def run_inference(inference_id):
inference = Inference.objects.get(id=inference_id)
print('task', inference.number_of_iterations, inference_id)
for ll in inference.log_likelihoods.all():
print('LL', ll.id, ll.name, ll.biomarker_type)
# create the mixin object to run the inference
inference_mixin = InferenceMixin(inference)
inference_mixin.run_inference()
# remove task id to indicate that we're finished
inference.task_id = None
inference.save()
def test_inference_results(self):
# create mixin object
inference_mixin = InferenceMixin(self.inference)
inference_mixin.run_inference()
chain = self.inference.chains.first()
chain_serializer = InferenceChainSerializer(chain)
data = chain_serializer.data
self.assertTrue('outputs' in data)
self.assertTrue(len(data['outputs']) > 0)
self.assertTrue('data' in data)
self.assertTrue('kde' in data['data'])
self.assertTrue(len(data['data']['kde']) > 0)
self.assertTrue('chain' in data['data'])
self.assertTrue(len(data['data']['chain']) > 0)
class TestInferenceMixinSingleOutputOptimisationCovariate(TestCase):
def setUp(self):
# ensure we've got nothing in the cache
cache._cache.flush_all()
self.project = Project.objects.get(
name='demo',
)
biomarker_type = BiomarkerType.objects.get(
name='Tumour volume',
dataset__name='lxf_control_growth'
)
model = PharmacodynamicModel.objects.get(
name='tumour_growth_inhibition_model_koch',
read_only=False,
)
self.inference = Inference.objects.create(
name='bob',
project=self.project,
max_number_of_iterations=10,
algorithm=Algorithm.objects.get(name='XNES'),
number_of_chains=3,
)
log_likelihood = LogLikelihood.objects.create(
variable=model.variables.first(),
inference=self.inference,
form=LogLikelihood.Form.MODEL
)
# remove all outputs except
output_names = [
'myokit.tumour_volume',
]
outputs = []
for output in log_likelihood.outputs.all():
if output.variable.qname in output_names:
output.parent.biomarker_type = biomarker_type
output.parent.observed = True
output.parent.save()
outputs.append(output.parent)
else:
for param in output.parent.parameters.all():
if param != output:
param.child.delete()
output.parent.delete()
self.outputs = outputs
# set uniform prior on everything, except
# we'll do a population prior on the first one
for i, param in enumerate(log_likelihood.parameters.all()):
if i == 0:
# first param is sampled from a normal distribution with a mean
# derived from subject body weight
body_weight_biomarker_type = BiomarkerType.objects.get(
name='Body weight',
dataset__name='lxf_control_growth'
)
param.child.biomarker_type = body_weight_biomarker_type
param.child.time_independent_data = True
param.child.form = LogLikelihood.Form.NORMAL
param.child.save()
body_weight_values, _, subjects = param.child.get_data()
param.length = len(subjects)
param.save()
# use a covariate to adjust the mean of the normal according to
# body weight
mean, sigma = param.child.get_noise_log_likelihoods()
mean.form = LogLikelihood.Form.EQUATION
mean.description = '1.0 if arg0 < 20 else 2.0'
mean.biomarker_type = body_weight_biomarker_type
mean.time_independent_data = True
mean.save()
body_weight = LogLikelihood.objects.create(
name='Body weight',
inference=self.inference,
form=LogLikelihood.Form.FIXED,
biomarker_type=body_weight_biomarker_type,
time_independent_data=True,
)
LogLikelihoodParameter.objects.create(
name='Body weight',
parent=mean,
child=body_weight,
parent_index=0,
length=len(subjects)
)
sigma.value = 0.01
sigma.save()
else:
param.set_uniform_prior(0.0, 2.0)
# 'run' inference to create copies of models
self.inference.run_inference(test=True)
# create mixin object
self.inference_mixin = InferenceMixin(self.inference)
def test_inference_runs(self):
# tests that inference runs and writes results to db
self.inference_mixin.run_inference()
chains = self.inference_mixin.inference.chains.all()
self.assertEqual(len(chains), 3)
priors = self.inference_mixin._priors
for chain in chains:
fun_res = chain.inference_function_results
f_vals = fun_res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(f_vals), 11)
for prior in priors:
