Predictions with PYMC, Dont work on test data set

1

Hey I am trying to set up a model, but I can not predict with it. The discussions before did not really helped me. I just wanna draw some samples from my posterior distribution. pm.set_data sets up my data correctly, but the posterior_predictive function works just on my training set. I briefed all possible mistakes, there are no invalid groups, both columns from my test dataset having the same length, but It wont work. Please help me It drives me crazy.

import pandas as pd
import numpy as np
import pymc3 as pm

# 
dtype_dict = {
    'FEATURE_1': 'float',  
    'CATEGORY': 'category',  
    'COUNT': 'int',  
    'TIME': 'int'  
}
usecols_list = list(dtype_dict.keys())
data = pd.read_csv(
    'path/to/your/data.csv',  
    sep=';',  
    header=0,
    dtype=dtype_dict,
    usecols=usecols_list
)

# Filter 
data = data[data['FEATURE_1'].isin([0, 1]) & data['FEATURE_1'].notna()]
data_prior_analysis = data[(data['TIME'] >= 1994) & (data['TIME'] < 1996)]
data_train = data[(data['TIME'] > =1996) & (data['TIME'] <= 1998)]
data_test = data[data['TIME'] > 1999]

# function Prior knowledge
def observing_prior_knowledge(df, count_col, category_col, feature_col):
    epsilon = 1e-8  

    selected_data = df[[count_col, category_col, feature_col]]
    category_stats = selected_data.groupby(category_col)[count_col].agg(['mean', 'median', 'std'])
    category_stats['log_mean'] = np.log(category_stats['mean'] + epsilon)
    category_stats['log_median'] = np.log(category_stats['median'] + epsilon)

    feature_stats = selected_data[feature_col].agg(['mean', 'median', 'std'])
    feature_stats['log_mean'] = np.log(feature_stats['mean'] + epsilon)
    feature_stats['log_median'] = np.log(feature_stats['median'] + epsilon)

    feature_df = pd.DataFrame(feature_stats).transpose()
    feature_df.index = ['FEATURE_STATS']

    result_df = pd.concat([category_stats, feature_df])
    return result_df

# prior knowledge
data_prior = observing_prior_knowledge(data_prior_analysis, 'COUNT', 'CATEGORY', 'FEATURE_1')

# preparing data train
cat_idx, groups = pd.factorize(data_train['CATEGORY'], sort=True)
feature_data = data_train['FEATURE_1'].values
counts = data_train['COUNT'].values

# Priors
group_means = data_prior.loc[groups, 'mean'].fillna(0).values
group_stds = data_prior.loc[groups, 'std'].fillna(1).values
beta_cont_mean = data_prior.loc['FEATURE_STATS', 'mean']
beta_cont_std = data_prior.loc['FEATURE_STATS', 'std']

# Model
coords = {"group": groups, "feature": feature_data}

with pm.Model(coords=coords) as frequency_model:
    category_data = pm.Data("CATEGORY", cat_idx, mutable=True)
    feature_data = pm.Data("FEATURE_DATA", feature_data, mutable=True)

    # Priors
    Intercept = pm.Normal("Intercept", mu=0, sigma=1)
    beta_cat = pm.Normal("beta_cat", mu=group_means, sigma=group_stds, dims="group")
    beta_cont = pm.Normal("beta_cont", mu=beta_cont_mean, sigma=beta_cont_std)

    # Model
    lambda_freq = pm.math.exp(Intercept + beta_cat[category_data] + beta_cont * feature_data)

    # Likelihood
    counts_obs = pm.Poisson("counts", mu=lambda_freq, observed=counts)

    # Sampling
    trace = pm.sample(1000, chains=4, tune=1000, return_inferencedata=True)
    posterior_predictive = pm.sample_posterior_predictive(trace, extend_inferencedata=True)

Here starts the Problem. It sets up my data correctly but it does not sample from the posterior distrubtion. I just run into the same error again and again.

# Test data
feature_data_test = data_test[data_test['TIME'] == 1999]['FEATURE_1'].values
category_test = data_test[data_test['TIME'] == 1999]['CATEGORY'].values

# 
cat_idx_test = pd.Categorical(category_test, categories=groups).codes

with frequency_model:
    pm.set_data({"CATEGORY": cat_idx_test, "FEATURE_DATA": feature_data_test})
    
    # Vorhersagen generieren
    posterior_predictive_test = pm.sample_posterior_predictive(trace, extend_inferencedata=True)

This is the error I recive:
File: c:\...\site-packages\pytensor\tensor\random\op.py, line 378, in RandomVariable.perform
    smpl_val = self.rng_fn(rng, *args + [size])

ValueError: Inputs values: [Generator(PCG64) at 0x253519E9D60, array([141352], dtype=int64), array([4], dtype=int64), 'not shown']

Outputs clients: [['output_1'], ['output_1']]```
2

Use triple backticks to format your code blocks ```

You should specify the shape of you observed variable explicitly, something like shape=lamba_freq.shape, otherwise it defaults to the shape of the observations and is invalid if you change the shape of mu before doing predictions

1 Like
3

I implemented your advise and it worked.
I also have the problem that the ppc_plot no longer works because of a too large array. Have you ever faced a similar issue?

4

Can you share the error message?