I am trying to model an ordinal predicted variable using PyMC3 based on the approach in chapter 23 of Doing Bayesian Data Analysis. I would like to determine a good starting value using find_MAP, but am receiving an optimization error.
The model:
import pymc3 as pm
import numpy as np
import theano
import theano.tensor as tt
# Some helper functions
def cdf(x, location=0, scale=1):
epsilon = np.array(1e-32, dtype=theano.config.floatX)
location = tt.cast(location, theano.config.floatX)
scale = tt.cast(scale, theano.config.floatX)
div = tt.sqrt(2 * scale ** 2 + epsilon)
div = tt.cast(div, theano.config.floatX)
erf_arg = (x - location) / div
return .5 * (1 + tt.erf(erf_arg + epsilon))
def percent_to_thresh(idx, vect):
return 5 * tt.sum(vect[:idx + 1]) + 1.5
def full_thresh(thresh):
idxs = tt.arange(thresh.shape[0] - 1)
thresh_mod, updates = theano.scan(fn=percent_to_thresh,
sequences=[idxs],
non_sequences=[thresh])
return tt.concatenate([[-1 * np.inf, 1.5], thresh_mod, [6.5, np.inf]])
def compute_ps(thresh, location, scale):
f_thresh = full_thresh(thresh)
return cdf(f_thresh[1:], location, scale) - cdf(f_thresh[:-1], location, scale)
# Generate data
real_ps = [0.05, 0.05, 0.1, 0.1, 0.2, 0.3, 0.2]
data = np.random.choice(7, size=1000, p=real_ps)
# Run model
with pm.Model() as model:
mu = pm.Normal('mu', mu=4, sd=3)
sigma = pm.Uniform('sigma', lower=0.1, upper=70)
thresh = pm.Dirichlet('thresh', a=np.ones(5))
cat_p = compute_ps(thresh, mu, sigma)
results = pm.Categorical('results', p=cat_p, observed=data)
with model:
start = pm.find_MAP()
trace = pm.sample(2000, start=start)
When running this, I receive the following error:
Applied interval-transform to sigma and added transformed sigma_interval_ to model.
Applied stickbreaking-transform to thresh and added transformed thresh_stickbreaking_ to model.
Traceback (most recent call last):
File "cm_net_log.v1-for_so.py", line 53, in <module>
start = pm.find_MAP()
File "/usr/local/lib/python3.5/site-packages/pymc3/tuning/starting.py", line 133, in find_MAP
specific_errors)
ValueError: Optimization error: max, logp or dlogp at max have non-finite values. Some values may be outside of distribution support. max: {'thresh_stickbreaking_': array([-1.04298465, -0.48661088, -0.84326554, -0.44833646]), 'sigma_interval_': array(-2.220446049250313e-16), 'mu': array(7.68422528308479)} logp: array(-3506.530143064723) dlogp: array([ 1.61013190e-06, nan, -6.73994118e-06,
-6.93873894e-06, 6.03358122e-06, 3.18954680e-06])Check that 1) you don't have hierarchical parameters, these will lead to points with infinite density. 2) your distribution logp's are properly specified. Specific issues:
My questions:
How can I determine why dlogp is nan at certain points?
Is there a different way that I can express this model to avoid dlogp being nan?
Also worth noting:
This model runs fine if I don't find_MAP and use a Metropolis sampler. However, I'd like to have the flexibility of using other samplers as this model becomes more complex.
I have a suspicion that the issue is due to the relationship between the thresholds and the normal distribution, but I don't know how to disentangle them for the optimization.
