In the model definition, I used the kernel_regularizer=tf.contrib.layers.l2_regularizer(scale=0.00001) into the tf.layers.conv2d() to regularize the convolution kernels in each convolution layer.
My question is: to compute the total loss of the whole network for some batch inputs, do we need to manually add the regularization loss as follows:
reg_losses = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)
reg_constant = 0.01 # Choose an appropriate one.
loss = my_normal_loss + reg_constant * sum(reg_losses)
And if yes, how to determine the reg_constant above? What's the relationship between the scale and the reg_constant? Thanks.
You are right.
You technically do not need reg_constant. You can control each layer regularization by the scale param, which can be the same for all layers. In this case you can just set reg_constant=1.
The only advantage of using reg_constant I see over scale, multiplying the regularization loss by reg_constant, is perhaps readability of your code.
If you're using a standard architecture I suggest to start with setting reg_constant=1 and set scale to some small scalar, say 0.001. If you have the resources, a better approach is to apply grid search to find the value that empirically minimizes your validation loss, i.e in [0.0001, 0.1].
If you suspect of a layer which should be specifically regularized, you can follow the first case only set that specific layer scale for a different value. Apply grid search like before only this time over the two different scale values.
Related
I am working on a Semantic segmentation project where I have to work on multiclass data which is highly imbalanced. I searched for optimizing it during training using the model.fit parameter and in that to use class_weights or sample_weights.
I can implement a following using a class_weight dictionary as
{ 0:1, 1:10,2:15 }
I also saw a method of updating weights in loss function
But at what point do these weights get updated?
If class_weights are used where will it get penalized? I already have a kernel_regularizer for each layer so if my classes have to be penalized based on my class weights then will it penalize the output of each layer y=Wx+b or only at the final layer?
Same if I use a weighted loss function will it get penalized only on the final layer before loss calculation or on each layer and then the final loss is calculated?
Any explanation on this would be very useful.
The class_weights you mentioned in your dictionary are there to account for your imbalanced data. They will never change, they are only there to increase the penalty for misclassified instances of minority classes (that way your network pays more attention to them and the gradients returned treat one 'Class2' instance as if it was 15 times more important than one 'Class0' instance).
The kernel_regularizer you mention resides at your loss function and penalizes large weight norms for weight matrices throughout the network (if you use kernel_regularizer = tf.keras.regularizers.l1(0.01) in a Dense layer, it only affects that layer). So that is a different weight that has nothing to do with classes, only with weights inside your network. Your eventual loss will be something like loss = Cross_entropy + a * norm(Weight_matrix) and that way the network will have as an additional task assigned to it to minimize the classification loss (cross entropy) while the weight norms remain low.
I am training an autoencoder DNN for a regression question. Need suggestions on how to improve the training process.
The total number of training sample is about ~100,000. I use Keras to fit the model, setting validation_split = 0.1. After training, I drew loss function change and got the following picture. As can be seen here, validation loss is unstable and mean values are very close to training loss.
My question is: based on this, what is the next step I should try to improve the training process?
[Edit on 1/26/2019]
The details of network architecture are as follows:
It has 1 latent layer of 50 nodes. The input and output layer have 1000 nodes,respectively. The activation of hidden layer is ReLU. Loss function is MSE. For optimizer, I use Adadelta with default parameter settings. I also tried to set lr=0.5, but got very similar results. Different features of the data have scaled between -10 and 10, with mean of 0.
By observing the graph provided, the network could not approximate the function which establishes a relation between the input and output.
If your features are too diverse. That one of them is large and others have a very small value, then you should normalize the feature vector. You can read more here.
For a better training and testing result, you can follow these tips,
Use a small network. A network with one hidden layer is enough.
Perform activations in the input as well as hidden layers. The output layer must have a linear function. Use ReLU activation function.
Prefer small learning rate like 0.001. Use RMSProp optimizer. It works fine on most regression problems.
If you are not using mean squared error function, use it.
Try slow and steady learning and not fast learning.
I am using 1x1 convolution in the deep network to reduce a feature x: Bx2CxHxW to BxCxHxW. I have three options:
x -> Conv (1x1) -> Batchnorm-->ReLU. Code will be output = ReLU(BN(Conv(x))). Reference resnet
x -> BN -> ReLU-> Conv. So the code will be output = Conv(ReLU(BN(x))) . Reference densenet
x-> Conv. The code is output = Conv(x)
Which one is most using for feature reduction? Why?
Since you are going to train your net end-to-end, whatever configuration you are using - the weights will be trained to accommodate them.
BatchNorm?
I guess the first question you need to ask yourself is do you want to use BatchNorm? If your net is deep and you are concerned with covariate shifts then you probably should have a BatchNorm -- and here goes option no. 3
BatchNorm first?
If your x is the output of another conv layer, than there's actually no difference between your first and second alternatives: your net is a cascade of ...-conv-bn-ReLU-conv-BN-ReLU-conv-... so it's only an "artificial" partitioning of the net into triplets of functions conv, bn, relu and up to the very first and last functions you can split things however you wish. Moreover, since Batch norm is a linear operation (scale + bias) it can be "folded" into an adjacent conv layer without changing the net, so you basically left with conv-relu pairs.
So, there's not really a big difference between the first two options you highlighted.
What else to consider?
Do you really need ReLU when changing dimension of features? You can think of the reducing dimensions as a linear mapping - decomposing the weights mapping to x into a lower rank matrix that ultimately maps into c dimensional space instead of 2c space. If you consider a linear mapping, then you might omit the ReLU altogether.
