dfm-python 0.5.2__tar.gz → 0.5.3__tar.gz

{dfm_python-0.5.2 → dfm_python-0.5.3}/PKG-INFO

@@ -1,12 +1,12 @@
 Metadata-Version: 2.4
 Name: dfm-python
-Version: 0.5.2
+Version: 0.5.3
 Summary: Dynamic Factor Model (DFM) estimation and nowcasting in Python
 Project-URL: Homepage, https://github.com/yourusername/dfm-python
 Project-URL: Documentation, https://github.com/yourusername/dfm-python#readme
 Project-URL: Repository, https://github.com/yourusername/dfm-python
 Project-URL: Issues, https://github.com/yourusername/dfm-python/issues
-Author: DFM Python Contributors
+Author: Minkey Chang
 License: MIT
 License-File: LICENSE
 Keywords: dfm,dynamic-factor-model,econometrics,forecasting,nowcasting,time-series

dfm_python-0.5.3/lightning_logs/ddfm/version_11/metrics.csv

@@ -0,0 +1,2 @@
+epoch,step,train_loss_epoch
+0,0,0.8447005748748779

dfm_python-0.5.3/lightning_logs/ddfm/version_12/metrics.csv

@@ -0,0 +1,2 @@
+epoch,step,train_loss_epoch
+0,0,0.8700895309448242

dfm_python-0.5.3/lightning_logs/dfm/version_210/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-88.56625366210938,0,3.0,-60569.14453125,60471.5859375
+5.0,1,-88.56625366210938,1,,,
+5.0,2,-88.56625366210938,2,,,
+5.0,3,-88.56625366210938,3,,,
+5.0,4,-88.56625366210938,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_211/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-91.18582916259766,0,3.0,-70.42988586425781,12.529343605041504
+5.0,1,-91.18582916259766,1,,,
+5.0,2,-91.18582916259766,2,,,
+5.0,3,-91.18582916259766,3,,,
+5.0,4,-91.18582916259766,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_212/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-88.56625366210938,0,3.0,-60569.14453125,60471.5859375
+5.0,1,-88.56625366210938,1,,,
+5.0,2,-88.56625366210938,2,,,
+5.0,3,-88.56625366210938,3,,,
+5.0,4,-88.56625366210938,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_213/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-91.18582916259766,0,3.0,-70.42988586425781,12.529343605041504
+5.0,1,-91.18582916259766,1,,,
+5.0,2,-91.18582916259766,2,,,
+5.0,3,-91.18582916259766,3,,,
+5.0,4,-91.18582916259766,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_214/metrics.csv

@@ -0,0 +1,2 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+38.0,0,1164.719482421875,0,19.5,-2481.600830078125,3379.77783203125

dfm_python-0.5.3/lightning_logs/dfm/version_215/metrics.csv

@@ -0,0 +1,2 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+11.0,0,-223.8029022216797,0,6.0,-225.65797424316406,6.871443271636963

dfm_python-0.5.3/lightning_logs/dfm/version_216/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-88.56625366210938,0,3.0,-60569.14453125,60471.5859375
+5.0,1,-88.56625366210938,1,,,
+5.0,2,-88.56625366210938,2,,,
+5.0,3,-88.56625366210938,3,,,
+5.0,4,-88.56625366210938,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_217/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-91.18582916259766,0,3.0,-70.42988586425781,12.529343605041504
+5.0,1,-91.18582916259766,1,,,
+5.0,2,-91.18582916259766,2,,,
+5.0,3,-91.18582916259766,3,,,
+5.0,4,-91.18582916259766,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_218/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-88.56625366210938,0,3.0,-60569.14453125,60471.5859375
+5.0,1,-88.56625366210938,1,,,
+5.0,2,-88.56625366210938,2,,,
+5.0,3,-88.56625366210938,3,,,
+5.0,4,-88.56625366210938,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_219/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-91.18582916259766,0,3.0,-70.42988586425781,12.529343605041504
+5.0,1,-91.18582916259766,1,,,
+5.0,2,-91.18582916259766,2,,,
+5.0,3,-91.18582916259766,3,,,
+5.0,4,-91.18582916259766,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_220/metrics.csv

