mae = MechanisticAutoEncoder(N_HIDDEN, (
os.path.join('data', f'{DATA}__{MODEL}__measurements.tsv'),
os.path.join('data', f'{DATA}__{MODEL}__conditions.tsv'),
os.path.join('data', f'{DATA}__{MODEL}__observables.tsv'),
), MODEL)
pretraining_problems = generate_per_sample_pretraining_problems(mae)
pretrained_samples = []
pretraindir = 'pretraining'
prefix = f'{mae.pathway_name}__{mae.data_name}'
for sample, importer in pretraining_problems.items():
problem = importer.create_problem()
model = importer.create_model()
result = pretrain(problem, pypesto.startpoint.uniform, 100)
output_prefix = f'{prefix}__{sample}'
store_and_plot_pretraining(result, pretraindir, output_prefix)
pretrained_samples.append(output_prefix + '.csv')
simulation = amici.petab_objective.simulate_petab(
importer.petab_problem,
model,
problem_parameters=dict(
zip(
problem.x_names,
result.optimize_result.list[0]['x'],
)),
scaled_parameters=True,
edatas=problem.objective._objectives[0].edatas)
problem.get_reduced_vector(np.asarray(problem.x_names),
problem.x_free_indices)
):
if xname.startswith(MODEL_FEATURE_PREFIX):
match = re.match(fr'{MODEL_FEATURE_PREFIX}([\w_]+)__'
r'([\w0-9]+)', xname)
par = match.group(1)
sample = match.group(2)
xs[istart, ix] = par_combo.loc[sample, par] - means[par]
else:
xs[istart, ix] = means[xname]
return xs
result = pretrain(problem, startpoints, 10,
subspace=fides.SubSpaceDim.TWO, maxiter=int(1e2))
store_and_plot_pretraining(result, pretraindir, output_prefix)
N_STARTS = 5
# plot residuals and pca of inputs for debugging purposes
data_dicts = []
fig_pca, axes_pca = plt.subplots(1, N_STARTS, figsize=(18.5, 10.5))
for ir, r in enumerate(result.optimize_result.list[:N_STARTS]):
rdatas = problem.objective._objectives[0](r['x'], return_dict=True)[
pypesto.objective.constants.RDATAS
]
inputs.rename(columns={0: 'value'}, inplace=True)
input_start = next(
ix for ix, name in enumerate(mae.pypesto_subproblem.x_names)
if name.startswith('INPUT_'))
pretrained_inputs = pd.pivot(
inputs, index='par', columns='sample').reindex([
re.match(pattern, name).group(1) for name in
mae.pypesto_subproblem.x_names[input_start:input_start +
mae.n_model_inputs]
])
pretrained_inputs.columns = [r[1] for r in pretrained_inputs.columns]
problem = generate_encoder_inflate_pretraining_problem(
mae, pretrained_inputs, pars)
result = pretrain(problem, pypesto.startpoint.uniform, 10, fatol=1e-4)
for r in result.optimize_result.list:
r['id'] += f'_{i_input}'
results.append(result)
# merge results
result = pypesto.Result(problem)
result.optimize_result.list = [
r for result in results for r in result.optimize_result.list
]
result.optimize_result.sort()
store_and_plot_pretraining(result, pretraindir, output_prefix)
# compute and plot residuals