id
stringlengths 3
8
| text
stringlengths 1
115k
|
|---|---|
st99968
|
In the following I am updating the wights in the resnet101 without a problem
base_ = ResNet(Bottleneck, [3, 4, 23, 3])
resnet_weights = torch.load('resnet101_caffe.pth')
base_.load_state_dict(resnet_weights, strict=False)
if i do
print(resnet_weights.keys())
i will have
odict_keys([‘conv1.weight’, ‘bn1.weight’, ‘bn1.bias’, ‘bn1.running_mean’, ‘bn1.running_var’, ‘layer1.0.conv1.weight’, ‘layer1.0.bn1.weight’, ‘layer1.0.bn1.bias’, ‘layer1.0.bn1.running_mean’, ‘layer1.0.bn1.running_var’, ‘layer1.0.conv2.weight’, ‘layer1.0.bn2.weight’, ‘layer1.0.bn2.bias’, ‘layer1.0.bn2.running_mean’, ‘layer1.0.bn2.running_var’, ‘layer1.0.conv3.weight’, ‘layer1.0.bn3.weight’,
.
.
.
my question is:
if i use
model = models.resnet101(pretrained = True)
which is from pytorch, how i can use the weights in this model to update my model in base_.load_state_dict(resnet_weights, strict=False)
like i did above? is it even possible?
|
st99969
|
Hi,
If the keys are the same, you can do the following:
model = models.resnet101(pretrained = True)
base_.load_state_dict(model.state_dict(), strict=False)
|
st99970
|
I see, thank you for your answer.
I have a follow up question, let say my model is model = models.resnet101(pretrained = True) and then I finetune it, and then want to save it and load it again.
Do you suggest to save and load it like:
torch.save(the_model.state_dict(), PATH)
the_model.load_state_dict(torch.load(PATH))
or you have other suggestion?
in link 3
it says to do
the_model = TheModelClass(*args, **kwargs)
first, but i dont get why and what it means by TheModelClass
|
st99971
|
Yes what you want to do here is fine.
What it means is that creation of the model and loading of the weights are two different things: You don’t save an nn.Module with all it’s weights. You save the weights on one side with .state_dict() and the module’s info on the other side by saving the arguments to create it.
|
st99972
|
I see! Thanks
albanD:
You don’t save an nn.Module with all it’s weights.
Well, can I save and load an nn.Module with all it’s weights ? if so, how?
|
st99973
|
Sorry it was not clear “You don’t” means “You should not”.
Saving python class instances like models breaks so easily in so many levels that you have a very high chance to end up with an object that you will never be able to load again
You can do it, but you shouldn’t so I won’t tell you how to do it
|
st99974
|
Creation of a LongTensor with size < 74240 yields a tensor filled with random long int noise. Above 74240 yields a tensor filled with uniform 0s. Any idea why this is?
Python 3.6.5 | packaged by conda-forge | (default, Apr 6 2018, 13:44:09)
Type 'copyright', 'credits' or 'license' for more information
IPython 6.4.0 -- An enhanced Interactive Python. Type '?' for help.
In [1]: import torch
In [2]: max(torch.LongTensor(74240))
tensor(9222708909033422584)
In [3]: max(torch.LongTensor(74241))
tensor(0)
|
st99975
|
Hi,
Creating a Tensor will not initialize the memory. It will contain whatever the memory contained when it was returned by the allocator.
In your case I guess that you have an allocator for small objects that don’t initialize memory and one for large objects that does zero-out all the values.
|
st99976
|
Hello,
latest pytoch, cuda 9 and cudnn7 installed using conda on linux (lubuntu 16.04, latest nvidiam driver).
I am running this code on a computer with 2x GTX 1080 Ti:
GitHub
tensorboy/pytorch_Realtime_Multi-Person_Pose_Estimation 6
This is a pytorch version of Realtime_Multi-Person_Pose_Estimation, origin code is here https://github.com/ZheC/Realtime_Multi-Person_Pose_Estimation - tensorboy/pytorch_Realtime_Multi-Person_Pose...
Previsously, I was running the imagenet example from the pytorch example.
For both of them, the RAM needed to run the training seems to be expensive according to the same kind of training using caffe or tensorflow - preventing it to be greedy on the RAM.
For instance, on the pose estimation, there is a burst of RAM either at end of an epoch either at the beginning of the evaluation step leading to a cuda error (no more memory). With 11Gb of RAM on each GPU, I can run the training only with batch size = 32. The script is using, within an epoch, only half of the available RAM.
For the imagenet example, same problem, I fixed the batch_size to 80 to be able to train a VGG19_bn. Another problem, when restarting from a check point, the script dies at the end of each epoch with cuda error: no more memory, even with the same batch size.
Can anyone help me understand this burst of RAM?
Thank you.
|
st99977
|
If it can help, here is the package version from my conda environment:
Name Version Build Channel
appdirs 1.4.3 py37h28b3542_0
asn1crypto 0.24.0 py37_0
attrs 18.2.0 py37h28b3542_0
automat 0.7.0 py37_0
backcall 0.1.0 py37_0
blas 1.0 mkl
bleach 2.1.4 py37_0
bzip2 1.0.6 h14c3975_5
ca-certificates 2018.03.07 0
cairo 1.14.12 h8948797_3
certifi 2018.8.24 py37_1
cffi 1.11.5 py37h9745a5d_0
chardet 3.0.4 py37_1
conda 4.5.11 py37_0
conda-env 2.6.0 1
constantly 15.1.0 py37h28b3542_0
cryptography 2.3.1 py37hc365091_0
cudatoolkit 9.0 h13b8566_0
cudnn 7.1.2 cuda9.0_0
cycler 0.10.0 py37_0
cython 0.28.5 py37hf484d3e_0
dbus 1.13.2 h714fa37_1
decorator 4.3.0 py37_0
easydict 1.8
entrypoints 0.2.3 py37_2
expat 2.2.5 he0dffb1_0
ffmpeg 4.0 hcdf2ecd_0
fontconfig 2.13.0 h9420a91_0
freeglut 3.0.0 hf484d3e_5
freetype 2.9.1 h8a8886c_0
future 0.16.0
glib 2.56.1 h000015b_0
gmp 6.1.2 h6c8ec71_1
graphite2 1.3.12 h23475e2_2
gst-plugins-base 1.14.0 hbbd80ab_1
gstreamer 1.14.0 hb453b48_1
harfbuzz 1.8.8 hffaf4a1_0
hdf5 1.10.2 hba1933b_1
html5lib 1.0.1 py37_0
hyperlink 18.0.0 py37_0
icu 58.2 h9c2bf20_1
idna 2.7 py37_0
incremental 17.5.0 py37_0
intel-openmp 2018.0.3 0
ipykernel 4.9.0 py37_0
ipython 6.5.0 py37_0
ipython_genutils 0.2.0 py37_0
ipywidgets 7.4.1 py37_0
jasper 2.0.14 h07fcdf6_1
jedi 0.12.1 py37_0
jinja2 2.10 py37_0
jpeg 9b h024ee3a_2
jsonschema 2.6.0 py37_0
jupyter 1.0.0 py37_5
jupyter_client 5.2.3 py37_0
jupyter_console 5.2.0 py37_1
jupyter_core 4.4.0 py37_0
kiwisolver 1.0.1 py37hf484d3e_0
libedit 3.1.20170329 h6b74fdf_2
libffi 3.2.1 hd88cf55_4
libgcc-ng 8.2.0 hdf63c60_1
libgfortran-ng 7.2.0 hdf63c60_3
libglu 9.0.0 hf484d3e_1
libopencv 3.4.2 hb342d67_1
libopus 1.2.1 hb9ed12e_0
libpng 1.6.34 hb9fc6fc_0
libsodium 1.0.16 h1bed415_0
libstdcxx-ng 8.2.0 hdf63c60_1
libtiff 4.0.9 he85c1e1_1
libuuid 1.0.3 h1bed415_2
libvpx 1.7.0 h439df22_0
libxcb 1.13 h1bed415_1
libxml2 2.9.8 h26e45fe_1
markupsafe 1.0 py37h14c3975_1
matplotlib 2.2.3 py37hb69df0a_0
mistune 0.8.3 py37h14c3975_1
mkl 2018.0.3 1
mkl_fft 1.0.4 py37h4414c95_1
mkl_random 1.0.1 py37h4414c95_1
nbconvert 5.3.1 py37_0
nbformat 4.4.0 py37_0
nccl 1.3.5 cuda9.0_0
ncurses 6.1 hf484d3e_0
ninja 1.8.2 py37h6bb024c_1
notebook 5.6.0 py37_0
numpy 1.15.1 py37h1d66e8a_0
numpy-base 1.15.1 py37h81de0dd_0
olefile 0.45.1 py37_0
opencv 3.4.2 py37h6fd60c2_1
openssl 1.0.2p h14c3975_0
pandoc 2.2.3.2 0
pandocfilters 1.4.2 py37_1
parso 0.3.1 py37_0
pcre 8.42 h439df22_0
pexpect 4.6.0 py37_0
pickleshare 0.7.4 py37_0
pillow 5.2.0 py37heded4f4_0
pip 10.0.1 py37_0
pixman 0.34.0 hceecf20_3
prometheus_client 0.3.1 py37h28b3542_0
prompt_toolkit 1.0.15 py37_0
protobuf 3.6.1
ptyprocess 0.6.0 py37_0
py-opencv 3.4.2 py37hb342d67_1
pyasn1 0.4.4 py37h28b3542_0
pyasn1-modules 0.2.2 py37_0
pycocotools 2.0.0
pycosat 0.6.3 py37h14c3975_0
pycparser 2.18 py37_1
pygments 2.2.0 py37_0
PyHamcrest 1.9.0
pyopenssl 18.0.0 py37_0
pyparsing 2.2.0 py37_1
pyqt 5.9.2 py37h22d08a2_0
pysocks 1.6.8 py37_0
python 3.7.0 hc3d631a_0
python-dateutil 2.7.3 py37_0
pytorch 0.4.1 py37ha74772b_0
pytz 2018.5 py37_0
pyyaml 3.13 py37h14c3975_0
pyzmq 17.1.2 py37h14c3975_0
qt 5.9.6 h52aff34_0
qtconsole 4.4.1 py37_0
readline 7.0 ha6073c6_4
requests 2.19.1 py37_0
ruamel_yaml 0.15.46 py37h14c3975_0
scipy 1.1.0 py37hfa4b5c9_1
send2trash 1.5.0 py37_0
service_identity 17.0.0 py37h28b3542_0
setuptools 40.0.0 py37_0
simplegeneric 0.8.1 py37_2
sip 4.19.8 py37hf484d3e_0
six 1.11.0 py37_1
sqlite 3.24.0 h84994c4_0
tensorboardX 1.4
terminado 0.8.1 py37_1
testpath 0.3.1 py37_0
tk 8.6.7 hc745277_3
torchvision 0.2.1 py37_1 pytorch
tornado 5.1 py37h14c3975_0
traitlets 4.3.2 py37_0
twisted 18.7.0 py37h14c3975_1
urllib3 1.23 py37_0
wcwidth 0.1.7 py37_0
webencodings 0.5.1 py37_1
wheel 0.31.1 py37_0
widgetsnbextension 3.4.1 py37_0
xz 5.2.4 h14c3975_4
yaml 0.1.7 had09818_2
zeromq 4.2.5 hf484d3e_1
zlib 1.2.11 ha838bed_2
zope 1.0 py37_1
zope.interface 4.5.0 py37h14c3975_0
|
st99978
|
My gueses are:
You don’t evaluate under a with torch.no_grad() environment.
