dfm-python 0.5.58__py3-none-any.whl → 0.5.59__py3-none-any.whl

dfm_python/functional/em.py

@@ -1233,7 +1233,7 @@ def em_step(
         _logger.info(f"    E-step: Running Kalman filter + smoother (T={T}, N={N}, m={m}, ~{ops_estimate:.1f}B ops)...")
         _logger.info(f"    E-step: Filter is fast (~1s), but smooth may take 5-7 minutes (bottleneck)")
     EZ, V_smooth, VVsmooth, loglik = kalman_filter.filter_and_smooth(X_masked)
-    
+
     # Keep float64 for numerical stability (prevents V_smooth accumulation overflow)
     # Do NOT convert to float32 - maintain float64 precision throughout M-step
     # Float64 is critical for large state spaces (m=183) and long time series (T=2135)

dfm_python/models/dfm.py

@@ -1319,7 +1319,33 @@ class DFM(BaseFactorModel):
         
         # Run filter and smooth on new data
         y_masked = np.ma.masked_invalid(data_new)
-        Z_new, V_smooth_new, _, _ = kalman_new.filter_and_smooth(y_masked)
+
+        # ------------------------------------------------------------------
+        # IMPORTANT (pykalman quirk):
+        # pykalman.KalmanFilter._parse_observations() transposes when input has
+        # shape (1, n_dim_obs), interpreting it as (n_dim_obs, 1). For our
+        # update() case, (1, N) means 1 timestep with N series. This transpose
+        # causes n_dim_obs=1 and breaks CholeskyKalmanFilter internals.
+        #
+        # Workaround: for a single new timestep, use filter_update() directly
+        # with a 1D observation vector (n_dim_obs,), avoiding _parse_observations.
+        # ------------------------------------------------------------------
+        if hasattr(data_new, "shape") and len(data_new.shape) == 2 and int(data_new.shape[0]) == 1:
+            pk = getattr(kalman_new, "_pykalman", None)
+            if pk is None:
+                raise ModelNotTrainedError("Kalman filter not initialized for update()")
+
+            # 1D observation vector for this timestep
+            obs = np.ma.masked_invalid(data_new[0])
+            next_mean, _next_cov = pk.filter_update(
+                filtered_state_mean=Z_last,
+                filtered_state_covariance=V_last,
+                observation=obs,
+            )
+            Z_new = np.asarray(next_mean, dtype=DEFAULT_DTYPE)[np.newaxis, :]
+            V_smooth_new = None
+        else:
+            Z_new, V_smooth_new, _, _ = kalman_new.filter_and_smooth(y_masked)
         
         # Update model state: append new factors and data
         # Concatenate new factors to existing result.Z
@@ -1509,6 +1535,18 @@ class DFM(BaseFactorModel):
         except (ModelNotInitializedError, AttributeError):
             pass
         target_series, target_indices = resolve_target_series(dataset, series_ids, result, self.__class__.__name__)
+
+        # Fallback: If no Dataset is attached, use config.target_series.
+        # This is needed for checkpoints used in external pipelines where dataset.pkl
+        # is missing or not attached to the model instance.
+        if (target_series is None or len(target_series) == 0 or target_indices is None) and getattr(self, "_config", None) is not None:
+            cfg_ts = getattr(self._config, "target_series", None)
+            if cfg_ts is not None:
+                cfg_ts_list = cfg_ts if isinstance(cfg_ts, list) else [cfg_ts]
+                cfg_indices = [series_ids.index(t) for t in cfg_ts_list if series_ids is not None and t in series_ids]
+
+                if cfg_indices:
+                    target_series, target_indices = cfg_ts_list, cfg_indices
         
         if target_series is None or len(target_series) == 0:
             raise ValueError(
@@ -1747,19 +1785,36 @@ class DFM(BaseFactorModel):
             raise FileNotFoundError(f"Checkpoint file not found: {path}")
         
         # Try loading with error handling for corrupted files
+        # NOTE: Our training pipeline typically saves with joblib (for speed/compression),
+        # but older checkpoints may be plain pickle. Support both.
+        checkpoint = None
         try:
-            with open(path, 'rb') as f:
-                checkpoint = pickle.load(f)
-        except (pickle.UnpicklingError, EOFError, ValueError) as e:
-            raise RuntimeError(
-                f"Failed to load checkpoint from {path}. "
-                f"The file may be corrupted. Error: {e}. "
-                f"Try re-training the model to generate a new checkpoint."
-            ) from e
+            import joblib  # local import to avoid hard dependency at import-time
+            checkpoint = joblib.load(path)
+        except Exception:
+            # Fallback to plain pickle
+            try:
+                with open(path, 'rb') as f:
+                    checkpoint = pickle.load(f)
+            except (pickle.UnpicklingError, EOFError, ValueError) as e:
+                raise RuntimeError(
+                    f"Failed to load checkpoint from {path}. "
+                    f"The file may be corrupted or saved in an unsupported format. Error: {e}. "
+                    f"Try re-training the model to generate a new checkpoint."
+                ) from e
         
