 4, batch_first=True)
def forward(self, x):
y, _ = self.mul(x, x, x)
return y
x = torch.randn(1, 240 * 3, 128)
my = MyModel()
from torch.utils.tensorboard import SummaryWriter
writer = SummaryWriter("./tensorboard_otameshi")
writer.add_graph(my, x)
writer.close()
错误是这样的。
TracingCheckError Traceback (most recent call last)
<ipython-input-1-c4876e5ab470> in <cell line: 18>()
16 from torch.utils.tensorboard import SummaryWriter
17 writer = SummaryWriter("./tensorboard_otameshi")
---> 18 writer.add_graph(my, x)
19 writer.close()
5 frames
/usr/local/lib/python3.10/dist-packages/torch/jit/_trace.py in _check_trace(check_inputs, func, traced_func, check_tolerance, strict, force_outplace, is_trace_module, _module_class, example_inputs_is_kwarg)
560 diag_info = graph_diagnostic_info()
561 if any(info is not None for info in diag_info):
--> 562 raise TracingCheckError(*diag_info)
563
564
TracingCheckError: Tracing failed sanity checks!
ERROR: Graphs differed across invocations!
Graph diff:
graph(%self.1 : __torch__.MyModel,
%x : Tensor):
%mul : __torch__.torch.nn.modules.activation.MultiheadAttention = prim::GetAttr[name="mul"](%self.1)
+ %4 : bool = prim::Constant[value=1](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1160:0
- %4 : NoneType = prim::Constant(), scope: __module.mul
? ^
+ %5 : NoneType = prim::Constant(), scope: __module.mul
? ^
- %5 : Tensor = prim::Constant[value={5.65685}](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5331:0
- %6 : int = prim::Constant[value=-2](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4767:0
- %7 : int = prim::Constant[value=3](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/_tensor.py:1189:0
- %8 : int = prim::Constant[value=-1](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4761:0
- %9 : str = prim::Constant[value="trunc"](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5208:0
- %10 : Tensor = prim::Constant[value={4}](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5208:0
- %11 : int = prim::Constant[value=2](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5164:0
- %12 : int = prim::Constant[value=0](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1183:0
? ^^ ^ ^^
+ %6 : int = prim::Constant[value=4](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1160:0
? ^ ^ ^^
- %13 : int = prim::Constant[value=1](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1183:0
? ^^ ^^
+ %7 : int = prim::Constant[value=128](), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1160:0
? ^ ++ ^^
%out_proj : __torch__.torch.nn.modules.linear.NonDynamicallyQuantizableLinear = prim::GetAttr[name="out_proj"](%mul)
%bias : Tensor = prim::GetAttr[name="bias"](%out_proj)
%out_proj.1 : __torch__.torch.nn.modules.linear.NonDynamicallyQuantizableLinear = prim::GetAttr[name="out_proj"](%mul)
%weight : Tensor = prim::GetAttr[name="weight"](%out_proj.1)
%in_proj_bias : Tensor = prim::GetAttr[name="in_proj_bias"](%mul)
%in_proj_weight : Tensor = prim::GetAttr[name="in_proj_weight"](%mul)
+ %14 : Tensor, %15 : Tensor = aten::_native_multi_head_attention(%x, %x, %x, %7, %6, %in_proj_weight, %in_proj_bias, %weight, %bias, %5, %4, %4, %5), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1160:0
- %query : Tensor = aten::transpose(%x, %13, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1183:0
- %21 : int = aten::size(%query, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5164:0
- %tgt_len : Tensor = prim::NumToTensor(%21), scope: __module.mul
- %23 : int = aten::size(%query, %13), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5164:0
- %bsz : Tensor = prim::NumToTensor(%23), scope: __module.mul
- %25 : int = aten::size(%query, %11), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5164:0
- %embed_dim : Tensor = prim::NumToTensor(%25), scope: __module.mul
- %head_dim : Tensor = aten::div(%embed_dim, %10, %9), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5208:0
- %28 : int = aten::Int(%head_dim), scope: __module.mul
- %29 : int = aten::Int(%head_dim), scope: __module.mul
- %30 : int = aten::Int(%head_dim), scope: __module.mul
- %31 : int = aten::size(%query, %8), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4761:0
- %32 : Tensor = aten::linear(%query, %in_proj_weight, %in_proj_bias), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4765:0
- %33 : int[] = prim::ListConstruct(%7, %31), scope: __module.mul