You are storing the tensor losses instead of their values (that is, you are missing calling the item() method of the tensor) keeping in memory the whole computation graph.
Check those two first and see if that’s the problem. At this moment I don’t have time to check out your code.
|
st99979
|
Thank you for your answer.
I already test the loss hypothesis. I did not test with torch.no_grad(). The training is currently running (but it will failed as the memory is more and more filled… ).
If someone has some other hypothesis about this problem, he is welcome to submit it before the next traning.
thank you.
|
st99980
|
Sorry for my late answer. Thus, the point:
I did not manage to make it work with ` torch.no_grad(). And actually, I want to compare with other runs thus I drop this idea.
I tried to add some “del” on tensors (as seen in pytorch document), no changes appeared. The amazing thing is that pytorch start with half of the GPU memory and finish to fill it after several epochs but I have no cuda errors running the training (at the end, the memory of both GPU is almost full). I still do not undestand what is hapenning… Is there any way to track all memory blocks in the GPU and where they were allocated?
Thank you.
PS: anyway, my run went fine, it is just a pity that with 2x1080 Ti GPUs, I can not run with bigger batch size…
|
st99981
|
I want to implement a framework like AlphaGo Zero. My MCTS needs to get values from the network very frequently and quickly. So is there a good communication mechanism could do this? I hear that I could convert the PyTorch into Caffe to use its C++ API, but is there another way?
|
st99982
|
Based on what I have been reading here, one can get L2 regularization by providing a value other than 0 to the optimizer through the argument weigh_decay.
Yet, one may implement a custom loss function like this one where the L2 regularization is already taken into account:
class AutoRec_Loss(torch.nn.Module):
def __init__(self):
super(AutoRec_Loss,self).__init__()
def forward(self,predicted_ratings, real_ratings, weights, reg_strength):
ratings_loss = torch.norm(real_ratings - predicted_ratings)
# L2 regularization
weights_regularization = (reg_strength/2)*torch.norm(weight)
return ratings_loss + weights_regularization
What would happen if I set a value other than 0 to the underlying optimizer given this loss function?
|
st99983
|
Hi,
It will just increase even more the l2 regularization. You loss would be loss = ratings_loss + weight_regularization + weight_decay * weight_norm.
|
st99984
|
Hi everyone.
I am implementing local reparameretization (https://papers.nips.cc/paper/5666-variational-dropout-and-the-local-reparameterization-trick.pdf 5) and realized that I need somehow a matrix that has the same vector parameter row-wise. Supposing a layer with 512 neurons.
If I code this:
bias=nn.Parameter(torch.zeros(512,)).repeat(batch,1)
If i now sample from this bias matrix, does pytorch (when performing backward) know that each row is the same parameter?
What I want to do is avoid this:
bias=nn.Parameter(torch.zeros(512,))
for i in range(batch):
bias.sample()
|
st99985
|
Hi,
Yes it will work.
Be careful though that if you do bias=nn.Parameter(torch.zeros(512,)).repeat(batch,1), the python vairable bias will not contain the original Parameter object and is not the tensor that will be optimized.
You would need to do:
bias_for_optimizer = nn.Parameter(torch.zeros(512,))
bias = bias_for_optimizer.repeat(batch, 1)
|
st99986
|
Because in the one liner, the variable that you get is not a leaf tensor and so will not have it’s gradient saved. Meaning you won’t be able to optimize it:
import torch
from torch import nn
a = nn.Parameter(torch.rand(10))
b = a.repeat(2)
b.sum().backward()
print("b is leaf: ",b.is_leaf) # False
print("b.grad: ", b.grad) # None
print("a is leaf: ",a.is_leaf) # True
print("a.grad: ", a.grad) # some gradients
a = nn.Parameter(torch.rand(10)).repeat(2)
a.sum().backward()
print("a is leaf: ",a.is_leaf) # False
print("a.grad: ", a.grad) # None
|
st99987
|
The problem with this approach is that the resulting tensor bias is not a parameter:
bias_for_optim=nn.Parameter(torch.zeros(topology[idx+1],).cuda())
bias=bias_for_optim.repeat(batch,1)
print type(bias)
and cannot be registered in the module. Shoud I put into nn.Parameter again after repeat?
|
st99988
|
That’s my point,
You should register bias_for_optim in the module and for the optimization.
The bias=bias_for_optim.repeat(batch,1) should only be done during the forward pass.
|
st99989
|
Ah okei, I wanted to avoid calling repeat in each forward step for efficience. However I suppose .repeat() is good optimized (at least would be much quicker than looping at python level)
Thanks @albanD
|
st99990
|
Hi,
You can consider that repeat is for free litterally. It changes 2 numbers on the cpu memory. You should not worry about it
|
st99991
|
Yes, I was talking more about memory reservation, however pytorch pool memory allocator should not have problems at this level, my intention was to have all memory allocated to avoid the typical problems that arise when you allocate and deallocate dynamically.
|
st99992
|
Hi!
I have following problem during PyTorch compilation on Linux (I’ve been able to successfully compile it on macOS):
Install the project...
