Open selimbek opened 6 years ago
Strange, perhaps you used a different Tensorflow version for earlier models and this could have, for some reason, cause an incompatibility. The 4.18 build used either https://github.com/getnamo/tensorflow-ue4/releases/tag/0.5.0, https://github.com/getnamo/tensorflow-ue4/releases/tag/0.6.0, or https://github.com/getnamo/tensorflow-ue4/releases/tag/0.7.0 so the matching tensorflow version is between 1.4-1.6. Let's try to downgrade the 4.20 tensorflow version.
Plugins\UnrealEnginePython\Binaries\Win64\Lib\site-packages
Let me know if that solves it, if not see what your logs say under Saved/Logs
Hello,
I get a similar issue with UE 4.21.0. I successfully load my model using model_from_json() and the prediction works fine with default weights. However, when I install h5py to load my trained weights using load_weights(), UE crash.
@getnamo I tried to follow your method but UE crash with tensorflow 1.5 & 1.6. For "tensorflow": "1.4.0", I got the following error in the log: LogPython: Error: module 'enum' has no attribute 'IntFlag'
Any updates on how to solve this issue?
Thanks!
EDIT: I get the following errors in the crash reporter:
Access violation - code c0000005 (first/second chance not available)
_conv_cp36_win_amd64
_conv_cp36_win_amd64
python36
... //repeat many times python36
python36
UE4Editor_UnrealEnginePython
UE4Editor_Engine!AActor::InitializeComponents() [d:\build\++ue4\sync\engine\source\runtime\engine\private\actor.cpp:4315]
UE4Editor_Engine!ULevel::RouteActorInitialize() [d:\build\++ue4\sync\engine\source\runtime\engine\private\level.cpp:1822]
UE4Editor_Engine!UWorld::InitializeActorsForPlay() [d:\build\++ue4\sync\engine\source\runtime\engine\private\world.cpp:3642]
UE4Editor_Engine!UGameInstance::StartPlayInEditorGameInstance() [d:\build\++ue4\sync\engine\source\runtime\engine\private\gameinstance.cpp:376]
UE4Editor_UnrealEd!UEditorEngine::CreatePIEGameInstance() [d:\build\++ue4\sync\engine\source\editor\unrealed\private\playlevel.cpp:3332]
UE4Editor_UnrealEd!UEditorEngine::PlayInEditor() [d:\build\++ue4\sync\engine\source\editor\unrealed\private\playlevel.cpp:2459]
UE4Editor_UnrealEd!UEditorEngine::StartQueuedPlayMapRequest() [d:\build\++ue4\sync\engine\source\editor\unrealed\private\playlevel.cpp:1270]
UE4Editor_UnrealEd!UEditorEngine::Tick() [d:\build\++ue4\sync\engine\source\editor\unrealed\private\editorengine.cpp:1640]
UE4Editor_UnrealEd!UUnrealEdEngine::Tick() [d:\build\++ue4\sync\engine\source\editor\unrealed\private\unrealedengine.cpp:403]
UE4Editor!FEngineLoop::Tick() [d:\build\++ue4\sync\engine\source\runtime\launch\private\launchengineloop.cpp:3699]
UE4Editor!GuardedMain() [d:\build\++ue4\sync\engine\source\runtime\launch\private\launch.cpp:174]
UE4Editor!GuardedMainWrapper() [d:\build\++ue4\sync\engine\source\runtime\launch\private\windows\launchwindows.cpp:145]
UE4Editor!WinMain() [d:\build\++ue4\sync\engine\source\runtime\launch\private\windows\launchwindows.cpp:276]
UE4Editor!__scrt_common_main_seh() [f:\dd\vctools\crt\vcstartup\src\startup\exe_common.inl:283]
kernel32
ntdll
I am facing the same error currently with UE4.21.0
try downgrading h5py to e.g. 2.7 in https://github.com/getnamo/tensorflow-ue4/blob/master/Content/Scripts/upymodule.json
Hi,
I used this pluggin with Unreal Engine 4.18 where I was able to load keras models with the load_weights function. This one use h5py module so it was installed as well. And everything worked fine :)
Today i tryed to migrate my work to UE 4.20.2 and tensorflow-ue4 v0.9.0. Python modules seems to be correctly installed, and basic examples works fine. But when I install h5py module and run my project, the engine crash instantly (even if i dont import/use it in my python script) I tried to install manually keras and cython too... with no positive effects
This problem disapear when I uninstall h5py (but i can't load my weights anymore).
Have you an idea to solve this issue, or have you some suggestions to save/load keras models without using h5py dependencie ?
Thank you in advance, --Selim