See fast RCNN SVD trick for an example.
Training fully convolutional nerworks (FCNs) for pixelwise semantic segmentation is very memory intensive. So we often use batchsize=1 for traing FCNs. However, when we finetune the pretrained networks with BatchNorm (BN) layers, batchsize=1 doesn't make sense for the BN layers. So, how to handle the BN layers?
Some options:
delete the BN layers (merge the BN layers with the preceding layers for the pretrained model)
Freeze the parameters and statistics of the BN layers
....
which is better and any demo for implementation in pytorch/tf/caffe?
Having only one element will make the batch normalization zero if epsilon is non-zero (variance is zero, mean will be same as input).
Its better to delete the BN layers from the network and try the activation function SELU (scaled exponential linear units). This is from the paper 'Self normalizing neural networks' (SNNs).
Quote from the paper:
While batch normalization requires explicit normalization, neuron
activations of SNNs automatically converge towards zero mean and
unit variance. The activation function of SNNs are “scaled
exponential linear units” (SELUs), which induce self-normalizing
properties.
The SELU is defined as:
def selu(x, name="selu"):
alpha = 1.6732632423543772848170429916717
scale = 1.0507009873554804934193349852946
return scale * tf.where(x >= 0.0, x, alpha * tf.nn.elu(x))
Batch Normalization was introduced to reduce the internal covariate shift of the input feature maps. Due to change of parameters of each layer after every optimization steps, input distribution of a layer also changes, this slow down the model convergence. By using Batch Normalization we can normalize the input distribution irrespective of the batch_size (whether batch_size =1 or larger).
BN normalizes the input distribution
For convolutional network input for intermediate layer is 4D tensor. [batch_size, width, height, num_filters]. Normalization effect all the feature maps.
delete the BN layers (merge the BN layers with the preceding layers for the pretrained model)
This may further slow down the training step and convergence mayn't be achieved.
Freeze the parameters and statistics of the BN layers
Sometime the input data distribution for retrain/finetune, may vary significantly from the original data used to train the pretrained model used for initialization, Due to which your model may end-up in non-optimal solution.
According to my experiments in PyTorch, if convolutional layer before the BN outputs more than one value (i.e. 1 x feat_nb x height x width, where height > 1 or width > 1), then the BN still works fine even when the batch size is equal to one. However, I suspect that in this case the variance estimate might be very biased since all samples that are used for variance calculation come from the same image. Therefore in my case I still decided to use small batch.
The effective batch size over convolutional layer
I think the CNN-relative section (Section 3.2) in the BN original paper could help. From the point of view of the authors, it should be OK to use batch size = 1 for convolutional layers. The "effective batch size" for convolutional layer actually is batch_size * image_height * image_width.
I do not have an exact answer, but here are my thoughts:
networks with BatchNorm (BN) layers, batchsize=1 doesn't make sense
for the BN layers
The main motivation of BN is to fix the distribution (mean/variance) of the input in the batch. In my opinion, having one element this does not make sense. Judging from the paper
you will need to calculate the mean and the variance for 1 element, which does not make sense.
You can always just remove BN but are you sure you can't afford at least 16 elements in the batch?
My observation is in contrary with Stephan's: using PyTorch on a similar input batch x feat_nb x height x width, where height > 1 or width > 1, I found adding BatchNorm after the last conv and before the last non-linear (sigmoid) actually hurts the accuracy by a big margin. Still trying to make sense out of it..
(batch size = 8)
I'm using TensorFlow for a multi-target regression problem. Specifically, in a convolutional network with pixel-wise labeling with the input being an image and the label being a "heat-map" where each pixel has a float value. More specifically, the ground truth labeling for each pixel is lower bounded by zero, and, while technically having no upper bound, usually gets no larger than 1e-2.
Without batch normalization, the network is able to give a reasonable heat-map prediction. With batch normalization, the network takes much long to get to reasonable loss value, and the best it does is making every pixel the average value. This is using the tf.contrib.layers conv2d and batch_norm methods, with the batch_norm being passed to the conv2d's normalization_fn (or not in the case of no batch normalization). I had briefly tried batch normalization on another (single value) regression network, and had trouble then as well (though, I hadn't tested that as extensively). Is there a problem using batch normalization on regression problems in general? Is there a common solution?
If not, what could be some causes batch normalization failing on such an application? I've attempted a variety of initializations, learning rates, etc. I would expect the final layer (which of course does not use batch normalization) could use weights to scale the output of the penultimate layer to the appropriate regression values. Failing that, I removed batch norm from that layer, but with no improvement. I've attempted a small classification problem using batch normalization and saw no problem there, so it seems reasonable that it could be due somehow to the nature of the regression problem, but I don't know how that could cause such a drastic difference. Is batch normalization known to have trouble on regression problems?
I believe your issue is in the labels. Batch norm will scale all input values between 0 and 1. If the labels are not scaled to a similar range the task will be more difficult. This is because it requires the NN to learn values of a different scale.
By removing the batch norm from the penultimate layer, the task may be improved slightly, but you are still requiring an NN layer to learn to downscale values of its input while subsequently normalizing back to the range 0 - 1 (opposite to your objective).
To solve this problem, apply a 0 - 1 scaler to the labels such that your upper bound is no longer 1e-2. During inference, transform the predictions back with the same function to get the actual prediction.