Skip to content
COMIC: This is the code repo of our TMM2019 work titled "COMIC: Towards a Compact Image Captioning Model with Attention".
Branch: master
Clone or download
Type Name Latest commit message Commit time
Failed to load latest commit information.
common Update Jun 11, 2019
datasets/preprocessing description update Jun 13, 2019
src checkpoint path fix Jun 14, 2019
.gitattributes Create .gitattributes Jun 13, 2019
.gitignore add pretrained, cleanup, bug fix Jun 3, 2019
LICENSE Update LICENSE Jun 10, 2019 description update Jun 13, 2019
TMM.png Add files via upload Jun 3, 2019

COMIC: Towards a Compact Image Captioning Model with Attention

Updated on 12 June 2019 (Self-Critical Sequence Training (SCST))

Updated on 06 June 2019 (Pre-trained model repo)

Released on 03 June 2019.


This is the code repo of our TMM 2019 work titled "COMIC: Towards A Compact Image Captioning Model with Attention". In this paper, we tackle the problem of compactness of image captioning models which is hitherto unexplored. We showed competitive results on both MS-COCO and InstaPIC-1.1M datasets despite having an embedding vocabularly size that is 39x-99x smaller.

Some pre-trained model checkpoints are available at this repo.


If you find this repository useful for your research or work, please cite:

  title={COMIC: Towards A Compact Image Captioning Model with Attention},
  author={Tan, Jia Huei and Chan, Chee Seng and Chuah, Joon Huang},
  journal={IEEE Transactions on Multimedia},
  year={in Press},


  • tensorflow 1.9.0
  • python 2.7
  • java 1.8.0
  • tqdm >= 4.24.0
  • Pillow >= 3.1.2
  • packaging >= 19.0
  • requests >= 2.18.4

Running the code

More examples are given in

First setup

Run ./src/ This will download the required Stanford models and run all the dataset pre-processing scripts.

Training models

The training scheme is as follows:

  1. Start with decoder mode (freezing the CNN)
  2. Followed by cnn_finetune mode
  3. Finally, scst mode


for mode in 'decoder' 'cnn_finetune' 'scst'
    python  \
        --train_mode ${mode}

# InstaPIC
for mode in 'decoder' 'cnn_finetune' 'scst'
    python  \
        --train_mode ${mode}  \
        --dataset_file_pattern 'insta_{}_v25595_s15'


for mode in 'decoder' 'cnn_finetune' 'scst'
    python  \
        --train_mode ${mode}  \
        --token_type 'word'  \
        --cnn_fm_projection 'none'  \
        --attn_num_heads 1

# InstaPIC
for mode in 'decoder' 'cnn_finetune' 'scst'
    python  \
        --train_mode ${mode}  \
        --dataset_file_pattern 'insta_{}_v25595_s15'  \
        --token_type 'word'  \
        --cnn_fm_projection 'none'  \
        --attn_num_heads 1


Just point to the directory containing the checkpoints. Model configurations are loaded from config.pkl.

python  \
	--infer_checkpoints_dir 'mscoco/word_add_softmax_h8_tie_lstm_run_01'

# InstaPIC
python  \
	--infer_checkpoints_dir 'insta/word_add_softmax_h8_ind_lstm_run_01'  \
	--annotations_file 'insta_testval_raw.json'

Avoid re-downloading datasets

Re-downloading can be avoided by:

  1. Editing
  2. Providing the path to the directory containing the dataset files
python --dataset_dir /path/to/coco/dataset
python --dataset_dir /path/to/insta/dataset

In the same way, both and accept alternative dataset paths.

python --dataset_dir /path/to/dataset
python --dataset_dir /path/to/dataset

This code assumes the following dataset directory structures:


+-- captions
|   +-- {folder and files generated by}
+-- test2014
|   +-- {image files}
+-- train2014
|   +-- {image files}
+-- val2014
    +-- {image files}


+-- captions
|   +-- {folder and files generated by}
+-- images
|   +-- {image files}
+-- json
    +-- insta-caption-test1.json
    +-- insta-caption-train.json

Project structure

+-- common
|   +-- {shared libraries and utility functions}
+-- datasets
|   +-- preprocessing
|       +-- {dataset pre-processing scripts}
+-- pretrained
|   +-- {pre-trained checkpoints for some COMIC models. Details are provided in a separate README.}
+-- src
    +-- {main scripts}

Differences compared to our TMM paper

To match the settings as described in our paper, set the legacy argument of to True (the default is False). This will override some of the provided arguments.

