sounakdey / doodle2search Goto Github PK
View Code? Open in Web Editor NEWDoodle to Search: Practical Zero Shot Sketch Based Image Retrieval
Home Page: https://sounakdey.github.io/doodle2search.github.io/
Doodle to Search: Practical Zero Shot Sketch Based Image Retrieval
Home Page: https://sounakdey.github.io/doodle2search.github.io/
Respected Authors,
Your paper is great. I have two question to ask you.
The first one is that after training on Sketchy-Extend dataset, I can't reproduce your result. Your mAP result is 0.3691 in the paper, but I only got 0.0589. Can you publish your pretrained model?
The second one is I can't produce sketchy_semantic_label by word2Vec, it told me "word 'XXX' is not in vocabulary"。Can you help me deal with this problem?
Thank you!
How to create word vector for the quickdraw dataset?
Is there some tricks to normalize them? like in #10
Hello,
When I tried to train the model, it told me that I lacked the ’class_labels.txt‘ file and couldn't find it. I think youmay forgot to upload it. I wish you can upload it or tell me how to generate it. I wish you can reply to me.
Thank you
Hi, how much memory did you use when you tested the quickdraw dataset? I have 60GB memory but it is still not enough for testing quickdraw.
Even if the network is unblocked, the data set cannot be obtained;I tried to run the .sh file, but it didn't work once; I directly used URL stitching and searched on Google, and an error was prompted when the data was downloaded halfway.So, how can I get the dataset downloaded by the .sh file?
For some reason, I have difficult in download dataset sketchy and TU-Berlin with download_datasets. But, I have had the two datasets.
So,could you only provide me the txt file such as /dataset/Sketchy\class_labels.txt and all others in Sketchy and TU-Berlin.
My e-mail is [email protected]
Thanks.
Thanks for your great work,I am wondering why don't you use the ID loss?Looking forward to your response,Thank You!
Hi, I am very interested in the QuickDraw-Extended dataset. And I want to try it in a RNN framework. Do you have the sketch data in vector format rather than raster?
Can you provide a source folder like this? It's a little difficult for me to implement those
Hi, I think you are MAP calculation is wrong, especially MAP@200. You have ignored the total number of related documents in the MAP@200 calculation. That is why your MAP@200 value is big. I hope I am wrong but it seems you are wrong. You can check the MAP calculation of this paper that you have not cited: https://github.com/qliu24/SAKE
As I got through the code, and tried to run the experiments. One of the key problems is to produce the class semantic vector using the word2vec or other NLP method. But the problem is that there are some class words that do not exist in the google-news-300. Even there are some simimlar ones, but I can not find how to convert these similar ones into google-news-300 format. For example, I can not find the "axe" in google-news-300 in any form. Although the authors provided the genereted semantic labels (in word2vec), we still confuse on how to generate it.
@sounakdey , do you please give us a hint on how did you deal with these problem? Thank you very much.
Hi, thanks for providing the code. It is easy to understand. However, there is some errors. You have renamed and removed some files/folders. However, the code is based on the old format. Can you fix this?
Do you guys have any plan on releasing the pretrained models?
Excellent work by the way.
After reach the end of line in file src/train.py. The program still run, but the CPU and GPU usages are both 0%. The memory was not released.
Hi, I can not download the dataset from the link you provided, could you update the link? Thanks.
Thanks for your contribution. Can you please add the license for the repo.
In th code 'test.py' at line 108, the Presision@200 is computed as:
# Precision@200 means at the place 200th
precision_200 = np.mean(sort_str_sim[:, 200])
As the sort_str_sim
is actually the sorted version for a tensor with [n_sketches, n_images, n_clssses], it has the true label information on sketches and images. But it has no information on the precision and recall. The number should be computed using average_precision_score().
Does anyone know how to compute the Precision@200 value?
The class eiffel_tower has two different spellings in the sketch and image dataset, which are eiffel_tower and effiel_tower.
