Image File Extension

[SURABHI-GUPTA]

@rosinality I have a doubt related to the image format used for training the model. As FFHQ is a collection of high dimensional PNG images, does training with LFW dataset, images are JPG of size 160X160 affects compression ratio/or performance of your model?

[rosinality]

It will be depend on jpeg compression rates, but I don't think it will affect the quality much.

[SURABHI-GUPTA]

@rosinality My original image and the reconstructed image using the model have no difference in their size. It doesn't seem that image is getting compressed. Is there anything I am misunderstanding?

[rosinality]

Latent code is compressed representation of images.

[SURABHI-GUPTA]

@rosinality okay. I want to check the size of the latent map for top and bottom. How to save it ?

[rosinality]

You don't need save it. Size of latent codes is 32x32 & 64x64 with 512 discrete codes, so size of latent codes will be 9 (32 32 + 64 * 64) bits.

[SURABHI-GUPTA]

@rosinality If I want to check the final compressed size to evaluate my model with another compression algo, what will be the size? Will it be 9 (32 32 + 64 * 64) bits ?

[rosinality]

Of course it will be reduced if you uses additional compression.

[SURABHI-GUPTA]

@rosinality thanks for clearing my doubts.

[SURABHI-GUPTA]

@rosinality So, what I understand from the above discussion is,

For any given image of size 256X256, with the size of latent codes 32x32 & 64x64 with 512 discrete codes, I will always end up with a compressed size of 9 (32 32 + 64 * 64) = 46080 bits and bbp = 46080/(256X256) = 0.7. Am I correct ?

[rosinality]

It is upper bound of bits. If model uses less than 512 latent codes, like 250, or the number of latent codes need for the specific images is less than 512, than it could be reduced. But it is depend on how you defines the size of the latent codebooks. (Upper bound, trained model specific, or image specific.)

[SURABHI-GUPTA]

@rosinality so what we define as the number of embedding vectors, K=512, it is also the upper bound? How can we find out how many latent codes are used? It is important to know that for calculating the number of bits .

[rosinality]

Yes, 512 will be upper bound. You can do it like this for the number of latent codes actually used.

_, _, _, id_t, id_b = vqvae.encode(img)
# this will corresponds to the number of distinct latent codes used in top/bottom latent map
torch.unique(id_t).shape, torch.unique(id_b).shape 

[SURABHI-GUPTA]

@rosinality although the number of unique values is less than 512, the number of bits required is still 9 bits and hence the image will be of 46080 bits only, right ? Okay, let's say I want to make my top latent map of size 16X16 rather than 32X32, how to do that ? Also, I tried to look at the loss vs epoch curve, it doesn't seem to reduce smoothly. At some points, there is sudden increase in loss and then decrease.

[rosinality]

If the number of uniques codes is less than 256, then 8 bits will be enough. And if you want to reduce latent map sizes, then you can just add downsample & upsample layers. Also, VQ-VAE training seems like that inherently not very smooth. It maybe because it is constructed on discretization and k-means clustering.

[SURABHI-GUPTA]

@rosinality I tried with K=128, then also the number of unique codes is less than 128. So, if I reduce the K value, the number of unique codes also decreases in the same ratio. If I talk about testing the model, then as you mentioned that if the number of uniques codes is less than 256, then 8 bits will be enough, so it means these 8 bits will be counted for calculating the size, although during training there could be more than 256 unique codes with K=512.

[SURABHI-GUPTA]

@rosinality How can we calculate the bpp of the original/uncompressed image? When we talk about this formula, compression ratio=S_uncompressed/S_compressed, what do we exactly mean by S_uncompressed here? I am confused with file size that we get using "os.path.getsize(full_path)" or its wXhXcX8 ?

[rosinality]

Isn't bpp corresponds to bits per pixel? Then it would be (total bits for the image) / (# pixel of the image). If we think (R, G, B) pair is 1 pixel then denominator would be 256 256, if we consider it independently, then it would be 3 256 * 256.

If you want to calculate length of uncompressed bit sequences, then it would be 3 256 256 * 8. Actual size of images will not be accurate as it is generally compressed, and it would contain headers.

[SURABHI-GUPTA]

@rosinality sorry for the mistake, I am talking about CR, compression ratio=S_uncompressed/S_compressed. How to find out S_uncompressed for calculating compression ratio? For an RGB image of 160X160 pixels, I calculated 'os.path.getsize(full_image_path)', and I get 3426 B as size. I didn't understand how to get this value using your above calculation.

[rosinality]

I think you need to check how to calculate CR as I don't know that problem much. Anyway, uncompressed RGB image need to use 3 8 H * W bits. As I noted generally image files are already compressed so actual file sizes will not corresponds to this. (It will be close if the format of image is BMP, though.)

[SURABHI-GUPTA]

@rosinality btw I am talking about this image.

[rosinality]

Yes, and it is jpeg compressed. You would find that size of it will be close to 160 x 160 x 3 bytes if you convert it to bmp.

[SURABHI-GUPTA]

@rosinality okay... I got some clarity now.. thank you for the quick response.

[SURABHI-GUPTA]

@rosinality I have some query regarding your code. qt, qb, di, _id_t, _id_b = vqvae.encode(img_t)

When I try reconstructions using quantized embedding (qb) and quantized latent map using codebook (_id_b), it should be same, right? However I get different results.

decoded_sample_bot = vqvae.decode(qt*0, qb).detach()[0].squeeze(0) decoded_sample_bot = vqvae.decode_code(id1_top, _id_b).detach()[0].squeeze(0), here id1_top is array of 1s

[SURABHI-GUPTA]

@rosinality please help. For bottom code, I am getting different reconstructions. With decoded_sample_bot = vqvae.decode_code(id1_top, _id_b).detach()[0].squeeze(0), I get

but with decoded_sample_bot = vqvae.decode(qt*0, qb).detach()[0].squeeze(0), I get

[rosinality]

qt * 0 and embeddings from id 1 (id1_top) will not be same.

[SURABHI-GUPTA]

@rosinality considering id1_top=torch.zeros([1,10,10],dtype=torch.long), it also implies that for bottom we are making top as 0. how can we get the correct result?

[rosinality]

No, it is not. As id (code) is mapped with embeddings, if you uses code filled with zeros, then it will be mapped to latent maps that filled with 0th embeddings. (which is nonzero.) Definitely it is different from latent maps that filled with zeros.

[SURABHI-GUPTA]

@rosinality yes, that makes sense, but for top latent embedding, I get correct reconstruction.

[rosinality]

I don't know it is fair to say it is correct reconstructions. Anyway zero latent code != zero latent embedding.