Regarding question 2: I expressed the model for the ordinal predicted variable (single group) differently; I used the Theano #as_op decorator for a function that calculates probabilities for the outcomes. That also explains why I cannot use find_MAP() or gradient based samplers: Theano cannot calculate a gradient for the custom function. (http://pymc-devs.github.io/pymc3/notebooks/getting_started.html#Arbitrary-deterministics)
# Number of outcomes
nYlevels = df.Y.cat.categories.size
thresh = [k + .5 for k in range(1, nYlevels)]
thresh_obs = np.ma.asarray(thresh)
thresh_obs[1:-1] = np.ma.masked
#as_op(itypes=[tt.dvector, tt.dscalar, tt.dscalar], otypes=[tt.dvector])
def outcome_probabilities(theta, mu, sigma):
out = np.empty(nYlevels)
n = norm(loc=mu, scale=sigma)
out[0] = n.cdf(theta[0])
out[1] = np.max([0, n.cdf(theta[1]) - n.cdf(theta[0])])
out[2] = np.max([0, n.cdf(theta[2]) - n.cdf(theta[1])])
out[3] = np.max([0, n.cdf(theta[3]) - n.cdf(theta[2])])
out[4] = np.max([0, n.cdf(theta[4]) - n.cdf(theta[3])])
out[5] = np.max([0, n.cdf(theta[5]) - n.cdf(theta[4])])
out[6] = 1 - n.cdf(theta[5])
return out
with pm.Model() as ordinal_model_single:
theta = pm.Normal('theta', mu=thresh, tau=np.repeat(.5**2, len(thresh)),
shape=len(thresh), observed=thresh_obs, testval=thresh[1:-1])
mu = pm.Normal('mu', mu=nYlevels/2.0, tau=1.0/(nYlevels**2))
sigma = pm.Uniform('sigma', nYlevels/1000.0, nYlevels*10.0)
pr = outcome_probabilities(theta, mu, sigma)
y = pm.Categorical('y', pr, observed=df.Y.cat.codes.as_matrix())
http://nbviewer.jupyter.org/github/JWarmenhoven/DBDA-python/blob/master/Notebooks/Chapter%2023.ipynb
Related
I am trying to implement a custom loss function in Tensorflow 2.4 using the Keras backend.
The loss function is a ranking loss; I found the following paper with a somewhat log-likelihood loss: Chen et al. Single-Image Depth Perception in the Wild.
Similarly, I wanted to sample some (in this case 50) points from an image to compare the relative order between ground-truth and predicted depth maps using the NYU-Depth dataset. Being a fan of Numpy, I started working with that but came to the following exception:
ValueError: No gradients provided for any variable: [...]
I have learned that this is caused by the arguments not being filled when calling the loss function but instead, a C function is compiled which is then used later. So while I know the dimensions of my tensors (4, 480, 640, 1), I cannot work with the data as wanted and have to use the keras.backend functions on top so that in the end (if I understood correctly), there is supposed to be a path between the input tensors from the TF graph and the output tensor, which has to provide a gradient.
So my question now is: Is this a feasible loss function within keras?
I have already tried a few ideas and different approaches with different variations of my original code, which was something like:
def ranking_loss_function(y_true, y_pred):
# Chen et al. loss
y_true_np = K.eval(y_true)
y_pred_np = K.eval(y_pred)
if y_true_np.shape[0] != None:
num_sample_points = 50
total_samples = num_sample_points ** 2
err_list = [0 for x in range(y_true_np.shape[0])]
for i in range(y_true_np.shape[0]):
sample_points = create_random_samples(y_true, y_pred, num_sample_points)
for x1, y1 in sample_points:
for x2, y2 in sample_points:
if y_true[i][x1][y1] > y_true[i][x2][y2]:
#image_relation_true = 1
err_list[i] += np.log(1 + np.exp(-1 * y_pred[i][x1][y1] + y_pred[i][x2][y2]))
elif y_true[i][x1][y1] < y_true[i][x2][y2]:
#image_relation_true = -1
err_list[i] += np.log(1 + np.exp(y_pred[i][x1][y1] - y_pred[i][x2][y2]))
else:
#image_relation_true = 0
err_list[i] += np.square(y_pred[i][x1][y1] - y_pred[i][x2][y2])
err_list = np.divide(err_list, total_samples)
return K.constant(err_list)
As you can probably tell, the main idea was to first create the sample points and then based on the existing relation between them in y_true/y_pred continue with the corresponding computation from the cited paper.
Can anyone help me and provide some more helpful information or tips on how to correctly implement this loss using keras.backend functions? Trying to include the ordinal relation information really confused me compared to standard regression losses.
