mahyarnajibi / fast-rcnn-torch Goto Github PK

when I run the qlua demo.lua, it says:
Unable to connect X11 server (continuing with -nographics)
THCudaCheck FAIL file=/home/wangjian/torch/extra/cutorch/init.c line=255 error=2 : out of memory
qlua: /home/wangjian/torch/install/share/lua/5.1/cudnn/find.lua:192: cuda runtime error (2) : out of memory at /home/wangjian/torch/extra/cutorch/init.c:255
stack traceback:
[C]: at 0x7f8da3db19c0
[C]: in function 'synchronizeAll'
/home/wangjian/torch/install/share/lua/5.1/cudnn/find.lua:192: in function 'reset'
/home/wangjian/torch/install/share/lua/5.1/cudnn/find.lua:554: in main chunk
[C]: in function 'require'
/home/wangjian/torch/install/share/lua/5.1/cudnn/init.lua:319: in main chunk
[C]: in function 'require'
./detection.lua:2: in main chunk
[C]: in function 'require'
demo.lua:3: in main chunk

both scripts download the same file: http://www.umiacs.umd.edu/~najibi/data

Hi, Mahyar,

Thanks very much for sharing this code. It is very helpful!

I am wondering whether you tried to train the AlexNet (without BN) from scratch. What performance can you get? Recently, I can only get ~34% after 40K iterations.

Looking forward to your reply

thanks

My Btech project is using Fast RCNN on GTSRB dataset to detect traffic signs but the annotations are in .txt file. Can you tell how to use this method for my dataset

In GeneralUtils:visualize_detections(), it is supposed to show the regressed bounding box of the most confident class, isn't it?

The actual code seems not to get the bounding box coordinates of the most confident class, but simply pick the first 4 coordinates (corresponding to the first class) in the output:

  local num_boxes = boxes_thresh:size(1)
  local widths  = boxes_thresh[{{},3}] - boxes_thresh[{{},1}]
  local heights = boxes_thresh[{{},4}] - boxes_thresh[{{},2}]

Although in practice, the first class regression results are quite close to the regression results of other classes.

Hi,
I'm trying to download the the pre-trained model and object proposal from the /script folder but the links in get_imgnet_models.sh, get_proposals.sh and get_frcnn_models are not available.
Then I tried to open the links from browser but of course they don't work (ERROR 403: Forbidden).

Are there other links to download that files?
Thanks

/Users/lihong/torch/install/bin/luajit: ./models/AlexNet/FRCNN.lua:60: attempt to index local 'opt' (a nil value)
stack traceback:
# # ./models/AlexNet/FRCNN.lua:60: in function <./models/AlexNet/FRCNN.lua:9>
./network/Net.lua:10: in function '__init'
/Users/lihong/torch/install/share/lua/5.1/torch/init.lua:91: in function </Users/lihong/torch/install/share/lua/5.1/torch/init.lua:87>
[C]: in function 'Net'
demo.lua:14: in main chunk
[C]: in function 'dofile'
...hong/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:145: in main chunk
[C]: at 0x010cc11bd0

Excuse me, how do I get the proposal for Mscoco?

Thanks for releasing the Fast-RCNN code based on torch. I would like to train a new detection model from VGG pre-trained one. But I noticed that you have only published Alex pre-trained model for now. Would you mind to share VGG pre-trained model? Thank you very much!

line 45: _db_name, should be self._db_name
136: _optimState should be self._optimState

Is it possible to train any Imagenet-like Data using this code base for fast-rcnn?
If so what are data preparation steps?

Hi，Mahyarnajibi，I want to train a model using 4 GPUs，and I want to know how to set these three parameters : GPU_ID, nthread, and img_per_batch？

Thanks and look forward to a reply！

If we replace AlexNet with a Resnet50 model for training a fast-RCNN network, how should we initialize the ROI pooling layer? I saw for VGG and AlexNet, it's as follows:
local ROIPooling = detection.ROIPooling(6,6):setSpatialScale(1/16)
How should we calculate the spatial scale for a ResNet50?

Hi, when I run ./scripts/get_proposals.sh I get an Error 404: Not found. It looks like the link https://people.eecs.berkeley.edu/~rbg/fast-rcnn-data/selective_search_data.tgz
is dead. Is there any other source?

THCudaCheck FAIL file=/tmp/luarocks_cutorch-scm-1-2572/cutorch/lib/THC/THCTensorRandom.cu line=20 error=2 : out of memory
qlua: cuda runtime error (2) : out of memory at /tmp/luarocks_cutorch-scm-1-2572/cutorch/lib/THC/THCTensorRandom.cu:20
stack traceback:
[C]: at 0x010a734870
[C]: at 0x1739c460
[C]: in function 'require'
/Users/hbzhang/torch/install/share/lua/5.1/cutorch/init.lua:2: in main chunk
[C]: in function 'require'
/Users/hbzhang/torch/install/share/lua/5.1/cudnn/init.lua:1: in main chunk
[C]: in function 'require'
./detection.lua:2: in main chunk
[C]: in function 'require'
demo.lua:3: in main chunk

Get this error. How I can fix it? Thank you!

Hi ,
Can anyone tell where is the means average precision in the code???
thanks

Hi,
thanks for the reply in the previous issue, but now I have another question.
I'm trying alexnet and the training seems to be working but when I should save the net (in SequentialTrainer.lua) in this way:

network:save(net_path,self._roi_means,self._roi_stds)

I have this error:

THCudaCheck FAIL file=/tmp/luarocks_cutorch-scm-1-T1qml2/cutorch/lib/THC/generic/THCStorage.cu line=66 error=2 : out of memory
/home/torch/install/bin/luajit: /home/torch/install/share/lua/5.1/torch/File.lua:351: cuda runtime error (2) : out of memory at /tmp/luarocks_cutorch-scm-1-T1qml2/cutorch/lib/THC/generic/THCStorage.cu:66
stack traceback:
[C]: in function 'read'
/home/torch/install/share/lua/5.1/torch/File.lua:351: in function </home/torch/install/share/lua/5.1/torch/File.lua:245>
[C]: in function 'read'
/home/torch/install/share/lua/5.1/torch/File.lua:351: in function 'readObject'
/home/torch/install/share/lua/5.1/torch/File.lua:369: in function 'readObject'
/home/torch/install/share/lua/5.1/nn/Module.lua:193: in function 'read'
/home/torch/install/share/lua/5.1/torch/File.lua:351: in function 'readObject'
/home/torch/install/share/lua/5.1/nn/Module.lua:141: in function 'clone'
./network/Net.lua:125: in function 'save'
./train/SequentialTrainer.lua:150: in function '_trainBatch'
./train/SequentialTrainer.lua:97: in function 'train'
./network/NetworkWrapper.lua:40: in function 'trainNetwork'
main_train.lua:48: in main chunk
[C]: in function 'dofile'
.../torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk
[C]: at 0x00405d50

The problem seem to be in this line in Net.lua:

tmp_regressor = self.regressor:clone()

but I don't know why.
I tried to reduce the size of the dataset thinking it was a GPU problem but the error persists.