Version
0.8.0
On which installation method(s) does this occur?
source
Describe the issue
Graphcast has the following input coords
input_coords = OrderedDict(
{
"batch": np.empty(0),
"time": np.empty(0),
"lead_time": np.array(
[
np.timedelta64(-6, "h"),
np.timedelta64(0, "h"),
]
),
"variable": np.array(VARIABLES),
"lat": np.linspace(-90, 90, 721, endpoint=True),
"lon": np.linspace(0, 360, 1440, endpoint=False),
}
)
Which would indicate that the time dimension is free but in reality the model errors (under the hood in evaluation) if more than a single timestamp is passed. I would suggest adding a simple loop over the timestamps passed in to alleviate this issue.
Version
0.8.0
On which installation method(s) does this occur?
source
Describe the issue
Graphcast has the following input coords
Which would indicate that the
timedimension is free but in reality the model errors (under the hood in evaluation) if more than a single timestamp is passed. I would suggest adding a simple loop over the timestamps passed in to alleviate this issue.