I can't get TensorFlow RELU activations (neither tf.nn.relu nor tf.nn.relu6) working without NaN values for activations and weights killing my training runs.
I believe I'm following all the right general advice. For example I initialize my weights with
weights = tf.Variable(tf.truncated_normal(w_dims, stddev=0.1))
biases = tf.Variable(tf.constant(0.1 if neuron_fn in [tf.nn.relu, tf.nn.relu6] else 0.0, shape=b_dims))
and use a slow training rate, e.g.,
tf.train.MomentumOptimizer(0.02, momentum=0.5).minimize(cross_entropy_loss)
But any network of appreciable depth results in NaN for cost and and at least some weights (at least in the summary histograms for them). In fact, the cost is often NaN right from the start (before training).
I seem to have these issues even when I use L2 (about 0.001) regularization, and dropout (about 50%).
Is there some parameter or setting that I should adjust to avoid these issues? I'm at a loss as to where to even begin looking, so any suggestions would be appreciated!
Following He et. al (as suggested in lejlot's comment), initializing the weights of the l-th layer to a zero-mean Gaussian distribution with standard deviation
where nl is the flattened length of the the input vector or
stddev=np.sqrt(2 / np.prod(input_tensor.get_shape().as_list()[1:]))
results in weights that generally do not diverge.
If you use a softmax classifier at the top of your network, try to make the initial weights of the layer just below the softmax very small (e.g. std=1e-4). This makes the initial distribution of outputs of the network very soft (high temperature), and helps ensure that the first few steps of your optimization are not too large and numerically unstable.
Have you tried gradient clipping and/or a smaller learning rate?
Basically, you will need to process your gradients before applying them, as follows (from tf docs, mostly):
# Replace this with what follows
# opt = tf.train.MomentumOptimizer(0.02, momentum=0.5).minimize(cross_entropy_loss)
# Create an optimizer.
opt = tf.train.MomentumOptimizer(learning_rate=0.001, momentum=0.5)
# Compute the gradients for a list of variables.
grads_and_vars = opt.compute_gradients(cross_entropy_loss, tf.trainable_variables())
# grads_and_vars is a list of tuples (gradient, variable). Do whatever you
# need to the 'gradient' part, for example cap them, etc.
capped_grads_and_vars = [(tf.clip_by_value(gv[0], -5., 5.), gv[1]) for gv in grads_and_vars]
# Ask the optimizer to apply the capped gradients.
opt = opt.apply_gradients(capped_grads_and_vars)
Also, the discussion in this question might help.
Related
I am training a unsupervised NN model and for some reason, after exactly one epoch (80 steps), model stops learning.
]
Do you have any idea why it might happen and what should I do to prevent it?
This is more info about my NN:
I have a deep NN that tries to solve an optimization problem. My loss function is customized and it is my objective function in the optimization problem.
So if my optimization problems is min f(x) ==> loss, now in my DNN loss = f(x). I have 64 input, 64 output, 3 layers in between :
self.l1 = nn.Linear(input_size, hidden_size)
self.relu1 = nn.LeakyReLU()
self.BN1 = nn.BatchNorm1d(hidden_size)
and last layer is:
self.l5 = nn.Linear(hidden_size, output_size)
self.tan5 = nn.Tanh()
self.BN5 = nn.BatchNorm1d(output_size)
to scale my network.
with more layers and nodes(doubles: 8 layers each 200 nodes), I can get a little more progress toward lower error, but again after 100 steps training error becomes flat!
The symptom is that the training loss stops being improved relatively early. Suppose that your problem is learnable at all, there are many reasons for the for this behavior. Following are most relavant:
Improper preprocessing of input: Neural network prefers input with
zero mean. E.g., if the input is all positive, it will restrict the
weights to be updated in the same direction, which may not be
desirable (https://youtu.be/gYpoJMlgyXA).
Therefore, you may want to subtract the mean from all the images (e.g., subtract 127.5 from each of the 3 channels). Scaling to make unit standard deviation in each channel may also be helpful.
