goals.txt

Dual Contradistinctive Generative Autoencoder
https://arxiv.org/abs/2011.10063

Qualitatively, CIFAR-10 Samples will be reproduced like in Figure 3 (b) of the paper. 
Quantitatively, from Table 2, FID score of DC-VAE Sampling (ours) line on CIFAR-10 dataset will be reproduced.
Since we don't have much computational resources, we chose a paper with published results at lowest resolution as possible in the first place. Since there is an ambiguity on training time
If we cannot match the required(unknown) training iterations in a reasonable time, we may get 10-15% worse results

—— version 1 submission —— 

- We did not have changed any of our goals.
- We are able to reach as low as 56 FID score in Sampling at CIFAR-10 dataset while in the paper 18 FID is reported.
- if you could not reproduce the results that you have targeted: what is your future work plan (discuss what is missing in your implementation / what might be buggy / etc.)?
The answer to this qustion is at the last paragraph of Challenges and Discussion part in the main.ipynb

—— version 2 submission —— 

- We did not have changed any of our goals.
- We are able to reach as low as 24 FID score in Sampling at CIFAR-10 dataset while in the paper 18 FID is reported.
- We could not exactly reach our goals, however the results we obtain are very close to the reported FID score. As we explained in the challanges part in version 1
"looking at the mentioned papers and the number of negative samples (8096) used in training, the original paper adapts the contrastive loss proposed in 
Momentum Contrast for Unsupervised Visual Representation Learning." We read the MoCo paper and understand how they calculate the contrastive loss with a 
very high number of negative samples such as 8096, however, we decided not to implement it since the implementation would be very challenging and 
stays out of the scope of our course. To increase the performance, the number of negative samples must be increased. To achieve that, we increased our batch size, 
since number of negative samples in our implementation equals to batch_size - 1. This change improved the FID score from 29 to 24.