When using the default arguments, the differences compared to our TMM paper are:

  • Attention map dropout is set to 0.1 instead of none
  • RNN init method is changed to x_{t=-1} = W_I * CNN(I) from h_{t=-1} = W_I tanh (LN (I_{embed} ))
  • Changed training scheme (learning rate, ADAM epsilon)

Changes that can be enabled:

  • CNN fine-tuning via the train_mode flag. Model is initialised using the last training checkpoint of RNN training.
  • RNN variational dropout [arxiv] [tf]
  • Context layer (linear projection after attention)
  • Self-Critical Sequence Training (SCST) [arxiv]

Performance on MS-COCO (using Inception-V1 and LSTM)

Default mode Decoder params. BLEU-1 BLEU-4 CIDEr SPICE
Baseline 12.7 M 0.716 0.311 0.937 0.174
COMIC-256 4.3 M 0.713 0.308 0.944 0.176
(+ CNN fine-tune) 0.729 0.328 1.001 0.185
(+ SCST ^) 0.753 0.344 1.050 0.190

^ SCST using beam search sampling strategy as described in this paper.

Legacy mode Decoder params. BLEU-1 BLEU-4 CIDEr SPICE
Baseline 12.2 M 0.707 0.300 0.906 0.169
(0.701) (0.296) (0.885) (0.167)
COMIC-256 4.0 M 0.711 0.302 0.913 0.170
(0.706) (0.292) (0.881) (0.164)

Note that scores in brackets () indicate figures stated in our TMM paper. The differences are due to reimplementation from TF-1.2.

Please see above for info on downloading checkpoints of the models listed above.

Main arguments

  • Main:
    • train_mode: The training regime. Choices are decoder, cnn_finetune, scst.
    • token_type: Language model. Choices are word, radix, char.
    • legacy: If True, will match settings as described in paper.
  • CNN:
    • cnn_name: CNN model name.
    • cnn_input_size: CNN input size.
    • cnn_fm_attention: End point name of feature map for attention.
    • cnn_fm_projection: Feature map projection method. Choices are none, independent, tied.
  • RNN:
    • rnn_name: Type of RNN. Choices are LSTM, LN_LSTM, GRU.
    • rnn_size: Number of RNN units.
    • rnn_word_size: Size of word embedding.
    • rnn_init_method: RNN init method. Choices are project_hidden, first_input.
    • rnn_recurr_dropout: If True, enable variational recurrent dropout.
  • Attention:
    • attn_num_heads: Number of attention heads.
    • attn_context_layer: If True, add linear projection after multi-head attention.
    • attn_alignment_method: Alignment / composition method. Choices are add_LN, dot.
    • attn_probability_fn: Attention map probability function. Choices are softmax, sigmoid.
  • SCST:
    • scst_beam_size: The beam size for SCST sampling.
    • scst_weight_ciderD: The weight for CIDEr-D metric during SCST training.
    • scst_weight_bleu: The weight for BLEU metrics during SCST training.

  • Main:
    • infer_set: The split to perform inference on. Choices are test, valid, coco_test, coco_valid. coco_test and coco_valid are for inferencing on the whole test2014 and val2014 sets respectively. These are used for MS-COCO online server evaluation.
    • infer_checkpoints_dir: Directory containing the checkpoint files.
    • infer_checkpoints: Checkpoint numbers to be evaluated. Comma-separated.
    • annotations_file: Annotations / reference file for calculating scores.
  • Inference parameters:
    • infer_beam_size: Beam size of beam search. Pass 1 for greedy search.
    • infer_length_penalty_weight: Length penalty weight used in beam search.
    • infer_max_length: Maximum caption length allowed during inference.
    • batch_size_infer: Inference batch size for parallelism.

Microsoft COCO Caption Evaluation

This code uses the standard coco-caption code with SPICE metric [Link to repo].

To perform online server evaluation:

  1. Infer on coco_test (test2014), rename the JSON output file to captions_test2014__results.json.
  2. Infer on coco_valid (val2014), rename the JSON output file to captions_val2014__results.json.
  3. Zip the files and submit.


Thanks to the developers of:


Suggestions and opinions (both positive and negative) are greatly welcomed. Please contact the authors by sending an email to tan.jia.huei at or cs.chan at

License and Copyright

The project is open source under BSD-3 license (see the LICENSE file).

© 2019 Center of Image and Signal Processing, Faculty of Computer Science and Information Technology, University of Malaya.

You can’t perform that action at this time.