@@ -0,0 +1,2 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+38.0,0,1164.719482421875,0,19.5,-2481.600830078125,3379.77783203125

dfm_python-0.5.3/lightning_logs/dfm/version_221/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-88.56625366210938,0,3.0,-60569.14453125,60471.5859375
+5.0,1,-88.56625366210938,1,,,
+5.0,2,-88.56625366210938,2,,,
+5.0,3,-88.56625366210938,3,,,
+5.0,4,-88.56625366210938,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_222/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-91.18582916259766,0,3.0,-70.42988586425781,12.529343605041504
+5.0,1,-91.18582916259766,1,,,
+5.0,2,-91.18582916259766,2,,,
+5.0,3,-91.18582916259766,3,,,
+5.0,4,-91.18582916259766,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_223/metrics.csv

@@ -0,0 +1,2 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+11.0,0,-223.8029022216797,0,6.0,-225.65797424316406,6.871443271636963

dfm_python-0.5.3/lightning_logs/dfm/version_224/metrics.csv

@@ -0,0 +1,2 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+38.0,0,1164.719482421875,0,19.5,-2481.600830078125,3379.77783203125

dfm_python-0.5.3/lightning_logs/dfm/version_225/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-88.56625366210938,0,3.0,-60569.14453125,60471.5859375
+5.0,1,-88.56625366210938,1,,,
+5.0,2,-88.56625366210938,2,,,
+5.0,3,-88.56625366210938,3,,,
+5.0,4,-88.56625366210938,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_226/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-91.18582916259766,0,3.0,-70.42988586425781,12.529343605041504
+5.0,1,-91.18582916259766,1,,,
+5.0,2,-91.18582916259766,2,,,
+5.0,3,-91.18582916259766,3,,,
+5.0,4,-91.18582916259766,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_227/metrics.csv

@@ -0,0 +1,2 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+11.0,0,-223.8029022216797,0,6.0,-225.65797424316406,6.871443271636963

dfm_python-0.5.3/lightning_logs/dfm/version_228/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-88.56625366210938,0,3.0,-60569.14453125,60471.5859375
+5.0,1,-88.56625366210938,1,,,
+5.0,2,-88.56625366210938,2,,,
+5.0,3,-88.56625366210938,3,,,
+5.0,4,-88.56625366210938,4,,,

dfm_python-0.5.3/lightning_logs/dfm/version_229/metrics.csv

@@ -0,0 +1,6 @@
+em_iteration_epoch,epoch,loglik_epoch,step,train/em_iteration,train/loglik,train/loglik_change
+5.0,0,-91.18582916259766,0,3.0,-70.42988586425781,12.529343605041504
+5.0,1,-91.18582916259766,1,,,
+5.0,2,-91.18582916259766,2,,,
+5.0,3,-91.18582916259766,3,,,
+5.0,4,-91.18582916259766,4,,,

{dfm_python-0.5.2 → dfm_python-0.5.3}/pyproject.toml

@@ -1,12 +1,12 @@
 [project]
 name = "dfm-python"
-version = "0.5.2"
+version = "0.5.3"
 description = "Dynamic Factor Model (DFM) estimation and nowcasting in Python"
 readme = "README.md"
 requires-python = ">=3.10"
 license = {text = "MIT"}
 authors = [
-    {name = "DFM Python Contributors"}
+    {name = "Minkey Chang"}
 ]
 keywords = ["dfm", "dynamic-factor-model", "nowcasting", "forecasting", "time-series", "econometrics"]
 classifiers = [

