# Tag Info

0

You've specified 10 filters in a 2d convolution, each of size $3\times 3$ so you have $3 \times 3 \times 10=90$ trainable parameters. You have 1d data, but you're using a 2d convolution. Perhaps this is a typo and you meant Conv1D?

2

What loss function are you using? When you want to perfectly overfit, you should use L2 and not L1 loss. The reason for this is that the derivative for the L2 loss is $\frac{d l^2} {d l} = 2l$ and the derivative for the L1 loss is $\frac{dl^1}{d l} = 1$. This means that the gradient update for the L2 loss gets smaller as the loss value decreases, therefore ...

2

To the best of my knowledge, there is no consistent answer to the first question. It's like asking "is it better to use (3,3) kernel size or (5,5) kernel size?". As far as I know, the reason behind choosing either an even or odd kernel comes from trial and error in practical implementations and can vary from one case to the other. For the second ...

2

Depending on what you are trying to do with your CNN, regularization may indeed make sense. Pruning your network by regularization to make it sparse has two main advantages: It simplifies the network, making training and computation faster and easier; It prevents overfitting, and allows to make sure your network will generalize well on new data. An ...

0

Because the general approach in neural networks, especially in deep learning, is to let the NN learn to create useful features by themselves, instead of us engineering them. Especially in computer vision problems, the networks are made deeper (e.g. more layers) and wider (e.g. more filters) to learn more sophisticated representations of the input images.

0

I am also relatively unexperienced. I think crop just means using a subset of the image for training. It is a common way of data augmentation when you need more training data but only have a limited number of images. For a 448X448 image, you can randomly get a lot of different 224X224 cropped sub-images. They can be any position within the original image. As ...

0

I think the most important thing is to run your training for desired number of epochs, log training loss, validation loss, look at the plots of them and ask yourself what you don't like about those plots. It's highly unlikely that they'll look perfect on the first go (but if they did, congrats, you have optimal set of parameters). Most probably you'll ...

0

In the paper you referenced, the suggested test time behavior is to compute sample mean and variance for each feature using a large number of training images rather than using a running average. This block of code running_mean = momentum * running_mean + (1 - momentum) * sample_mean running_var = momentum * running_var + (1 - momentum) * sample_var ...

1

You have overfitted the training set. Try again with more data, or with some form of regularization, possibly including added noise.

Top 50 recent answers are included