Dataset preperation

1. Run ./data/attribute/prepro_attribute.py
   • check argument --input_filename # output file from shuffle_duplicate_remove.py
   • check argument --output_json, # resulting output_json file. this file finally used for train/val/test
   • check argument --output_h5 # resulting output_h5 file. this file also used for train/val/test
   • check argument image_dim, # default: 342

Training/val

1. Move to the directory, cd /path/to/the/git/repo.
2. Change options, vim opts/opt_attribute.lua
   • check argument input_h5 --output h5 file of prepro_attribute.py
   • check argument input_json --output json file of prepro_attribute.py
   • check argument torch_model --path to pretrained-image encoder network
   • check argument gpus = {1,2,3,4}
   • check argument for image-encoder image_size, crop_size, crop_jitter, flip_jitter = 342, 299, true, false
   • tune hyperparameters for the language-decoder (e.g. rnn_size, num_rnn_layers and drop_prob_lm use_bn='original' etc)
   • tune hyperparameters for optimisation (e.g. batch_size, optimizer, learning_rate etc)

Launching training/val scripts

1. Suppose you want to run on 4 gpu cards
   • check argument local gpus = {1,2,3,4} --in opts/opt_xxx.lua
   • CUDA_VISIBLE_DEVICES=0,1,2,3 luajit train.lua

Plotting learning curve

1. Both log-files for train/val are saved in your checkpoint path automatically.