Sure, you won’t be training high-resolution style GANs on it any time soon, but that’s mostly due to 8 GB of memory limitation. RTX3060Ti from NVIDIA is a mid-tier GPU that does decently for beginner to intermediate deep learning tasks. The only way around it is renting a GPU in the cloud, but that’s not the option we explored today. Still, if you need decent deep learning performance, then going for a custom desktop configuration is mandatory. Nothing comes close if we compare the compute power per wat. Parting WordsĪpple’s M1 chip is remarkable - no arguing there. We knew right from the start that M1 doesn’t stand a chance. Still, these results are more than decent for an ultralight laptop that wasn’t designed for data science in the first place. For the augmented dataset, the difference drops to 3X faster in favor of the dedicated GPU. RTX3060Ti is 10X faster per epoch when training transfer learning models on a non-augmented image dataset. The results look more realistic this time. Image 6 - Transfer learning model results in seconds (M1: 395.2 M1 augmented: 442.4 RTX3060Ti: 39.4 RTX3060Ti augmented: 143) (image by author) Here are the results for the transfer learning models: Both are roughly the same on the augmented dataset.īut who writes CNN models from scratch these days? Transfer learning is always recommended if you have limited data and your images aren’t highly specialized. On the non-augmented dataset, RTX3060Ti is 4.7X faster than the M1 MacBook. Don’t get me wrong, I expected RTX3060Ti to be faster overall, but I can’t reason why it’s running so slow on the augmented dataset. One thing is certain - these results are unexpected. Image 5 - Custom model results in seconds (M1: 106.2 M1 augmented: 133.4 RTX3060Ti: 22.6 RTX3060Ti augmented: 134.6) (image by author) Keep in mind that two models were trained, one with and one without data augmentation: We’ll now compare the average training time per epoch for both M1 and custom PC on the custom model architecture. RTX3060Ti - Data Science Benchmark Results Print( f 'Duration: ')įinally, let’s see the results of the benchmarks. Conv2D(filters = 32, kernel_size =( 3, 3), activation = 'relu'), MaxPool2D(pool_size =( 2, 2), padding = 'same'), # USED ON A TEST WITH DATA AUGMENTATION train_datagen = tf. Data loading # USED ON A TEST WITHOUT DATA AUGMENTATION train_datagen = tf. TensorFlow for Image Classification - Top 3 Prerequisites for Deep Learning Projects Refer to the following article for detailed instructions on how to organize and preprocess it: Long story short, you can use it for free. Cats dataset from Kaggle, which is licensed under the Creative Commons License.
0 Comments
Leave a Reply. |
AuthorWrite something about yourself. No need to be fancy, just an overview. ArchivesCategories |