<TensorType(float64, scalar)
D. K
I run a high dimensional tvp pvar with numerous endogenous variables . I get the error below. I have look at many places, but I did not find a solution. I would much appreciate it if you could help me
David
/Users/eva/anaconda3/lib/python3.11/site-packages/sklearn/impute/_iterative.py:801: ConvergenceWarning: [IterativeImputer] Early stopping criterion not reached.
warnings.warn(
Auto-assigning NUTS sampler...
Initializing NUTS using jitter+adapt_diag...
You can find the C code in this temporary file: /var/folders/p_/4shh5wgx7vn_sp4_77tp6c7m0000gn/T/theano_compilation_error_0s9f3yfr
Traceback (most recent call last):
File ~/anaconda3/lib/python3.11/site-packages/theano/tensor/basic.py:184 in as_tensor_variable
x = [extract_constants(i) for i in x]
File ~/anaconda3/lib/python3.11/site-packages/theano/tensor/basic.py:184 in <listcomp>
x = [extract_constants(i) for i in x]
File ~/anaconda3/lib/python3.11/site-packages/theano/tensor/basic.py:179 in extract_constants
raise TypeError
TypeError
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
Cell In[2], line 1
model.fit_tvp_pvar(num_iterations=50000, burn=20000, tune=10000, cores=4, imputation_method='iterative', imputer_max_iter=1000)
Cell In[1], line 130 in fit_tvp_pvar
trace = pm.sample(num_iterations, tune=tune, cores=cores, return_inferencedata=True)
File ~/anaconda3/lib/python3.11/site-packages/pymc3/sampling.py:496 in sample
start_, step = init_nuts(
File ~/anaconda3/lib/python3.11/site-packages/pymc3/sampling.py:2187 in init_nuts
step = pm.NUTS(potential=potential, model=model, **kwargs)
File ~/anaconda3/lib/python3.11/site-packages/pymc3/step_methods/hmc/nuts.py:168 in __init__
super().__init__(vars, **kwargs)
File ~/anaconda3/lib/python3.11/site-packages/pymc3/step_methods/hmc/base_hmc.py:88 in __init__
super().__init__(vars, blocked=blocked, model=model, dtype=dtype, **theano_kwargs)
File ~/anaconda3/lib/python3.11/site-packages/pymc3/step_methods/arraystep.py:254 in __init__
func = model.logp_dlogp_function(vars, dtype=dtype, **theano_kwargs)
File ~/anaconda3/lib/python3.11/site-packages/pymc3/model.py:1001 in logp_dlogp_function
costs = [self.logpt]
File ~/anaconda3/lib/python3.11/site-packages/pymc3/model.py:1011 in logpt
logp = tt.sum([tt.sum(factor) for factor in factors])
File ~/anaconda3/lib/python3.11/site-packages/theano/tensor/basic.py:3221 in sum
out = elemwise.Sum(axis=axis, dtype=dtype, acc_dtype=acc_dtype)(input)
File ~/anaconda3/lib/python3.11/site-packages/theano/graph/op.py:250 in __call__
node = self.make_node(*inputs, **kwargs)
File ~/anaconda3/lib/python3.11/site-packages/theano/tensor/elemwise.py:1925 in make_node
input = as_tensor_variable(input)
File ~/anaconda3/lib/python3.11/site-packages/theano/tensor/basic.py:194 in as_tensor_variable
return theano.tensor.opt.MakeVector(dtype)(*x)
File ~/anaconda3/lib/python3.11/site-packages/theano/graph/op.py:253 in __call__
compute_test_value(node)
File ~/anaconda3/lib/python3.11/site-packages/theano/graph/op.py:126 in compute_test_value
thunk = node.op.make_thunk(node, storage_map, compute_map, no_recycling=[])
File ~/anaconda3/lib/python3.11/site-packages/theano/graph/op.py:634 in make_thunk
return self.make_c_thunk(node, storage_map, compute_map, no_recycling)
File ~/anaconda3/lib/python3.11/site-packages/theano/graph/op.py:600 in make_c_thunk
outputs = cl.make_thunk(
File ~/anaconda3/lib/python3.11/site-packages/theano/link/c/basic.py:1203 in make_thunk
cthunk, module, in_storage, out_storage, error_storage = self.__compile__(
File ~/anaconda3/lib/python3.11/site-packages/theano/link/c/basic.py:1138 in __compile__
thunk, module = self.cthunk_factory(
File ~/anaconda3/lib/python3.11/site-packages/theano/link/c/basic.py:1634 in cthunk_factory