-- Install configuration: "Release"
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/share/cmake/Gloo/GlooConfig.cmake
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/share/cmake/Gloo/GlooConfigVersion.cmake
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/share/cmake/Gloo/GlooTargets.cmake
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/share/cmake/Gloo/GlooTargets-release.cmake
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libgloo.a
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libgloo_builder.a
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libgloo_cuda.a
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/config.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/algorithm.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/allgather_ring.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/allreduce_halving_doubling.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/allreduce_bcube.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/allreduce_local.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/allreduce_ring.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/allreduce_ring_chunked.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/barrier.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/barrier_all_to_all.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/barrier_all_to_one.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/broadcast_one_to_all.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/reduce_scatter.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/context.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/math.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/pairwise_exchange.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/types.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/common/common.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/common/error.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/common/linux.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/common/linux_devices.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/common/logging.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/common/string.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/rendezvous/file_store.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/rendezvous/hash_store.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/rendezvous/prefix_store.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/rendezvous/store.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/rendezvous/context.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/address.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/buffer.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/device.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/pair.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/tcp/address.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/tcp/buffer.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/tcp/device.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/transport/tcp/pair.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/allreduce_builder.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/broadcast_builder.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_allreduce_halving_doubling.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_allreduce_halving_doubling_pipelined.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_allreduce_local.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_allreduce_ring.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_allreduce_ring_chunked.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_broadcast_one_to_all.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_collectives_device.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_collectives_host.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_collectives_native.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_collectives_nccl.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_private.h
-- Installing: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/gloo/cuda_workspace.h
+ popd
~/Programs/pytorch/third_party ~/Programs/pytorch
++ uname
+ [[ Linux == \D\a\r\w\i\n ]]
+ popd
~/Programs/pytorch
+ for arg in '"$@"'
+ [[ THD == \n\c\c\l ]]
+ [[ THD == \g\l\o\o ]]
+ [[ THD == \c\a\f\f\e\2 ]]
+ [[ THD == \T\H\D ]]
+ pushd /macierz/home/155079jp/Programs/pytorch/torch/lib
~/Programs/pytorch/torch/lib ~/Programs/pytorch
+ build THD
+ mkdir -p build/THD
+ pushd build/THD
~/Programs/pytorch/torch/lib/build/THD ~/Programs/pytorch/torch/lib ~/Programs/pytorch
+ BUILD_C_FLAGS=
+ case $1 in
+ BUILD_C_FLAGS=' -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/TH" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THC" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THNN" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCUNN" -DOMPI_SKIP_MPICXX=1 -fexceptions'
+ cmake ../../THD -DCMAKE_MODULE_PATH=/macierz/home/155079jp/Programs/pytorch/cmake/Modules_CUDA_fix -DTorch_FOUND=1 -DCMAKE_INSTALL_PREFIX=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install '-DCMAKE_C_FLAGS= -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/TH" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THC" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THNN" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCUNN" -DOMPI_SKIP_MPICXX=1 -fexceptions ' '-DCMAKE_CXX_FLAGS= -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/TH" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THC" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THNN" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCUNN" -DOMPI_SKIP_MPICXX=1 -fexceptions -std=c++11 ' '-DCMAKE_EXE_LINKER_FLAGS=-L"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib" -Wl,-rpath,$ORIGIN ' '-DCMAKE_SHARED_LINKER_FLAGS=-L"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib" -Wl,-rpath,$ORIGIN ' -DCMAKE_INSTALL_LIBDIR=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib '-DCUDA_NVCC_FLAGS= -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/TH" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THC" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCS" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THNN" -I"/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/THCUNN" -DOMPI_SKIP_MPICXX=1' -DCUDA_DEVICE_DEBUG=0 -DCMAKE_PREFIX_PATH=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install '-Dcwrap_files=/macierz/home/155079jp/Programs/pytorch/torch/lib/ATen/Declarations.cwrap;/macierz/home/155079jp/Programs/pytorch/torch/lib/THNN/generic/THNN.h;/macierz/home/155079jp/Programs/pytorch/torch/lib/THCUNN/generic/THCUNN.h;/macierz/home/155079jp/Programs/pytorch/torch/lib/ATen/nn.yaml' -DTH_INCLUDE_PATH=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include -DTH_LIB_PATH=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib -DTH_LIBRARIES=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libTH.so -DCAFFE2_LIBRARIES=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libcaffe2.so -DTHNN_LIBRARIES=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libTHNN.so -DTHCUNN_LIBRARIES=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libTHCUNN.so -DTHS_LIBRARIES=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libTHS.so -DTHC_LIBRARIES=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libTHC.so -DTHCS_LIBRARIES=/macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libTHCS.so -DTH_SO_VERSION=1 -DTHC_SO_VERSION=1 -DTHNN_SO_VERSION=1 -DTHCUNN_SO_VERSION=1 -DTHD_SO_VERSION=1 -DUSE_CUDA=1 -DNO_NNPACK=0 -DNCCL_EXTERNAL=1 -Dnanopb_BUILD_GENERATOR=0 -DCMAKE_DEBUG_POSTFIX= -DCMAKE_BUILD_TYPE=Release -DCMAKE_EXPORT_COMPILE_COMMANDS=1
-- The C compiler identification is GNU 4.8.5
-- The CXX compiler identification is GNU 4.8.5
-- Check for working C compiler: /usr/bin/cc
-- Check for working C compiler: /usr/bin/cc -- works
-- Detecting C compiler ABI info
-- Detecting C compiler ABI info - done
-- Detecting C compile features
-- Detecting C compile features - done
-- Check for working CXX compiler: /usr/bin/c++
-- Check for working CXX compiler: /usr/bin/c++ -- works
-- Detecting CXX compiler ABI info
-- Detecting CXX compiler ABI info - done
-- Detecting CXX compile features
-- Detecting CXX compile features - done
-- Performing Test HAS_THREAD_LOCAL
-- Performing Test HAS_THREAD_LOCAL - Success
-- Found MPI_C: /usr/lib64/mpi/gcc/mvapich2/lib64/libmpi.so (found version "3.0")
-- Found MPI_CXX: /usr/lib64/mpi/gcc/mvapich2/lib64/libmpicxx.so (found version "3.0")
-- Found MPI: TRUE (found version "3.0")
-- Found Gloo: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include
-- Looking for pthread.h
-- Looking for pthread.h - found
-- Looking for pthread_create
-- Looking for pthread_create - not found
-- Looking for pthread_create in pthreads
-- Looking for pthread_create in pthreads - not found
-- Looking for pthread_create in pthread
-- Looking for pthread_create in pthread - found
-- Found Threads: TRUE
-- Caffe2: Found protobuf with new-style protobuf targets.
-- Caffe2: Protobuf version 3.5.0
-- Found CUDA: /usr/local/cuda (found suitable version "9.1", minimum required is "7.0")
-- Caffe2: CUDA detected: 9.1
-- Caffe2: CUDA nvcc is: /usr/local/cuda/bin/nvcc
-- Caffe2: CUDA toolkit directory: /usr/local/cuda
-- Caffe2: Header version is: 9.1
-- Found CUDNN: /usr/local/cuda/include
-- Found cuDNN: v7.0.5 (include: /usr/local/cuda/include, library: /usr/local/cuda/lib64/libcudnn.so.7)
-- Autodetected CUDA architecture(s): 6.1
-- Added CUDA NVCC flags for: -gencode;arch=compute_61,code=sm_61
-- Found CUDA: /usr/local/cuda (found suitable version "9.1", minimum required is "7.5")
-- Found NCCL: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include
-- Determining NCCL version from the header file: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include/nccl.h
-- Found NCCL (include: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/include, library: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libnccl.so)
-- MPI_LIBRARIES: /usr/lib64/mpi/gcc/mvapich2/lib64/libmpicxx.so;/usr/lib64/mpi/gcc/mvapich2/lib64/libmpi.so
-- Found Gloo, will compile with Gloo distributed backend
-- Building the gloo backend with TCP support only
-- NCCL_LIBRARIES: /macierz/home/155079jp/Programs/pytorch/torch/lib/tmp_install/lib/libnccl.so
-- Found NCCL, but the NCCL version is either not 2+ or not determinable, will not compile with NCCL distributed backend
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/Communication.cpp from the build
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/Generator.cpp from the build
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/Storage.cpp from the build
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/Tensor.cpp from the build
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/TensorCopy.cpp from the build
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/TensorLapack.cpp from the build
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/TensorMath.cpp from the build
-- Excluding /macierz/home/155079jp/Programs/pytorch/torch/lib/THD/master_worker/worker/dispatch/TensorRandom.cpp from the build
-- MPI_COMPILE_FLAGS: -fmessage-length=0;-funwind-tables;-fasynchronous-unwind-tables;-fstack-clash-protection
-- MPI_LINK_FLAGS: -Wl,-rpath -Wl,/usr/lib64/mpi/gcc/mvapich2/lib64 -Wl,--enable-new-dtags
-- Configuring done
-- Generating done
CMake Warning:
Manually-specified variables were not used by the project:
CMAKE_DEBUG_POSTFIX
CMAKE_INSTALL_LIBDIR
NCCL_EXTERNAL
NO_NNPACK
THCS_LIBRARIES
THCUNN_LIBRARIES
THCUNN_SO_VERSION
THC_LIBRARIES
THC_SO_VERSION
THD_SO_VERSION
THNN_LIBRARIES
THNN_SO_VERSION
THS_LIBRARIES
TH_INCLUDE_PATH
TH_LIBRARIES
TH_LIB_PATH
TH_SO_VERSION
Torch_FOUND
cwrap_files
nanopb_BUILD_GENERATOR
-- Build files have been written to: /macierz/home/155079jp/Programs/pytorch/torch/lib/build/THD
+ make install -j8
Scanning dependencies of target THD
[ 5%] Building CXX object CMakeFiles/THD.dir/base/ChannelUtils.cpp.o
c++: fatal error: no input files
compilation terminated.
/bin/sh: -funwind-tables: nie znaleziono polecenia
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
[ 11%] Building CXX object CMakeFiles/THD.dir/base/Cuda.cpp.o
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
CMakeFiles/THD.dir/build.make:62: polecenia dla obiektu 'CMakeFiles/THD.dir/base/ChannelUtils.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/ChannelUtils.cpp.o] Błąd 127
make[2]: *** Oczekiwanie na niezakończone zadania....
c++: fatal error: no input files
compilation terminated.
/bin/sh: -funwind-tables: nie znaleziono polecenia
[ 35%] Building CXX object CMakeFiles/THD.dir/base/RPCType.cpp.o
[ 35%] Building CXX object CMakeFiles/THD.dir/base/DataChannel.cpp.o
[ 35%] Building CXX object CMakeFiles/THD.dir/base/DataChannelRequest.cpp.o
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
CMakeFiles/THD.dir/build.make:75: polecenia dla obiektu 'CMakeFiles/THD.dir/base/Cuda.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/Cuda.cpp.o] Błąd 127
[ 47%] Building CXX object CMakeFiles/THD.dir/base/data_channels/DataChannelGloo.cpp.o
[ 47%] Building CXX object CMakeFiles/THD.dir/base/data_channels/DataChannelMPI.cpp.o
[ 47%] Building CXX object CMakeFiles/THD.dir/base/data_channels/DataChannelTCP.cpp.o
c++: fatal error: no input files
c++: fatal error: no input files
compilation terminated.
compilation terminated.
c++: fatal error: no input files
compilation terminated.
