HFAiLab/OpenCastKit

FourCastNet+

jafri opened this issue · 8 comments

jafri commented

Do you plan on releasing model weights / training script for FourCastNet+?

VachelHU commented

The model weights have been released in https://huggingface.co/hf-ai/OpenCastKit

jafri commented

Thank you @VachelHU, the backbone weights look good, will try them out soon

I see precipitation.pt is missing for FourCastNet+, is it because backbone.pt can be used to predict precipitation?

I'd also like to test out the GraphCast model, do you have sample inference code like you provided for FourCastNet?

Appreciate your work here, cool project!

msxyz commented

Could you please list names of 4 extra channels?

VachelHU commented

the sine and cosine of the local time of day, ref to https://arxiv.org/abs/2212.12794

JianxGao commented

Thanks to @VachelHU for providing the code and the names of extra channels. But, the sine and cosine of the local time of day is implemented in

OpenCastKit/data_factory/convert_era5_hourly.py

Lines 102 to 110 in 6e3f934

for i in range(len(valid_data['time'])):
cursor_time = cursor_time + timedelta(hours=6) * i
tmp_feats = fetch_time_features(cursor_time)
time_features.append(tmp_feats)
time_features = np.asarray(time_features)
Xt0 = np.concatenate([Xt0[:, 1:], time_features[:-2]], axis=-1)
Xt1 = np.concatenate([Xt1[:, 1:], time_features[1:-1]], axis=-1)
Yt = np.concatenate([Yt[:, 1:], time_features[2:]], axis=-1)

Now, the number of channel is 22, which is not equivalent to 24.
In https://www.high-flyer.cn/blog/opencast/, Fourcast+ is trained with "时间和地理信息". Could you please tell me the 2 extra channels? Sine or Cosine of latitude and longitude?

  • 👍1
oubahe commented

Thanks to @VachelHU for providing the code and the names of extra channels. But, the sine and cosine of the local time of day is implemented in

OpenCastKit/data_factory/convert_era5_hourly.py

Lines 102 to 110 in 6e3f934

for i in range(len(valid_data['time'])):
cursor_time = cursor_time + timedelta(hours=6) * i
tmp_feats = fetch_time_features(cursor_time)
time_features.append(tmp_feats)
time_features = np.asarray(time_features)
Xt0 = np.concatenate([Xt0[:, 1:], time_features[:-2]], axis=-1)
Xt1 = np.concatenate([Xt1[:, 1:], time_features[1:-1]], axis=-1)
Yt = np.concatenate([Yt[:, 1:], time_features[2:]], axis=-1)

Now, the number of channel is 22, which is not equivalent to 24. In https://www.high-flyer.cn/blog/opencast/, Fourcast+ is trained with "时间和地理信息". Could you please tell me the 2 extra channels? Sine or Cosine of latitude and longitude?

Hi, i got this problem too, what are 4 extra channels specifically？Suppose it contains sine and cosine of the local time of day, what are the other 2 channels about position? Hope you don't hesitate to enlighten me.

JianxGao commented

Thanks to @VachelHU for providing the code and the names of extra channels. But, the sine and cosine of the local time of day is implemented in

OpenCastKit/data_factory/convert_era5_hourly.py

Lines 102 to 110 in 6e3f934

for i in range(len(valid_data['time'])):
cursor_time = cursor_time + timedelta(hours=6) * i
tmp_feats = fetch_time_features(cursor_time)
time_features.append(tmp_feats)
time_features = np.asarray(time_features)
Xt0 = np.concatenate([Xt0[:, 1:], time_features[:-2]], axis=-1)
Xt1 = np.concatenate([Xt1[:, 1:], time_features[1:-1]], axis=-1)
Yt = np.concatenate([Yt[:, 1:], time_features[2:]], axis=-1)

Now, the number of channel is 22, which is not equivalent to 24. In https://www.high-flyer.cn/blog/opencast/, Fourcast+ is trained with "时间和地理信息". Could you please tell me the 2 extra channels? Sine or Cosine of latitude and longitude?

Hi, i got this problem too, what are 4 extra channels specifically？Suppose it contains sine and cosine of the local time of day, what are the other 2 channels about position? Hope you don't hesitate to enlighten me.

Now, do you have any ideas?🥹

oubahe commented

Thanks to @VachelHU for providing the code and the names of extra channels. But, the sine and cosine of the local time of day is implemented in

OpenCastKit/data_factory/convert_era5_hourly.py

Lines 102 to 110 in 6e3f934

for i in range(len(valid_data['time'])):
cursor_time = cursor_time + timedelta(hours=6) * i
tmp_feats = fetch_time_features(cursor_time)
time_features.append(tmp_feats)
time_features = np.asarray(time_features)
Xt0 = np.concatenate([Xt0[:, 1:], time_features[:-2]], axis=-1)
Xt1 = np.concatenate([Xt1[:, 1:], time_features[1:-1]], axis=-1)
Yt = np.concatenate([Yt[:, 1:], time_features[2:]], axis=-1)

Now, the number of channel is 22, which is not equivalent to 24. In https://www.high-flyer.cn/blog/opencast/, Fourcast+ is trained with "时间和地理信息". Could you please tell me the 2 extra channels? Sine or Cosine of latitude and longitude?

Hi, i got this problem too, what are 4 extra channels specifically？Suppose it contains sine and cosine of the local time of day, what are the other 2 channels about position? Hope you don't hesitate to enlighten me.

Now, do you have any ideas?🥹

Regrettably, i have no idea about 4 extra channels.