markdtw / vqa-winner-cvprw-2017 Goto Github PK

Hello we are using you code to preprocess the data and build a few models on top of and are running into some bugs. Don't know why this is not working for us.

https://github.com/SinghJasdeep/CS230-CS224N-VQA

Mainly issues in the loader.py file:

i_batch.append(self.i_feat[iid]) is actually a dictionary not an array of size (36,2048)?

Your help would be greatly appreciated Mark.

Hi, I try to add this dataset to train, and follow this paper's guide, use ' questions whose correct answers overlap the output vocabulary determined on the VQA v2 dataset'. But I got 970,000 questions or so, and it is much larger than 485,000 questions reported in the paper. Have any ideas?

Does this implementation get the performance reported by the paper?

Hi, How to visualize the attention

you mean the accuracy is 64.42% when training or validating? Why 'sigmoid multi-label classifier is also implemented but I can't train based on that'?

Thank you in advance!

I download the code and try to reimplement your score on VQA 2.0 set. Since my computer cannot support the whole training data, I split the - vqa_train_final.json and - coco_features.npy into 7 folds, each set of them grouped by imageid.(like vqa_train_final.0.json contains image ids : [1, 2, 3] the coco_features.0.npy contains image features of [1, 2, 3] and other sets doesnot have any data related to image [1, 2, 3]) I train the model in two way, one is loading the data from 0 to 6 in each epoch and repeat 50 times. the other is loading each data set training 50 epochs and then move on to the next data set.
However, both of them result in a low accuracy, 30% or so. the tokenized question, coco 36 features is downloaded from the link you described. what do you think might be the cause? Thanks

this is how I split the data

def split_images():
    list_train = os.listdir('G:/train2014/')
    list_train.remove('COCO_train2014_000000372405.jpg')
    ids = [int(f[15:27]) for f in list_train]
    length = int(len(ids)/7)+1
    ids_list = [ids[i:i + length] for i in range(0, len(ids), length)]
    for i in range(len(ids_list)):
        np.savetxt("split/imageIds.train." + str(i), ids_list[i], fmt='%d')

def split_json():
    train = json.load(open('vqa_train_final.json'))
    for i in range(7):
        ids = np.loadtxt("split/imageIds.train." + str(i)).astype(int)
        s = set(ids)
        data = []
        for j in range(len(train)):
            if train[j]['image_id'] in s:
                data.append(train[j])
        json.dump(data, open('split/vqa_train_final.json.' + str(i), 'w'))

for k in range(7):
    ids = np.loadtxt("split/imageIds.train." + str(k)).astype(int)
    s = set(ids)
    in_data = {}
    with open(infile, "rt") as tsv_in_file:
        reader = csv.DictReader(tsv_in_file, delimiter='\t', fieldnames = FIELDNAMES)
        i = 0
        for item in reader:
            i = i+1
            if i % 1000 == 0:
                print(k,i)
            try:
                data = {}
                data['image_id'] = int(item['image_id'])
                if data['image_id'] in s:
                    b = base64.decodestring(bytes(item['features'], encoding = "utf8"))
                    data['features'] = np.frombuffer(b, dtype=np.float32).reshape((36, -1))
                    in_data[data['image_id']] = data['features']
            except:
                print('error',item['image_id'])
        np.save('split/coco_features.npy.train.' + str(k), in_data)

Hello, thank you very much for your code, but I raised the following questions during the training, may I ask why？
Traceback (most recent call last):
File "main.py", line 153, in
main()
File "main.py", line 119, in main
train_loader, val_loader, vocab_size, num_classes = prepare_data(args)
File "/home/disk/lmc/vqa-winner-cvprw-2017/data_loader.py", line 72, in prepare_data
VQAv2(root=args.data_root, train=True),
File "/home/disk/lmc/vqa-winner-cvprw-2017/data_loader.py", line 38, in init
que[i] = word2idx[word]
IndexError: index 14 is out of bounds for axis 0 with size 14