kmnp / intentonomy Goto Github PK

Hi, may I know if we can get a way to get the hashtags used per image of the dataset?

Thank you for your work!
I want to see the contents of Intentonomy's data set, where is it available?
hope hear from you soon.

Is there a timeline to public the dataset?

1. Training cost. Using a single NVIDIA V100 GPU, how long can you finish the training process?

2. Do you plan to release the hashtags along with each image?

Hello there,

I have two minor questions about _get_CAM and get_CAM, in the Localizationloss class -- the part that has to be changed according to one's model definition. First, you set out_b = gcam(X, image_ids="")[1], but looking into GradCam I wonder if it shouldn't be just out_b = gcam(X) in general; this in fact returns a tensor of shape [bs, num_classes] as commented.

Later on, you set gcam_b = gcam.get(layer) (and norm_img as its normalization), a tensor wish shape [bs, 1, fmpH, fmpW]. As a few lines below, in get_CAM, you set CAM_masks[:, cls_idx, :, :] = self._get_CAM(model, X, cls_idx).squeeze(1), and since the masks' shape depend on the mask_length parameter (but not on fmpH or fmpW, of course), I wonder if also get_CAM won't depend on the model adopted. That is, wouldn't this part crack if fmpH or fmpW and mask_length don't match?

https://images.unsplash.com/photo-1496214469482-f7340639138f?ixid=eyJhcHBfaWQiOi0xMywib3BlbnNvdXJjZSI6dHJ1ZSwidHJhY2tpbmciOmZhbHNlfQ==&q=30

https://images.unsplash.com/photo-1550634176-3719964e95fa?ixid=eyJhcHBfaWQiOi0xMywib3BlbnNvdXJjZSI6dHJ1ZSwidHJhY2tpbmciOmZhbHNlfQ==&q=30

there are two images that can not be downloaded. So how can we deal with that?

Hello,

I want to use the Intentonomy dataset in my research, but the box urls to the JSON annotation files are no longer available. Are there any other ways I can obtain these annotations?

Thanks

Why in category_ids_softprob , some probability is more than 1 and some is less than 1.
Are they abandoned datas?

Could I get the code of the Intentonomy? Object/Context localization .