{dfm_python-0.5.2 → dfm_python-0.5.3}/src/dfm_python/encoder/vae.py

@@ -82,13 +82,27 @@ if _has_torch:
             # Build layers
             prev_dim = input_dim
             for hidden_dim in hidden_dims:
-                self.layers.append(nn.Linear(prev_dim, hidden_dim))
+                layer = nn.Linear(prev_dim, hidden_dim)
+                # Initialize weights using Xavier/Kaiming initialization for better training stability
+                # Use Kaiming for ReLU, Xavier for tanh/sigmoid
+                if activation == 'relu':
+                    nn.init.kaiming_normal_(layer.weight, mode='fan_in', nonlinearity='relu')
+                else:
+                    nn.init.xavier_normal_(layer.weight, gain=1.0)
+                # Initialize bias to small values
+                if layer.bias is not None:
+                    nn.init.constant_(layer.bias, 0.0)
+                self.layers.append(layer)
                 if use_batch_norm:
                     self.batch_norms.append(nn.BatchNorm1d(hidden_dim))
                 prev_dim = hidden_dim
             
-            # Output layer (no activation, linear)
+            # Output layer (linear, no activation)
+            # Use smaller initialization for output layer to prevent large initial factors
             self.output_layer = nn.Linear(prev_dim, output_dim)
+            nn.init.xavier_normal_(self.output_layer.weight, gain=0.1)  # Smaller gain for output
+            if self.output_layer.bias is not None:
+                nn.init.constant_(self.output_layer.bias, 0.0)
         
         def forward(self, x: torch.Tensor) -> torch.Tensor:
             """Forward pass through encoder.

{dfm_python-0.5.2 → dfm_python-0.5.3}/src/dfm_python/models/ddfm.py

@@ -308,6 +308,10 @@ class DDFM(BaseFactorModel):
         decay_learning_rate: bool = True,
         min_obs_pretrain: int = 50,
         mult_epoch_pretrain: int = 1,
+        loss_function: str = 'mse',
+        huber_delta: float = 1.0,
+        weight_decay: float = 0.0,
+        grad_clip_val: float = 1.0,
         **kwargs
     ):
         """Initialize DDFM instance.
@@ -349,6 +353,19 @@ class DDFM(BaseFactorModel):
         mult_epoch_pretrain : int, default 1
             Multiplier for number of epochs during pre-training
             Display progress every 'disp' iterations
+        loss_function : str, default 'mse'
+            Loss function for training ('mse', 'huber'). 
+            'mse': Mean squared error (default, matches original DDFM)
+            'huber': Huber loss (more robust to outliers)
+        huber_delta : float, default 1.0
+            Delta parameter for Huber loss (only used if loss_function='huber').
+            Controls the transition point between quadratic and linear regions.
+        weight_decay : float, default 0.0
+            Weight decay (L2 regularization) for optimizer. Helps prevent overfitting to linear features.
+            Recommended: 1e-5 to 1e-3 for deeper encoders or when encoder collapses to linear behavior.
+        grad_clip_val : float, default 1.0
+            Maximum gradient norm for gradient clipping. Prevents training instability.
+            Set to 0.0 to disable gradient clipping.
         seed : int, optional
             Random seed for reproducibility
         **kwargs
@@ -383,6 +400,22 @@ class DDFM(BaseFactorModel):
         self.decay_learning_rate = decay_learning_rate
         self.min_obs_pretrain = min_obs_pretrain
         self.mult_epoch_pretrain = mult_epoch_pretrain
+        self.loss_function = loss_function.lower()
+        self.huber_delta = huber_delta
+        self.weight_decay = weight_decay
+        self.grad_clip_val = grad_clip_val
+        
+        # Validate loss function
+        if self.loss_function not in ['mse', 'huber']:
+            raise ValueError(
+                f"DDFM initialization failed: loss_function must be 'mse' or 'huber', got '{loss_function}'"
+            )
+        
+        # Validate gradient clipping value
+        if self.grad_clip_val < 0.0:
+            raise ValueError(
+                f"DDFM initialization failed: grad_clip_val must be >= 0.0, got {grad_clip_val}"
+            )
         
         # Determine number of factors
         # DDFM does not use block structure - num_factors is specified directly
@@ -586,7 +619,9 @@ class DDFM(BaseFactorModel):
         
         # Clip input data to prevent extreme values that cause NaN
         # Clip to reasonable range: -10 to 10 standard deviations
-        data_clipped = torch.clamp(data, min=-10.0, max=10.0)
+        # For deeper networks, use slightly tighter clipping to improve stability
+        clip_range = 8.0 if len(self.encoder_layers) > 2 else 10.0
+        data_clipped = torch.clamp(data, min=-clip_range, max=clip_range)
         
