Our overall approach centers around the Bottom-Up and Top-Down Attention model, as designed by Anderson et al. We used this framework as a starting point for further experimentation, implementing, in addition to various hyperparameter tunings, two additional model architectures. First, we reduced the complexity of Bottom-Up and Top-Down by considering only a simple LSTM architecture. Then, taking inspiration from the Transformer architecture, we implement a non-recurrent model which does not need to keep track of an internal state across time. Our results are comparable to the author’s implementation of the Bottom- Up and Top Down Attention model. Our code serves as a baseline for future experiments which are done using the Pytorch framework.

Our implementation can be improved using some more hyperparameter tuning and using some tricks like gradient clipping and using ReLU instead of tanh. We hope our model serves as the baseline for future experiments.

We use the Adam optimizer with a learning rate of 0.0001 and teacher forcing during training with a batch size of 100. The training was completed with Nvidia P40 GPUs in approximately 8 GPU hours which is significantly less than author’s 18 GPU hours on Titan X GPUs. During testing, a beam width of 5 was found to be the most effective.

Ablation study:

Machine configuration used for testing: Nvidia P40 GPUs card with 24GB memory (though a machine with lesser memory would work just fine)

We use the Karpathy splits as described in Deep visual-semantic alignments for generating image descriptions.. The Bottom-Up image features are used directly from here. Please refer to this repo for clarifications. The annotations are downloaded from the COCO website (2014 train val annotations). All the models have been trained from scratch.

The code takes around 8 hours to train on the karpathy train split.

What things you need to install the software and how to install them

Software used:

1. Pytorch 0.4.1
2. Python 2.7

Dependencies: Create a conda environment using the captioning_env.yml file. Use: conda env create -f captioning_env.yml

If you are not using conda as a package manager, refer to the yml file and install the libraries manually.

0. Data

• We use image features generated by a resnet101 Faster RCNN network directly. These are available here as mentioned previously. Convert the tsv file to hdf5 files based on [this] (https://github.com/hengyuan-hu/bottom-up-attention-vqa/blob/master/tools/detection_features_converter.py) code. Now generate the karpathy splits using [this] (https://github.com/njchoma/transformer_image_caption/blob/master/src/data_helpers/create_ks_splits_h5.py) code.

• For the annotations, download from the COCO website with the link mentioned earlier and run [this] (https://github.com/njchoma/transformer_image_caption/blob/master/src/data_helpers/create_ks_splits1.py) script to generate the karpathy splits.

• We will be using [dataloader_ks.py] (https://github.com/njchoma/transformer_image_caption/blob/master/src/data_helpers/data_loader_ks.py) script which will handle the data loading part.

1. Training

• Edit the main.sh file by changing the path variables and activating the appropriate conda environment. Run the same script with the appropriate arguments. The arguments have been listed in the src/utils_experiment.py file.

• After the model has been trained, run src/evaluate_test.py

2. Evaluation

• For calculating the metrics such as BLEU, run [this] (https://github.com/njchoma/transformer_image_caption/blob/master/src/eval/coco_caption/eval.py) script with the appropriate test annotations (pkl file) and the json file generated by the model.

This project is licensed under the MIT License - see the LICENSE file for details

1. Nicholas Choma (New York University)
2. Omkar Damle (New York University)

Note: Equal contribution from both contributors.

This code was produced as a part of my course project at New York University. I would like to thank Prof. Fergus for his guidance and providing access to the GPUs.

References:

1. Code for metrics evaluation was borrowed from https://github.com/tylin/coco-caption
2. Code for converting the image features from tsv to hdf5 is based on https://github.com/hengyuan-hu/bottom-up-attention-vqa