234 points by tensorflow_user 1 year ago flag hide 35 comments
user20 1 year ago next
What kind of hardware did you use to train the model?
original_poster 1 year ago next
@user20 I used a cloud service to train the model. It was a machine with 4 GPUs and 16GB of memory. It took about 3 hours to train the final model.
user1 1 year ago prev next
Interesting project! I've been playing around with similar projects recently.
mod2 1 year ago next
@user1 that's great to hear! what kind of results have you been getting?
user3 1 year ago prev next
Nice work, could you open source the code somewhere?
original_poster 1 year ago next
@user3 I plan to, I just need to clean it up a bit first.
user5 1 year ago prev next
This is very cool! I am planning a project that uses the same technology for a different use case
user6 1 year ago next
@user5 I'd love to hear about it!
user7 1 year ago prev next
could you explain how the convolutional neural networks are working under the hood here?
user8 1 year ago prev next
how long did it take to train?
user9 1 year ago prev next
Did you consider using a different type of neural network for this task? why or why not?
original_poster 1 year ago next
@user9 I did experiment with other types of networks, but the convolutional neural network seemed to work best. I think it is because I am using image data, so the spatial characteristics of the convolutional neural network are useful for identifying patterns within the images
user10 1 year ago prev next
Have you considered using this for OCR?
original_poster 1 year ago next
@user10 Not directly, but I can see how it could be used for that purpose. I will definitely look into it!
user11 1 year ago prev next
This is incredible, I love the way you've explained the process in the post!
original_poster 1 year ago next
@user11 Thank you! I wanted to make sure that it was understandable for people who might not be familiar with the concepts.
user12 1 year ago prev next
Great job, really like this kind of post. Would love to see more in the future!
original_poster 1 year ago next
@user12 Thanks, that's the kind of response I was hoping for! I plan to make more posts like this as I continue to learn.
user13 1 year ago prev next
Have you looked into this type of work with GAN networks?
original_poster 1 year ago next
@user13 I haven't yet, but I will definitely check it out. It is a really interesting use case for GANs.
user14 1 year ago prev next
This is fascinating, I really like how you explained the architecture. What are your next steps?
original_poster 1 year ago next
@user14 My next steps are to try to improve the accuracy of the model and see how well it generalizes to other datasets. I also would like to try it out on more unstructured data to see how well the process translates.
user15 1 year ago prev next
Have you considered using a different optimization algorithm?
original_poster 1 year ago next
@user15 I have tried several optimization algorithms, but I had the best results with stochastic gradient descent. I will continue to experiment with different algorithms to see if I can improve the performance further.
user16 1 year ago prev next
What tools are you using to setup the nas?
original_poster 1 year ago next
@user16 I am using keras to set up the neural architecture search. It has a lot of functionality that makes it easy to create and test different architectures.
user17 1 year ago prev next
what preprocessing did you do to the data before training?
original_poster 1 year ago next
@user17 I did a few different steps to preprocess the data. I resized the images to a standard size and then normalized the pixel values. I also used data augmentation to increase the size of the training dataset and help the model generalize better.
user18 1 year ago prev next
how did you handle class imbalance in the dataset?
original_poster 1 year ago next
@user18 I handled class imbalance by using a technique called weighted cross entropy loss. This assigns different weights to the different classes during training to make sure that the model is not biased towards the majority class. I set the weights based on the class distribution in the training set.
user19 1 year ago prev next
very cool work, I am trying to do something similar for my company. will definitely use this as a reference
user21 1 year ago prev next
what criteria did you use to evaluate the model?
original_poster 1 year ago next
@user21 I used several criteria to evaluate the model. I looked at the accuracy and the loss of the validation set during training to make sure that the model was not overfitting. I also calculated the precision, recall, and f1 score for each class after training was finished. It is important to look at all of these metrics to get a full picture of the performance of the model.
user22 1 year ago prev next
Have you experimented with the dropout rate during training? I've found that it can greatly affect the performance of the model.
original_poster 1 year ago next
@user22 Yes, I experimented with the dropout rate and found that a rate of 0.5 worked best for my model. However, the optimal dropout rate can vary depending on the specific architecture and dataset. It is definitely something worth playing around with to see if it improves the performance.