         # Forward pass
         reconstructed = self.forward(data_clipped)
@@ -603,8 +638,23 @@ class DDFM(BaseFactorModel):
         # Compute loss with missing data masking
         mask = torch.isfinite(target)
         target_clean = torch.where(mask, target, torch.zeros_like(target))
-        squared_diff = (target_clean - reconstructed) ** 2
-        loss = torch.sum(squared_diff * mask) / (torch.sum(mask) + 1e-8)
+        
+        # Use specified loss function
+        if self.loss_function == 'huber':
+            # Huber loss: more robust to outliers
+            # L_delta(a) = 0.5 * a^2 if |a| <= delta, else delta * (|a| - 0.5 * delta)
+            diff = target_clean - reconstructed
+            abs_diff = torch.abs(diff)
+            huber_loss = torch.where(
+                abs_diff <= self.huber_delta,
+                0.5 * diff ** 2,
+                self.huber_delta * (abs_diff - 0.5 * self.huber_delta)
+            )
+            loss = torch.sum(huber_loss * mask) / (torch.sum(mask) + 1e-8)
+        else:
+            # MSE loss (default)
+            squared_diff = (target_clean - reconstructed) ** 2
+            loss = torch.sum(squared_diff * mask) / (torch.sum(mask) + 1e-8)
         
         # Handle NaN/Inf in loss
         if not torch.isfinite(loss):
@@ -614,6 +664,9 @@ class DDFM(BaseFactorModel):
         # Log metrics
         self.log('train_loss', loss, on_step=True, on_epoch=True, prog_bar=True)
         
+        # Note: Gradient clipping is handled automatically by Lightning trainer if gradient_clip_val is set
+        # The grad_clip_val parameter is used in pre_train() and MCMC training for manual training loops
+        
         return loss
     
     
@@ -697,7 +750,8 @@ class DDFM(BaseFactorModel):
         
         optimizer = torch.optim.Adam(
             list(self.encoder.parameters()) + list(self.decoder.parameters()),
-            lr=self.learning_rate
+            lr=self.learning_rate,
+            weight_decay=self.weight_decay
         )
         
         if self.decay_learning_rate:
@@ -750,7 +804,8 @@ class DDFM(BaseFactorModel):
         
         optimizer = torch.optim.Adam(
             list(self.encoder.parameters()) + list(self.decoder.parameters()),
-            lr=lr
+            lr=lr,
+            weight_decay=self.weight_decay
         )
         
         return optimizer
@@ -842,7 +897,8 @@ class DDFM(BaseFactorModel):
         # Create optimizer for pre-training
         optimizer = torch.optim.Adam(
             list(self.encoder.parameters()) + list(self.decoder.parameters()),
-            lr=self.learning_rate
+            lr=self.learning_rate,
+            weight_decay=self.weight_decay
         )
         
         # Pre-train for epochs * mult_epoch_pretrain
@@ -889,10 +945,11 @@ class DDFM(BaseFactorModel):
                 loss.backward()
                 
                 # Gradient clipping for stability
-                torch.nn.utils.clip_grad_norm_(
-                    list(self.encoder.parameters()) + list(self.decoder.parameters()),
-                    max_norm=1.0
-                )
+                if self.grad_clip_val > 0.0:
+                    torch.nn.utils.clip_grad_norm_(
+                        list(self.encoder.parameters()) + list(self.decoder.parameters()),
+                        max_norm=self.grad_clip_val
+                    )
                 
                 optimizer.step()
                 

{dfm_python-0.5.2 → dfm_python-0.5.3}/src/dfm_python/models/dfm.py

@@ -23,11 +23,15 @@ from ..config import (
     MergedConfigSource,
 )
 from ..config.results import DFMResult, FitParams
+from ..config.utils import get_agg_structure, get_tent_weights, FREQUENCY_HIERARCHY, TENT_WEIGHTS_LOOKUP
 from ..logger import get_logger
 from ..ssm.em import EMAlgorithm, EMStepParams
 from ..ssm.kalman import KalmanFilter
 from .base import BaseFactorModel
 
