VQA-Keras-Visual-Question-Answering
VQA-Keras-Visual-Question-Answering copied to clipboard
anantzoid
Overfitting problem
Hi, I implemented the model proposed in the VQA paper the same as what you’ve done and also I used processed features provided by VT Vision Lab. Unfortunately, we ran into an overfitting problem. Could you help us to solve this problem?
This is training configuration:
DROPOUT_RATE = 0.5
EMBEDDING_DIM = 300
EPOCHS = 50
BATCH_SIZE = 256
SEQ_LENGTH = 26
EMBEDDING_TYPE = Glove.6b.300d.en
Image_Features = VGG19
Train_size = 215359
Val_size = 121512
optimizer='rmsprop'
loss='categorical_crossentropy'
As @maryamhashemi said, we faced this problem.After scrutinizing, We've done 2 actions :
- Setting learning rate for rmsprop
- L2 normalizng of VGG19's output features . These 2 actions solved over fitting problem in our task.Now,our accuracy is about 46 percent. Could you explain that is there any thing else to improve this accuracy ?
I have the same issue. also there is an issue when I set mode to Val
for _ in annotations['annotations']: KeyError: 'annotations'
I think the error occurred because val annotation path is set to question file not annotation file.
but when I set it to annotation file, I got this error
abs_val_y = [ques_annotations[ques_id]['multiple_choice_answer_idx'] for ques_id in ques_data['ques tion_id_test']] KeyError: 3506232
Can anyone help me please to solve this error
