rizquuula / cdcgan-zooanimals

cDCGAN - ZooAnimals

Implementation of Conditional Deep Concolutional Generative Adverserial Networks (cDCGAN) on Zoo Animals Dataset

Variables

Terms used in this repository

NC # image channel
IMG_SIZE # image size for training
BATCH_SIZE # training batch size
NZ # num of latent space
NGF # num generator filters
NDF # num discriminator filters
LR # learning rate
BETA1 # beta1 for Adam optimizer
EPOCHS # num of epochs
NGPU # num GPUs
S_SIZE # number of image each category

Dataset Zoo Animals

Used data

Category	Num of Image
cat	5588
dog	5720
penguin	5026
zebra	4068
giraffe	1902
elephant	4112
sheep	5220
turtle	1670

We don't use all of these data, we will sample it using S_SIZE variable

Architecture

cDCGAN architecture used in this research

General architecture

Generator model

Discriminator model

Design Experiments

There are 3 experiments conducted in this research

Finding best S_SIZE, NGF, and NDF

ID	S_SIZE	NGF	NDF
1	500	32	32
2	500	64	64
3	500	128	128
4	500	64	128
5	500	128	64
6	1500	32	32
7	1500	64	64
8	1500	128	128
9	1500	64	128
10	1500	128	64

Results

Trained on Google Colab free version, Tesla T4, 1 GPU, 15GB GPU memory

ID	NGF	NDF	S_SIZE	Training Time (s)	Training Time (H:MM:SS)
1	32	32	500	586.0322123	0:09:46
2	64	64	500	1021.844561	0:17:02
3	64	128	500	2790.996404	0:46:31
4	128	64	500	1729.198487	0:28:49
5	128	128	500	3557.511505	0:59:18
6	32	32	1500	1838.337111	0:30:38
7	64	64	1500	2909.312079	0:48:29
8	64	128	1500	9269.438068	2:34:29
9	128	64	1500	4830.129345	1:20:30
10	128	128	1500	11109.74546	3:05:09

Full results available on Google Drive: https://drive.google.com/drive/folders/1qJ4igSl1s2oKNhAF-xq557lUwqv65cIw?usp=sharing

Finding best LR

We get this as the best model from previous experiment.

ID	NGF	NDF	S_SIZE	Training Time (s)	Training Time (H:MM:SS)
10	128	128	1500	11109.74546	3:05:09

We conduct experiment with LR of 2e-2, 2e-3, 2e-4, and 1e-4, and get 2e-4 as the best LR.

Full code and experiment available on Kaggle: https://www.kaggle.com/code/linkgish/conditional-image-generation-cdcgan-lr-exp

Finding best NZ

We use our best results from previous experiment, and trying to get best NZ size with NZ value of 10, 20, 100, and 200. This experiment conclude with 20 as the best NZ.

Full code and experiment available on Kaggle: https://www.kaggle.com/linkgish/conditional-image-generation-cdcgan-nz-exp

Experiment Conclusion

Based on several experiments and analyzes carried out, this study resulted in the following conclusions.

• Increasing the amount of data, namely the use of S_SIZE=1500 produces better results than S_SIZE=1500.
• The unbalanced use of NGF and NDF gives unfavorable results, and in the architecture used in this study the role of the amount of NDF is more dominant than NGF. It is proven by the experiment NGF=64, NDF=128 which gives better results than NGF=128, NDF=64.
• The results of the experiments for S_SIZE, NGF, and NDF values obtained the best results for S_SIZE values of 1500, NGF=128, and NDF=128. Experimental results to determine the best LR, at the 100th epoch, the results showed that LR 2e-4 produces better images than 2e-2, 2e-3, and 1e-4. If the LR is too high, the model cannot converge, while if the LR is too low, the model will take a long time to converge.
• Experimental results for the amount of latent space (NZ) show that using NZ=20 obtains better results than 10, 100, and 200. This indicates that the number of NZ affects the generator yield and it is necessary to know the most optimal NZ value.
• The best model produced in this experiment with an epoch of 100 and an image size of 64x64x3 is when using S_SIZE=1500, NGF=128, NDF=128, LR=2e-4, and NZ=20.

References

Daberger J., 2023. Zoo animals. Object detection dataset of animals using Pascal VOC labeling format - XML files. https://www.kaggle.com/datasets/jirka- daberger/zoo-animals

Jordan, M.I. and Mitchell, T.M., 2015. Machine learning: Trends, perspectives, and prospects. Science, 349(6245), pp.255-260.

Kostrikov, I., Yarats, D. and Fergus, R., 2020. Image augmentation is all you need: Regularizing deep reinforcement learning from pixels. arXiv preprint arXiv:2004.13649.

Gregor, K., Danihelka, I., Graves, A., Rezende, D. and Wierstra, D., 2015, June. Draw: A recurrent neural network for image generation. In International conference on machine learning (pp. 1462-1471). PMLR.

Goodfellow, I., Pouget-Abadie, J., Mirza, M., Xu, B., Warde-Farley, D., Ozair, S., Courville, A. and Bengio, Y., 2020. Generative adversarial networks. Communications of the ACM, 63(11), pp.139-144.

Mirza, M. and Osindero, S., 2014. Conditional generative adversarial nets. arXiv preprint arXiv:1411.1784.

Radford, A., Metz, L. and Chintala, S., 2015. Unsupervised representation learning with deep convolutional generative adversarial networks. arXiv preprint arXiv:1511.06434.

Mustapha, I.B., Hasan, S., Nabus, H. and Shamsuddin, S.M., 2022. Conditional deep convolutional generative adversarial networks for isolated handwritten arabic character generation. Arabian Journal for Science and Engineering, 47(2), pp.1309-1320.

Zhu, F., He, M. and Zheng, Z., 2020. Data augmentation using improved cDCGAN for plant vigor rating. Computers and Electronics in Agriculture, 175, p.105603.