/bin/sh: -funwind-tables: nie znaleziono polecenia
/bin/sh: -funwind-tables: nie znaleziono polecenia
/bin/sh: -funwind-tables: nie znaleziono polecenia
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
c++: fatal error: no input files
CMakeFiles/THD.dir/build.make:114: polecenia dla obiektu 'CMakeFiles/THD.dir/base/RPCType.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/RPCType.cpp.o] Błąd 127
compilation terminated.
CMakeFiles/THD.dir/build.make:101: polecenia dla obiektu 'CMakeFiles/THD.dir/base/DataChannelRequest.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/DataChannelRequest.cpp.o] Błąd 127
c++: fatal error: no input files
compilation terminated.
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
CMakeFiles/THD.dir/build.make:88: polecenia dla obiektu 'CMakeFiles/THD.dir/base/DataChannel.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/DataChannel.cpp.o] Błąd 127
c++: fatal error: no input files
compilation terminated.
/bin/sh: -funwind-tables: nie znaleziono polecenia
/bin/sh: -funwind-tables: nie znaleziono polecenia
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
/bin/sh: -funwind-tables: nie znaleziono polecenia
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
/bin/sh: -fasynchronous-unwind-tables: nie znaleziono polecenia
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
CMakeFiles/THD.dir/build.make:127: polecenia dla obiektu 'CMakeFiles/THD.dir/base/data_channels/DataChannelGloo.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/data_channels/DataChannelGloo.cpp.o] Błąd 127
CMakeFiles/THD.dir/build.make:140: polecenia dla obiektu 'CMakeFiles/THD.dir/base/data_channels/DataChannelMPI.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/data_channels/DataChannelMPI.cpp.o] Błąd 127
/bin/sh: -fstack-clash-protection: nie znaleziono polecenia
CMakeFiles/THD.dir/build.make:153: polecenia dla obiektu 'CMakeFiles/THD.dir/base/data_channels/DataChannelTCP.cpp.o' nie powiodły się
make[2]: *** [CMakeFiles/THD.dir/base/data_channels/DataChannelTCP.cpp.o] Błąd 127
CMakeFiles/Makefile2:67: polecenia dla obiektu 'CMakeFiles/THD.dir/all' nie powiodły się
make[1]: *** [CMakeFiles/THD.dir/all] Błąd 2
Makefile:129: polecenia dla obiektu 'all' nie powiodły się
make: *** [all] Błąd 2
Failed to run 'bash tools/build_pytorch_libs.sh --use-cuda --use-nnpack --use-mkldnn nccl caffe2 nanopb libshm gloo THD c10d'
Sorry for polish language, it’s uni PC, I can’t change that (or can I? Do you know how?). I can translate if you don’t understand something.
Versions:
OS: openSUSE Leap 42.3
GCC: 4.8.5
conda: 4.5.11
Python: 3.6.5
CUDA: 9.1
cuDNN: 7.0
I work on NFS drive if it changes anything.
I’ve executed those commands before installation:
export CMAKE_PREFIX_PATH="$(dirname $(which conda))/../"
conda install numpy pyyaml mkl mkl-include setuptools cmake cffi typing
conda install -c mingfeima mkldnn
conda install -c pytorch magma-cuda91
I’ve changed magma-cuda80 to magma-cuda91.
I was looking through the Internet and couldn’t find any help. Do someone know what is going on?
Thanks for any help!
|
st99993
|
Solved by albanD in post #2
Hi,
It does look weird indeed.
It looks like it splits the compilation command line in the middle and so the compiler does not have input files specified, and it tries to execute the compiler arguments as bash methods.
Do you by any change used a windows computer to get the code before putting i…
|
st99994
|
Hi,
It does look weird indeed.
It looks like it splits the compilation command line in the middle and so the compiler does not have input files specified, and it tries to execute the compiler arguments as bash methods.
Do you by any change used a windows computer to get the code before putting it on the machine where you compile?
Will you need the distributed package? If not, try compiling with NO_DISTRIBUTED=1 as MPI might be the problem here.
If that does not help, could you run make in verbose mode directly within /macierz/home/155079jp/Programs/pytorch/torch/lib/build/THD to see what is the exact command that it tries to execute please? That would help us fix this problem.
|
st99995
|
Do you by any change used a windows computer to get the code before putting it on the machine where you compile?
Nope.
Will you need the distributed package? If not, try compiling with NO_DISTRIBUTED=1 as MPI might be the problem here.
Yea, I don’t think I’ll be trying to distribute training anytime soon. I set this variable and it compiled and installed fine this time. Tanks a lot for help!
|
st99996
|
Hello,
When should we use repeat instead of expanse?
Apart from the case where we want to call an inplace function?
|
st99997
|
expand does not allocate new memories, which means if you do
x = torch.tensor([[1], [2], [3]])
expand_x = x.expand([3,4])
x[0, 0] = 4
you will get expand_x
tensor([[4, 4, 4, 4],
[2, 2, 2, 2],
[3, 3, 3, 3]])
If you use repeat_x = x.repeat(1, 4), changing x will not affect repeat_x.
|
st99998
|
If they have the same number of layers, I know that How to load part of pre trained model? 156 works.
However, if my new model has m+n layers and my old model has m layers, pytorch will complain about missing layers. How should I load the model then?
|
st99999
|
The pre-trained model is loaded as a OrderedDict by calling torch.load(), you can then extract weights from the dictionary and do what you want. For example, in your case, you could get your model’s state_dict, then assign weights to layers of interests and load the dict using model.load_state_dict().
|
st115000
|
x = np.array([1.8507 ,-2.7324])
y = np.array([0.9722 , 0.4470 , 1.0000 , 0.0000 , 0.0000 , 0.0000])
x = torch.from_numpy(x)
y = torch.from_numpy(y)
z = x.unsqueeze(0)*y.unsqueeze(1)
print(z)
or
x = np.array([1.8507 ,-2.7324])
y = np.array([0.9722 , 0.4470 , 1.0000 , 0.0000 , 0.0000 , 0.0000])
x = torch.from_numpy(x)
y = torch.from_numpy(y)
z = x*y.unsqueeze(1)
print(z)
both work on my computer. I prefer the former as it explicit specifies the tensor shape.
|
st115001
|
@chenchr: Thanks for your suggestion. I could solve the problem using Smith’s method. But this alternative solution also helped me to understand performance of “unsqueeze” better. I’m still new to pytorch. So taking sometime to grab the new concepts.
|
st115002
|
I’m training an object detection model in pytorch on Ubuntu 16.04 with a TitanX Pascal gpu. In the middle of training (after several thousand training iterations e.g. 70k), the training crashes with the message:
[libprotobuf FATAL google/protobuf/wire_format.cc:830] CHECK failed: (output->ByteCount()) == (expected_endpoint): : Protocol message serialized to a size different from what was originally expected. Perhaps it was modified by another thread during serialization?
terminate called after throwing an instance of 'google::protobuf::FatalException'
what(): CHECK failed: (output->ByteCount()) == (expected_endpoint): : Protocol message serialized to a size different from what was originally expected. Perhaps it was modified by another thread during serialization?
Command terminated by signal 6
I’ve searched, but couldn’t find anyone experiencing similar errors. I’m using tensorboard-pytorch to visualize the training artifacts. Any ideas on how to resolve this issue?
|
st115003
|
can you check if by any chance you are running out of memory or disk space?
It looks like one of those errors where it tried to write X bytes, but could only write bytes less than X due to some unknown constraints.
|
st115004
|
It’s training now, and the gpu is only using 7293MiB / 12189MiB. Since the crash happens almost at random, it’s difficult to pin down. I’m currently scaling the shortest image side to 800 pixels, but could try reducing that to 600 pixels and seeing whether the issue arises. My batch size is already 1 image.
|
st115005
|
I just noticed that I had a zombie process taking up some memory, so this may have been the culprit. After killing the zombie process, my gpu memory is down to 4228MiB / 12189MiB. Hopefully this was the issue.
|
st115006
|
The error was actually related to tensorboard. I removed the code portion that writes tfevents and the training completed just fine.
|
st115007
|
I found some features like upsampling_nearest are in the github but not in the conda package. Is there a timeline when the conda package will be updated?
|
st115008
|
the conda package was updated yesterday evening with the 0.1.7 release which has the upsampling_nearest available
|
st115009
|
trypag:
conda update pytorch torchvision -c soumith
Oh, so the packages are not sent to the default channels, likewise it was happening for luarocks in Torch.
I would suggest, therefore, to run this command, in order to automate the process.
conda config --add channels soumith
[which] adds the new channel [soumith] to the top of the channel list, making it the highest priority.
|
st115010
|
How long does it usually take to update the packages?