res = chain.inference_results.filter(log_likelihood=prior)
p_vals = res.order_by('iteration').values_list(
'value', flat=True
)
length = prior.get_total_length()
self.assertEqual(len(p_vals), 11 * length)
# don't test for inference log_posterior(param) = fn since they won't
# because of the way the 'best' params are picked
inference = self.inference_mixin.inference
self.assertTrue(inference.time_elapsed > 0)
self.assertTrue(inference.number_of_function_evals > 0)
class TestInferenceMixinSingleOutputOptimisationPopulation(TestCase):
def setUp(self):
# ensure we've got nothing in the cache
cache._cache.flush_all()
self.project = Project.objects.get(
name='demo',
)
biomarker_type = BiomarkerType.objects.get(
name='Tumour volume',
dataset__name='lxf_control_growth'
)
model = PharmacodynamicModel.objects.get(
name='tumour_growth_inhibition_model_koch',
read_only=False,
)
self.inference = Inference.objects.create(
name='bob',
project=self.project,
max_number_of_iterations=10,
algorithm=Algorithm.objects.get(name='XNES'),
number_of_chains=3,
)
log_likelihood = LogLikelihood.objects.create(
variable=model.variables.first(),
inference=self.inference,
form=LogLikelihood.Form.MODEL
)
# remove all outputs except
output_names = [
'myokit.tumour_volume',
]
outputs = []
for output in log_likelihood.outputs.all():
if output.variable.qname in output_names:
output.parent.biomarker_type = biomarker_type
output.parent.observed = True
output.parent.save()
outputs.append(output.parent)
else:
for param in output.parent.parameters.all():
if param != output:
param.child.delete()
output.parent.delete()
self.outputs = outputs
# set uniform prior on everything, except
# we'll do a population prior on the first one
for i, param in enumerate(log_likelihood.parameters.all()):
if i == 0:
param.set_uniform_prior(
0.0, 2.0, biomarker_type=biomarker_type
)
else:
param.set_uniform_prior(0.0, 2.0)
# 'run' inference to create copies of models
self.inference.run_inference(test=True)
# create mixin object
self.inference_mixin = InferenceMixin(self.inference)
def test_inference_runs(self):
# tests that inference runs and writes results to db
self.inference_mixin.run_inference()
chains = self.inference_mixin.inference.chains.all()
self.assertEqual(len(chains), 3)
priors = self.inference_mixin._priors
for chain in chains:
fun_res = chain.inference_function_results
f_vals = fun_res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(f_vals), 11)
for prior in priors:
res = chain.inference_results.filter(log_likelihood=prior)
p_vals = res.order_by('iteration').values_list(
'value', flat=True
)
length = prior.get_total_length()
self.assertEqual(len(p_vals), 11 * length)
# don't test for inference log_posterior(param) = fn since they won't
# because of the way the 'best' params are picked
inference = self.inference_mixin.inference
self.assertTrue(inference.time_elapsed > 0)
self.assertTrue(inference.number_of_function_evals > 0)
def test_inference_can_be_restarted(self):
self.inference.chains.all().delete()
self.inference.chains.set([
InferenceChain.objects.create(inference=self.inference)
for i in range(self.inference.number_of_chains)
])
self.inference.refresh_from_db()
for chain in self.inference.chains.all():
chain.inference_function_results.add(
InferenceFunctionResult.objects.create(
chain=chain,
iteration=0,
value=1.4
),
InferenceFunctionResult.objects.create(
chain=chain,
iteration=1,
value=2.4
)
)
for prior in self.inference_mixin._priors:
chain.inference_results.add(
InferenceResult.objects.create(
chain=chain,
log_likelihood=prior,
iteration=0,
value=0.5,
),
InferenceResult.objects.create(
chain=chain,
log_likelihood=prior,
iteration=1,
value=0.4,
)
)
chain.refresh_from_db()
self.inference.number_of_iterations = 1
self.inference.save()
# tests that inference runs and writes results to db
inference_mixin = InferenceMixin(self.inference)
inference_mixin.run_inference()
chains = inference_mixin.inference.chains.all()
self.assertEqual(len(chains), 3)
for chain in chains:
priors = inference_mixin._priors