+# Frequency to integer mapping for tensor conversion
+_FREQ_TO_INT = {'d': 1, 'w': 2, 'm': 3, 'q': 4, 'sa': 5, 'a': 6}
+
 if TYPE_CHECKING:
     from omegaconf import DictConfig
     from ..lightning import DFMDataModule
@@ -90,6 +94,7 @@ class DFM(BaseFactorModel):
         nan_method: int = 2,
         nan_k: int = 3,
         tent_weights_dict: Optional[dict] = None,
+        mixed_freq: bool = False,
         **kwargs
     ):
         """Initialize DFM instance.
@@ -108,6 +113,12 @@ class DFM(BaseFactorModel):
             Missing data handling method
         nan_k : int, default 3
             Spline interpolation order
+        tent_weights_dict : dict, optional
+            [DEPRECATED] Optional tent weights to control aggregation. Use `mixed_freq` parameter instead.
+            Kept for backward compatibility only.
+        mixed_freq : bool, default False
+            If True, use tent kernels for mixed-frequency data. If False, treat all series as clock frequency.
+            When True, raises ValueError if any frequency pair is not in TENT_WEIGHTS_LOOKUP.
         **kwargs
             Additional arguments passed to BaseFactorModel
         """
@@ -120,8 +131,20 @@ class DFM(BaseFactorModel):
         self.max_iter = max_iter
         self.nan_method = nan_method
         self.nan_k = nan_k
-        # Optional tent weights to control aggregation (e.g., disable tent by passing {'q':[1.0]})
-        self.tent_weights_dict = tent_weights_dict
+        self.mixed_freq = mixed_freq
+        
+        # Legacy tent_weights_dict parameter (deprecated - use mixed_freq instead)
+        # Only used when mixed_freq=False for backward compatibility
+        self._legacy_tent_weights_dict = tent_weights_dict
+        
+        # Mixed frequency parameters (set during initialize_from_data)
+        self._em_R_mat = None
+        self._em_q = None
+        self._em_nQ = 0
+        self._em_tent_weights_dict = None
+        self._em_frequencies = None
+        self._em_i_idio = None
+        self._em_idio_chain_lengths = None
         
         # Determine number of factors
         if num_factors is None:
@@ -181,6 +204,53 @@ class DFM(BaseFactorModel):
         # Parameters will be initialized during fit_em() or first training step
         pass
     
+    def _create_em_step_params(
+        self, 
+        y: torch.Tensor, 
+        device: torch.device, 
+        dtype: torch.dtype
+    ) -> EMStepParams:
+        """Create EM step parameters using stored mixed frequency parameters.
+        
+        Parameters
+        ----------
+        y : torch.Tensor
+            Data tensor (N x T)
+        device : torch.device
+            Device for tensors
+        dtype : torch.dtype
+            Data type for tensors
+            
+        Returns
+        -------
+        EMStepParams
+            EM step parameters
+        """
+        clock = getattr(self.config, 'clock', 'm')
+        N = y.shape[0]
+        
+        return EMStepParams(
+            y=y,
+            A=self.A,
+            C=self.C,
+            Q=self.Q,
+            R=self.R,
+            Z_0=self.Z_0,
+            V_0=self.V_0,
+            r=self.r.to(device),
+            p=self.p,
+            R_mat=self._em_R_mat,
+            q=self._em_q,
+            nQ=self._em_nQ,
+            i_idio=self._em_i_idio if self._em_i_idio is not None else torch.ones(N, device=device, dtype=dtype),
+            blocks=self.blocks.to(device),
+            tent_weights_dict=self._em_tent_weights_dict,
+            clock=clock,
+            frequencies=self._em_frequencies,
+            idio_chain_lengths=self._em_idio_chain_lengths if self._em_idio_chain_lengths is not None else torch.zeros(N, device=device, dtype=dtype),
+            config=self.config
+        )
+    
     def initialize_from_data(self, X: torch.Tensor) -> None:
         """Initialize parameters from data using PCA and OLS.
         
