Testing model: attempt to call method 'double' (a nil value)

[marielvanstav]

Hi there! I'm trying to test my own dataset on the pre-trained model, by running: th main.lua -ngpus 1 -test 1 -dataset path/to/data.h5 -ingrW2V /path/to/vocab.bin -semantic 0 -loadsnap im2recipe_model.t7

I get the following output:

{ resnet_model : "/var/scratch/mvstaver/im2recipe_code/data/vision/resnet-50.t7" extract : "test" dispfreq : 1000 batchSize : 150 maxImgs : 5 stDim : 1024 imstore : 256 imfeatDim : 2048 niters : -1 valfreq : 10000 gpu : 1 loadsnap : "im2recipe_model.t7" test : 1 numClasses : 1048 patience : 3 net : "resnet" nworkers : 4 caffemodel : "/var/scratch/mvstaver/im2recipe_code/data/vision/VGG_ILSVRC_16_layers.caffemodel" irnnDim : 300 embDim : 1024 cosw : 0.98 seed : 4242 semantic : 0 maxIngrs : 20 ingrW2VDim : 300 clsw : 0.01 dec_lr : 1 imsize : 224 n_layer_trijoint : 7 iter_swap : -1 mismatchFreq : 0.8 remove : 2 dataset : "/var/scratch/mvstaver/data/jamie_oliver/data.h5" proto : "/var/scratch/mvstaver/im2recipe_code/data/vision/VGG_ILSVRC_16_layers_deploy.prototxt" ngpus : 1 freeze_first : "vision" lr : 0.0001 ingrW2V : "/var/scratch/mvstaver/data/recipe1m/vocab.bin" maxSeqlen : 20 optim : "adam" srnnDim : 1024 snapfile : "snaps/resnet_reg" backend : "cudnn" rundesc : "" nRNNs : 1 } Loading model from im2recipe_model.t7
loading data
data loaded Testing model
3000
/home/mvstaver/torch/install/bin/luajit: ...mvstaver/torch/install/share/lua/5.1/threads/threads.lua:179: [thread 1 endcallback] /var/scratch/mvstaver/im2recipe_code/drivers/test.lua:80: attempt to call method 'double' (a nil value) stack traceback: /var/scratch/mvstaver/im2recipe_code/drivers/test.lua:80: in function </var/scratch/mvstaver/im2recipe_code/drivers/test.lua:40> [C]: in function 'xpcall' ...mvstaver/torch/install/share/lua/5.1/threads/threads.lua:174: in function 'dojob' ...mvstaver/torch/install/share/lua/5.1/threads/threads.lua:223: in function 'addjob' /var/scratch/mvstaver/im2recipe_code/drivers/test.lua:38: in function </var/scratch/mvstaver/im2recipe_code/drivers/test.lua:24> /var/scratch/mvstaver/im2recipe_code/drivers/init.lua:47: in function </var/scratch/mvstaver/im2recipe_code/drivers/init.lua:45> /var/scratch/mvstaver/im2recipe_code/drivers/init.lua:43: in function 'test' main.lua:50: in main chunk [C]: in function 'dofile' ...aver/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk [C]: at 0x004064f0 stack traceback: [C]: in function 'error' ...mvstaver/torch/install/share/lua/5.1/threads/threads.lua:179: in function 'dojob' ...mvstaver/torch/install/share/lua/5.1/threads/threads.lua:223: in function 'addjob' /var/scratch/mvstaver/im2recipe_code/drivers/test.lua:38: in function </var/scratch/mvstaver/im2recipe_code/drivers/test.lua:24> /var/scratch/mvstaver/im2recipe_code/drivers/init.lua:47: in function </var/scratch/mvstaver/im2recipe_code/drivers/init.lua:45> /var/scratch/mvstaver/im2recipe_code/drivers/init.lua:43: in function 'test' main.lua:50: in main chunk [C]: in function 'dofile' ...aver/torch/install/lib/luarocks/rocks/trepl/scm-1/bin/th:150: in main chunk [C]: at 0x004064f0

Does anyone have any idea what could cause this error?

Thanks!

[nhynes]

According to the second line of the stack trace, it looks like batchImgEmbs isn't a Tensor because it's nil. You should ensure that your data are formatted in the same way as the original data.

[marielvanstav]

Thanks for the quick reply! You are right, torch.type(embs) returns table instead of Tensor. When I run in with the default -semantic 1, it does work. Is it possible to test the pre-trained model with -semantic 0 (even though it has been trained with -semantic 1)?

[nhynes]

possible to test the pre-trained model with -semantic 0

It is, but you'd have to do some model surgery to remove the semantic branch and extract just the RNN and CNN.

[marielvanstav]

Okay, thanks! One more question: I want to test other datasets on the pre-trained model. When making the HDF5 files for these datasets, am I right to assume that I should use the word2vec model (and the corresponding vocabulary in vocab.txt) that has been trained on the Recipe1M data, or do I need to train a new word2vec model for each dataset? Thanks for sharing your great work!

[nhynes]

do I need to train a new word2vec model for each dataset?

If you want to use the pre-trained model, you'll have to use the pre-trained word2vec since that's what the later layers of the model expect. Of course, if your test dataset is sufficiently large and different from Recipe1M, you can always try fine tuning the pre-trained model.

[marielvanstav]

Clear! Thanks again!