fun_res = chain.inference_function_results
f_vals = fun_res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(f_vals), 11)
p_vals_all = []
for prior in priors:
res = chain.inference_results.filter(log_likelihood=prior)
p_vals = res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(p_vals), 11)
p_vals_all.append(p_vals)
iterations = res.order_by('iteration').values_list(
'iteration', flat=True
)
expected = list(range(11))
self.assertTrue(np.array_equal(iterations, expected))
class TestInferenceMixinSingleOutputOptimisation(TestCase):
def setUp(self):
# ensure we've got nothing in the cache
cache._cache.flush_all()
self.project = Project.objects.get(
name='demo',
)
biomarker_type = BiomarkerType.objects.get(
name='Tumour volume',
dataset__name='lxf_control_growth'
)
model = PharmacodynamicModel.objects.get(
name='tumour_growth_inhibition_model_koch',
read_only=False,
)
self.inference = Inference.objects.create(
name='bob',
project=self.project,
max_number_of_iterations=10,
algorithm=Algorithm.objects.get(name='XNES'),
number_of_chains=3,
)
log_likelihood = LogLikelihood.objects.create(
variable=model.variables.first(),
inference=self.inference,
form=LogLikelihood.Form.MODEL
)
# remove all outputs except
output_names = [
'myokit.tumour_volume',
]
outputs = []
for output in log_likelihood.outputs.all():
if output.variable.qname in output_names:
output.parent.biomarker_type = biomarker_type
output.parent.observed = True
output.parent.save()
outputs.append(output.parent)
else:
for param in output.parent.parameters.all():
if param != output:
param.child.delete()
output.parent.delete()
self.outputs = outputs
# set uniform prior on everything, except amounts
for param in log_likelihood.parameters.all():
param.set_uniform_prior(0.0, 2.0)
# 'run' inference to create copies of models
self.inference.run_inference(test=True)
# create mixin object
self.inference_mixin = InferenceMixin(self.inference)
def test_inference_runs(self):
# tests that inference runs and writes results to db
self.inference_mixin.run_inference()
chains = self.inference_mixin.inference.chains.all()
self.assertEqual(len(chains), 3)
priors = self.inference_mixin._priors
for chain in chains:
fun_res = chain.inference_function_results
f_vals = fun_res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(f_vals), 11)
p_vals_all = []
for prior in priors:
res = chain.inference_results.filter(log_likelihood=prior)
p_vals = res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(p_vals), 11)
p_vals_all.append(p_vals)
iterations = res.order_by('iteration').values_list(
'iteration', flat=True
)
expected = list(range(11))
self.assertTrue(np.array_equal(iterations, expected))
# don't test for inference log_posterior(param) = fn since they won't
# because of the way the 'best' params are picked
inference = self.inference_mixin.inference
self.assertTrue(inference.time_elapsed > 0)
self.assertTrue(inference.number_of_function_evals > 0)
def test_inference_can_be_restarted(self):
self.inference.chains.all().delete()
self.inference.chains.set([
InferenceChain.objects.create(inference=self.inference)
for i in range(self.inference.number_of_chains)
])
self.inference.refresh_from_db()
for chain in self.inference.chains.all():
chain.inference_function_results.add(
InferenceFunctionResult.objects.create(
chain=chain,
iteration=0,
value=1.4
),
InferenceFunctionResult.objects.create(
chain=chain,
iteration=1,
value=2.4
)
)
for prior in self.inference_mixin._priors:
chain.inference_results.add(
InferenceResult.objects.create(
chain=chain,
log_likelihood=prior,
iteration=0,
value=0.5,
),
InferenceResult.objects.create(
chain=chain,
log_likelihood=prior,
iteration=1,
value=0.4,
)
)
chain.refresh_from_db()
self.inference.number_of_iterations = 1
self.inference.save()
# tests that inference runs and writes results to db
inference_mixin = InferenceMixin(self.inference)
inference_mixin.run_inference()
chains = inference_mixin.inference.chains.all()
self.assertEqual(len(chains), 3)
for chain in chains:
priors = inference_mixin._priors
fun_res = chain.inference_function_results
