search

rajat503 / zero-shot-classification

Zero shot classification using shared text and image embeddings in TensorFlow
23 stars 9 forks source link

Unclear about data files #1

Open dbalaram opened 6 years ago

dbalaram commented 6 years ago

hey, Thanks for the code. I was just wondering whether you have pretrained models

rajat503 commented 6 years ago

Hi, I don't have any pre-trained model, but training it doesn't take much time on GPU. It should work pretty much straight away, given you have CIFAR dataset in the directory.

On Fri, Nov 10, 2017, 1:41 PM dbalaram notifications@github.com wrote:

hey, Thanks for the code. I was just wondering whether you have pretrained models

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AEj3eUEyfzXgsSp-TS-X0Zxq6krN-pwjks5s1AUUgaJpZM4QZNSg .

dbalaram commented 6 years ago

Hey,

Thanks for the quick reply. I'm actually using this code for my project baseline. It would be super helpful if you could tell me the order in which to run all the files.

Thanks a lot!

Regards, Dhanashree Balaram Department of Electrical and Computer Engineering Carnegie Mellon University

On Fri, Nov 10, 2017 at 3:23 AM Rajat Agarwal notifications@github.com wrote:

Hi, I don't have any pre-trained model, but training it doesn't take much time on GPU. It should work pretty much straight away, given you have CIFAR dataset in the directory.

On Fri, Nov 10, 2017, 1:41 PM dbalaram notifications@github.com wrote:

hey, Thanks for the code. I was just wondering whether you have pretrained models

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AEj3eUEyfzXgsSp-TS- X0Zxq6krN-pwjks5s1AUUgaJpZM4QZNSg .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/issues/1#issuecomment-343407249, or mute the thread https://github.com/notifications/unsubscribe-auth/AeoK9fzS0ZHy18x6UI8mkKiWmpMlupfBks5s1AgXgaJpZM4QZNSg .

rajat503 commented 6 years ago

read_data_cifar10.py starts the training and computes the accuracy, but I haven't implemented checkpointing to save the model. You'll need to have the GLoVe embeddings (available on the Stanford website) and the CIFAR-10 dataset in the directory.

On Fri, Nov 10, 2017, 1:57 PM dbalaram notifications@github.com wrote:

Hey,

Thanks for the quick reply. I'm actually using this code for my project baseline. It would be super helpful if you could tell me the order in which to run all the files.

Thanks a lot!

Regards, Dhanashree Balaram Department of Electrical and Computer Engineering Carnegie Mellon University

On Fri, Nov 10, 2017 at 3:23 AM Rajat Agarwal notifications@github.com wrote:

Hi, I don't have any pre-trained model, but training it doesn't take much time on GPU. It should work pretty much straight away, given you have CIFAR dataset in the directory.

On Fri, Nov 10, 2017, 1:41 PM dbalaram notifications@github.com wrote:

hey, Thanks for the code. I was just wondering whether you have pretrained models

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/issues/1, or mute the thread https://github.com/notifications/unsubscribe-auth/AEj3eUEyfzXgsSp-TS- X0Zxq6krN-pwjks5s1AUUgaJpZM4QZNSg .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub < https://github.com/rajat503/zero-shot-classification/issues/1#issuecomment-343407249 , or mute the thread < https://github.com/notifications/unsubscribe-auth/AeoK9fzS0ZHy18x6UI8mkKiWmpMlupfBks5s1AgXgaJpZM4QZNSg

.

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/issues/1#issuecomment-343408359, or mute the thread https://github.com/notifications/unsubscribe-auth/AEj3eVz0flmu-mFgjOu6VkjZGKaRlOLfks5s1AkLgaJpZM4QZNSg .

dbalaram commented 6 years ago

Hey,

So I ran the code and the loss seems to increase. (fluctuate mostly). Where do you think the error might be?

On Fri, Nov 10, 2017 at 3:40 AM, Rajat Agarwal notifications@github.com wrote:

read_data_cifar10.py starts the training and computes the accuracy, but I haven't implemented checkpointing to save the model. You'll need to have the GLoVe embeddings (available on the Stanford website) and the CIFAR-10 dataset in the directory.

On Fri, Nov 10, 2017, 1:57 PM dbalaram notifications@github.com wrote:

Hey,

Thanks for the quick reply. I'm actually using this code for my project baseline. It would be super helpful if you could tell me the order in which to run all the files.

Thanks a lot!

Regards, Dhanashree Balaram Department of Electrical and Computer Engineering Carnegie Mellon University

On Fri, Nov 10, 2017 at 3:23 AM Rajat Agarwal notifications@github.com wrote:

Hi, I don't have any pre-trained model, but training it doesn't take much time on GPU. It should work pretty much straight away, given you have CIFAR dataset in the directory.

On Fri, Nov 10, 2017, 1:41 PM dbalaram notifications@github.com wrote:

hey, Thanks for the code. I was just wondering whether you have pretrained models

— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/issues/1, or mute the thread https://github.com/notifications/unsubscribe- auth/AEj3eUEyfzXgsSp-TS- X0Zxq6krN-pwjks5s1AUUgaJpZM4QZNSg .

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub < https://github.com/rajat503/zero-shot-classification/ issues/1#issuecomment-343407249 , or mute the thread < https://github.com/notifications/unsubscribe-auth/ AeoK9fzS0ZHy18x6UI8mkKiWmpMlupfBks5s1AgXgaJpZM4QZNSg

.

— You are receiving this because you commented.

Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/ issues/1#issuecomment-343408359, or mute the thread https://github.com/notifications/unsubscribe-auth/AEj3eVz0flmu- mFgjOu6VkjZGKaRlOLfks5s1AkLgaJpZM4QZNSg

.

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/rajat503/zero-shot-classification/issues/1#issuecomment-343411680, or mute the thread https://github.com/notifications/unsubscribe-auth/AeoK9XEL12iU-cnNuU5C82-6Bohnhuz3ks5s1AwPgaJpZM4QZNSg .

rajat503 commented 6 years ago

A constant increase in loss seems unlikely and I don't remember this happening for me. L2 loss can have some variations initially but it eventually converges. Can you try the classification based model instead of the regression model? Just comment and uncomment the last line in read_data_cifar10.py to switch between the two models.

rajat503 commented 6 years ago

Although, if I remember correctly, the regression model worked better. Should have documented everything somewhere and written better quality code :(