@@ -190,22 +260,100 @@ class DFM(BaseFactorModel):
             Standardized data (T x N)
         """
         opt_nan = {'method': self.nan_method, 'k': self.nan_k}
+        clock = getattr(self.config, 'clock', 'm')
         
-        # Use self.em.initialize_parameters() with direct tensor operations (no CPU transfers)
+        # Handle mixed_freq parameter
+        if self.mixed_freq:
+            # Use tent kernels for mixed-frequency data
+            agg_structure = get_agg_structure(self.config, clock=clock)
+            
+            # Validate that all required frequency pairs are in TENT_WEIGHTS_LOOKUP
+            frequencies_list = [s.frequency for s in self.config.series]
+            frequencies_set = set(frequencies_list)
+            clock_hierarchy = FREQUENCY_HIERARCHY.get(clock, 3)
+            
+            missing_pairs = []
+            for freq in frequencies_set:
+                freq_hierarchy = FREQUENCY_HIERARCHY.get(freq, 3)
+                if freq_hierarchy > clock_hierarchy:
+                    # This frequency is slower than clock, needs tent kernel
+                    tent_w = get_tent_weights(freq, clock)
+                    if tent_w is None:
+                        missing_pairs.append((freq, clock))
+            
+            if missing_pairs:
+                raise ValueError(
+                    f"mixed_freq=True but the following frequency pairs are not in TENT_WEIGHTS_LOOKUP: {missing_pairs}. "
+                    f"Available pairs: {list(TENT_WEIGHTS_LOOKUP.keys())}. "
+                    f"Either add the missing pairs to TENT_WEIGHTS_LOOKUP or set mixed_freq=False."
+                )
+            
+            # Convert tent_weights to torch tensors
+            tent_weights_dict = {k: torch.tensor(v, dtype=torch.float32, device=X.device) 
+                                for k, v in agg_structure['tent_weights'].items()}
+            
+            # Get R_mat and q from first structure (if any)
+            R_mat = None
+            q = None
+            if agg_structure['structures']:
+                first_structure = list(agg_structure['structures'].values())[0]
+                R_mat = torch.tensor(first_structure[0], dtype=torch.float32, device=X.device)
+                q = torch.tensor(first_structure[1], dtype=torch.float32, device=X.device)
+            
+            # Create frequencies array
+            frequencies_array = np.array(frequencies_list, dtype=object)
+            
+            # Count slower-frequency series
+            nQ = sum(1 for freq in frequencies_list 
+                    if FREQUENCY_HIERARCHY.get(freq, 3) > clock_hierarchy)
+            
+            # Compute i_idio (1 for clock frequency, 0 for slower frequencies)
+            i_idio = torch.tensor([1 if freq == clock else 0 for freq in frequencies_list], 
+                                 dtype=torch.float32, device=X.device)
+        else:
+            # Treat all series as clock frequency (unified frequency)
+            tent_weights_dict = self._legacy_tent_weights_dict
+            R_mat = None
+            q = None
+            frequencies_array = None
+            nQ = 0
+            i_idio = torch.ones(X.shape[1], dtype=torch.float32, device=X.device)
+        
+        # Convert frequencies to torch tensor
+        frequencies_tensor = None
+        if frequencies_array is not None:
+            frequencies_tensor = torch.tensor(
+                [_FREQ_TO_INT.get(f, 3) for f in frequencies_array], 
+                dtype=torch.int32, 
+                device=X.device
+            )
+        
+        # Store for reuse in EM steps
+        self._em_R_mat = R_mat
+        self._em_q = q
+        self._em_nQ = nQ
+        self._em_tent_weights_dict = tent_weights_dict
+        self._em_frequencies = frequencies_tensor
+        self._em_i_idio = i_idio
+        
+        # Idiosyncratic chain lengths (currently unused, set to zeros)
+        self._em_idio_chain_lengths = torch.zeros(X.shape[1], dtype=torch.int32, device=X.device)
+        
+        # Initialize parameters using EM algorithm
         A, C, Q, R, Z_0, V_0 = self.em.initialize_parameters(
             X,
             r=self.r.to(X.device),
             p=self.p,
             blocks=self.blocks.to(X.device),
             opt_nan=opt_nan,
-            R_mat=None,
-            q=None,
-            nQ=0,
-            i_idio=None,
-            clock=getattr(self.config, 'clock', 'm'),
-            tent_weights_dict=self.tent_weights_dict,
-            frequencies=None,
-            idio_chain_lengths=None,
+            R_mat=R_mat,
+            q=q,
+            nQ=nQ,
+            i_idio=i_idio,
+            clock=clock,
+            tent_weights_dict=tent_weights_dict,
+            frequencies=frequencies_tensor,
+            idio_chain_lengths=self._em_idio_chain_lengths,
             config=self.config
         )
         