EDIT: Just in case this causes confusion: create_random_samples() just creates 50 random sample points (x, y) coordinate pairs based on the shape[1] and shape[2] of y_true (image width and height)
EDIT(2): After finding this variation on GitHub, I have tried out a variation using only TF functions to retrieve data from the tensors and compute the output. The adjusted and probably more correct version still throws the same exception though:
def ranking_loss_function(y_true, y_pred):
#In the Wild ranking loss
y_true_np = K.eval(y_true)
y_pred_np = K.eval(y_pred)
if y_true_np.shape[0] != None:
num_sample_points = 50
total_samples = num_sample_points ** 2
bs = y_true_np.shape[0]
w = y_true_np.shape[1]
h = y_true_np.shape[2]
total_samples = total_samples * bs
num_pairs = tf.constant([total_samples], dtype=tf.float32)
output = tf.Variable(0.0)
for i in range(bs):
sample_points = create_random_samples(y_true, y_pred, num_sample_points)
for x1, y1 in sample_points:
for x2, y2 in sample_points:
y_true_sq = tf.squeeze(y_true)
y_pred_sq = tf.squeeze(y_pred)
d1_t = tf.slice(y_true_sq, [i, x1, y1], [1, 1, 1])
d2_t = tf.slice(y_true_sq, [i, x2, y2], [1, 1, 1])
d1_p = tf.slice(y_pred_sq, [i, x1, y1], [1, 1, 1])
d2_p = tf.slice(y_pred_sq, [i, x2, y2], [1, 1, 1])
d1_t_sq = tf.squeeze(d1_t)
d2_t_sq = tf.squeeze(d2_t)
d1_p_sq = tf.squeeze(d1_p)
d2_p_sq = tf.squeeze(d2_p)
if d1_t_sq > d2_t_sq:
# --> Image relation = 1
output.assign_add(tf.math.log(1 + tf.math.exp(-1 * d1_p_sq + d2_p_sq)))
elif d1_t_sq < d2_t_sq:
# --> Image relation = -1
output.assign_add(tf.math.log(1 + tf.math.exp(d1_p_sq - d2_p_sq)))
else:
output.assign_add(tf.math.square(d1_p_sq - d2_p_sq))
return output/num_pairs
EDIT(3): This is the code for create_random_samples():
(FYI: Because it was weird to get the shape from y_true in this case, I first proceeded to hard-code it here as I know it for the dataset which I am currently using.)
def create_random_samples(y_true, y_pred, num_points=50):
y_true_shape = (4, 480, 640, 1)
y_pred_shape = (4, 480, 640, 1)
if y_true_shape[0] != None:
num_samples = num_points
population = [(x, y) for x in range(y_true_shape[1]) for y in range(y_true_shape[2])]
sample_points = random.sample(population, num_samples)
return sample_points
I have the following code, which basically tries to fit a simple regression model using tensorflow probability. The code runs without error, but the MCMC sampler doesn't seem to be doing anything in that it returns a trace of the initial states.
import tensorflow.compat.v2 as tf
import tensorflow_probability as tfp
from tensorflow_probability import distributions as tfdimport warnings
tf.enable_v2_behavior()
plt.style.use("ggplot")
warnings.filterwarnings('ignore')
ru=4
N=102
N = 102 # number of data points
t = np.linspace(0, 4*np.pi, N)
data = 3+np.sin(t+0.001) + 0.5 + np.random.randn(N)
media_1 = ((data-min(data))/(max(data)-min(data)) ) #+ np.random.normal(0,.05, N)
y = np.repeat(ru, N) + np.random.normal(.3,.01,N) * media_1 + np.random.normal(0, .005, N)
# model
model = tfd.JointDistributionNamed(dict(
beta1 = tfd.Normal(0,1) ,
intercept = tfd.Normal(0,5 ) ,
var = tfd.Normal(0.05, 0.0005) ,
y = lambda intercept,beta1,var:
tfd.Independent(tfd.Normal(loc=intercept + beta1 * media_1, scale=var),
reinterpreted_batch_ndims=1
)
))
def target_log_prob_fn(intercept, beta1, var):
return model.log_prob({'intercept':intercept, 'beta1':beta1, 'var':var, 'y': y })
s = model.sample()
init_states = [ tf.fill([1], s['intercept'].numpy(), name='init_intercept'),
tf.fill([1], s['beta1'].numpy(), name='init_beta1'),
tf.fill([1], s['var'].numpy(), name='init_var'),]