f_vals = fun_res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(f_vals), 11)
p_vals_all = []
for prior in priors:
res = chain.inference_results.filter(log_likelihood=prior)
p_vals = res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(p_vals), 11)
p_vals_all.append(p_vals)
iterations = res.order_by('iteration').values_list(
'iteration', flat=True
)
expected = list(range(11))
self.assertTrue(np.array_equal(iterations, expected))
class TestInferenceMixinSingleOutputSampling(TestCase):
def setUp(self):
# ensure we've got nothing in the cache
cache._cache.flush_all()
project = Project.objects.get(
name='demo',
)
biomarker_type = BiomarkerType.objects.get(
name='Tumour volume',
dataset__name='lxf_control_growth'
)
model = PharmacodynamicModel.objects.get(
name='tumour_growth_inhibition_model_koch',
read_only=False,
)
self.inference = Inference.objects.create(
name='bob',
project=project,
max_number_of_iterations=10,
algorithm=Algorithm.objects.get(name='Haario-Bardenet'),
)
log_likelihood = LogLikelihood.objects.create(
variable=model.variables.first(),
inference=self.inference,
form=LogLikelihood.Form.MODEL
)
# remove all outputs except
output_names = [
'myokit.tumour_volume',
]
outputs = []
for output in log_likelihood.outputs.all():
if output.variable.qname in output_names:
output.parent.biomarker_type = biomarker_type
output.parent.observed = True
output.parent.save()
outputs.append(output.parent)
else:
for param in output.parent.parameters.all():
if param != output:
param.child.delete()
output.parent.delete()
# set uniform prior on everything, except amounts
for param in log_likelihood.parameters.all():
param.set_uniform_prior(0.0, 2.0)
# 'run' inference to create copies of models
self.inference.run_inference(test=True)
# create mixin object
self.inference_mixin = InferenceMixin(self.inference)
def test_objective_functions(self):
# Test log-posterior
log_posterior = self.inference_mixin._pints_log_posterior
log_posterior(
log_posterior.to_search([1.3, 0.5, 1.1, 0.9, 1.2, 1])
)
def test_inference_runs(self):
# tests that inference runs and writes results to db
self.inference_mixin.run_inference()
chains = self.inference_mixin.inference.chains.all()
self.assertEqual(len(chains), 4)
for chain in chains:
priors = self.inference_mixin._priors
fun_res = chain.inference_function_results
f_vals = fun_res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(f_vals), 11)
p_vals_all = []
for prior in priors:
res = chain.inference_results.filter(log_likelihood=prior)
p_vals = res.order_by('iteration').values_list(
'value', flat=True
)
self.assertEqual(len(p_vals), 11)
p_vals_all.append(p_vals)
iterations = res.order_by('iteration').values_list(
'iteration', flat=True
)
expected = list(range(11))
self.assertTrue(np.array_equal(iterations, expected))
# transpose list of lists
p_vals_all = list(map(list, zip(*p_vals_all)))
fn = self.inference_mixin._pints_log_posterior
for idx, params in enumerate(p_vals_all):
self.assertTrue(
abs(fn(fn.to_search(params)) - f_vals[idx]) < 0.01
)
output_res = chain.inference_output_results.all()
self.assertTrue(len(output_res), 28)
inference = self.inference_mixin.inference
self.assertTrue(inference.time_elapsed > 0)
self.assertTrue(inference.number_of_function_evals > 0)
def test_usecase2(self):
def faux_test_file(url, ending='.csv'):
tempname, _ = urlretrieve(url)
file = File(open(tempname, 'rb'))
file.name = 'test' + ending
return file
project = Project.objects.get(name='demo', )
# upload dataset
BASE_URL_DATASETS = 'https://raw.githubusercontent.com/pkpdapp-team/pkpdapp-datafiles/main/usecase2/' # noqa: E501
file = faux_test_file(BASE_URL_DATASETS + 'PKPD_UseCase_Abx.csv',
'.csv')
dataset = Dataset.objects.create(
name='usecase2',
project=project,
)
response = self.client.put(
'/api/dataset/{}/csv/'.format(dataset.id),
data={'csv': file},
)
self.assertEquals(response.status_code, status.HTTP_200_OK)
dataset.refresh_from_db()
for bt in dataset.biomarker_types.all():
print('have bt', bt.name, bt.display_unit.symbol,
bt.display_time_unit.symbol)