There remains an s.sh file in the images/cactus
Respected Authors,
I am unable to run the code for the sketchy dataset. It mentioned class labels file is missing so i created a labels file for the dataset. I am unable to comprehend the error in the code. It would be really nice if you can let me know what can I do to reproduce the results.
python3 src/train.py sketchy_extend --data_path /data/anurag/doodledata/sketchy
Parameters: Namespace(attn=False, batch_size=20, data_path='/data/anurag/doodledata/sketchy', dataset='sketchy_extend', decay=0.0005, early_stop=20, emb_size=256, epochs=1000, exp_idf=None, gamma=0.1, grl_lambda=0.5, learning_rate=0.0001, load=None, log=None, log_interval=20, momentum=0.9, ngpu=1, nopretrain=True, plot=False, prefetch=2, save=None, schedule=[], seed=42, w_domain=1, w_semantic=1, w_triplet=1)
Prepare data
Traceback (most recent call last):
File "src/train.py", line 287, in <module>
main()
File "src/train.py", line 99, in main
train_data, [valid_sk_data, valid_im_data], [test_sk_data, test_im_data], dict_class = load_data(args, transform)
File "/data/anurag/codes/baselines/doodle2search/src/data/generator_train.py", line 18, in load_data
return Sketchy_Extended(args, transform)
File "/data/anurag/codes/baselines/doodle2search/src/data/sketchy_extended.py", line 35, in Sketchy_Extended
class_emb = create_class_embeddings(list_class, args.dataset)
File "/data/anurag/codes/baselines/doodle2search/src/data/class_word2vec.py", line 9, in create_class_embeddings
model = Word2Vec(google_300_corpus)
File "/home/anurag/anaconda3/lib/python3.6/site-packages/gensim/models/word2vec.py", line 783, in __init__
fast_version=FAST_VERSION)
File "/home/anurag/anaconda3/lib/python3.6/site-packages/gensim/models/base_any2vec.py", line 759, in __init__
self.build_vocab(sentences=sentences, corpus_file=corpus_file, trim_rule=trim_rule)
File "/home/anurag/anaconda3/lib/python3.6/site-packages/gensim/models/base_any2vec.py", line 936, in build_vocab
sentences=sentences, corpus_file=corpus_file, progress_per=progress_per, trim_rule=trim_rule)
File "/home/anurag/anaconda3/lib/python3.6/site-packages/gensim/models/word2vec.py", line 1592, in scan_vocab
total_words, corpus_count = self._scan_vocab(sentences, progress_per, trim_rule)
File "/home/anurag/anaconda3/lib/python3.6/site-packages/gensim/models/word2vec.py", line 1561, in _scan_vocab
for sentence_no, sentence in enumerate(sentences):
File "/home/anurag/anaconda3/lib/python3.6/site-packages/gensim/models/keyedvectors.py", line 355, in __getitem__
return vstack([self.get_vector(entity) for entity in entities])
TypeError: 'int' object is not iterable
doodle2search/src/data/class_word2vec.py
Line 13 in 07680c4
When I run 'bash download_datasets.sh', it happened.
doodle2search/src/data/quickdraw_extended.py
Lines 89 to 92 in 07680c4
A declarative, efficient, and flexible JavaScript library for building user interfaces.
🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.
TypeScript is a superset of JavaScript that compiles to clean JavaScript output.
An Open Source Machine Learning Framework for Everyone
The Web framework for perfectionists with deadlines.
A PHP framework for web artisans
Bring data to life with SVG, Canvas and HTML. 📊📈🎉
JavaScript (JS) is a lightweight interpreted programming language with first-class functions.
Some thing interesting about web. New door for the world.
A server is a program made to process requests and deliver data to clients.
Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.
Some thing interesting about visualization, use data art
Some thing interesting about game, make everyone happy.
We are working to build community through open source technology. NB: members must have two-factor auth.
Open source projects and samples from Microsoft.
Google ❤️ Open Source for everyone.
Alibaba Open Source for everyone
Data-Driven Documents codes.
China tencent open source team.