Generalization ability of the network: The network is not complicated
or deep enough for the task.
This is very easy to check. You can train the network on just a few
images (says from 3 to 10). The network should be able to overfit the
data and drives the loss to almost 0. If it is not the case, you may
have to add more layers such as using more than 1 Dense layer.
Another good idea is to used pre-trained weights (in applications of Keras documentation). You may adjust the Dense layers at the top to fit with your problem.
Improper weight initialization. Improper weight initialization can
prevent the network from converging (https://youtu.be/gYpoJMlgyXA,
the same video as before).
For the ReLU activation, you may want to use He initialization
instead of the default Glorot initialiation. I find that this may be
necessary sometimes but not always.
Lastly, you can use debugging tools for Keras such as keras-vis, keplr-io, deep-viz-keras. They are very useful to open the blackbox of convolutional networks.
I faced the same problem then I followed the following:
After going through a blog post, I managed to determine that my problem resulted from the encoding of my labels. Originally I had them as one-hot encodings which looked like [[0, 1], [1, 0], [1, 0]] and in the blog post they were in the format [0 1 0 0 1]. Changing my labels to this and using binary crossentropy has gotten my model to work properly. Thanks to Ngoc Anh Huynh and rafaelvalle!
I have implemented Autoencoder using Keras that takes 112*112*3 neurons as input and 100 neurons as the compressed/encoded state. I want to find the neurons out of these 100 that learns the important features. So far i have calculated eigen values(e) and eigen vectors(v) using the following steps. And i found out that around first 30 values of (e) is greater than 0. Does that mean the first 30 modes are the important ones? Is there any other method that could find the important neurons?
Thanks in Advance
x_enc = enc_model.predict(x_train, batch_size=BATCH_SIZE) # shape (3156,100)
x_mean = np.mean(x_enc, axis=0) # shape (100,)
x_stds = np.std(x_enc, axis=0) # shape (100,)
x_cov = np.cov((x_enc - x_mean).T) # shape (100,100)
e, v = np.linalg.eig(x_cov) # shape (100,) and (100,100) respectively
I don't know if the approach you are using will actually give you any useful results since the way the network learns and what it exactly learns aren't known, I suggest you use a different kind of autoencoder, that automatically learns disentangled representations of the data in a latent space, this way you can be sure that all the parameters you find are actually contributing to the representation of your data. check this article
I have 38 variables, like oxygen, temperature, pressure, etc and have a task to determine the total yield produced every day from these variables. When I calculate the regression coefficients and intercept value, they seem to be abnormal and very high (Impractical). For example, if 'temperature' coefficient was found to be +375.456, I could not give a meaning to them saying an increase in one unit in temperature would increase yield by 375.456g. That's impractical in my scenario. However, the prediction accuracy seems right. I would like to know, how to interpret these huge intercept( -5341.27355) and huge beta values shown below. One other important point is that I removed multicolinear columns and also, I am not scaling the variables/normalizing them because I need beta coefficients to have meaning such that I could say, increase in temperature by one unit increases yield by 10g or so. Your inputs are highly appreciated!
modl.intercept_
Out[375]: -5341.27354961415
modl.coef_
Out[376]:
array([ 1.38096017e+00, -7.62388829e+00, 5.64611255e+00, 2.26124164e-01,
4.21908571e-01, 4.50695302e-01, -8.15167717e-01, 1.82390184e+00,
-3.32849969e+02, 3.31942553e+02, 3.58830763e+02, -2.05076898e-01,
-3.06404757e+02, 7.86012402e+00, 3.21339318e+02, -7.00817205e-01,
-1.09676321e+04, 1.91481734e+00, 6.02929848e+01, 8.33731416e+00,
-6.23433431e+01, -1.88442804e+00, 6.86526274e+00, -6.76103795e+01,
-1.11406021e+02, 2.48270706e+02, 2.94836048e+01, 1.00279016e+02,
1.42906659e-02, -2.13019683e-03, -6.71427100e+02, -2.03158515e+02,
9.32094007e-03, 5.56457014e+01, -2.91724945e+00, 4.78691176e-01,
8.78121854e+00, -4.93696073e+00])
It's very unlikely that all of these variables are linearly correlated, so I would suggest that you have a look at simple non-linear regression techniques, such as Decision Trees or Kernel Ridge Regression. These are however more difficult to interpret.