num_results = 5000
num_burnin_steps = 3000
# Improve performance by tracing the sampler using `tf.function`
# and compiling it using XLA.
#tf.function(autograph=False, experimental_compile=True)
def do_sampling():
return tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=init_states,
kernel=tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=target_log_prob_fn,
step_size=0.1,
num_leapfrog_steps=3)
)
states, kernel_results = do_sampling()
The trace that is returned in states is exactly the same as the initial values in initial_states... Any ideas?
I can confirm that this MCMC sampler is not mixing by printing the acceptance rate with a snippet I found here
print("Acceptance rate:", kernel_results.is_accepted.numpy().mean())
That same page provides some hints about how to make your HMC kernel adaptive, which means it will automatically reduce the step size if too many proposed steps are rejected (and increase it if too many are accepted, too):
# Apply a simple step size adaptation during burnin
#tf.function
def do_sampling():
adaptive_kernel = tfp.mcmc.SimpleStepSizeAdaptation(
tfp.mcmc.HamiltonianMonteCarlo(
target_log_prob_fn=target_log_prob_fn,
step_size=0.1,
num_leapfrog_steps=3),
num_adaptation_steps=int(.8 * num_burnin_steps),
target_accept_prob=np.float64(.65))
return tfp.mcmc.sample_chain(
num_results=num_results,
num_burnin_steps=num_burnin_steps,
current_state=init_states,
kernel=adaptive_kernel,
trace_fn=lambda cs, kr: kr)
samples, kernel_results = do_sampling()
print("Acceptance rate:", kernel_results.inner_results.is_accepted.numpy().mean())
This produces a non-zero acceptance rate for me.
I am trying to implement a Bayesian network and solve a regression problem using PYMC3. In my model, I have a fair coin as the parent node. If the parent node is H, the child node selects the normal distribution N(5,0.2); if T, the child selects N(0,0.5). Here is an illustration of my network.
To simulate this network, I generated a sample dataset and tried doing Bayesian regression using the code below. Currently, the model does regression only for the child node as if the parent node does not exist. I would greatly appreciate it if anyone can let me know how to implement the conditional probability P(D|C). Ultimately, I am interested in finding the probability distribution for mu1 and mu2. Thank you!
# Generate data for coin flip P(C) and store in c1
theta_real = 0.5 # unkown value in a real experiment
n_sample = 10
c1 = bernoulli.rvs(p=theta_real, size=n_sample)
# Generate data for normal distribution P(D|C) and store in d1
np.random.seed(123)
mu1 = 0
sigma1 = 0.5
mu2 = 5
sigma2 = 0.2
d1 = []
for index, item in enumerate(c1):
if item == 0:
d1.extend(normal(mu1, sigma1, 1))
else:
d1.extend(normal(mu2, sigma2, 1))
# I start building PYMC3 model here
c1_tensor = theano.shared(np.array(c1))
d1_tensor = theano.shared(np.array(d1))
with pm.Model() as model:
# define prior for c1. I am not sure how to do this.
#c1_present = pm.Categorical('c1',observed=c1_tensor)
# how do I incorporate P(D | C)
mu_prior = pm.Normal('mu', mu=2, sd=2, shape=1)
sigma_prior = pm.HalfNormal('sigma', sd=2, shape=1)
y_likelihood = pm.Normal('y', mu=mu_prior, sd=sigma_prior, observed=d1_tensor)
You could use the Dirichlet distribution as a prior for the coin toss and NormalMixture as the prior of the two Gaussians. In the following snippet I changed the fairness of the coin and increased the number of coin tosses, but you could adjust these in any way want:
import numpy as np
import pymc3 as pm
from scipy.stats import bernoulli
# Generate data for coin flip P(C) and store in c1
theta_real = 0.2 # unkown value in a real experiment
n_sample = 2000
c1 = bernoulli.rvs(p=theta_real, size=n_sample)
# Generate data for normal distribution P(D|C) and store in d1
np.random.seed(123)
mu1 = 0
sigma1 = 0.5
mu2 = 5
sigma2 = 0.2
d1 = []
for index, item in enumerate(c1):
if item == 0:
d1.extend(np.random.normal(mu1, sigma1, 1))
else:
d1.extend(np.random.normal(mu2, sigma2, 1))
with pm.Model() as model:
w = pm.Dirichlet('p', a=np.ones(2))
mu = pm.Normal('mu', 0, 20, shape=2)
sigma = np.array([0.5,0.2])
pm.NormalMixture('like',w=w,mu=mu,sigma=sigma,observed=np.array(d1))
trace = pm.sample()
pm.summary(trace)
This will give you the following:
mean sd mc_error hpd_2.5 hpd_97.5 n_eff Rhat
mu__0 4.981222 0.023900 0.000491 4.935044 5.027420 2643.052184 0.999637
mu__1 -0.007660 0.004946 0.000095 -0.017388 0.001576 2481.146286 1.000312
p__0 0.213976 0.009393 0.000167 0.195602 0.231803 2245.905021 0.999302
p__1 0.786024 0.009393 0.000167 0.768197 0.804398 2245.905021 0.999302
The parameters are recovered nicely as you can also see from the traceplots:
The above implementation will give you the posterior of theta_real, mu1 and mu2 but I could not get convergence when I added sigma1 and sigma2 as parameters to be estimated by the data (even though the prior was quite narrow):
with pm.Model() as model:
w = pm.Dirichlet('p', a=np.ones(2))
mu = pm.Normal('mu', 0, 20, shape=2)
sigma = pm.HalfNormal('sigma', sd=2, shape=2)
pm.NormalMixture('like',w=w,mu=mu,sigma=sigma,observed=np.array(d1))
trace = pm.sample()
print(pm.summary(trace))
Auto-assigning NUTS sampler...
Initializing NUTS using jitter+adapt_diag...
Multiprocess sampling (4 chains in 4 jobs)
NUTS: [sigma, mu, p]
Sampling 4 chains: 100%|██████████| 4000/4000 [00:10<00:00, 395.57draws/s]
The acceptance probability does not match the target. It is 0.883057127209148, but should be close to 0.8. Try to increase the number of tuning steps.
The gelman-rubin statistic is larger than 1.4 for some parameters. The sampler did not converge.
The estimated number of effective samples is smaller than 200 for some parameters.
mean sd mc_error ... hpd_97.5 n_eff Rhat
mu__0 1.244021 2.165433 0.216540 ... 5.005507 2.002049 212.596596
mu__1 3.743879 2.165122 0.216510 ... 5.012067 2.002040 235.750129
p__0 0.643069 0.248630 0.024846 ... 0.803369 2.004185 30.966189
p__1 0.356931 0.248630 0.024846 ... 0.798632 2.004185 30.966189
sigma__0 0.416207 0.125435 0.012517 ... 0.504110 2.009031 17.333177
sigma__1 0.271763 0.125539 0.012533 ... 0.497208 2.007779 19.217223
[6 rows x 7 columns]
Based on that you most likely will need to reparametrize if you also wanted to estimate the two standard deviations from this data.
