Integration with LXMERT

[johntiger1]

If I want to use this repo to extract RCNN image features to train LXMERT, how can I do that? Do I just dump the features from

# Show the boxes, labels, and features
pred = instances.to('cpu')
v = Visualizer(im[:, :, :], MetadataCatalog.get("vg"), scale=1.2)
v = v.draw_instance_predictions(pred)
showarray(v.get_image()[:, :, ::-1])
print('instances:\n', instances)
print()
print('boxes:\n', instances.pred_boxes)
print()
print('Shape of features:\n', features.shape)

(from https://github.com/airsplay/py-bottom-up-attention/blob/master/demo/demo_feature_extraction_attr.ipynb)

into a .tsv file?

Btw, what is the difference between with and without attributes? Thanks!

[airsplay]

Yea; It would work well (at least in my test).

But the NMS approach would be the best to use this one:

py-bottom-up-attention/demo/detectron2_mscoco_proposal_maxnms.py

Lines 54 to 65 in 834fa8b

	# Select max scores
	max_scores, max_classes = scores.max(1) # R x C --> R
	num_objs = boxes.size(0)
	boxes = boxes.view(-1, 4)
	idxs = torch.arange(num_objs).cuda() * num_bbox_reg_classes + max_classes
	max_boxes = boxes[idxs] # Select max boxes according to the max scores.
	# Apply NMS
	keep = nms(max_boxes, max_scores, nms_thresh)
	if topk_per_image >= 0:
	keep = keep[:topk_per_image]
	boxes, scores = max_boxes[keep], max_scores[keep]

[johntiger1]

Thank you, I will try the Non Maximal Suppression. But, just curious, does this mean that other SOTA recurrent vision models could be used too in the future? rCNN is now several years old, I was wondering if you experimented with more modern vision models, and perhaps can get better performance

[airsplay]

Hmmm... This code does not provide a training, just the weight converted. from the original CAFFE weight.

You could try this and switch the backbone:
https://github.com/MILVLG/bottom-up-attention.pytorch

[yezhengli-Mr9]

If I want to use this repo to extract RCNN image features to train LXMERT, how can I do that? Do I just dump the features from

# Show the boxes, labels, and features
pred = instances.to('cpu')
v = Visualizer(im[:, :, :], MetadataCatalog.get("vg"), scale=1.2)
v = v.draw_instance_predictions(pred)
showarray(v.get_image()[:, :, ::-1])
print('instances:\n', instances)
print()
print('boxes:\n', instances.pred_boxes)
print()
print('Shape of features:\n', features.shape)

(from https://github.com/airsplay/py-bottom-up-attention/blob/master/demo/demo_feature_extraction_attr.ipynb)

into a .tsv file?

Btw, what is the difference between with and without attributes? Thanks!

Hi @johntiger1, before I finish coding my project:

How long does it take to extract NLPR2's 107,292 images when LXMERT takes around 5 to 6 hours for the training split and 1 to 2 hours for the valid and test splits?

Would you mind taking a time estimate? Thanks.

[yezhengli-Mr9]

If I want to use this repo to extract RCNN image features to train LXMERT, how can I do that? Do I just dump the features from

# Show the boxes, labels, and features
pred = instances.to('cpu')
v = Visualizer(im[:, :, :], MetadataCatalog.get("vg"), scale=1.2)
v = v.draw_instance_predictions(pred)
showarray(v.get_image()[:, :, ::-1])
print('instances:\n', instances)
print()
print('boxes:\n', instances.pred_boxes)
print()
print('Shape of features:\n', features.shape)

(from https://github.com/airsplay/py-bottom-up-attention/blob/master/demo/demo_feature_extraction_attr.ipynb)
into a .tsv file?
Btw, what is the difference between with and without attributes? Thanks!

Hi @johntiger1, before I finish coding my project:

How long does it take to extract NLPR2's 107,292 images when LXMERT takes around 5 to 6 hours for the training split and 1 to 2 hours for the valid and test splits?

Would you mind taking a time estimate? Thanks.

Hi @johntiger1 , I get my solution for this question of time estimate and summarize it here. Thanks anyway.