ruotianluo / Context-aware-ZSR

hi ，When I download the pretrained imagenet-model according to the README.md, the connection fails all the time. Is the URL in it changed?

interseting. Maybe the code breaks.

you can still download by enter in the url: https://docs.google.com/uc?export=download&id=1wHSvusQ1CiEMc5Nx5R8adqoHQjIDWXl1

you can replace id with other ids in that python file.

Thank you, there is another question, in which folder do I need to put the model downloaded according to the link? Download pretrained model from link in README.md

a new directory called pretrained_model under the repo dir.

Sorry, I mean you reproduce the model in your paper, not pretrained imagenet-model，Is for example this we_relt_geo_sc.pth file .

a dir called pretrained.
Checkout scripts/reproduce for clue.

thank you very much！

Hello, please disturb you again，When I run the test file, it shows that there is no the tagging_eval.pkl file,Does this file need to be downloaded? Where is it if downloaded? Thank you！

no. it's the output of tools/test_net.py. i may put the path wrong. do you mind search a little bit where the tagging_eval.pkl is?

sorry，i didn't find where the file was.

Can you show me the full error log. And if possible can you run line by line in the script; and also what's your version of pytorch.

Do you have to use cuda to run? pytorch version is 1.0

Yes, cuda is needed.

i know ，thank you very much！

hi,When I run the test file （bash scripts/test/we_infer.sh).I got the above error.These two parameters load_ckpt and load_detectron in the run command are None,Can you look at it for me？

You should run scripts/reproduce

Hi,Do you need both python2 and python3? I'm installing python 3. Do commands like（In python 2 python lib/datasets/vg/convert_from_bansal.py） need to be executed?

I used a conda environment to have both py2 and py3

I completed （ Convert from bansal train test split ）using python2，but later compilation fails,My machine is ubantu18 cuda 9.0 gcc have 4.8 5.4 6 7,gcc version i have tried，Can you show it to me?

sorry ,gcc have 4.8 5.5 6 7

There is no change in this part compared to the original Detectron.pytorch. I suggest you go there and search for the solution.

Can you tell me which version of cuda and GCC you are using?

gcc4.9.4 and cuda10

hi,I want to ask where this weight is downloaded？

Hi, I didn't provide that weight; I only provided the weight after conversion.

When I run this test file, I can't find the file. Can I ask if this file can be downloaded?

Have you trained the model by yourself?

no ,I am not going to train, I want to use the trained ones directly

https://github.com/ruotianluo/Context-aware-ZSR#to-reproduce-numbers-in-the-paper

You need to run the scripts in scripts/reproduce.

Does everything need to be running? I have run the two inside

It depends on what you want?

I want to run this

that's gcn_infer and sync_infer

I have run gcn_infer in scripts/reproduce，But when I run this bash scripts/test/detection_gcn.sh ，Still without this weight ckpt/model_final_gcn_wn.pth

Sorry, my bad, now I understand.
replace the load_ckpt in detection_gcn with the load_ckpt in gcn_infer
same for sync.

Oh thanks，This is the result of my running bash scripts/test/we_infer.sh，Can you tell me what these numbers represent?

It corresponds to Table 1 GCN+Context row.

thanks,I want to put a picture now, and then use this zero sample to identify what I have seen and what I have n’t seen. Which file do I need to run?

Not sure what you mean.

How can the 7794 pictures in the test set be modified into my own pictures?

I haven't done this before. My guess is you need to create a coco annotation type json file for your own images.

If you want to test a single image. You may need to dig around. It shouldn't be too complicated, but it may take a few hours.

If I only want to test one picture, what do I need to modify?

Try if you can extend from this function:

Context-aware-ZSR/lib/core/test_engine.py

Line 222 in 12c1d08

def test_net(

But you still need to be able to run edgebox.

hi，Can this file be replaced? Is it necessary to install matlab to run this?

Unfortunately, it's necessary

oh,thanks,Is it okay to follow this instruction?

Yes.

Now need to install matlab, and then install this Matlab Toolbox?

I think so.

hi,This score is not very clear, why some are greater than 1?

I think it’s because the scores after the Crf is not normalized Ruotian Luo

…

On Apr 17, 2020, at 9:23 AM, buxpeng ***@***.***> wrote: hi,This score is not very clear, why some are greater than 1? — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

Ok，I tested a picture. The category_id in the generated bbox_vgbansalone_test_results.json is the last category detected with zero ZSR, right?How does this category_id correspond to the real label?

It should be in some json file in the datasets folder.

Ok, can you tell me where you can normalize the value after Crf?

I am not sure. In fact from https://github.com/ruotianluo/Context-aware-ZSR/blob/master/lib/modeling/rel_heads.py#L135, it seems the scores should be somewhat normalized.

interseting. Maybe the code breaks.

you can still download by enter in the url: https://docs.google.com/uc?export=download&id=1wHSvusQ1CiEMc5Nx5R8adqoHQjIDWXl1

you can replace id with other ids in that python file.

Hello, could you please refresh this download link? Best wishes.

interseting. Maybe the code breaks.
you can still download by enter in the url: https://docs.google.com/uc?export=download&id=1wHSvusQ1CiEMc5Nx5R8adqoHQjIDWXl1
you can replace id with other ids in that python file.

Hello, could you please refresh this download link? Best wishes.

https://drive.google.com/drive/folders/0B7fNdx_jAqhtdXR1NUN4NkZwS00?resourcekey=0-4-sXYhJKBsRD8sohsPfDvA&usp=sharing this folder should have all the weights.

Unable to download pretrained imagenet-model