junhyukoh / icml2016-minecraft Goto Github PK

The recurrent structure does not work, which I mean the drqn, rmqn, frmqn architectures all fail to run.
It throws an error as below:

/home/nina/tools/torch/install/bin/luajit: /home/nina/tools/torch/install/share/lua/5.1/nn/Linear.lua:66: invalid arguments: CudaTensor number CudaTensor number FloatTensor CudaTensor
expected arguments: CudaTensor~2D [CudaTensor~2D] [float] CudaTensor~2D CudaTensor~2D | CudaTensor~2D float [CudaTensor~2D] float CudaTensor~2D CudaTensor~2D
stack traceback:
[C]: in function 'addmm'
/home/nina/tools/torch/install/share/lua/5.1/nn/Linear.lua:66: in function 'func'
...na/tools/torch/install/share/lua/5.1/nngraph/gmodule.lua:345: in function 'neteval'
...na/tools/torch/install/share/lua/5.1/nngraph/gmodule.lua:380: in function 'forward'
./algorithm/NeuralQLearner.lua:174: in function 'getQUpdate'
./algorithm/NeuralQLearner.lua:220: in function 'qLearnMinibatch'
./algorithm/NeuralQLearner.lua:299: in function 'perceive'
train_mon.lua:102: in main chunk
[C]: in function 'dofile'
...ools/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:145: in main chunk
[C]: at 0x00406670

What is the reason for this? Could you please fix this?

Hi Junhyuk,
I'm trying to run the code and getting an error. Could you help me to resolve it?

My setup: Ubuntu 16.04, CUDA 8.0, CUDNN 5.1, Torch7 commit 0219027e6c4644a0ba5c5bf137c989a0a8c9e01b

Error log:
~/projects/icml2016-minecraft$ ./train_imaze frmqn 0
-save_name IMaze_frmqn -env IMaze -agent_params lr=0.0005,hist_len=12,replay_memory=50000,network="frmqn",state_dim=3072,minibatch_size=32 -steps 15000000 -gpu 0 -seed 1 -test_env IMazeTest -test_hist_len 50
Torch Threads: 4
Using GPU device id: 0
Torch Seed: 1
CUTorch Seed: 1791095845
Searching for available Minecraft instance..
Connected to Port: 30000
Loading IMaze
Initiailzed Minecraft.
Screen Size: 32x32
Creating Agent Network from frmqn
Warning: cudnn.convert does not work with nngraph yet. Ignoring nn.gModuleSet up Torch using these options:
eval_steps 10000
seed 1
save_name IMaze_frmqn
verbose 2
network
tensorType torch.FloatTensor
saveNetworkParams true
test_env IMazeTest
gpu 1
eval_freq 100000
env_params {}
agent_params {
hist_len : 12
verbose : 2
network : "frmqn"
minibatch_size : 32
replay_memory : 50000
gpu : 0
state_dim : 3072
lr : 0.0005
}
env IMaze
framework environment.mcwrap
agent NeuralQLearner
threads 4
actrep 1
test_hist_len 50
steps 15000000
save_freq 100000
port 0
Searching for available Minecraft instance..
Connected to Port: 30001
Loading IMazeTest
Initiailzed Minecraft.
Screen Size: 32x32
Creating Agent Network from frmqn
Warning: cudnn.convert does not work with nngraph yet. Ignoring nn.gModule [..../home/nsavinov/torch/install/bin/luajit: ...nsavinov/torch/install/share/lua/5.1/cudnn/Pointwise.lua:13: Non-contiguous inputs not supported yet
stack traceback:
[C]: in function 'assert'
...nsavinov/torch/install/share/lua/5.1/cudnn/Pointwise.lua:13: in function 'createIODescriptors'
...nsavinov/torch/install/share/lua/5.1/cudnn/Pointwise.lua:41: in function 'func'
...nsavinov/torch/install/share/lua/5.1/nngraph/gmodule.lua:345: in function 'neteval'
...nsavinov/torch/install/share/lua/5.1/nngraph/gmodule.lua:380: in function 'forward'
./algorithm/NeuralQLearner.lua:182: in function 'getQUpdate'
./algorithm/NeuralQLearner.lua:228: in function 'qLearnMinibatch'
./algorithm/NeuralQLearner.lua:307: in function 'perceive'
train.lua:123: in main chunk
[C]: in function 'dofile'
...inov/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50

Hi!

First of all, great work! It really amazed me, great idea and very good paper.
My team and I currently work on problem of planning in imagination. We want to use AlphaZero search capabilities to solve problems like random maze exploration in 3D environments, where you have to learn from raw observations without access to simulator (you need to learn dynamics model too). For state representation and dynamics model learning we want to use other great work World Models.

To make long story short, we would need exactly your environment, but we use Python. Would you allow us to port it into my RL framework HumbleRL on MIT licence? Maybe do you know some easy way to use it in Python? Do you have some tips?

Thanks for your time!