This answer is to supplement #balleveryday's answer, which suggests the Gaussian Mixture Model, but had some trouble getting the symmetry breaking to work. Admittedly, the symmetry breaking in the official example is done in the context of Metropolis-Hastings sampling, whereas I think NUTS might be a little more sensitive to encountering impossible values (not sure). Here's what worked for me:
import numpy as np
import pymc3 as pm
from scipy.stats import bernoulli
import theano.tensor as tt
# everything should reproduce
np.random.seed(123)
n_sample = 2000
# Generate data for coin flip P(C) and store in c1
theta_real = 0.2 # unknown value in a real experiment
c1 = bernoulli.rvs(p=theta_real, size=n_sample)
# Generate data for normal distribution P(D|C) and store in d1
mu1, mu2 = 0, 5
sigma1, sigma2 = 0.5, 0.2
d1 = np.empty_like(c1, dtype=np.float64)
d1[c1 == 0] = np.random.normal(mu1, sigma1, np.sum(c1 == 0))
d1[c1 == 1] = np.random.normal(mu2, sigma2, np.sum(c1 == 1))
with pm.Model() as gmm_asym:
# mixture vector
w = pm.Dirichlet('p', a=np.ones(2))
# Gaussian parameters (testval helps start off ordered)
mu = pm.Normal('mu', 0, 20, shape=2, testval=[-10, 10])
sigma = pm.HalfNormal('sigma', sd=2, shape=2)
# break symmetry, forcing mu[0] < mu[1]
order_means_potential = pm.Potential('order_means_potential',
tt.switch(mu[1] - mu[0] < 0, -np.inf, 0))
# observed
pm.NormalMixture('like', w=w, mu=mu, sigma=sigma, observed=d1)
# reproducible sampling
tr_gmm_asym = pm.sample(tune=2000, target_accept=0.9, random_seed=20191121)
This produces samples with the statistics
mean sd mc_error hpd_2.5 hpd_97.5 n_eff Rhat
mu__0 0.004549 0.011975 0.000226 -0.017398 0.029375 2425.487301 0.999916
mu__1 5.007663 0.008993 0.000166 4.989247 5.024692 2181.134002 0.999563
p__0 0.789983 0.009091 0.000188 0.773059 0.808062 2417.356539 0.999788
p__1 0.210017 0.009091 0.000188 0.191938 0.226941 2417.356539 0.999788
sigma__0 0.497322 0.009103 0.000186 0.480394 0.515867 2227.397854 0.999358
sigma__1 0.191310 0.006633 0.000141 0.178924 0.204859 2286.817037 0.999614
and the traces
I am trying to train an autoencoder NN (3 layers - 2 visible, 1 hidden) using numpy and scipy for the MNIST digits images dataset. The implementation is based on the notation given here Below is my code:
def autoencoder_cost_and_grad(theta, visible_size, hidden_size, lambda_, data):
"""
The input theta is a 1-dimensional array because scipy.optimize.minimize expects
the parameters being optimized to be a 1d array.
First convert theta from a 1d array to the (W1, W2, b1, b2)
matrix/vector format, so that this follows the notation convention of the
lecture notes and tutorial.
You must compute the:
cost : scalar representing the overall cost J(theta)
grad : array representing the corresponding gradient of each element of theta
"""
training_size = data.shape[1]
# unroll theta to get (W1,W2,b1,b2) #
W1 = theta[0:hidden_size*visible_size]
W1 = W1.reshape(hidden_size,visible_size)
W2 = theta[hidden_size*visible_size:2*hidden_size*visible_size]
W2 = W2.reshape(visible_size,hidden_size)
b1 = theta[2*hidden_size*visible_size:2*hidden_size*visible_size + hidden_size]
b2 = theta[2*hidden_size*visible_size + hidden_size: 2*hidden_size*visible_size + hidden_size + visible_size]
#feedforward pass
a_l1 = data
z_l2 = W1.dot(a_l1) + numpy.tile(b1,(training_size,1)).T
a_l2 = sigmoid(z_l2)
z_l3 = W2.dot(a_l2) + numpy.tile(b2,(training_size,1)).T
a_l3 = sigmoid(z_l3)
#backprop
delta_l3 = numpy.multiply(-(data-a_l3),numpy.multiply(a_l3,1-a_l3))
delta_l2 = numpy.multiply(W2.T.dot(delta_l3),
numpy.multiply(a_l2, 1 - a_l2))
b2_derivative = numpy.sum(delta_l3,axis=1)/training_size
b1_derivative = numpy.sum(delta_l2,axis=1)/training_size
W2_derivative = numpy.dot(delta_l3,a_l2.T)/training_size + lambda_*W2
#print(W2_derivative.shape)
W1_derivative = numpy.dot(delta_l2,a_l1.T)/training_size + lambda_*W1
W1_derivative = W1_derivative.reshape(hidden_size*visible_size)
W2_derivative = W2_derivative.reshape(visible_size*hidden_size)
b1_derivative = b1_derivative.reshape(hidden_size)
b2_derivative = b2_derivative.reshape(visible_size)
grad = numpy.concatenate((W1_derivative,W2_derivative,b1_derivative,b2_derivative))
cost = 0.5*numpy.sum((data-a_l3)**2)/training_size + 0.5*lambda_*(numpy.sum(W1**2) + numpy.sum(W2**2))
return cost,grad
I have also implemented a function to estimate the numerical gradient and verify the correctness of my implementation (below).