@@ -267,28 +415,8 @@ class DFM(BaseFactorModel):
         # EM expects y as (N x T), but data is (T x N)
         y = data.T  # (N x T)
         
-        # Create EM step parameters
-        em_params = EMStepParams(
-            y=y,
-            A=self.A,
-            C=self.C,
-            Q=self.Q,
-            R=self.R,
-            Z_0=self.Z_0,
-            V_0=self.V_0,
-            r=self.r.to(y.device),
-            p=self.p,
-            R_mat=None,
-            q=None,
-            nQ=0,
-            i_idio=torch.ones(y.shape[0], device=y.device, dtype=y.dtype),
-            blocks=self.blocks.to(y.device),
-            tent_weights_dict=self.tent_weights_dict,
-            clock=getattr(self.config, 'clock', 'm'),
-            frequencies=None,
-            idio_chain_lengths=torch.zeros(y.shape[0], device=y.device, dtype=y.dtype),
-            config=self.config
-        )
+        # Create EM step parameters using stored mixed frequency parameters
+        em_params = self._create_em_step_params(y, y.device, y.dtype)
         
         # Perform EM step - use self.em(...) instead of em_step(...)
         C_new, R_new, A_new, Q_new, Z_0_new, V_0_new, loglik = self.em(em_params)
@@ -390,28 +518,8 @@ class DFM(BaseFactorModel):
         
         # EM loop
         while num_iter < self.max_iter and not converged:
-            # Create EM step parameters
-            em_params = EMStepParams(
-                y=y,
-                A=self.A,
-                C=self.C,
-                Q=self.Q,
-                R=self.R,
-                Z_0=self.Z_0,
-                V_0=self.V_0,
-                r=self.r.to(device),
-                p=self.p,
-                R_mat=None,
-                q=None,
-                nQ=0,
-                i_idio=torch.ones(y.shape[0], device=device, dtype=dtype),
-                blocks=self.blocks.to(device),
-                tent_weights_dict={},
-                clock=getattr(self.config, 'clock', 'm'),
-                frequencies=None,
-                idio_chain_lengths=torch.zeros(y.shape[0], device=device, dtype=dtype),
-                config=self.config
-            )
+            # Create EM step parameters using stored mixed frequency parameters
+            em_params = self._create_em_step_params(y, device, dtype)
             
             # Perform EM step - use self.em(...) instead of em_step(...)
             C_new, R_new, A_new, Q_new, Z_0_new, V_0_new, loglik = self.em(em_params)

{dfm_python-0.5.2 → dfm_python-0.5.3}/src/dfm_python/models/mcmc.py

@@ -347,10 +347,11 @@ class DDFMMCMCTrainer:
                     loss = torch.sum(squared_diff) / (torch.sum(mask) + 1e-8)
                     loss.backward()
                     # Gradient clipping to prevent NaN and improve stability
-                    torch.nn.utils.clip_grad_norm_(
-                        list(self.model.encoder.parameters()) + list(self.model.decoder.parameters()),
-                        max_norm=1.0
-                    )
+                    if self.model.grad_clip_val > 0.0:
+                        torch.nn.utils.clip_grad_norm_(
+                            list(self.model.encoder.parameters()) + list(self.model.decoder.parameters()),
+                            max_norm=self.model.grad_clip_val
+                        )
                     optimizer.step()
                 # Extract factors from this sample
                 x_sample_tensor = torch.tensor(x_sim_den[i, :, :], device=device, dtype=dtype)