Michael Shieh

Hi, we used random images from the whole ImageNet and didn't keep the class balanced. We used this [script](https://github.com/tensorflow/models/blob/master/research/inception/inception/data/build_imagenet_data.py) to preprocess the ImageNet and saved the data into 1024 files....

Yes, that's correct. We use ResNet-50 as our baseline model. The code of our baseline was obtained [here](https://github.com/tensorflow/tpu/tree/master/models/official/resnet).

Updated the requirements section in the readme. Thanks!

Hi, you can use scripts/preprocess.sh to generate the supervised data by setting sup_size to your desired data size and set use_equal_split to True. I uploaded the hyperparameters for all data...

Hi, to tune hyperparameters, we split part of the training as the validation. For example, when we used 4,000 examples, we used 3,200 for training and 800 for validation. After...

Hi Roy, We tried to use multiple GPU by using MirroredStrategy, but the overhead is very high. With 4 GPUs, the speedup on using 1 GPU is far lower than...

Hi, This is probably due to a vocabulary mismatch. We used nltk.sent_tokenize and nltk.word_tokenize as our tokenizers, which have some inherent randomness. Could you try to retrain the model using...

Sorry for the late response. The loss plateaus in the middle phase of the training but will gradually improve at the later stage. Did you train the model for 100...

Hi, only the human performance is based on the 3000 sampled questions. All the models‘ performance is measured on the whole test set.

Hi, Sorry for the late reply. You would need to run the src/classifier/prepare_data.py to process the raw inputs.