Hi!
Appreciate all help that contributes to solving this.
UPDATE: Still a problem though! Found the solution which practically feels like a bug. I am required to have the same input dimensions as the training data which is 1000 samples. Even if my test data is only 3 samples or 100 samples. The hold-out dimensions must be strictly like the training. No less, no more. Why?
TLDR; Input dimension mismatch for new data
I’ve come a far with PyMC and by gradually increasing the complexity of each iteration of model building up until this point.
To clarify:
pm.sample_numpyro_nuts()workspm.sample_posterior_predictive()works in isolationpm.sample_posterior_predictive()doesn’t work with new/unseen/hold-out data
The structure of my model is Global → Second Level → Third level → Output. See image.
Here is the code. Good to know is that I have a total of 1000 samples trained on. Excuse the different variable names as this is an artifact from the iterative model building. It has no effect on the sampling.
coords dictionary
coords = {
'loc': df.index.values,
'branch': np.unique(branch_code),
'city': np.unique(city_code),
'productline': np.unique(productline_code),
'branch_city': np.arange(len(unique_combo_branch_city)),
'city_productline': np.arange(len(unique_combo_city_productline))
}
Model below
with pm.Model(coords=coords) as multilevel_model_ts3:
# Indices
branch_cty_idx = pm.MutableData("branch_idx", branch_city_idx, dims="loc")
city_productline_idx = pm.MutableData("city_idx", city_product_idx, dims="loc")
y_mut = pm.MutableData("y_mut", y, dims="loc")
t_mut = pm.MutableData("t_mut", t, dims="loc")
tax_mut = pm.MutableData("tax_mut", tax, dims="loc")
ewallet_mut = pm.MutableData("ewallet_mut", ewallet_code, dims="loc")
cash_mut = pm.MutableData("cash_mut", cash_code, dims="loc")
# Global hyperparameters
g_mu = pm.Normal("g_mu", mu=0, sigma=0.01)
g_sd = pm.HalfNormal("g_sd", sigma=0.01)
# Global intercept
a = pm.Normal("alpha", mu=15, sigma=.001)
# Branch level intercept
branch_intercept = pm.Normal("branch_intercept", mu=15, sigma=0.1, dims="branch")
# Country-level parameters
mu_country = pm.Normal("mu_branch", mu=g_mu, sigma=g_sd, dims="branch")
sigma_country = pm.HalfNormal("sigma_branch", sigma=0.1, dims="branch")
# City level intercept
city_intercept = pm.Normal("city_intercept", mu=15, sigma=.1, dims="branch_city")
# City-level parameters
mu_city = pm.Normal("mu_city",
mu=mu_country[unique_combo_branch_city[:, 0]],
sigma=sigma_country[unique_combo_branch_city[:, 0]],
dims="branch_city")
sigma_city = pm.HalfNormal("sigma_city", sigma=1, dims="branch_city")
# City-Article level parameters
mu_city_article = pm.Normal("mu_city_article",
mu=mu_city[city_product_to_branch],
sigma=sigma_city[city_product_to_branch],
dims="city_productline")
# Time effect at branch-city level
beta_time_branch_city = pm.Normal("beta_time_branch_city", mu=0, sigma=.01, dims="branch_city")
# Time effect
beta_time_city_productline = pm.Normal("beta_time_city_productline", mu=0, sigma=5, dims="city_productline")
# Tax 5% predictor
kappa_tax_branch_city = pm.Normal("kappa_tax_branch_city", mu=0, sigma=2, dims="branch_city")
kappa_tax_city_productline = pm.Normal("kappa_tax_city_productline", mu=kappa_tax_branch_city[city_product_to_branch], sigma=2, dims="city_productline")
