Merge pull request #210 from google-research/rajat_dev

updating benchmarks to point to v2.0

experiments/extended_benchmarks/run_timesfm.py

@@ -54,7 +54,10 @@ dataset_names = [
     "hospital",
 ]
 
-context_dict = {
+
+context_dict_v2 = {}
+
+context_dict_v1 = {
     "cif_2016": 32,
     "tourism_yearly": 64,
     "covid_deaths": 64,
@@ -71,7 +74,7 @@ context_dict = {
     "m4_yearly": 64,
 }
 
-_MODEL_PATH = flags.DEFINE_string("model_path", "google/timesfm-1.0-200m",
+_MODEL_PATH = flags.DEFINE_string("model_path", "google/timesfm-2.0-500m-jax",
                                   "Path to model")
 _BATCH_SIZE = flags.DEFINE_integer("batch_size", 64, "Batch size")
 _HORIZON = flags.DEFINE_integer("horizon", 128, "Horizon")
@@ -84,14 +87,27 @@ QUANTILES = list(np.arange(1, 10) / 10.0)
 
 def main():
   results_list = []
+  model_path = _MODEL_PATH.value
+  num_layers = 20
+  max_context_len = 512
+  use_positional_embedding = True
+  context_dict = context_dict_v1
+  if "2.0" in model_path:
+    num_layers = 50
+    use_positional_embedding = False
+    max_context_len = 2048
+    context_dict = context_dict_v2
+
   tfm = timesfm.TimesFm(
       hparams=timesfm.TimesFmHparams(
-          backend=_BACKEND.value,
+          backend="gpu",
-          per_core_batch_size=_BATCH_SIZE.value,
+          per_core_batch_size=32,
-          horizon_len=_HORIZON.value,
+          horizon_len=128,
+          num_layers=num_layers,
+          context_len=max_context_len,
+          use_positional_embedding=use_positional_embedding,
       ),
-      checkpoint=timesfm.TimesFmCheckpoint(
+      checkpoint=timesfm.TimesFmCheckpoint(huggingface_repo_id=model_path),
-          huggingface_repo_id=_MODEL_PATH.value),
   )
   run_id = np.random.randint(100000)
   model_name = "timesfm"
@@ -102,7 +118,8 @@ def main():
     if dataset in context_dict:
       context_len = context_dict[dataset]
     else:
-      context_len = 512
+      context_len = max_context_len
+
     train_df = exp.train_df
     freq = exp.freq
     init_time = time.time()
@@ -113,6 +130,7 @@ def main():
         model_name=model_name,
         forecast_context_len=context_len,
         num_jobs=_NUM_JOBS.value,
+        normalize=True,
     )
     total_time = time.time() - init_time
     time_df = pd.DataFrame({"time": [total_time], "model": model_name})

experiments/long_horizon_benchmarks/run_eval.py

@@ -32,8 +32,8 @@ _BATCH_SIZE = flags.DEFINE_integer("batch_size", 64,
                                    "Batch size for the randomly sampled batch")
 _DATASET = flags.DEFINE_string("dataset", "etth1", "The name of the dataset.")
 
-_MODEL_PATH = flags.DEFINE_string("model_path", "./timesfm_q10_20240501",
+_MODEL_PATH = flags.DEFINE_string("model_path", "google/timesfm-2.0-500m-jax",
-                                  "The name of the dataset.")
+                                  "The name of the model.")
 _DATETIME_COL = flags.DEFINE_string("datetime_col", "date",
                                     "Column having datetime.")
 _NUM_COV_COLS = flags.DEFINE_list("num_cov_cols", None,
@@ -43,7 +43,7 @@ _CAT_COV_COLS = flags.DEFINE_list("cat_cov_cols", None,
 _TS_COLS = flags.DEFINE_list("ts_cols", None, "Columns of time-series features")
 _NORMALIZE = flags.DEFINE_bool("normalize", True,
                                "normalize data for eval or not")
-_CONTEXT_LEN = flags.DEFINE_integer("context_len", 512,
+_CONTEXT_LEN = flags.DEFINE_integer("context_len", 2048,
                                     "Length of the context window")
 _PRED_LEN = flags.DEFINE_integer("pred_len", 96, "prediction length.")
 _BACKEND = flags.DEFINE_string("backend", "gpu", "backend to use")
@@ -177,9 +177,12 @@ def eval():
   else:
     model = timesfm.TimesFm(
         hparams=timesfm.TimesFmHparams(
-            backend=_BACKEND.value,
+            backend="gpu",
-            per_core_batch_size=_BATCH_SIZE.value,
+            per_core_batch_size=32,
-            horizon_len=_PRED_LEN.value,
+            horizon_len=128,
+            num_layers=50,
+            context_len=_CONTEXT_LEN.value,
+            use_positional_embedding=False,
         ),
         checkpoint=timesfm.TimesFmCheckpoint(huggingface_repo_id=model_path),
     )

src/timesfm/timesfm_base.py

@@ -636,6 +636,7 @@ class TimesFmBase:
       model_name: str = "timesfm",
       window_size: int | None = None,
       num_jobs: int = 1,
+      normalize: bool = False,
       verbose: bool = True,
   ) -> pd.DataFrame:
     """Forecasts on a list of time series.
@@ -654,6 +655,7 @@ class TimesFmBase:
       window_size: window size of trend + residual decomposition. If None then
         we do not do decomposition.
       num_jobs: number of parallel processes to use for dataframe processing.
+      normalize: normalize context before forecasting or not.
       verbose: output model states in terminal.
 
     Returns:
@@ -698,6 +700,7 @@ class TimesFmBase:
     freq_inps = [freq_map(freq)] * len(new_inputs)
     _, full_forecast = self.forecast(new_inputs,
                                      freq=freq_inps,
+                                     normalize=normalize,
                                      window_size=window_size)
     if verbose:
       print("Finished forecasting.")