Going back to your issue, these high weights might well be due to there being some high amount of correlation between the variables, or that you simply don't have very much training data.
If you instead of linear regression use Lasso Regression, the solution is biased away from high regression coefficients, and the fit will likely improve as well.
A small example on how to do this in scikit-learn, including cross validation of the regularization hyper-parameter:
from sklearn.linear_model LassoCV
# Make up some data
n_samples = 100
n_features = 5
X = np.random.random((n_samples, n_features))
# Make y linear dependent on the features
y = np.sum(np.random.random((1,n_features)) * X, axis=1)
model = LassoCV(cv=5, n_alphas=100, fit_intercept=True)
model.fit(X,y)
print(model.intercept_)
If you have a linear regression, the formula looks like this (y= target, x= features inputs):
y= x1*b1 +x2*b2 + x3*b3 + x4*b4...+ c
where b1,b2,b3,b4... are your modl.coef_. AS you already realized one of your bigges number is 3.319+02 = 331 and the intercept is also quite big with -5431.
As you already mentioned the coeffiecient variables means how much the target variable changes, if the coeffiecient feature changes with 1 unit and all others features are constant.
so for your interpretation, the higher the absoult coeffienct, the higher the influence of your analysis. But it is important to note that the model is using a lot of high coefficient, that means your model is not depending only of one variable
I'm building Kmeans in pytorch using gradient descent on centroid locations, instead of expectation-maximisation. Loss is the sum of square distances of each point to its nearest centroid. To identify which centroid is nearest to each point, I use argmin, which is not differentiable everywhere. However, pytorch is still able to backprop and update weights (centroid locations), giving similar performance to sklearn kmeans on the data.
Any ideas how this is working, or how I can figure this out within pytorch? Discussion on pytorch github suggests argmax is not differentiable: https://github.com/pytorch/pytorch/issues/1339.
Example code below (on random pts):
import numpy as np
import torch
num_pts, batch_size, n_dims, num_clusters, lr = 1000, 100, 200, 20, 1e-5
# generate random points
vector = torch.from_numpy(np.random.rand(num_pts, n_dims)).float()
# randomly pick starting centroids
idx = np.random.choice(num_pts, size=num_clusters)
kmean_centroids = vector[idx][:,None,:] # [num_clusters,1,n_dims]
kmean_centroids = torch.tensor(kmean_centroids, requires_grad=True)
for t in range(4001):
# get batch
idx = np.random.choice(num_pts, size=batch_size)
vector_batch = vector[idx]
distances = vector_batch - kmean_centroids # [num_clusters, #pts, #dims]
distances = torch.sum(distances**2, dim=2) # [num_clusters, #pts]
# argmin
membership = torch.min(distances, 0)[1] # [#pts]
# cluster distances
cluster_loss = 0
for i in range(num_clusters):
subset = torch.transpose(distances,0,1)[membership==i]
if len(subset)!=0: # to prevent NaN
cluster_loss += torch.sum(subset[:,i])
cluster_loss.backward()
print(cluster_loss.item())
with torch.no_grad():
kmean_centroids -= lr * kmean_centroids.grad
kmean_centroids.grad.zero_()
As alvas noted in the comments, argmax is not differentiable. However, once you compute it and assign each datapoint to a cluster, the derivative of loss with respect to the location of these clusters is well-defined. This is what your algorithm does.