def compute_gradient_numerical_estimate(J, theta, epsilon=0.0001):
"""
:param J: a loss (cost) function that computes the real-valued loss given parameters and data
:param theta: array of parameters
:param epsilon: amount to vary each parameter in order to estimate
the gradient by numerical difference
:return: array of numerical gradient estimate
"""
gradient = numpy.zeros(theta.shape)
eps_vector = numpy.zeros(theta.shape)
for i in range(0,theta.size):
eps_vector[i] = epsilon
cost1,grad1 = J(theta+eps_vector)
cost2,grad2 = J(theta-eps_vector)
gradient[i] = (cost1 - cost2)/(2*epsilon)
eps_vector[i] = 0
return gradient
The norm of the difference between the numerical estimate and the one computed by the function is around 6.87165125021e-09 which seems to be acceptable. My main problem seems to be to get the gradient descent algorithm "L-BGFGS-B" working using the scipy.optimize.minimize function as below:
# theta is the 1-D array of(W1,W2,b1,b2)
J = lambda x: utils.autoencoder_cost_and_grad(theta, visible_size, hidden_size, lambda_, patches_train)
options_ = {'maxiter': 4000, 'disp': False}
result = scipy.optimize.minimize(J, theta, method='L-BFGS-B', jac=True, options=options_)
I get the below output from this:
scipy.optimize.minimize() details:
fun: 90.802022224079778
hess_inv: <16474x16474 LbfgsInvHessProduct with dtype=float64>
jac: array([ -6.83667742e-06, -2.74886002e-06, -3.23531941e-06, ...,
1.22425735e-01, 1.23425062e-01, 1.28091250e-01])
message: b'ABNORMAL_TERMINATION_IN_LNSRCH'
nfev: 21
nit: 0
status: 2
success: False
x: array([-0.06836677, -0.0274886 , -0.03235319, ..., 0. ,
0. , 0. ])
Now, this post seems to indicate that the error could mean that the gradient function implementation could be wrong? But my numerical gradient estimate seems to confirm that my implementation is correct. I have tried varying the initial weights by using a uniform distribution as specified here but the problem still persists. Is there anything wrong with my backprop implementation?
Turns out the issue was a syntax error (very silly) with this line:
J = lambda x: utils.autoencoder_cost_and_grad(theta, visible_size, hidden_size, lambda_, patches_train)
I don't even have the lambda parameter x in the function declaration. So the theta array wasn't even being passed whenever J was being invoked.
This fixed it:
J = lambda x: utils.autoencoder_cost_and_grad(x, visible_size, hidden_size, lambda_, patches_train)
I've implemented the Bayesian Probabilistic Matrix Factorization algorithm using pymc3 in Python. I also implemented it's precursor, Probabilistic Matrix Factorization (PMF). See my previous question for a reference to the data used here.
