I want to implement a kernel ridge regression that also works within MLJ. Moreover, I want to have the option to use either feature vectors or a predefined kernel matrix as in Python sklearn.
When I run this code
const MMI = MLJModelInterface
MMI.#mlj_model mutable struct KRRModel <: MLJModelInterface.Deterministic
mu::Float64 = 1::(_ > 0)
kernel::String = "linear"
end
function MMI.fit(m::KRRModel,verbosity::Int,K,y)
K = MLJBase.matrix(K)
fitresult = inv(K+m.mu*I)*y
cache = nothing
report = nothing
return (fitresult,cache,report)
end
N = 10
K = randn(N,N)
K = K*K
a = randn(N)
y = K*a + 0.2*randn(N)
m = KRRModel()
kregressor = machine(m,K,y)
cv = CV(; nfolds=6, shuffle=nothing, rng=nothing)
evaluate!(kregressor, resampling=cv, measure=rms, verbosity=1)
the evaluate! function evaluates the machine on different subsets of rows of K. Due to the Representer Theorem, a kernel ridge regression has a number of nonzero coefficients equal to the number of samples. Hence, a reduced size matrix K[train_rows,train_rows] can be used instead of K[train_rows,:].
To denote I'm using a kernel matrix I'd set m.kernel = "" . How do I make evaluate! select the columns as well as the rows to form a smaller matrix when m.kernel = ""?
This is my first time using MLJ and I'd like to make as few modifications as possible.
Quoting the answer I got on the Julia Discourse from #ablaom
The intended use of evaluate! is to estimate the generalisation error
associated with some supervised learning model, by subsampling
observations, as in cross-validation, a common use-case. I’m afraid
there is no natural way for evaluate! do feature subsampling.
https://alan-turing-institute.github.io/MLJ.jl/dev/evaluating_model_performance/
FYI: There is a version of kernel regression implementing the MLJ
model interface, namely kernel partial least squares regression from
the package GitHub - lalvim/PartialLeastSquaresRegressor.jl:
Implementation of a Partial Least Squares Regressor 2 .
Related
I'm using sklearn for SVR (regression) using an RBF kernel. I'm want to know how the inference is done under the hood. I thought it was a function of the support vectors, function mean, and gamma, but it appears I'm missing one aspect (probably some scaling based on how close 2 points are.
Here is "my Equation" that I've tried in the graph's below.
out = mean
for vect in vectors:
out = out + (vect.y - mean) * math.exp(-(vect.x - x) ** 2 * gamma)
When I do just 2 points spaced away, my equation matches what skLearn reports with svr.predict.
With 3 training points and 2 close together, my equation does not match what svr.predict gives:
Given the support vectors, gamma, and mean, and anything else needed, what is the equation for SVR inference with RBF kernel? Can those be obtained from the sklearn svr class?
The equation that works for me using sklearn library and SVR inference with RBF kernel is as follows with python code:
# x and y is already defined, and is the training data for the SVR
svr = svm.SVR(kernel="rbf", C=C, gamma=gamma, epsilon=epsilon, tol=tol)
svr.fit(x,y)
vectors = []
for i in svr.support_:
vectors.append([x[i][0], y[i]])
out = svr._intercept_[0]
for vect, coef in zip(vectors, svr._dual_coef_[0]):
out = out + coef * math.exp(-(vect[0] - x) ** 2 * gamma)
I found that svr._intercept_[0] contains the y offset for the function.
I found that svr._dual_coef_[0] contains the coefficients to multiply each of the exponentials by.
I found that svr.support_ contains the indexes of the elements in your training set used as the support vectors.
I realize I'm accessing what is intended to be accessed within the svr class only, however, I don't see an official API method for accessing these variables, and this is working for me for now.
I'm running tensorflow 2.1 and tensorflow_probability 0.9. I have fit a Structural Time Series Model with a seasonal component. I am using code from the Tensorflow Probability Structural Time Series Probability example:
Tensorflow Github.
In the example there is a great plot where the decomposition is visualised:
# Get the distributions over component outputs from the posterior marginals on
# training data, and from the forecast model.
component_dists = sts.decompose_by_component(
demand_model,
observed_time_series=demand_training_data,
parameter_samples=q_samples_demand_)
forecast_component_dists = sts.decompose_forecast_by_component(
demand_model,
forecast_dist=demand_forecast_dist,
parameter_samples=q_samples_demand_)
demand_component_means_, demand_component_stddevs_ = (
{k.name: c.mean() for k, c in component_dists.items()},
{k.name: c.stddev() for k, c in component_dists.items()})
(
demand_forecast_component_means_,
demand_forecast_component_stddevs_
) = (
{k.name: c.mean() for k, c in forecast_component_dists.items()},
{k.name: c.stddev() for k, c in forecast_component_dists.items()}
)
When using a trend component, is it possible to decompose and visualise both:
trend/_level_scale & trend/_slope_scale
I have tried many permutations to extract the nested element of the trend component with no luck.