         # Use checkpoint config if not provided
         if config is None:
             config = checkpoint.get('config')
+
+        # If target_series was stored separately in the checkpoint (common in our pipeline),
+        # inject it into config so predict() can resolve targets without a Dataset attached.
+        try:
+            ckpt_target_series = checkpoint.get('target_series')
+            if config is not None and getattr(config, 'target_series', None) is None and ckpt_target_series is not None:
+                config.target_series = ckpt_target_series if isinstance(ckpt_target_series, list) else [ckpt_target_series]
+        except Exception:
+            pass
         
         # Get model state (new format preferred, fallback to old format)
         model_state = checkpoint.get('model_state')

dfm_python/ssm/kalman.py

@@ -204,7 +204,7 @@ class DFMKalmanFilter:
                 "DFMKalmanFilter parameters not initialized. "
                 "Call update_parameters() first."
             )
-        
+
         # Get filtered states first (needed for smoother)
         transition_offsets = getattr(self._pykalman, 'transition_offsets', None)
         observation_offsets = getattr(self._pykalman, 'observation_offsets', None)
@@ -241,7 +241,7 @@ class DFMKalmanFilter:
         
         # Run filter
         predicted_state_means, predicted_state_covariances, _, filtered_state_means, filtered_state_covariances = run_filter()
-        
+
         filter_time = time_module.time() - filter_start
         _logger.info(f"    Filter: Completed in {filter_time:.2f}s ({filter_time/T*1000:.2f}ms/timestep)")
         
@@ -266,7 +266,7 @@ class DFMKalmanFilter:
                 # This is much cheaper than full eigendecomposition (O(m²) vs O(m³))
                 from ..config.constants import MIN_EIGENVALUE
                 regularization = max(1e-6, MIN_EIGENVALUE * 100)  # 1e-4 for stability
-                
+
                 _logger.info(f"    Smooth: Stabilizing {len(predicted_state_covariances)} covariance matrices "
                             f"(regularization={regularization:.2e})")
                 

{dfm_python-0.5.58.dist-info → dfm_python-0.5.59.dist-info}/METADATA

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: dfm-python
-Version: 0.5.58
+Version: 0.5.59
 Summary: Dynamic Factor Model (DFM) estimation and forecasting in Python
 Project-URL: Homepage, https://github.com/yourusername/dfm-python
 Project-URL: Documentation, https://github.com/yourusername/dfm-python#readme
@@ -18,6 +18,7 @@ Classifier: Programming Language :: Python :: 3.11
 Classifier: Programming Language :: Python :: 3.12
 Classifier: Topic :: Scientific/Engineering
 Requires-Python: >=3.10
+Requires-Dist: build>=1.4.0
 Requires-Dist: einops>=0.8.1
 Requires-Dist: hydra-core>=1.3.2
 Requires-Dist: matplotlib>=3.5.0

