|
1 | 1 | # ACE: AI2 Climate Emulator
|
2 | 2 | This repo contains the inference code accompanying "ACE: A fast, skillful learned global atmospheric model for climate prediction" ([arxiv:2310.02074](https://arxiv.org/abs/2310.02074)).
|
3 | 3 |
|
4 |
| -## DISCLAIMER |
5 |
| -This is rapidly changing research software. We make no guarantees of maintaining backwards compatibility. |
6 |
| - |
7 | 4 | ## Quickstart
|
8 | 5 |
|
9 | 6 | ### 1. Install
|
@@ -35,27 +32,12 @@ enough data to span the desired prediction period. The checkpoint is also availa
|
35 | 32 | `gs://ai2cm-public-requester-pays/2023-11-29-ai2-climate-emulator-v1/checkpoints/ace_ckpt.tar`.
|
36 | 33 |
|
37 | 34 | ### 3. Update configuration and run
|
38 |
| -Update the paths in the [example config](examples/config-inference.yaml). Then in the |
| 35 | +Update the paths in the [example config](fme/docs/inference-config.yaml). Then in the |
39 | 36 | `fme` conda environment, run inference with:
|
40 | 37 | ```
|
41 |
| -python -m fme.fcn_training.inference.inference examples/config-inference.yaml |
| 38 | +python -m fme.ace.inference fme/docs/inference-config.yaml |
42 | 39 | ```
|
43 | 40 |
|
44 |
| -## Configuration options |
45 |
| -See the `InferenceConfig` class in [this file](fme/fme/fcn_training/inference/inference.py) for |
46 |
| -description of configuration options. The [example config](examples/config-inference.yaml) |
47 |
| -shows some useful defaults for performing a 400-step simulation (100 days, with the 6-hour time step). |
48 |
| - |
49 |
| -## Performance |
50 |
| -While inference can be performed without a GPU, it may be very slow in that case. In addition, |
51 |
| -I/O performance is critical for fast inference due to loading of forcing data and target data |
52 |
| -during inference. |
53 |
| - |
54 |
| -## Analyzing output |
55 |
| -Various climate performance metrics are computed online by the inference code. These can be viewed via |
56 |
| -[wandb](https://wandb.ai) by setting `logging.log_to_wandb` to true and updating `logging.entity` |
57 |
| -to your wandb entity. Additionally, raw output data is saved to netCDF by the inference code. |
58 |
| - |
59 | 41 | ## Available datasets
|
60 | 42 | Two versions of the dataset described in [arxiv:2310.02074](https://arxiv.org/abs/2310.02074)
|
61 | 43 | are available:
|
|
0 commit comments