This commit 7 from 4 days ago still seems not to be included.
|
st115011
|
I am currently trying to train a 3-layer LSTM for a classification task. The input sequence has variable length,so I padded every sequence with zero to the longest one within the minibatch and the padded label is set to -1 which will be ignore in the loss calculation. When I train LSTM with batch_size=1, it works well, the cross entropy loss decreases and the training classification accuracy increases. The problem is when I set batch_size >1, e.g. batch_size=8, the loss decreases while the accuracy do not increase. Could anyone help me to figure out why ?
Some related code is as follows:
class Model(nn.Module):
def __init__(self, args):
super(Model, self).__init__()
self.args = args
self.n_d = args.feadim
self.n_cell=args.hidnum
self.depth = args.depth
self.drop = nn.Dropout(args.dropout)
self.n_V = args.statenum
if args.lstm:
self.rnn = nn.LSTM(self.n_d, self.n_cell,
self.depth,
dropout = args.rnn_dropout,
batch_first = True
)
else:
pass
self.output_layer = nn.Linear(self.n_cell, self.n_V)
def forward(self, x, hidden,lens):
rnnout, hidden = self.rnn(x, hidden)
output = self.drop(rnnout)
output = output.view(-1, output.size(2))
output = self.output_layer(output)
return output, hidden
def train_model(epoch, model, train_reader):
model.train()
args = model.args
batch_size = args.batch_size
total_loss = 0.0
criterion = nn.CrossEntropyLoss(size_average=False,ignore_index=-1)
hidden = model.init_hidden(batch_size)
i=0
running_acc=0
total_frame=0
while True:
feat,label,length = train_reader.load_next_nstreams()
if length is None or label.shape[0]<args.batch_size:
break
else:
x, y = Variable(torch.from_numpy(feat)).cuda(), Variable(torch.from_numpy(label).long()).cuda()
hidden = model.init_hidden(batch_size)
hidden = (Variable(hidden[0].data), Variable(hidden[1].data)) if args.lstm else Variable(hidden.data)
model.zero_grad()
output, hidden = model(x, hidden,length)
assert x.size(0) == batch_size
loss = criterion(output, y.view(-1))
_,predict = torch.max(output,1)
correct = (predict == y).sum()
loss.backward()
total_loss += loss.data[0]
running_acc += correct.data[0]
total_frame += sum(length)
i+=1
if i%10 == 0:
sys.stdout.write(“time:{}, Epoch={},trbatch={},loss={:.4f},tracc={:.4f}\n”.format(datetime.now(),epoch,i,total_loss/total_frame,
running_acc*1.0/total_frame))
sys.stdout.flush()
|
st115012
|
hi Pan-Zhou,
I am not sure of exactly why you are seeing this behavior. If you pin it down, I would love to know.
Some easy things to try:
increase / decrease learning rate, see what happens
print out the min/max values of the weights of the network over learning, as well as the norm.
check if somehow weights are becoming NaN.
|
st115013
|
thanks for your advise. I do tuning the learning rate and founds it helps a little. I use about 150 hours speech feathure training a 3 layer lstm with 400 cell each. and set batch_size=4. Here is the log information and weights norm after each epoch.
Epoch=0 lr=2.0000 train_loss=3.6674 dev_loss=2.9469 tracc=0.0777 validacc=0.0789 [58.3999m]
Epoch=1 lr=2.0000 train_loss=3.1931 dev_loss=2.8304 tracc=0.0786 validacc=0.0787 [57.8110m]
Epoch=2 lr=2.0000 train_loss=3.0930 dev_loss=2.7190 tracc=0.0793 validacc=0.0786 [58.2592m]
p_norm: ['7', '23', '0', '0', '23', '23', '0', '0', '23', '23', '0', '0', '37', '0']
p_norm: ['71', '82', '20', '20', '82', '82', '17', '17', '82', '99', '16', '16', '144', '27']
p_norm: ['91', '104', '25', '25', '98', '104', '21', '21', '95', '119', '19', '19', '160', '30']
p_norm: ['97', '110', '27', '27', '106', '118', '22', '22', '102', '129', '19', '19', '166', '31']
In fact I use the same data and the same data io function to train 3 layer LSTM with tensorflow. I works well. traing loss and valid loss are:
End of epoch 0 with avg loss 3.66545295715 and accuracy 0.26187556982
End of epoch 1 with avg loss 2.78404808044 and accuracy 0.355499237776
End of epoch 2 with avg loss 2.55863642693 and accuracy 0.38808375597
End of epoch 3 with avg loss 2.42844891548 and accuracy 0.4079862535
End of epoch 4 with avg loss 2.33932137489 and accuracy 0.422125428915
End of epoch 5 with avg loss 2.20702433586 and accuracy 0.445332825184
End of epoch 6 with avg loss 2.12942314148 and accuracy 0.459314882755
End of epoch 7 with avg loss 2.08610677719 and accuracy 0.467183083296
End of epoch 8 with avg loss 2.06255722046 and accuracy 0.471532851458
End of epoch 9 with avg loss 2.04997444153 and accuracy 0.473860412836
epoch 0 valid split mean loss: 2.96507430077, accuracy: 0.331166476011
epoch 1 valid split mean loss: 2.6755862236, accuracy: 0.3697052598
epoch 2 valid split mean loss: 2.54053473473, accuracy: 0.389888346195
epoch 3 valid split mean loss: 2.47018957138, accuracy: 0.399514913559
epoch 4 valid split mean loss: 2.42790412903, accuracy: 0.40643504262
epoch 5 valid split mean loss: 2.35705971718, accuracy: 0.420234382153
epoch 6 valid split mean loss: 2.32587504387, accuracy: 0.426783770323
epoch 7 valid split mean loss: 2.31113815308, accuracy: 0.42946600914
epoch 8 valid split mean loss: 2.30379247665, accuracy: 0.430867373943
epoch 9 valid split mean loss: 2.29960465431, accuracy: 0.431604236364
|
st115014
|
Hello,
I want to use MultiProcessing to train several models simultaneously. Each thread (cpu core) update the parameters of its own model. Does PyTorch support such operations?
For example, firstly I create a model list that each element is a seperate model:
net_list=[PyTorch Net for _ in range(threads) ]
After then I use MultiPocessing to start each process to update these models. Will this work for PyTorch?
|
st115015
|
Why do you need multiprocessing for that? You can easily do a bash script for that
|
st115016
|
Because after optimizing these networks on different threads, the weights should be exchanged. Which is convenient to implement in a main process.
|
st115017
|
the official doc about the multiprocess 12 , but also i found some problem in it, here is the problem, if you have some idea, can you share it with me?
Thx.
|
st115018
|
Hi, I’m using pytorch on python 3.5.2. While attempting to use torch.multiprocessing.pool, I’m getting the following error.
Code snippet:
from torch.multiprocessing.pool import Pool
…
with Pool(processes=n_processes) as pool:
games = pool.map(self.play_game, range(n_processes))
…
Error:
Traceback (most recent call last):
File “rl_net.py”, line 188, in
agent.train(100000)
File “rl_net.py”, line 145, in train
with Pool(processes=n_processes) as pool:
File “/usr/lib/python3.5/multiprocessing/pool.py”, line 150, in init
self._setup_queues()
File “/home/cs234-gpu2/.env3.5/lib/python3.5/site-packages/torch/multiprocessing/pool.py”, line 23, in _setup_queues
self._inqueue = SimpleQueue()
TypeError: init() missing 1 required keyword-only argument: ‘ctx’
|
st115019
|
Use torch.multiprocessing.Pool instead of torch.multiprocessing.pool.Pool; the latter is a definition for Python 3.3 and under, and loaded into torch.multiprocessing.Pool – as stated in torch/multiprocessing/__init__.py.
|
st115020
|
Hi, I would like to know how you guys deal with the dropout in testing since the dropout rate should be set to 0 while testing. I directly set model.training = False while testing. Or is there other ways to handle this?
|
st115021
|
YongyiTang92:
Hi, I would like to know how you guys deal with the dropout in testing since the dropout rate should be set to 0 while testing. I directly set model.training = False while testing. Or is there other ways to handle this?
what do you mean for " the dropout rate should be set to 0 while testing"?
Is the dropout rate different between train and test?
|
st115022
|
Hi, I wonder if it is possible to instantiating modules in “init” as entries in a dictionary.
For example, is the following allowed?
class MyNet(nn.Module):
def __init__(self):
self.myModules = {}
self.myModule['dog'] = nn.Linear(4096, 300)
self.myModule['cat'] = nn.Linear(4096, 300)
self.myModule['flower'] = nn.Linear(4096, 300)
def forward(self, img):
...
This example is simple that it does not have to use dictionary, but what I actually need to do is something like following:
class MyNet(nn.Module):
def __init__(self, number_of_modules):
self.myModules = {}
for i in range(0, number_of_modules):
self.myModule[ 'dog' + str(i) ] = nn.Linear(4096, 300)
def forward(self, img):
...
