metnet3 pytorch

MetNet-3 - Pytorch

在 Pytorch 實現 MetNet 3（來自 Google Deepmind 的 SOTA 神經天氣模型）

模型架構相當不起眼。它基本上是一個帶有特定性能良好的視覺轉換器的 U 網。這篇論文最有趣的事情可能最終是第 4.3.2 節的損失縮放

欣賞

• StabilityAI、A16Z 開源 AI 資助計劃，以及？感謝慷慨的贊助，以及我的其他贊助商，為我提供了開源當前人工智慧研究的獨立性

安裝

$ pip install metnet3-pytorch

用法

 import torch
from metnet3_pytorch import MetNet3

metnet3 = MetNet3 (
    dim = 512 ,
    num_lead_times = 722 ,
    lead_time_embed_dim = 32 ,
    input_spatial_size = 624 ,
    attn_dim_head = 8 ,
    hrrr_channels = 617 ,
    input_2496_channels = 2 + 14 + 1 + 2 + 20 ,
    input_4996_channels = 16 + 1 ,
    precipitation_target_bins = dict (
        mrms_rate = 512 ,
        mrms_accumulation = 512 ,
    ),
    surface_target_bins = dict (
        omo_temperature = 256 ,
        omo_dew_point = 256 ,
        omo_wind_speed = 256 ,
        omo_wind_component_x = 256 ,
        omo_wind_component_y = 256 ,
        omo_wind_direction = 180
    ),
    hrrr_loss_weight = 10 ,
    hrrr_norm_strategy = 'sync_batchnorm' ,  # this would use a sync batchnorm to normalize the input hrrr and target, without having to precalculate the mean and variance of the hrrr dataset per channel
    hrrr_norm_statistics = None             # you can also also set `hrrr_norm_strategy = "precalculated"` and pass in the mean and variance as shape `(2, 617)` through this keyword argument
)

# inputs

lead_times = torch . randint ( 0 , 722 , ( 2 ,))
hrrr_input_2496 = torch . randn (( 2 , 617 , 624 , 624 ))
hrrr_stale_state = torch . randn (( 2 , 1 , 624 , 624 ))
input_2496 = torch . randn (( 2 , 39 , 624 , 624 ))
input_4996 = torch . randn (( 2 , 17 , 624 , 624 ))

# targets

precipitation_targets = dict (
    mrms_rate = torch . randint ( 0 , 512 , ( 2 , 512 , 512 )),
    mrms_accumulation = torch . randint ( 0 , 512 , ( 2 , 512 , 512 )),
)

surface_targets = dict (
    omo_temperature = torch . randint ( 0 , 256 , ( 2 , 128 , 128 )),
    omo_dew_point = torch . randint ( 0 , 256 , ( 2 , 128 , 128 )),
    omo_wind_speed = torch . randint ( 0 , 256 , ( 2 , 128 , 128 )),
    omo_wind_component_x = torch . randint ( 0 , 256 , ( 2 , 128 , 128 )),
    omo_wind_component_y = torch . randint ( 0 , 256 , ( 2 , 128 , 128 )),
    omo_wind_direction = torch . randint ( 0 , 180 , ( 2 , 128 , 128 ))
)

hrrr_target = torch . randn ( 2 , 617 , 128 , 128 )

total_loss , loss_breakdown = metnet3 (
    lead_times = lead_times ,
    hrrr_input_2496 = hrrr_input_2496 ,
    hrrr_stale_state = hrrr_stale_state ,
    input_2496 = input_2496 ,
    input_4996 = input_4996 ,
    precipitation_targets = precipitation_targets ,
    surface_targets = surface_targets ,
    hrrr_target = hrrr_target
)

total_loss . backward ()

# after much training from above, you can predict as follows

metnet3 . eval ()

surface_preds , hrrr_pred , precipitation_preds = metnet3 (
    lead_times = lead_times ,
    hrrr_input_2496 = hrrr_input_2496 ,
    hrrr_stale_state = hrrr_stale_state ,
    input_2496 = input_2496 ,
    input_4996 = input_4996 ,
)


# Dict[str, Tensor], Tensor, Dict[str, Tensor] 

托多

• 計算出所有交叉熵和 MSE 損失

• 透過追蹤訓練期間目標的運行平均值和方差，自動處理 HRRR 所有通道的歸一化（使用同步批量歸一化作為 hack）

• 允許研究者傳入自己的 HRRR 歸一化變量

• 建構規範的所有輸入，也確保 hrrr 輸入標準化，提供對 hrrr 預測進行非標準化的選項

• 確保模型可以輕鬆保存和加載，並以不同的方式處理 hrrr 範數

• 弄清楚拓撲嵌入，請諮詢神經天氣研究人員

引文

 @article { Andrychowicz2023DeepLF ,
    title   = { Deep Learning for Day Forecasts from Sparse Observations } ,
    author  = { Marcin Andrychowicz and Lasse Espeholt and Di Li and Samier Merchant and Alexander Merose and Fred Zyda and Shreya Agrawal and Nal Kalchbrenner } ,
    journal = { ArXiv } ,
    year    = { 2023 } ,
    volume  = { abs/2306.06079 } ,
    url     = { https://api.semanticscholar.org/CorpusID:259129311 }
}

 @inproceedings { ElNouby2021XCiTCI ,
    title   = { XCiT: Cross-Covariance Image Transformers } ,
    author  = { Alaaeldin El-Nouby and Hugo Touvron and Mathilde Caron and Piotr Bojanowski and Matthijs Douze and Armand Joulin and Ivan Laptev and Natalia Neverova and Gabriel Synnaeve and Jakob Verbeek and Herv{'e} J{'e}gou } ,
    booktitle = { Neural Information Processing Systems } ,
    year    = { 2021 } ,
    url     = { https://api.semanticscholar.org/CorpusID:235458262 }
}