How to use .npz files from ResNet feature extractor

[pratyushmore]

Hi Rakshith,

Thank you for putting up your code! It is extremely helpful.

I used the link (https://github.com/akirafukui/vqa-mcb/tree/master/preprocess) in your README to extract ResNet features for use in the adversarial training.

The extractor gave me a large number of .npz files. Your example makes it seem as though everything should be in one file. I am thus wondering how I should specify to the program which files to use as image features.

[rakshithShetty]

Hi,

I usually prefer to concat all features into one large npy file (Nxfeat dimensions). But current code also supports list of numpy files. Just create a text file with list of feature files and name it as .npzl . If the extension of the input feature file is is '.npzl', the data provider parses it as a list of numpy files. Take a look at line 666 in imagernn/data_provider.py and the function called there.

[pratyushmore]

Hi Rakshit,

Thank you for your response. Essentially, I just need to create a list of strings specifying all the npz files?

Also, should this include the train, and val files? Or should these be in three separate files?

[pratyushmore]

I also see the auxiliary input files. What do these entail? I do not quite understand how to create them.

[pratyushmore]

I tried running without the aux file, setting evaluation to image features. Got the following error:

Traceback (most recent call last):
  File "train_adversarial_caption_gen_v2.py", line 663, in <module>
    main(params)
  File "train_adversarial_caption_gen_v2.py", line 241, in main
    generator.model_th, params, xI = xI, xAux = xAux)
  File "/home/pmore/code/captionGAN/imagernn/lstm_generatorTheano.py", line 349, in build_prediction_model
    accLogProb, Idx, wOut_emb, updates, seq_lengths = self.lstm_advers_gen_layer(tparams, embImg, xAuxEmb, options, prefix=options['generator'])
  File "/home/pmore/code/captionGAN/imagernn/lstm_generatorTheano.py", line 592, in lstm_advers_gen_layer
    tparams[_p(prefix,'W_aux')]), n_samp,axis=0)
KeyError: 'lstm_W_aux'

[pratyushmore]

@rakshithShetty Was just wondering if you had seen these. I am working on a research paper for which I would like to cite your paper. Would appreciate your help in running your model.

[rakshithShetty]

Hi, Sorry for the delayed response. I use the image_feature_file and aux_inp_file to provide two different features for the captioning LSTM. In our paper we used two features, CNN features extracted from resnet and object detection features from Faster RCNN (fasterRcnn_clasDetFEat80.npy in the example command). I have now uploaded this feature file to the google drive since it's a small file. You can use it if you wish. (https://drive.google.com/drive/folders/0B76QzqVJdOJ5TV9FMjhpVmlsTFE) To see the details of how this feature is extracted refer to section 3.1 of (https://dl.acm.org/citation.cfm?id=2983571)

Since I have been using both feature files recently some parts of the code has not been guarded properly for the absence of the aux input. That's the error you are seeing.

[rakshithShetty]

Between, I have pushed the faster rcnn codebase I used to extract this feature vector here. It only has bare-bones usage instructions for now.

https://github.com/rakshithShetty/py-faster-rcnn-featextract

[pratyushmore]

Hi Rakshith.

Thank you so much. I set the feature file to the file you uploaded. And set the aux inp file to the resnet features.

I get the following error:

Traceback (most recent call last):
  File "train_adversarial_caption_gen_v2.py", line 663, in <module>
    main(params)
  File "train_adversarial_caption_gen_v2.py", line 187, in main
    dp = getDataProvider(params)
  File "/home/pmore/code/captionGAN/imagernn/data_provider.py", line 597, in getDataProvider
    return BasicDataProvider(params)
  File "/home/pmore/code/captionGAN/imagernn/data_provider.py", line 94, in __init__
    imgIdSet = allIdSet)
  File "/home/pmore/code/captionGAN/imagernn/data_provider.py", line 629, in loadFromLbls
    features, hdf5Flag = readFeaturesFromFile(features_path, idxes = feat_load_list)
  File "/home/pmore/code/captionGAN/imagernn/data_provider.py", line 814, in readFeaturesFromFile
    features, hdf5_flag = loadSingleFeat(filename, idxes, mat_new_ver)
  File "/home/pmore/code/captionGAN/imagernn/data_provider.py", line 797, in loadSingleFeat
    return features, hdf5_flag
UnboundLocalError: local variable 'features' referenced before assignment

Any idea as to why this is happening?

[pratyushmore]

Update:

I used the faster_RCNN features as the aux input. I used the ResNet features as feature file input. Also used disk_feature = 1.

After these changes, data reading seems to work.

I am now getting the following error:

Traceback (most recent call last):
  File "train_adversarial_caption_gen_v2.py", line 663, in <module>
    main(params)
  File "train_adversarial_caption_gen_v2.py", line 399, in main
    tsc_max, tsc_mean, tsc_min = eval_gen_samps(f_gen_only, dp, params, misc, params['rev_eval'], **trackMetargs)
  File "train_adversarial_caption_gen_v2.py", line 47, in eval_gen_samps
    g_out = gen_func(*inp)
  File "/home/pmore/anaconda2/envs/caption_gan/lib/python2.7/site-packages/theano/compile/function_module.py", line 917, in __call__
    storage_map=getattr(self.fn, 'storage_map', None))
  File "/home/pmore/anaconda2/envs/caption_gan/lib/python2.7/site-packages/theano/gof/link.py", line 325, in raise_with_op
    reraise(exc_type, exc_value, exc_trace)
  File "/home/pmore/anaconda2/envs/caption_gan/lib/python2.7/site-packages/theano/compile/function_module.py", line 903, in __call__
    self.fn() if output_subset is None else\
  File "pygpu/blas.pyx", line 47, in pygpu.blas.pygpu_blas_rgemm
pygpu.gpuarray.GpuArrayException: ('mismatched shapes', 2)
Apply node that caused the error: GpuDot22(GpuFromHost<None>.0, WIemb_aux)
Toposort index: 23
Inputs types: [GpuArrayType<None>(float32, matrix), GpuArrayType<None>(float32, matrix)]
Inputs shapes: [(100, 80), (2048, 512)]
Inputs strides: [(320, 4), (2048, 4)]
Inputs values: ['not shown', 'not shown']
Outputs clients: [[GpuElemwise{Add}[(0, 0)]<gpuarray>(GpuDot22.0, InplaceGpuDimShuffle{x,0}.0)]]

Backtrace when the node is created(use Theano flag traceback.limit=N to make it longer):
  File "train_adversarial_caption_gen_v2.py", line 663, in <module>
    main(params)
  File "train_adversarial_caption_gen_v2.py", line 241, in main
    generator.model_th, params, xI = xI, xAux = xAux)
  File "/home/pmore/code/captionGAN/imagernn/lstm_generatorTheano.py", line 338, in build_prediction_model
    xAuxEmb = tensor.dot(xAux,tparams['WIemb_aux']) + tparams['b_Img_aux']
  File "train_adversarial_caption_gen_v2.py", line 663, in <module>
    main(params)
  File "train_adversarial_caption_gen_v2.py", line 241, in main
    generator.model_th, params, xI = xI, xAux = xAux)
  File "/home/pmore/code/captionGAN/imagernn/lstm_generatorTheano.py", line 338, in build_prediction_model
    xAuxEmb = tensor.dot(xAux,tparams['WIemb_aux']) + tparams['b_Img_aux']

[Noahsark]

Hi, Rakshith, May I ask if the order of features in .npz file has to be exactly same with the order in your "labels.txt" mapping file?

Thanks.