{dfm_python-0.5.58.dist-info → dfm_python-0.5.59.dist-info}/RECORD

@@ -22,7 +22,7 @@ dfm_python/encoder/pca.py,sha256=C1U0bGOVdr-HQaJYD57MyA7RLaEr70LTXWgmo6P-sUU,105
 dfm_python/encoder/simple_autoencoder.py,sha256=9y5C6b3MEgNO3oZ0DjA2aX_AnBcX6eqG_9zFJPgHAjk,10717
 dfm_python/functional/__init__.py,sha256=J9EYUqQnstmY0Wq-g8v3Z3We01gOW99x6M9MxPcSp4w,329
 dfm_python/functional/dfm_block.py,sha256=lifzLA0fHudj-tnqCK4kMy09zVUtlJMPb2uw635yH4c,21006
-dfm_python/functional/em.py,sha256=WGZEuftelHgknSqi1Kb1GTo6rWDGWOkxjvcj1siYfCs,64168
+dfm_python/functional/em.py,sha256=4U6ftX8PWnMdc_ODYuPSaVVD6rqY2ykkV5zBBdO9zU8,64164
 dfm_python/logger/__init__.py,sha256=ZsUPp5Gd_FTWvgtBweOMafxc14hZRSOdfG8yaYskRgI,1622
 dfm_python/logger/inference_logger.py,sha256=t6K1Di-QkqIvcbSUGW1sCeWD84ZTgoC27oB2FWaY1SI,10314
 dfm_python/logger/logger.py,sha256=drTD5o4VPb34YKjux7MaEnI_k5a93AOlhQciMZVTkAE,10379
@@ -30,7 +30,7 @@ dfm_python/logger/train_logger.py,sha256=BGtlJt-FP3WC4u_oGpfNrIYctIXamkB1gVyMhhn
 dfm_python/models/__init__.py,sha256=NHv79jHILrnJG4Auuo3fspi1FtLLedWfLS9GK9WZx6A,552
 dfm_python/models/base.py,sha256=igYjuGpcEsaDbsRbz3_8mvYkYuMKGZYdMqVJodylg0E,11024
 dfm_python/models/ddfm.py,sha256=AjjcCdyFo4LJSfDZjMVd_FNa0ggmI6VnvhZhkFb_tLU,50700
-dfm_python/models/dfm.py,sha256=uGG4m0cOK6cjzSA2-9YjRpf-rtMIx_6rMvyfRg_rh14,81261
+dfm_python/models/dfm.py,sha256=36_Z6WabQC48SaCEBu7jSyU8IMSORCv1UanlsVUACVM,84407
 dfm_python/numeric/builder.py,sha256=zTaZFNqR0B7rb5hPkJs4I7beUYPMu44K9AaTIJniD8s,5852
 dfm_python/numeric/estimator.py,sha256=OPe7tAiTn0Jj_bL1ywOe47vJ37oeSMC2AtAX2FxLe5s,47534
 dfm_python/numeric/stability.py,sha256=PIR2XvxHGQ4uD1QuIkGGKuNmkfAasQdhoZlH2BWt56U,43922
@@ -38,7 +38,7 @@ dfm_python/numeric/statistic.py,sha256=Y9rrv18OKBiholy-R5Bz9tzGCkaziMn0Jk4Ca_KS0
 dfm_python/numeric/tent.py,sha256=tsmmYfj3UbV8caVUpeP-7dzIvslmTVl6p-l5X5--AqU,12229
 dfm_python/numeric/validator.py,sha256=FPgVC19PsWRgx40S9tDG0x0ccgiAUhnMFyf-IcfxOtA,39927
 dfm_python/ssm/__init__.py,sha256=AmmiJPlAy1bctM9AVso7eWKdPMrUbCfAZQHgqZv-KLY,784
-dfm_python/ssm/kalman.py,sha256=jJcY2AhFJGeELC5pbSRbT5uDN2mR7rC_sV44kNH-CMI,31771
+dfm_python/ssm/kalman.py,sha256=uwkMzUDqQOrhTszxDXalQP3xl0856o00Uizeiry6_2U,31739
 dfm_python/utils/__init__.py,sha256=WLczOBMPw-_VIvuzoCGHouf-S5dcjhrUnnaTDfNouzQ,5382
 dfm_python/utils/checkpoint.py,sha256=npdYsNg74KE7wUtU25Wuqs846oXfwYxUJcEzKnh-k6o,11201
 dfm_python/utils/errors.py,sha256=BKMjYldXjdFg-O_pIBo1-WITInNXvNrHk-6IyFHMf-k,1861
@@ -47,6 +47,6 @@ dfm_python/utils/loss.py,sha256=1KcYwFJmY9l-Uh6Sk-dFpJ2kgMfm0u41Gmgu-5NzaAs,3878
 dfm_python/utils/metric.py,sha256=AC_TnJ3eIXiF_maNNexSHzXpLv5Hrf1W88X5pJx9C-Q,15327
 dfm_python/utils/misc.py,sha256=NprRqONFlqYE_LbIbPE6G0CDbYNGXKikaoH80SpVY4E,9596
 dfm_python/utils/validation.py,sha256=WGosO4AY7HnvUNpzAptzRWHXeu75Uzv-qiZjHSFgn8M,1437
-dfm_python-0.5.58.dist-info/METADATA,sha256=qMnsGZT2apMLpxsvv_iTjXR4Zc2P5JCiDwTvc9NNIIc,13103
-dfm_python-0.5.58.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
-dfm_python-0.5.58.dist-info/RECORD,,
+dfm_python-0.5.59.dist-info/METADATA,sha256=wtA3kxn3Smr8abJUD8bi_sxxuBRlnOp9EJY_MrXIvVk,13131
+dfm_python-0.5.59.dist-info/WHEEL,sha256=WLgqFyCfm_KASv4WHyYy0P3pM_m7J5L9k2skdKLirC8,87
+dfm_python-0.5.59.dist-info/RECORD,,