I'm having trouble drawing MCMC samples using the NUTS sampler. I initialize the model parameters using the MAP from PMF, and the hyperparameters using Gaussian random draws sprinkled around 0. However, I get a PositiveDefiniteError when setting up the step object for the sampler. I've verified that the MAP estimate from PMF is reasonable, so I expect it has something to do with the way the hyperparameters are being initialized. Here is the PMF model:
import pymc3 as pm
import numpy as np
import pandas as pd
import theano
import scipy as sp
data = pd.read_csv('jester-dense-subset-100x20.csv')
n, m = data.shape
test_size = m / 10
train_size = m - test_size
train = data.copy()
train.ix[:,train_size:] = np.nan # remove test set data
train[train.isnull()] = train.mean().mean() # mean value imputation
train = train.values
test = data.copy()
test.ix[:,:train_size] = np.nan # remove train set data
test = test.values
# Low precision reflects uncertainty; prevents overfitting
alpha_u = alpha_v = 1/np.var(train)
alpha = np.ones((n,m)) * 2 # fixed precision for likelihood function
dim = 10 # dimensionality
# Specify the model.
with pm.Model() as pmf:
pmf_U = pm.MvNormal('U', mu=0, tau=alpha_u * np.eye(dim),
shape=(n, dim), testval=np.random.randn(n, dim)*.01)
pmf_V = pm.MvNormal('V', mu=0, tau=alpha_v * np.eye(dim),
shape=(m, dim), testval=np.random.randn(m, dim)*.01)
pmf_R = pm.Normal('R', mu=theano.tensor.dot(pmf_U, pmf_V.T),
tau=alpha, observed=train)
# Find mode of posterior using optimization
start = pm.find_MAP(fmin=sp.optimize.fmin_powell)
And here is BPMF:
n, m = data.shape
dim = 10 # dimensionality
beta_0 = 1 # scaling factor for lambdas; unclear on its use
alpha = np.ones((n,m)) * 2 # fixed precision for likelihood function
logging.info('building the BPMF model')
std = .05 # how much noise to use for model initialization
with pm.Model() as bpmf:
# Specify user feature matrix
lambda_u = pm.Wishart(
'lambda_u', n=dim, V=np.eye(dim), shape=(dim, dim),
testval=np.random.randn(dim, dim) * std)
mu_u = pm.Normal(
'mu_u', mu=0, tau=beta_0 * lambda_u, shape=dim,
testval=np.random.randn(dim) * std)
U = pm.MvNormal(
'U', mu=mu_u, tau=lambda_u, shape=(n, dim),
testval=np.random.randn(n, dim) * std)
# Specify item feature matrix
lambda_v = pm.Wishart(
'lambda_v', n=dim, V=np.eye(dim), shape=(dim, dim),
testval=np.random.randn(dim, dim) * std)
mu_v = pm.Normal(
'mu_v', mu=0, tau=beta_0 * lambda_v, shape=dim,
testval=np.random.randn(dim) * std)
V = pm.MvNormal(
'V', mu=mu_v, tau=lambda_v, shape=(m, dim),
testval=np.random.randn(m, dim) * std)
# Specify rating likelihood function
R = pm.Normal(
'R', mu=theano.tensor.dot(U, V.T), tau=alpha,
observed=train)
# `start` is the start dictionary obtained from running find_MAP for PMF.
for key in bpmf.test_point:
if key not in start:
start[key] = bpmf.test_point[key]
with bpmf:
step = pm.NUTS(scaling=start)
At the last line, I get the following error:
PositiveDefiniteError: Scaling is not positive definite. Simple check failed. Diagonal contains negatives. Check indexes [ 0 2 ... 2206 2207 ]
As I understand it, I can't use find_MAP with models that have hyperpriors like BPMF. This is why I'm attempting to initialize with the MAP values from PMF, which uses point estimates for the parameters on U and V rather than parameterized hyperpriors.
Unfortunately the Wishart distribution is non-functional. I recently added a warning here: https://github.com/pymc-devs/pymc3/commit/642f63973ec9f807fb6e55a0fc4b31bdfa1f261e
See here for more discussions on this tricky distribution: https://github.com/pymc-devs/pymc3/issues/538
You could confirm that that's the source by fixing the covariance matrix. If that's the case, I'd try using the JKL prior distribution: https://github.com/pymc-devs/pymc3/blob/master/pymc3/examples/LKJ_correlation.py