Thanks for your time in advance.
We didn't write a separate STS interface for this, but you can access the posterior on latent states (in this case, both the level and slope) by directly querying the underlying state-space model for its marginal means and covariances:
ssm = model.make_state_space_model(
num_timesteps=num_timesteps,
param_vals=parameter_samples)
posterior_means, posterior_covs = (
ssm.posterior_marginals(observed_time_series))
You should also be able to draw samples from the joint posterior by running ssm.posterior_sample(observed_time_series, num_samples).
It looks like there's currently a glitch when drawing posterior samples from a model with no batch shape (Could not find valid device for node. Node:{{node Reshape}}): while we fix that, it should work to add an artificial batch dimension as a workaround:
ssm.posterior_sample(observed_time_series[tf.newaxis, ...], num_samples).
I have recently started experimenting with OneClassSVM ( using Sklearn ) for unsupervised learning and I followed
this example .
I apologize for the silly questions But I’m a bit confused about two things :
Should I train my svm on both regular example case as well as the outliers , or the training is on regular examples only ?
Which of labels predicted by the OSVM and represent outliers is it 1 or -1
Once again i apologize for those questions but for some reason i cannot find this documented anyware
As this example you reference is about novelty-detection, the docs say:
novelty detection:
The training data is not polluted by outliers, and we are interested in detecting anomalies in new observations.
Meaning: you should train on regular examples only.
The approach is based on:
Schölkopf, Bernhard, et al. "Estimating the support of a high-dimensional distribution." Neural computation 13.7 (2001): 1443-1471.
Extract:
Suppose you are given some data set drawn from an underlying probability distribution P and you want to estimate a “simple” subset S of input space such that the probability that a test point drawn from P lies outside of S equals some a priori specied value between 0 and 1.
We propose a method to approach this problem by trying to estimate a function f that is positive on S and negative on the complement.
The above docs also say:
Inliers are labeled 1, while outliers are labeled -1.
This can also be seen in your example code, extracted:
# Generate some regular novel observations
X = 0.3 * np.random.randn(20, 2)
X_test = np.r_[X + 2, X - 2]
...
# all regular = inliers (defined above)
y_pred_test = clf.predict(X_test)
...
# -1 = outlier <-> error as assumed to be inlier
n_error_test = y_pred_test[y_pred_test == -1].size
The general idea I am trying to realize is a seq2seq-model (taken from the translate.py-example in the models, based on the seq2seq-class). This trains well.
Furthermore I am using the hidden state of the rnn after all the encoding is done, right before decoding starts (I call it the “hidden state at end of encoding”). I use this hidden state at end of encoding to feed it into a further sub-graph which I call “prices” (see below). The training gradients of this sub-graph backprop not only through this additional sub-graph, but also back into the encoder-part of the rnn (which is what I want and need).
The plan is to add more such sub-graph to the hidden state at end of encoding, as I want to analyze the input phrases in a variety of ways.
Now during training when I evaluate and train both sub-graphs (encoder+prices AND encoder+decoder) at the same time, the net does NOT converge. However, if I train by executing the training in the following way (pseudo-code):
if global_step % 10 == 0:
execute-the-price-training_code
else:
execute-the-decoder-training_code
So I am not training both sub-graphs simultaneously. Now it does converge, but the encoder+decoder-part converges MUCH slower than if I ONLY train this part and never train the prices-sub-graph.
My question is: I should be able to train both sub-graphs simultaneously. But probably I have to rescale the gradients flowing back into the hidden state at end of encoding. Here we get the gradients from the prices sub-graph AND from the decoder-sub-graph. How should this rescaling be done. I didnt find any papers describing such an undertaking, but maybe I am searching with the wrong keywords.
