enas
enas copied to clipboard
melodyguan
How to transfer to the larger data, like 300*300?
I tried to use ENAS to design architecture on my dataset which is much larger than CIFAR10. If I set the image size to 224*224, then raise ValueError("GraphDef cannot be larger than 2GB."). Is it possible to use placeholder to solve this problem and does anyone succeed? Thanks.
For classification architecture search the repository loads all CIFAR10 data to the memory. This is OK for small datasets, but for larger ones its a problem - thus the error you get. You need to change the data input pipeline to tf.data to be able to run larger datasets. See: https://www.tensorflow.org/guide/datasets
@Nadav1 Thank you so much. I've solved the problem and put the code in marsggbo/enas. But now I have another question, i.e. how to use multiple GPUs. Because the code of ENAS is so complex to me, there are many components, such as controller, child models. So I don't know how to sync the gradients of each GPUs, and I'm not sure whether I need to train controller in parallel. Looking forward to your suggestion. Thanks!
@marsggbo I downloaded your code provided above and tried to run it again for cifar10 dataset before running code for my own dataset for which I was getting the ValueError("GraphDef cannot be larger than 2GB.") with original ENAS. For cifar 10 dataset I am getting loss value as nan.
@ManasiPat Did you modify the code in *sh file. You may need to assign--output_classes=10 instead of 5.
@marsggbo Thanks. I changed that, now I am not getting nan however I am getting my training accuracy to be 0 , 2, or 1
ok got it. Its number of correct samples out of the batch. Let me run it for a while to see if I get the results as in the original paper. Thanks.
@marsggbo I am working on age prediction using facial images. Treating it as a classification problem, where I have 101 classes. So, to feed my dataset, I have changed the data_utils file and changed the output_classes=101 and data_path to point at my dataset in the script file. It should work with my data right? Or do I need to make any other changes? I have 1,37,480 training images and 17,186 for validation and testing.
@marsggbo I ran the modified code for cifar-10, however it doesn't reproduce the results which we get without the modifications. How you tried runing your code for cifar-10 to see if you can reproduce the results?
@ManasiPat Hi, I'm also working on age prediction using facial images. Do you success with ENAS?