# upload pd model
with urllib.request.urlopen(BASE_URL_DATASETS + 'ABx_updated3.xml',
timeout=5) as f:
sbml_string = codecs.decode(f.read(), 'utf-8')
pd_model = PharmacodynamicModel.objects.create(
name='usecase2-pd',
project=project,
)
response = self.client.put(
'/api/pharmacodynamic/{}/sbml/'.format(pd_model.id),
data={'sbml': sbml_string},
)
self.assertEquals(response.status_code, status.HTTP_200_OK)
pd_model.refresh_from_db()
# create pkpd model
pk_model = PharmacokineticModel.objects\
.get(name='three_compartment_pk_model')
protocol = Protocol.objects.get(
subjects__dataset=dataset,
subjects__id_in_dataset=1,
)
pkpd_model = DosedPharmacokineticModel.objects.create(
name='usecase2',
pk_model=pk_model,
pd_model=pd_model,
dose_compartment='central',
protocol=protocol,
)
PkpdMapping.objects.create(
pkpd_model=pkpd_model,
pk_variable=pkpd_model.variables.get(
qname='myokit.scaled_drug_c_concentration', ),
pd_variable=pkpd_model.variables.get(
qname='PD.Drug_concentration', ),
)
pkpd_variables = {
'dose.absorption_rate': (0.6, 6.0, 60.0),
'myokit.clearance': (0.003, 0.03, 0.3),
'central.size': (0.0024, 0.024, 0.24),
'myokit.k_peripheral1': (0.0001, 0.001, 0.01),
'peripheral_1.size': (0.001, 0.01, 0.1),
'myokit.k_peripheral2': (0.000003, 0.00003, 0.0003),
'peripheral_2.size': (0.00007, 0.0007, 0.007),
'myokit.drug_c_scale_factor': (0.0014, 0.014, 0.14),
'PD.KNetgrowth': (0.05, 0.5, 5.0),
'PD.tvbmax': (1.35E+08, 1.35E+09, 1.35E+10),
'PD.Kmax': (0.15, 1.5, 15.0),
'PD.EC50k': (0.01, 0.1, 1.0),
'PD.gamma': (0.4, 4.0, 40.0),
'PD.beta': (0.2, 2.0, 20.0),
'PD.tau': (0.015, 0.15, 1.5),
'PD.Kdeath': (0.005, 0.5, 5.0),
'PD.Ksr_max': (0.1, 1.0, 10.0),
}
for qname, values in pkpd_variables.items():
print(qname, values)
var = pkpd_model.variables.get(qname=qname)
var.lower_bound, var.default_value, var.upper_bound = values
var.save()
CFU = dataset.biomarker_types.get(name='CFU')
demodrug_concentration = dataset.biomarker_types.get(
name='DemoDrug Concentration')
Ptotal = pkpd_model.variables.get(qname='Ptotal.size')
drug_c_concentration = pkpd_model.variables.get(
qname='central.drug_c_concentration')
drug_c_concentration.unit = demodrug_concentration.display_unit
drug_c_concentration.save()
fit_parameters = {
qname: values
for qname, values in pkpd_variables.items()
if qname != 'myokit.drug_c_scale_factor'
}
data = {
# Inference parameters
'name':
"usecase2",
'project':
self.project.id,
'algorithm':
Algorithm.objects.get(name='PSO').id,
'initialization_strategy':
'D',
'number_of_chains':
4,
'max_number_of_iterations':
11,
'burn_in':
0,
# Model
'model': {
'form': 'PK',
'id': pkpd_model.id
},
'dataset':
dataset.id,
# Model parameters
'parameters': [{
'name': qname,
'form': 'U',
'parameters': [values[0], values[2]],
} for qname, values in fit_parameters.items()] +
[{
'name': 't.size',
'form': 'F',
'parameters': ['1'],
}, {
'name': 'central.drug_c_amount',
'form': 'F',
'parameters': ['0'],
}],
# output
'observations': [
{
'model': drug_c_concentration.qname,
'biomarker': demodrug_concentration.name,
'noise_form': 'N',
'noise_param_form': 'F',
'parameters': [10],
},
{
'model': Ptotal.qname,
'biomarker': CFU.name,
'noise_form': 'N',
'noise_param_form': 'F',
'parameters': [1e6],
},
]
}
response = self.client.post("/api/inference/wizard",
data,
format='json')
response_data = response.data
self.assertEqual(response.status_code, status.HTTP_200_OK)
# check that there are no equations,
# the strings should be converted to float
for ll in response_data['log_likelihoods']:
self.assertNotEqual(ll['form'], 'E',
'{} is an Equation!'.format(ll['name']))
inference = Inference.objects.get(id=response_data['id'])
inference_mixin = InferenceMixin(inference)
log_posterior = inference_mixin._pints_log_posterior
# pymc3_model = log_posterior._model
# graph = pymc3.model_graph.model_to_graphviz(pymc3_model)
# graph.render(directory='test', view=True)
eval_pt = [
fit_parameters[name][1]
for name in log_posterior.parameter_names()
]
log_posterior(log_posterior.to_search(eval_pt))
inference_mixin.run_inference()