The modules are of same shape, but I don’t want them to share weights.
|
st115023
|
def __init__():
.....
for i in range(0, number_of_modules):
setattr(self,'dog' + str(i) ,nn.Linear(4096, 300))
def forward(self,input):
dog1 = self.dog1(input)
dog2 = self.dog2(input)
dog3 = getattr(self,'dog3')(input)
or
def __init__():
.....
modules = []
for i in range(0, number_of_modules):
modules .append( nn.Linear(4096, 300))
self.modules = nn.ModuleList(modules)
def forward(self,input):
outputs = []
for model in self.modules:
outputs.append(model(input))
also,try to format your code like:
```Python
your code here
```
|
st115024
|
Hello,
I’m seeking suggestions w.r.t the best practice to monitor the “dead” neuron ratio during training process in pytorch. The goal is to probably set this as one of the early-stop signals so that I could abandon a certain model when I see that number goes up too high say 30% etc.
Thanks,
|
st115025
|
If you do weight decay, you can do this: model.target_layer.weight.data.var(1) to get the variance of the output units’ weight.
If you don’t do weight decay, and have to check the unit activation, look up Module.register_forward_hook 132. This allows you to check the output of intermediate layers quite conveniently.
|
st115026
|
torch.save({
‘epoch’: epoch,
‘model’: net,
‘model_state_dict’: net.state_dict(),
‘best_mean_iu’: meanIU_best,
}, os.path.join(model_path, ‘checkpoint.pth.tar’))
i save model like this.
checkpoint = torch.load(‘checkpoint.pth.tar’)
net = torch.load(checkpoint[‘model’])
but i try to load model from checkpoint, it would appear error like this:
Traceback (most recent call last):
File “/home/liuyf/DenseNet_clockwork/CamVid_DenseNet/camvid_train.py”, line 35, in
net = torch.load(checkpoint[‘model’])
File “/usr/local/lib/python2.7/dist-packages/torch/serialization.py”, line 231, in load
return _load(f, map_location, pickle_module)
File “/usr/local/lib/python2.7/dist-packages/torch/serialization.py”, line 364, in _load
return legacy_load(f)
File “/usr/local/lib/python2.7/dist-packages/torch/serialization.py”, line 299, in legacy_load
with closing(tarfile.open(fileobj=f, mode=‘r:’, format=tarfile.PAX_FORMAT)) as tar,
File “/usr/lib/python2.7/tarfile.py”, line 1691, in open
return func(name, filemode, fileobj, **kwargs)
File “/usr/lib/python2.7/tarfile.py”, line 1721, in taropen
return cls(name, mode, fileobj, **kwargs)
File “/usr/lib/python2.7/tarfile.py”, line 1579, in init
self.offset = self.fileobj.tell()
File “/usr/local/lib/python2.7/dist-packages/torch/nn/modules/module.py”, line 262, in getattr
type(self).name, name))
AttributeError: ‘DataParallel’ object has no attribute ‘tell’
|
st115027
|
Any preferences; do people generally favor one over the other? If so, why?
crayon repo 132
tensorboard-pytorch repo 182
|
st115028
|
Hi,
As one of the person that originally created crayon, I do not actively use it anymore (because I don’t need this kind of visualization). And so there are no new features added to it by the original authors.
I don’t know what is the status of tensorboard-pytorch though.
|
st115029
|
I would suggest tensorboard-pytorch. It’s easier to use especially if you used tensorboard before.
|
st115030
|
albanD:
I do not actively use it anymore (because I don’t need this kind of visualization).
Could you elaborate; are there other visualization tools you or @ruotianluo recommend?
|
st115031
|
I just don’t have any loss to plot. So I just don’t plot anything. I don’t use anything else.
|
st115032
|
I usually just use pyplot.plot in Jupyter. That flickers a bit when I clear the output abd redraw but works quite welk for me.
I also have something in bokeh. That is a bit more appealing conceptually because it supports updates, but I found that it was a bit less convenient.
Here is an example with pyplot 133.
Best regards
Thomas
|
st115033
|
I am following the tutotial for transfer learning
http://pytorch.org/tutorials/beginner/transfer_learning_tutorial.html 73
I wish to train on a custom dataset, which cannot be cropped as it will result in relevant data being lost.
224x224 is too small for my use case
Maybe I could resize my data to 480x640
But I would prefer not to alter the images.
When I try to train the model I get an error on size mismatch.
It seems the implementation of the model only allows for images which are 224 x 224.
Is this correct?
Looking at the model
github.com
pytorch/vision/blob/master/torchvision/models/resnet.py 46
import torch.nn as nn
import math
import torch.utils.model_zoo as model_zoo
__all__ = ['ResNet', 'resnet18', 'resnet34', 'resnet50', 'resnet101',
'resnet152']
model_urls = {
'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth',
'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth',
'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth',
'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth',
'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth',
}
def conv3x3(in_planes, out_planes, stride=1):
"""3x3 convolution with padding"""
This file has been truncated. show original
vs the torch versions
There is a single kernel size 7
and single input to AvgPool
Which suggests that the input must be square
32*7=224
In torch
github.com
facebook/fb.resnet.torch/blob/master/models/resnet.lua 12
--
-- Copyright (c) 2016, Facebook, Inc.
-- All rights reserved.
--
-- This source code is licensed under the BSD-style license found in the
-- LICENSE file in the root directory of this source tree. An additional grant
-- of patent rights can be found in the PATENTS file in the same directory.
--
-- The ResNet model definition
--
local nn = require 'nn'
require 'cunn'
local Convolution = cudnn.SpatialConvolution
local Avg = cudnn.SpatialAveragePooling
local ReLU = cudnn.ReLU
local Max = nn.SpatialMaxPooling
local SBatchNorm = nn.SpatialBatchNormalization
This file has been truncated. show original
model:add(Convolution(3,64,7,7,2,2,3,3))
in pytorch
self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3,
bias=False)
If I change the torch model to model:add(Convolution(3,64,15,20,2,2,3,3))
It will at least allow me to train with 480*640 images…although it will not allow me to fine tune a pretrained model
So I basically have 3 questions
To train on different image sizes can the pretrained models be used?
Do all images in training have to be the same size (I thought fully convolutional networks would allow any input size …this training works with tensorflow and inception-v3)?
How do I fine tune a model with images which are not 224*224?
|
st115034
|
Thanks for the pointer.
Used
model.avgpool = nn.AdaptiveAvgPool2d(1)
To get this to work
|
st115035
|
I try to replicate the net2net 4 torch code in pytorch and here where I stuck. It uses m.output to reach the last feedforward state to compute BatchNorm statistics but as far as I see, it is not possible in Pytorch. Do you have any other alternative method that you might like to suggest? especially @smth
|
st115036
|
this is not possible in pytorch. In pytorch, you might want to implement net2net as a utility function, that the user explicitly uses in their program. It’ll take weights of a Conv2d layer (or Linear layer) and return a new Conv2d layer that’s wider, or a Sequential of 2 Conv2d layers.
|
st115037
|
Hi, I just found that calling index_select() on long columns (dim=1) are much slower than doing it on long rows (dim=0). In fact, for large enough matrices, it’s much faster to first transpose the matrix and call index_select() on rows!
I wonder if it’s a known performance issue and if there’s any way to mitigate the problem?
Here’s the code I used to test. (I’m using Python 2.7.12, PyTorch 0.2.0.post1, and GTX 1080). It builds a 256*N matrix (for various N) and rearranges its rows. Then it does the same for columns.
import time
import numpy as np
import torch
DIM = 256
idxs = np.random.permutation(DIM)
idxs = torch.LongTensor(idxs).cuda()
def do_test(dim, sz, trans=False):
if dim == 0:
# Rearrange 256 rows, each size 'sz'.
A = torch.cuda.FloatTensor(DIM, sz)
else:
# Rearrange 256 columns, each size 'sz'.
A = torch.cuda.FloatTensor(sz, DIM)
A.uniform_(-1.0, 1.0)
B = A.new(A.shape)
torch.cuda.synchronize()
T0 = time.time()
for step in xrange(10):
if trans:
T = A.t().clone()
T.index_select(dim=1-dim, index=idxs, out=B)
else:
A.index_select(dim=dim, index=idxs, out=B)
torch.cuda.synchronize()
T1 = time.time()
print ' %6d : Elapsed %.3f ms' % (sz, (T1 - T0) / 10 * 1000.0)
sizes = [100, 200, 500, 1000, 2000, 5000, 10000, 20000, 50000, 100000]
print 'Rearranging rows:'
for sz in sizes: do_test(0, sz)
print 'Rearranging columns:'
for sz in sizes: do_test(1, sz)
print 'Rearranging columns (with transpose):'
for sz in sizes: do_test(1, sz, True)
Result:
Rearranging rows:
100 : Elapsed 0.051 ms
200 : Elapsed 0.049 ms
500 : Elapsed 0.056 ms
1000 : Elapsed 0.066 ms
2000 : Elapsed 0.084 ms
5000 : Elapsed 0.142 ms
10000 : Elapsed 0.235 ms
20000 : Elapsed 0.419 ms
50000 : Elapsed 0.999 ms
100000 : Elapsed 1.937 ms
Rearranging columns:
100 : Elapsed 0.048 ms
200 : Elapsed 0.051 ms
500 : Elapsed 0.065 ms
1000 : Elapsed 0.111 ms
2000 : Elapsed 0.298 ms
5000 : Elapsed 1.166 ms
10000 : Elapsed 2.953 ms
20000 : Elapsed 7.699 ms
50000 : Elapsed 22.267 ms
100000 : Elapsed 44.207 ms
Rearranging columns (with transpose):
100 : Elapsed 0.051 ms
200 : Elapsed 0.054 ms
500 : Elapsed 0.065 ms
1000 : Elapsed 0.081 ms
2000 : Elapsed 0.118 ms
5000 : Elapsed 0.236 ms
10000 : Elapsed 0.489 ms
20000 : Elapsed 1.472 ms
50000 : Elapsed 8.336 ms
100000 : Elapsed 14.236 ms
As you can see, index_select(dim=1) is much slower than dim=0: more importantly, it grows faster than O(N): N=100000 is about 400 times slower than N=1000, and first transposing the matrix is about three times faster (for N=100000), even considering the time spent on transposing.