- %34 : Tensor = aten::unflatten(%32, %8, %33), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/_tensor.py:1189:0
- %35 : Tensor = aten::unsqueeze(%34, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4767:0
- %36 : Tensor = aten::transpose(%35, %12, %6), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4767:0
- %37 : Tensor = aten::squeeze(%36, %6), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4767:0
- %proj : Tensor = aten::contiguous(%37, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4767:0
- %q.1 : Tensor = aten::select(%proj, %12, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4768:0
- %k.1 : Tensor = aten::select(%proj, %12, %13), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4768:0
- %v.1 : Tensor = aten::select(%proj, %12, %11), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:4768:0
- %42 : Tensor = aten::mul(%bsz, %10), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5277:0
- %43 : int = aten::Int(%42), scope: __module.mul
- %44 : int[] = prim::ListConstruct(%21, %43, %30), scope: __module.mul
- %45 : Tensor = aten::view(%q.1, %44), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5277:0
- %q : Tensor = aten::transpose(%45, %12, %13), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5277:0
- %47 : int = aten::size(%k.1, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5279:0
- %48 : Tensor = aten::mul(%bsz, %10), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5279:0
- %49 : int = aten::Int(%48), scope: __module.mul
- %50 : int[] = prim::ListConstruct(%47, %49, %29), scope: __module.mul
- %51 : Tensor = aten::view(%k.1, %50), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5279:0
- %k : Tensor = aten::transpose(%51, %12, %13), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5279:0
- %53 : int = aten::size(%v.1, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5288:0
- %54 : Tensor = aten::mul(%bsz, %10), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5288:0
- %55 : int = aten::Int(%54), scope: __module.mul
- %56 : int[] = prim::ListConstruct(%53, %55, %28), scope: __module.mul
- %57 : Tensor = aten::view(%v.1, %56), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5288:0
- %v : Tensor = aten::transpose(%57, %12, %13), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5288:0
- %q_scaled : Tensor = aten::div(%q, %5), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5331:0
- %60 : Tensor = aten::transpose(%k, %6, %8), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5338:0
- %input : Tensor = aten::bmm(%q_scaled, %60), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5338:0
- %attn_output_weights.1 : Tensor = aten::softmax(%input, %8, %4), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:1843:0
- %attn_output.1 : Tensor = aten::bmm(%attn_output_weights.1, %v), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5343:0
- %64 : Tensor = aten::transpose(%attn_output.1, %12, %13), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5345:0
- %65 : Tensor = aten::contiguous(%64, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5345:0
- %66 : Tensor = aten::mul(%tgt_len, %bsz), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5345:0
- %67 : int = aten::Int(%66), scope: __module.mul
- %68 : int[] = prim::ListConstruct(%67, %25), scope: __module.mul
- %attn_output.3 : Tensor = aten::view(%65, %68), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5345:0
- %attn_output.5 : Tensor = aten::linear(%attn_output.3, %weight, %bias), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5346:0
- %71 : int = aten::size(%attn_output.5, %13), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5347:0
- %72 : int[] = prim::ListConstruct(%21, %23, %71), scope: __module.mul
- %attn_output : Tensor = aten::view(%attn_output.5, %72), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/functional.py:5347:0
- %74 : Tensor = aten::transpose(%attn_output, %13, %12), scope: __module.mul # /usr/local/lib/python3.10/dist-packages/torch/nn/modules/activation.py:1217:0
- return (%74)
? ^
+ return (%14)
? ^
First diverging operator:
Node diff:
- %mul : __torch__.torch.nn.modules.activation.MultiheadAttention = prim::GetAttr[name="mul"](%self.1)
+ %mul : __torch__.torch.nn.modules.activation.___torch_mangle_1.MultiheadAttention = prim::GetAttr[name="mul"](%self.1)
? ++++++++++++++++++
我能做什么？
看起来像是jit的错误。但我对此一无所知。
有人能解决这个问题吗？我也遇到过类似的事情，但是不知道怎么解决？有没有大佬给我出个主意？