Here is the training-part of the code:
This is the (almost original) training-op-preparation:
if not forward_only:
self.gradient_norms = []
self.updates = []
opt = tf.train.AdadeltaOptimizer(self.learning_rate)
for bucket_id in xrange(len(buckets)):
tf.scalar_summary("seq2seq loss", self.losses[bucket_id])
gradients = tf.gradients(self.losses[bucket_id], var_list_seq2seq)
clipped_gradients, norm = tf.clip_by_global_norm(gradients, max_gradient_norm)
self.gradient_norms.append(norm)
self.updates.append(opt.apply_gradients(zip(clipped_gradients, var_list_seq2seq), global_step=self.global_step))
Now, additionally, I am running a second sub-graph that takes the hidden state at end of encoding as input:
with tf.name_scope('prices') as scope:
#First layer
W_price_first_layer = tf.Variable(tf.random_normal([num_layers*size, self.prices_hidden_layer_size], stddev=0.35), name="W_price_first_layer")
B_price_first_layer = tf.Variable(tf.zeros([self.prices_hidden_layer_size]), name="B_price_first_layer")
self.output_price_first_layer = tf.add(tf.matmul(self.hidden_state, W_price_first_layer), B_price_first_layer)
self.activation_price_first_layer = tf.nn.sigmoid(self.output_price_first_layer)
#self.activation_price_first_layer = tf.nn.Relu(self.output_price_first_layer)
#Second layer to softmax (price ranges)
W_price = tf.Variable(tf.random_normal([self.prices_hidden_layer_size, self.prices_bit_size], stddev=0.35), name="W_price")
W_price_t = tf.transpose(W_price)
B_price = tf.Variable(tf.zeros([self.prices_bit_size]), name="B_price")
self.output_price_second_layer = tf.add(tf.matmul(self.activation_price_first_layer, W_price),B_price)
self.price_prediction = tf.nn.softmax(self.output_price_second_layer)
self.label_price = tf.placeholder(tf.int32, shape=[self.batch_size], name="price_label")
#Remember the prices trainables
var_list_prices = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES, "prices")
var_list_all = tf.trainable_variables()
#Backprop
self.loss_price = tf.nn.sparse_softmax_cross_entropy_with_logits(self.output_price_second_layer, self.label_price)
self.loss_price_scalar = tf.reduce_mean(self.loss_price)
self.optimizer_price = tf.train.AdadeltaOptimizer(self.learning_rate_prices)
self.training_op_price = self.optimizer_price.minimize(self.loss_price, var_list=var_list_all)
Thx a bunch
I expect that running two optimizers simultaneously will lead to inconsistent gradient updates on the common variables, and this might be causing your training not to converge.
Instead, if you add the scalar loss from each sub-network to the "losses collection" (e.g. via tf.contrib.losses.add_loss() or tf.add_to_collection(tf.GraphKeys.LOSSES, ...), you can use tf.contrib.losses.get_total_loss() to get a single loss value that can be passed to a single standard TensorFlow tf.train.Optimizer subclass. TensorFlow will derive the appropriate back-prop computation for your split network.
The get_total_loss() method simply computes an unweighted sum of the values that have been added to the losses collection. I'm not familiar with the literature on how or if you should scale these values, but you can use any arbitrary (differentiable) TensorFlow expression to combine the losses and pass the result to a single optimizer.
Does it help in classifying better if I add linear, non-linear combinatinos of the existing features ? For example does it help to add mean, variance as new features computed from the existing features ? I believe that it definitely depends on the classification algorithm as in the case of PCA, the algorithm by itself generates new features which are orthogonal to each other and are linear combinations of the input features. But how does it effect in the case of decision tree based classifiers or others ?
Yes, combination of existing features can give new features and help for classification. Moreover, combination of the feature with itself (e.g. polynomial from the feature) can be used as this additional data to be used during classification.
As an example, consider logistic regression classifier with such linear formula as its core:
g(x, y) = 1*x + 2*y
Imagine, that you have 2 observations:
x = 6; y = 1
x = 3; y = 2.5
In both cases g() will be equal to 8. If observations belong to different classes, you have no possibility to distinguish them. But let's add one more variable (feature) z, which is combination of the previous 2 features - z = x * y:
g(x, y, z) = 1*x + 2*y + 0.5*z
Now for same observations we have:
x = 6; y = 1; z = 6 * 1 = 6 ==> g() = 11
x = 3; y = 2.5; z = 3 * 2.5 = 7.5 ==> g() = 11.75
So now we get 2 different points and can distinguish between 2 observations.
Polynomial features (x^2, x^3, y^2, etc.) do not give additional points, but instead change the graph of the function. For example, g(x) = a0 + a1*x is a line, while g(x) = a0 + a1*x + a2*x^2 is parabola and thus can fit data much more closely.
In general, it's always better to have more features. Unless you have very predictive features (i.e. they allow for perfect separation of the classes to predict) already, I would always recommend adding more features. In practice, many classification algorithms (and in particular decision tree inducers) select the best features for their purposes anyway.
There are open-source Python libraries that automate feature creation / combination:
We can automate polynomial feature creations with sklearn.
We can automatically create spline features with sklearn.
We can combine features mathematically with Feature-engine. With MathFeatures we combine feature groups, and with RelativeFeatures we combine feature pairs.