Does anyone know what’s going on here?
|
st115038
|
How to understand the backward() in stochastic functions 11 ?
e.g. For Normal distribution, grad_mean = -(output - mean)/std**2, however why it is following this formula ? Is it a derivative of Gaussian PDF ? The forward pass only use output = mean + std*eps where eps ~ N(0, 1), so the gradient w.r.t. mean should be identity ?
|
st115039
|
Gradient formulas are based on Simple Statistical Gradient-Following
Algorithms for Connectionist Reinforcement Learning, available at
http://incompleteideas.net/sutton/williams-92.pdf 10
https://github.com/pytorch/pytorch/blob/master/torch/autograd/_functions/stochastic.py#L3-L5 16
|
st115040
|
Hello Everyone,
My neural network is optimized in the way that the gradients are obtained by a certain algorithm. To pass the gradients to the network, I firstly clear the gradients by:
(1), for i in list(net.parameters()):
i.grad=None
(2), opt.zero_grad()
I assume that both methods do the same thing and have the same effect, is my understanding right?
After clearing the gradients, I pass the gradients ( the calculated gradients are in list type, and each element is an numpy array that has the same shape with the network element ) to the network by:
(1),for i in list(net.parameters()):
i.grad=Variable(torch.from_numpy(GRADIENT_ARRAY))
After passing the gradients, I use optimizer.step() to update the parameters. Does calling optimizer.step() do the same thing as updating the parameters manually?
|
st115041
|
yes, optimizer.step() does the update rule: x.data -= x.grad * learning_rate being a simple example of an update rule.
|
st115042
|
class EncoderRNN(nn.Module):
def __init__(self, input_size, hidden_size, n_layers=1):
super(EncoderRNN, self).__init__()
self.n_layers = n_layers
self.hidden_size = hidden_size
self.embedding = nn.Embedding(input_size, hidden_size)
self.gru = nn.GRU(hidden_size, hidden_size)
def forward(self, input, hidden):
embedded = self.embedding(input).view(1, 1, -1)
output = embedded
for i in range(self.n_layers):
output, hidden = self.gru(output, hidden)
return output, hidden
def initHidden(self):
result = Variable(torch.zeros(1, 1, self.hidden_size))
if use_cuda:
return result.cuda()
else:
return result
|
st115043
|
One GRU for many layers: the weights will be shared across all layers
Use n_layers instead: each layer will learn different weights.
|
st115044
|
Thanks for your response. Which one is more preferred if i need to use multiple layers?
|
st115045
|
look at the source code of the model definition (i.e. the python class definition)
|
st115046
|
Why are there two bias terms in RNNCell when they are pointwise added together? Wouldn’t this be equivalent to using one bias and doubling its gradient? Although I’m not sure if doubling the gradient is the desired behavior…
|
st115047
|
Solved by smth in post #2
Yes, it’d be equivalent to just learn 1 bias term. I guess it’s just convention to learn two bias terms for an Elman cell (or we just implemented it exactly as the formula says, rather than thinking this through).
Here’s the relevant code that I double-checked https://github.com/pytorch/pytorch/blo…
|
st115048
|
Yes, it’d be equivalent to just learn 1 bias term. I guess it’s just convention to learn two bias terms for an Elman cell (or we just implemented it exactly as the formula says, rather than thinking this through).
Here’s the relevant code that I double-checked https://github.com/pytorch/pytorch/blob/master/torch/nn/_functions/rnn.py#L14 12
|
st115049
|
a = Variable(torch.LongTensor(torch.rand(2, 3)))
The error:
TypeError Traceback (most recent call last)
<ipython-input-12-f8802b707ca7> in <module>()
----> 1 a = Variable(torch.LongTensor(torch.rand(2, 3)))
TypeError: torch.LongTensor constructor received an invalid combination of arguments - got (torch.FloatTensor), but expected one of:
* no arguments
* (int ...)
didn't match because some of the arguments have invalid types: (torch.FloatTensor)
* (torch.LongTensor viewed_tensor)
didn't match because some of the arguments have invalid types: (torch.FloatTensor)
* (torch.Size size)
didn't match because some of the arguments have invalid types: (torch.FloatTensor)
* (torch.LongStorage data)
didn't match because some of the arguments have invalid types: (torch.FloatTensor)
* (Sequence data)
didn't match because some of the arguments have invalid types: (torch.FloatTensor)
Why is not there a conversion?Is it just a code style, or just a little bug?
|
st115050
|
I’m not sure why you are trying to Construct a LongTensor from a FloatTensor (we do not support this in the constructor).
Instead, a much simpler and equivalent way:
a = Variable(torch.rand(2, 3).long())
|
st115051
|
Hi,
I am trying to train a model by using GPU. I can create simple tensors and do operations on them with CUDA. However, when I tried to build a complex model, it raises an exception “CUDNN_STATUS_NOT_INITIALIZED”.
raise CuDNNError(status)
CuDNNError: 1: b'CUDNN_STATUS_NOT_INITIALIZED'
CUDNN_STATUS_NOT_INITIALIZED\
Traceback (most recent call last):
File "<ipython-input-33-b62d175cac47>", line 1, in <module>
CUDNN_STATUS_NOT_INITIALIZED
NameError: name 'CUDNN_STATUS_NOT_INITIALIZED' is not defined
I did some research, similar problem reported in Tensorflow discussions. Few people reported that it is a memory issue, if you can limit TF to use fraction of GPU, it solves the problem. See the link: TF Discussion 19 . I actually once get a memory related exception but I can’t reproduce it.
How do we do that in PyTorch?
Thank you
|
st115052
|
Hi!
So…this might be a silly question, but where is matmul?
As can be seen from the code below from my Python interpreter,
mm works fine (as well as @, which might or might not be the same as mm or matmul , I am not sure),
but matmul doesn’t seem to exist in neither torch or a Tensor object.
What’s going on?!
>>> import torch
>>> a=torch.randn(5,6)
>>> b=torch.randn(6,7)
>>> a.mm(b)
1.3938 1.3466 1.8738 2.9177 -2.7334 1.9803 0.1643
1.2277 1.2948 2.2676 1.6977 -2.8532 5.0795 2.0144
-1.9988 0.2808 -1.6006 -2.8685 0.5934 0.1643 0.1560
-0.6365 -1.3311 -4.0025 -2.5772 0.5418 -2.0688 1.9729
-0.3097 0.3623 1.6439 0.3341 1.5335 -2.8216 -2.1900
[torch.FloatTensor of size 5x7]
>>> a.matmul(b)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: 'torch.FloatTensor' object has no attribute 'matmul'
>>> torch.matmul(b)
Traceback (most recent call last):
File "<stdin>", line 1, in <module>
AttributeError: module 'torch' has no attribute 'matmul'
|
st115053
|
0.1.12_2
Thanks, it was only a few months from the point when I installed it.
I guess things are moving super fast
|
st115054
|
I’m implementing a library for training paragraph vector models as proposed by Q. V. Le et al. (Distributed Representations of Sentences and Documents).
The code is available on GitHub: https://github.com/inejc/paragraph-vectors 410
I would appreciate any kind of feedback. Contributions in any form are also more than welcome (I have already opened some issues regarding future work).
|
st115055
|
Hi all,
I use PyTorch version 0.2.0_4 and get an IndexError which I cannot explain:
print("X:", x.size())
print("TYPE:", type(self.neuron_map[k]))
gives
X: torch.Size([25, 8])
TYPE: <class 'list'>
Now
x[:, self.neuron_map[k]]
results in
IndexError: When performing advanced indexing the indexing objects must be LongTensors or convertible to LongTensors
I cannot understand why this happens and I have no idea how to fix this. Any help appreciated.
|
st115056
|
can you do:
print(self.neuron_map[k]), I’m curious of it’s contents.
Also try:
x[:, torch.LongTensor(self.neuron_map[k])]
|
st115057
|
print("INDS:", self.neuron_map[k])
results in:
INDS: [0, 1]
Then,
inds = torch.LongTensor(self.neuron_map[k])
runs into
RuntimeError: tried to construct a tensor from a int sequence, but found an item of type numpy.int64 at index (0)
I actually found a workaround:
inds = np.array(self.neuron_map[k], dtype=np.int64)
inds = torch.LongTensor(inds)
nn_list.append(self.linears[k](x[:, inds]))
I actually have an additional question. The reason, I am splitting the tensor is to apply linear units (like in last posted code line). For the result, i use:
x_out = torch.cat(nn_list, 1)
How efficient is this, as compared to manually implement an autograd.Function (forward and backward)?
|
st115058
|
it should be pretty efficient if x[:, inds] is large enough. the matrix multiply will prob dominate the cost.
