90 points by musicae 7 months ago flag hide 6 comments
johnsmith 7 months ago next
Great work! I've been working on something similar, and I'm impressed with your accuracy. Do you mind sharing your dataset? It could help me train my model more effectively.
moredata 7 months ago next
@johnsmith Sure, I have a link to my dataset in my profile description for anybody who is interested in improving their own models. Please feel free to check it out!
janedoe 7 months ago prev next
This is fantastic! I've been interested in music recognition and am curious, how long did it take you to train your TensorFlow model? Was it GPU-accelerated?
gputech 7 months ago next
@janedoe The training took around a day on a Nvidia V100 with TensorFlow 2.4. I found it crucial to use GPU-acceleration for decent training times.
artsyengineer 7 months ago prev next
Incredible project! Have you experimented with different neural network architectures or CNNs for this task? How does the model perform with non-standard paper or penmanship?
cnnwiz 7 months ago next
@artsyengineer Yes, I tried several configurations, including CNNs. The final convolutional base consists of 5 convolutional layers, followed by batch normalization and ReLU activation. It performed better when dealing with various paper textures, but there's still room for improvement with penmanship irregularities.