README

The code contains the following files:

train.py - contains the main training loop. Includes an argparser to enable choice of hyperparameters (model architecture - for example ResNet depth and decoder type, epochs, learning rate, etc. )
eval.py -  to generate the output json files for uploading to the evaluation server
utils.py - utility functions for PyTorch
vocab.py - to build the vocabulary
models.py - contains code for different models, and an implementation of beam search
preprocess_glove.py - to preprocess data and convert to glove embedding form
test_coco_eval.py - to calculate evaluation metrics on the validation set 

The shell commands to be used are shown below:

Training:
python train.py \
	--models_dir sat_resnet50_lstm1_e512_h512_b256_glove \
	--vocab_path vocab2014.pkl \
	--image_root train2014/ \
	--captions_json annotations/captions_train2014.json \
	--embed_size 300 \
	--rnn_type lstm \
	--resnet_size 50 \
	--glove_embed_path glove_embeddings.pkl

Evaluation:
python eval.py \ 
    --vocab_path vocab2014.pkl 
    --image_root val2014 
    --eval_ckpt_path sat_resnet50_lstm1_e512_h512_b256_glove/model-after-epoch-19.ckpt 
    --glove_embed_path glove_embeddings.pkl 
    --batch_size 512 
    --num_workers 8 
    --embed_size 300 
    --hidden_size 512 
    --rnn_type lstm 
    --resnet_size 50 
    --caption_maxlen 25 
    --results_json_path sat_resnet50_lstm1_e512_h512_b256_glove_model-after-epoch-19.json