Writing a batched matrix multiply by hand is not easy to do efficiently.
|
st115059
|
Hi,
I’m just starting with pytorch, so starting the models from the basic. So I was implementing the numpy model into pytorch. Following is the code I was trying.
import torch
import numpy as np
import pandas as pd
admissions = pd.read_csv('https://stats.idre.ucla.edu/stat/data/binary.csv')
# Make dummy variables for rank
data = pd.concat([admissions, pd.get_dummies(admissions['rank'], prefix='rank')], axis=1)
data = data.drop('rank', axis=1)
# Standarize features
for field in ['gre', 'gpa']:
mean, std = data[field].mean(), data[field].std()
data.loc[:, field] = (data[field] - mean) / std
# Split off random 10% of the data for testing
np.random.seed(21)
sample = np.random.choice(data.index, size=int(len(data) * 0.9), replace=False)
data, test_data = data.ix[sample], data.drop(sample)
# Split into features and targets
features, targets = data.drop('admit', axis=1), data['admit']
features_test, targets_test = test_data.drop('admit', axis=1), test_data['admit']
dtype = torch.FloatTensor
m = torch.nn.Sigmoid()
n_hidden = 2
epochs = 10
learnrate = 0.005
n_records, n_features = features.shape
last_loss = None
weights_input_hidden = torch.randn(n_features, n_hidden).type(dtype)
weights_hidden_output = torch.randn(n_hidden).type(dtype)
for e in range(epochs):
del_w_input_hidden = torch.from_numpy(np.zeros(weights_input_hidden.size())).type(dtype)
del_w_hidden_output = torch.from_numpy(np.zeros(weights_hidden_output.size())).type(dtype)
for x, y in zip(features.values, targets):
hidden_input = torch.mm(x, weights_input_hidden)
hidden_output = m(hidden_input)
output = m(torch.mm(hidden_output, weights_hidden_output))
error = y - output
output_error_term = error * output * (1 - output)
hidden_error = torch.mm(output_error_term, weights_hidden_output)
hidden_error_term = hidden_error * hidden_output * (1 - hidden_output)
del_w_hidden_output += output_error_term * hidden_output
del_w_input_hidden += hidden_error_term * x[:, None]
weights_input_hidden += learnrate * del_w_input_hidden / n_records
weights_hidden_output += learnrate * del_w_hidden_output / n_records
if e % (epochs / 10) == 0:
hidden_output = m(torch.mm(x, weights_input_hidden))
out = m(np.dot(hidden_output,
weights_hidden_output))
loss = np.mean((out - targets) ** 2)
if last_loss and last_loss < loss:
print("Train loss: ", loss, " WARNING - Loss Increasing")
else:
print("Train loss: ", loss)
last_loss = loss
hidden = m(torch.mm(features_test, weights_input_hidden))
out = m(torch.mm(hidden, weights_hidden_output))
predictions = out > 0.5
accuracy = np.mean(predictions == targets_test)
print("Prediction accuracy: {:.3f}".format(accuracy))
The error I’m getting is the following:
Traceback (most recent call last):
File “pytorch_tutorial.py”, line 50, in
hidden_input = torch.mm(x, weights_input_hidden)
TypeError: torch.mm received an invalid combination of arguments - got (numpy.ndarray, torch.FloatTensor), but expected one of:
(torch.SparseFloatTensor mat1, torch.FloatTensor mat2)
didn’t match because some of the arguments have invalid types: (!numpy.ndarray!, torch.FloatTensor)
(torch.FloatTensor source, torch.FloatTensor mat2)
didn’t match because some of the arguments have invalid types: (!numpy.ndarray!, torch.FloatTensor)
I’m not getting how to convert the “x” into “torch.FloatTensor”.
If someone can please guide me, how to resolve the issue.
Edit:
For comparison I’m putting the numpy code as well.
def sigmoid(x):
return 1 / (1 + np.exp(-x))
n_hidden = 2
epochs = 10
learnrate = 0.005
n_records, n_features = features.shape
last_loss = None
weights_input_hidden = np.random.normal(scale=1 / n_features ** .5,
size=(n_features, n_hidden))
weights_hidden_output = np.random.normal(scale=1 / n_features ** .5,
size=n_hidden)
for e in range(epochs):
del_w_input_hidden = np.zeros(weights_input_hidden.shape)
del_w_hidden_output = np.zeros(weights_hidden_output.shape)
for x, y in zip(features.values, targets):
hidden_input = np.dot(x, weights_input_hidden)
hidden_output = sigmoid(hidden_input)
output = sigmoid(np.dot(hidden_output, weights_hidden_output))
error = y - output
output_error_term = error * output * (1 - output)
hidden_error = np.dot(output_error_term, weights_hidden_output)
hidden_error_term = hidden_error * hidden_output * (1 - hidden_output)
del_w_hidden_output += output_error_term * hidden_output
del_w_input_hidden += hidden_error_term * x[:, None]
weights_input_hidden += learnrate * del_w_input_hidden / n_records
weights_hidden_output += learnrate * del_w_hidden_output / n_records
if e % (epochs / 10) == 0:
hidden_output = sigmoid(np.dot(x, weights_input_hidden))
out = sigmoid(np.dot(hidden_output,
weights_hidden_output))
loss = np.mean((out - targets) ** 2)
if last_loss and last_loss < loss:
print("Train loss: ", loss, " WARNING - Loss Increasing")
else:
print("Train loss: ", loss)
last_loss = loss
hidden = sigmoid(np.dot(features_test, weights_input_hidden))
out = sigmoid(np.dot(hidden, weights_hidden_output))
predictions = out > 0.5
accuracy = np.mean(predictions == targets_test)
print("Prediction accuracy: {:.3f}".format(accuracy))
Thank you!
|
st115060
|
Going from numpy to pytorch tensors and back is very simple.
Pytorch tensor to Numpy array:
numpy_array = pytorch_tensor.numpy()
Numpy array to Pytorch tensor:
pytorch_tensor = torch.from_numpy(numpy_array)
More info here. 4
|
st115061
|
dhpollack:
x = torch.from_numpy(x).float()
@dhpollack: Thanks for your reply. But after implementing your suggestion, I’m getting the following error:
Traceback (most recent call last):
File “pytorch_tutorial.py”, line 50, in
hidden_input = torch.mm(x, weights_input_hidden)
RuntimeError: matrices expected, got 1D, 2D tensors at d:\downloads\pytorch-master-1\torch\lib\th\generic/THTensorMath.c:1233
This code works perfectly when I’m running using numpy. So during conversion to pytorch I’m making some mistake.
|
st115062
|
What is happening is that numpy is more lenient with regards to vector/matrix multiplication than pytorch. So you need to make one or both of the tensors into 2 dimensional tensors rather than 1d. You should look at the size of each tensor (x.size(), weights_input_hidden.size()), you’ll find one or both have just one dimension. To add dummy dimensions use any (but not all!) of the following:
x = x.unsqueeze(0)
x.unsqueeze_(0)
x = x.view(1, -1).contiguous()
|
st115063
|
Full example of going from Numpy to PyTorch for binary classification:
github.com
QuantScientist/Deep-Learning-Boot-Camp/blob/master/day02-PyTORCH-and-PyCUDA/PyTorch/18-PyTorch-NUMER.AI-Binary-Classification-BCELoss-0.691839667509 .ipynb 8
{
"cells": [
{
"cell_type": "markdown",
"metadata": {
"slideshow": {
"slide_type": "slide"
}
},
"source": [
"# Deep Learning Bootcamp November 2017, GPU Computing for Data Scientists\n",
"\n",
"<img src=\"../images/bcamp.png\" align=\"center\">\n",
"\n",
"## 18 PyTorch NUMER.AI Deep Learning Binary Classification using BCELoss \n",
"\n",
"Web: https://www.meetup.com/Tel-Aviv-Deep-Learning-Bootcamp/events/241762893/\n",
"\n",
"Notebooks: <a href=\"https://github.com/QuantScientist/Data-Science-PyCUDA-GPU\"> On GitHub</a>\n",
"\n",
This file has been truncated. show original
|
st115064
|
@QuantScientist: Thanks for sharing the link. I already checked your link and it’s a wonderful presentation. But I wanted a simpler conversion, which I already did. But yours is next level of complexity I’ll try.
|
st115065
|
Hi guys,
I’m new one to pytorch and cuda.(BTW, pytorch is quite friendly to new ones )
And I’m just confused when i’m reading these codes:
`net = torch.nn.DataParallel(net)`
`net = net.cuda()`
I only know that cuda() and DataParallel() has something to do with GPU and parallel computation etc.
But what’s the difference in those two lines? What have they done respectively?
Also, what happens if the second line is removed?
Thanks in advance!
|
st115066
|
DataParallel is a line to use multiple GPUs for processing your model.
.cuda() is just putting your model on GPU.
|
st115067
|
I am trying to train siamese network for sentence similarity task. i am using same lstm with pack_padded_sequence to two sentences and getting the norm difference between the two final output of two sequences as similarity and finding the error with actual similarity score and backpropagating. after some time (in first epoch only) gradients are becoming very low and then they are becoming nan.
|
Subsets and Splits
No community queries yet
The top public SQL queries from the community will appear here once available.