# Payment predictor
ewallet_payment_branch_city = pm.Normal("ewallet_payment_branch_city", mu=0, sigma=5, dims="branch_city")
cash_payment_branch_city = pm.Normal("cash_payment_branch_city", mu=0, sigma=5, dims="branch_city")
intercepts = a + branch_intercept[branch_cty_idx] + city_intercept[branch_cty_idx] + mu_city_article[city_productline_idx]
# Expected value
mu = pm.Deterministic("mu",
intercepts +
beta_time_branch_city[branch_cty_idx] * t_mut +
beta_time_city_productline[city_productline_idx] * t_mut +
kappa_tax_branch_city[branch_cty_idx] * tax_mut +
kappa_tax_city_productline[city_productline_idx] * tax_mut +
ewallet_payment_branch_city[branch_cty_idx] * ewallet_mut +
cash_payment_branch_city[branch_cty_idx] * cash_mut,
dims="loc")
# Observation noise
sigma = pm.HalfNormal("sigma", sigma=1)
# Likelihood
obs = pm.LogNormal("obs", mu=mu, sigma=sigma, observed=y_mut, dims="loc")
Sampling below
with multilevel_model_ts3:
trace_ts2 = sample_numpyro_nuts(draws=1000, tune=1000, target_accept=0.95)
Posterior predictive below
with multilevel_model_ts3:
posterior_predictive_ts2 = pm.sample_posterior_predictive(trace_ts2, var_names=["obs"], random_seed=1990)
Up until this part everything works smoothly. Then when executing below an input mismatch error arises.
with multilevel_model_ts3:
pm.set_data({"y_mut": [0,0,0],
"t_mut": [0, .3, .5],
"tax_mut": [0, .3, .5],
"ewallet_mut": [0, 1, 0],
"cash_mut": [0, 0, 1]},
coords={"loc": [1000, 1001, 1002],
"branch": [0, 0, 1],
"branch_city": [0, 0, 1],
"city": [0, 1, 0],
"productline": [3, 3, 5],
"city_productline": [10, 11, 15]})
pp = pm.sample_posterior_predictive(trace_ts2, var_names=["obs"], random_seed=1990 )
The error
ValueError: Input dimension mismatch: (input[%i].shape[%i] = %lld, input[%i].shape[%i] = %lld)
Apply node that caused the error: Composite{(i2 + i3 + i4 + i5 + (i13 * i12) + (i11 * i12) + (i10 * i9) + (i8 * i9) + (i6 * i7) + (i0 * i1))}(AdvancedSubtensor1.0, cash_mut, ExpandDims{axis=0}.0, AdvancedSubtensor1.0, AdvancedSubtensor1.0, AdvancedSubtensor1.0, AdvancedSubtensor1.0, ewallet_mut, AdvancedSubtensor1.0, tax_mut, AdvancedSubtensor1.0, AdvancedSubtensor1.0, t_mut, AdvancedSubtensor1.0)
Toposort index: 11
Inputs types: [TensorType(float64, shape=(None,)), TensorType(int32, shape=(None,)), TensorType(float64, shape=(1,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(int32, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,)), TensorType(float64, shape=(None,))]
Inputs shapes: [(1000,), (3,), (1,), (1000,), (1000,), (1000,), (1000,), (3,), (1000,), (3,), (1000,), (1000,), (3,), (1000,)]
Inputs strides: [(8,), (4,), (8,), (8,), (8,), (8,), (8,), (4,), (8,), (8,), (8,), (8,), (8,), (8,)]
Inputs values: ['not shown', array([0, 0, 1]), array([14.99939276]), 'not shown', 'not shown', 'not shown', 'not shown', array([0, 1, 0]), 'not shown', array([0. , 0.3, 0.5]), 'not shown', 'not shown', array([0. , 0.3, 0.5]), 'not shown']
Outputs clients: [[lognormal_rv{"(),()->()"}(RNG(<Generator(PCG64) at 0x2AB45632C00>), MakeVector{dtype='int64'}.0, mu, ExpandDims{axis=0}.0)]]
I have followed pymc.model.core.set_data — PyMC dev documentation and General API quickstart — PyMC example gallery but I get the mismatch error.