Why does it work? If you had only one cluster (so that the argmax operation didn't matter), your loss function would be quadratic, with minimum at the mean of the data points. Now with multiple clusters, you can see that your loss function is piecewise (in higher dimensions think volumewise) quadratic - for any set of centroids [C1, C2, C3, ...] each data point is assigned to some centroid CN and the loss is locally quadratic. The extent of this locality is given by all alternative centroids [C1', C2', C3', ...] for which the assignment coming from argmax remains the same; within this region the argmax can be treated as a constant, rather than a function and thus the derivative of loss is well-defined.
Now, in reality, it's unlikely you can treat argmax as constant, but you can still treat the naive "argmax-is-a-constant" derivative as pointing approximately towards a minimum, because the majority of data points are likely to indeed belong to the same cluster between iterations. And once you get close enough to a local minimum such that the points no longer change their assignments, the process can converge to a minimum.
Another, more theoretical way to look at it is that you're doing an approximation of expectation maximization. Normally, you would have the "compute assignments" step, which is mirrored by argmax, and the "minimize" step which boils down to finding the minimizing cluster centers given the current assignments. The minimum is given by d(loss)/d([C1, C2, ...]) == 0, which for a quadratic loss is given analytically by the means of data points within each cluster. In your implementation, you're solving the same equation but with a gradient descent step. In fact, if you used a 2nd order (Newton) update scheme instead of 1st order gradient descent, you would be implicitly reproducing exactly the baseline EM scheme.
Imagine this:
t = torch.tensor([-0.0627, 0.1373, 0.0616, -1.7994, 0.8853,
-0.0656, 1.0034, 0.6974, -0.2919, -0.0456])
torch.argmax(t).item() # outputs 6
We increase t[0] for some, δ close to 0, will this update the argmax? It will not, so we are dealing with 0 gradients, all the time. Just ignore this layer, or assume it is frozen.
The same is for argmin, or any other function where the dependent variable is in discrete steps.
I am trying to produce a mathematical operation selection nn model, which is based on the scalar input. The operation is selected based on the softmax result which is produce by the nn. Then this operation has to be applied to the scalar input in order to produce the final output. So far I’ve come up with applying argmax and onehot on the softmax output in order to produce a mask which then is applied on the concated values matrix from all the possible operations to be performed (as show in the pseudo code below). The issue is that neither argmax nor onehot appears to be differentiable. I am new to this, so any would be highly appreciated. Thanks in advance.
#perform softmax
logits = tf.matmul(current_input, W) + b
softmax = tf.nn.softmax(logits)
#perform all possible operations on the input
op_1_val = tf_op_1(current_input)
op_2_val = tf_op_2(current_input)
op_3_val = tf_op_2(current_input)
values = tf.concat([op_1_val, op_2_val, op_3_val], 1)
#create a mask
argmax = tf.argmax(softmax, 1)
mask = tf.one_hot(argmax, num_of_operations)
#produce the input, by masking out those operation results which have not been selected
output = values * mask
I believe that this is not possible. This is similar to Hard Attention described in this paper. Hard attention is used in Image captioning to allow the model to focus only on a certain part of the image at each step. Hard attention is not differentiable but there are 2 ways to go around this:
1- Use Reinforcement Learning (RL): RL is made to train models that makes decisions. Even though, the loss function won't back-propagate any gradients to the softmax used for the decision, you can use RL techniques to optimize the decision. For a simplified example, you can consider the loss as penalty, and send to the node, with the maximum value in the softmax layer, a policy gradient proportional to the penalty in order to decrease the score of the decision if it was bad (results in a high loss).
2- Use something like soft attention: instead of picking only one operation, mix them with weights based on the softmax. so instead of:
output = values * mask
Use:
output = values * softmax
Now, the operations will converge down to zero based on how much the softmax will not select them. This is easier to train compared to RL but it won't work if you must completely remove the non-selected operations from the final result (set them to zero completely).
This is another answer that talks about Hard and Soft attention that you may find helpful: https://stackoverflow.com/a/35852153/6938290