module = get_module_cache().module_from_key(key=key, lnk=self)
File ~/anaconda3/lib/python3.11/site-packages/theano/link/c/cmodule.py:1191 in module_from_key
module = lnk.compile_cmodule(location)
File ~/anaconda3/lib/python3.11/site-packages/theano/link/c/basic.py:1543 in compile_cmodule
module = c_compiler.compile_str(
File ~/anaconda3/lib/python3.11/site-packages/theano/link/c/cmodule.py:2546 in compile_str
raise Exception(
Exception: ('Compilation failed (return status=1): /Users/ev/.theano/compiledir_macOS-14.5-arm64-arm-64bit-arm-3.11.8-64/tmpenzneiyk/mod.cpp:72300:32: fatal error: bracket nesting level exceeded maximum of 256. if (!PyErr_Occurred()) {. ^. /Users/ev/.theano/compiledir_macOS-14.5-arm64-arm-64bit-arm-3.11.8-64/tmpenzneiyk/mod.cpp:72300:32: note: use -fbracket-depth=N to increase maximum nesting level. 1 error generated.. ', "FunctionGraph(MakeVector{dtype='float64'}(<TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64, scalar)>, <TensorType(float64,
my code is
def fit_tvp_pvar(self, num_iterations=1000, burn=500, tune=500, cores=1, delay=1, imputation_method='iterative', imputer_max_iter=100):
"""Fit the TVP-PVAR model using NUTS sampling"""
data = self.data[self.vars].values
if imputation_method == 'iterative':
# Iterative Imputer
imputer = IterativeImputer(max_iter=imputer_max_iter, random_state=self.seed)
data = imputer.fit_transform(data)
elif imputation_method == 'kalman':
# Kalman Filter Imputation
data = kalman_imputation(data, max_iter=imputer_max_iter, seed=self.seed)
elif imputation_method == 'knn':
# KNN Imputer
imputer = KNNImputer()
data = imputer.fit_transform(data)
valid_mask = ~np.isnan(data) & ~np.isinf(data) # Mask for valid values
self.n_vars = data.shape[1] # Set self.n_vars based on data
self.n_obs = data.shape[0]
with pm.Model() as tvp_pvar_model:
sd_dist = pm.HalfNormal.dist(sigma=1.0) # Standard deviation distribution
sd_vals = pm.HalfNormal('sd_vals', sigma=1, shape=(self.n_vars,))
alphas = pm.Normal('alphas', mu=0, sigma=1, shape=(self.n_obs, self.n_vars)) + 1e-6 # Add a small constant
phis = horseshoe_prior('phis', shape=(self.n_obs - self.p, self.p, self.n_vars, self.n_vars)) + 1e-6 # Add a small constant
# Define a Cholesky-factor covariance matrix
cov_matrix_chol = pm.LKJCholeskyCov('cov_matrix_chol', n=self.n_vars, eta=2, sd_dist=sd_dist)
# Compute diagonal matrix from sd_vals
diag_sd_vals = tt.diag(sd_vals)
# Compute covariance matrix
cov_matrix = diag_sd_vals.dot(cov_matrix_chol.dot(cov_matrix_chol.T)).dot(diag_sd_vals)
chol_cov = pm.Deterministic('chol_cov', tt.slinalg.cholesky(cov_matrix))
# Likelihood
for t in range(self.p, self.n_obs):
y_t = data[t]
y_lags = np.column_stack([data[t-i-1] for i in range(self.p)]) # Adjust index for lagged values
y_lags = y_lags.T.reshape(self.p, self.n_vars, 1) # Reshape y_lags for correct multiplication
mean_t = alphas[t] + tt.sum(phis[t-self.p] * y_lags, axis=1).T # Corrected matrix multiplication
delta = y_t - mean_t
valid_obs = valid_mask[t, :] # Check valid observations at time t
# Create valid_indices for indexing
valid_indices = valid_obs.nonzero()[0]
# Ensure chol_cov is not flattened unnecessarily
chol_cov_reshaped = tt.reshape(chol_cov, (self.n_vars, self.n_vars))
# Ensure delta is not reshaped unnecessarily
pm.MvNormal(f'y_{t}', mu=mean_t[valid_indices], chol=chol_cov_reshaped[valid_indices][:, valid_indices], observed=delta[valid_indices])
# Introduce a delay between iterations to avoid timeout issues
iteration_delay = 1 # Delay in seconds
for iteration in range(num_iterations):
time.sleep(iteration_delay)
trace = pm.sample(num_iterations, tune=tune, cores=cores, return_inferencedata=True)
trace = az.sel(trace, draw=slice(burn, None))
self.trace = trace