use .sizes instead of .dims to avoid warning and future deprecation

graphcast/autoregressive.py

@@ -112,7 +112,7 @@ def _validate_targets_and_forcings(self, targets, forcings):
                        f'forcings, which isn\'t allowed: {overlap}')
 
   def _update_inputs(self, inputs, next_frame):
-    num_inputs = inputs.dims['time']
+    num_inputs = inputs.sizes['time']
 
     predicted_or_forced_inputs = next_frame[list(inputs.keys())]
 
@@ -199,7 +199,7 @@ def one_step_prediction(inputs, scan_variables):
       return next_inputs, flat_pred
 
     if self._gradient_checkpointing:
-      scan_length = targets_template.dims['time']
+      scan_length = targets_template.sizes['time']
       if scan_length <= 1:
         logging.warning(
             'Skipping gradient checkpointing for sequence length of 1')

graphcast/rollout.py

@@ -124,7 +124,7 @@ def chunked_prediction_generator(
   if "datetime" in forcings.coords:
     del forcings.coords["datetime"]
 
-  num_target_steps = targets_template.dims["time"]
+  num_target_steps = targets_template.sizes["time"]
   num_chunks, remainder = divmod(num_target_steps, num_steps_per_chunk)
   if remainder != 0:
     raise ValueError(
@@ -202,7 +202,7 @@ def _get_next_inputs(
   next_inputs = next_frame[next_inputs_keys]
 
   # Apply concatenate next frame with inputs, crop what we don't need.
-  num_inputs = prev_inputs.dims["time"]
+  num_inputs = prev_inputs.sizes["time"]
   return (
       xarray.concat(
           [prev_inputs, next_inputs], dim="time", data_vars="different")