TypeError in model with shape(16,8,2)

Questions
1

Hi,
I’m trying to model an experiment where participant have to rate the age of different avatar.
Data from the experiment has the shape (16,8,2). I’m encountering a type error using a bernouilli distribution with shape(1, 8, 1) I get the error :
TypeError: Cannot convert Type TensorType(int64, 3D) (of Variable is_troll_shared__) into Type TensorType(int64, (True, False, True)). You can try to manually convert is_troll_shared__ into a TensorType(int64, (True, False, True)).

import pymc3 as pm
import numpy as np

# A study on avatar perceived age. Participants have to rate the age of each avatar twice

# 16 avatars to rate
# 8 participants
# Each participant rate the same avatar twice (2)
num_avatar = 16
num_participant = 8
exposure_time = 2


age_obs_data = np.random.randint(0, 100, size=(num_avatar, num_participant, exposure_time)) # fake data

with pm.Model() as model3:
    # Age
    ## priors
    μ_avatars_age = pm.Uniform('μ_avatars_age', 0, 100, shape=(num_avatar, 1, 1))
    σ_avatars_age = pm.HalfCauchy('σ_avatars_age', 25, shape=(num_avatar, 1, 1))
    
    
    μ_participant_bias = pm.Normal('μ_participant_bias', 0, 10, shape=(1, num_participant, 1))
    
    # Troll detection: they rate the age of the avatar at random instead of trying to identify the true age
    μ_troll  = pm.Uniform('μ_troll', 0, 100)
    σ_troll = pm.Uniform('σ_troll', 0, 10)
    p_troll  = pm.Uniform('p_troll', 0, 1, shape=(1, num_participant, 1)) 
    is_troll = pm.Bernoulli('is_troll', p=p_troll, shape=(1, num_participant, 1))
    
    μ = pm.math.switch(is_troll > 0, μ_troll, μ_avatars_age + μ_participant_bias)
    σ = pm.math.switch(is_troll > 0, σ_troll, σ_avatars_age)
    
    obs_age = pm.Normal('obs_age', 
                        mu=μ_avatars_age + μ_participant_bias,
                        sd=σ_avatars_age,
                        shape=(num_avatar, num_participant, 1),
                        observed=age_obs_data,
                       )
    

    trace = pm.sample(draws=2000,)
2

Mixed effect model or hierarchical model is much easier to handle if you format your data into a long data table format. You can have a look at the doc eg http://docs.pymc.io/notebooks/GLM-hierarchical.html, @Jack_Caster’s repository https://github.com/JackCaster/GLM_with_PyMC3, my repository https://github.com/junpenglao/GLMM-in-Python as examples.

3

I’m observing a big drop in sampling performance using long form data, a factor of 10X.

Original code 150 it/s

import pymc3 as pm
import numpy as np

# A study on avatar perceived age. Participants have to rate the age of each avatar twice

# 16 avatars to rate
# 8 participants
# Each participant rate the same avatar twice (2)
num_avatar = 16
num_participant = 100
exposure_time = 2


age_obs_data = np.random.randint(0, 100, size=(num_avatar, num_participant, exposure_time)) # fake data

with pm.Model() as model1:
    # Age
    ## priors
    μ_avatars_age = pm.Uniform('μ_avatars_age', 0, 100, shape=(num_avatar, 1, 1))
    σ_avatars_age = pm.HalfCauchy('σ_avatars_age', 25, shape=(num_avatar, 1, 1))
    
    
    μ_participant_bias = pm.Normal('μ_participant_bias', 0, 10, shape=(1, num_participant, 1))
    
    
    obs_age = pm.Normal('obs_age', 
                        mu=μ_avatars_age + μ_participant_bias,
                        sd=σ_avatars_age,
                        shape=(num_avatar, num_participant, 1),
                        observed=age_obs_data,
                       )
    

    trace = pm.sample(draws=2000, cores=1)

long form code 20 it/s

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

# A study on avatar perceived age. Participants have to rate the age of each avatar twice

# 16 avatars to rate
# 8 participants
# Each participant rate the same avatar twice (2)
num_avatar = 16
num_participant = 100
exposure_time = 2


participant = np.repeat(np.arange(num_participant), num_avatar*exposure_time)
avatar = np.repeat(np.arange(num_avatar), num_participant*exposure_time)
data = np.random.randint(0, 100, size=(num_avatar*num_participant*exposure_time))


df = pd.DataFrame([participant, avatar, data], index=['participant', 'avatar', 'data']).T
df.participant = df.participant.astype('category')
df.avatar = df.avatar.astype('category')

participant_dm = np.asarray(patsy.dmatrix('0 + participant', data=df))
avatar_dm = np.asarray(patsy.dmatrix('0 + avatar', data=df))

with pm.Model() as model2:
    # Age
    ## priors
    μ_avatars_age = pm.math.dot(avatar_dm, pm.Uniform('μ_avatars_age', 0, 100, shape=num_avatar))
    σ_avatars_age = pm.math.dot(avatar_dm, pm.HalfCauchy('σ_avatars_age', 25, shape=num_avatar))
    
    μ_participant = pm.math.dot(participant_dm, pm.Normal('μ_participant', mu=0, sd=5, shape=num_participant))
        
    obs_age = pm.Normal('obs_age',
                        mu=μ_avatars_age + μ_participant,
                        sd=σ_avatars_age,
                        observed=df.data.values,
                        shape=len(df.data.values)
                       )
    
    trace = pm.sample(draws=2000, cores=1)

Is it normal or am I doing something wrong?