GEFS-HRRR Corrdiff Devs (#703)

* merge gefs-hrrr model

* update gefs-hrrr

* Update gefs_hrrr.py

* Update gefs_hrrr.yaml

edit stats path

* Update train.py

* Update gefs_hrrr.py

* Add unit test for GEFS Corrdiff regression loss and lead-time aware songunet

* Format so init prob_channel scalar factor by channel length

* Add docstrings and license for dataloader, formating in general

* Delete examples/generative/corrdiff/stats.json

* Update loss.py

* Update song_unet.py

* Update utils.py

* Update README.md

* Update README.md

* Update CHANGELOG.md

* Fixing generalization to unit test case for corrdiff utils

* Update unit test for fixing diffusion step signature

* Reformat with black

---------

Co-authored-by: Tao Ge <[email protected]>
Co-authored-by: Tao Ge <[email protected]>

Author: 3 people

CHANGELOG.md

@@ -18,6 +18,8 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 - Bistride Multiscale MeshGraphNet example.
 - FIGConvUNet model and example.
 - The Transolver model.
+- Incoporated CorrDiff-GEFS-HRRR model into CorrDiff, with lead-time aware SongUNet and
+  cross entropy loss.
 
 ### Changed
 

examples/generative/corrdiff/README.md

@@ -30,8 +30,11 @@ CorrDiff training is handled by `train.py` and controlled by YAML configuration
 ### Training the regression model
 To train the CorrDiff-Mini regression model, we use the main configuration file [config_training_mini_regression.yaml](conf/config_training_mini_regression.yaml). This includes the following components:
 * The HRRR-Mini dataset: [conf/dataset/hrrrmini.yaml](conf/dataset/hrrrmini.yaml)
+* The GEFS-HRRR dataset: [conf/dataset/hrrrmini.yaml](conf/dataset/gefs_hrrr.yaml)
 * The CorrDiff-Mini regression model: [conf/model/corrdiff_regression_mini.yaml](conf/model/corrdiff_regression_mini.yaml)
 * The CorrDiff-Mini regression training options: [conf/training/corrdiff_regression_mini.yaml](conf/training/corrdiff_regression_mini.yaml)
+* The CorrDiff-GEFS-HRRR regression training options: [conf/model/corrdiff_regression_mini.yaml](conf/training/config_training_gefs_regression.yaml)
+  
 To start the training, run:
 ```bash
 python train.py --config-name=config_training_mini_regression.yaml ++dataset.data_path=</path/to/dataset>/hrrr_mini_train.nc ++dataset.stats_path=</path/to/dataset>/stats.json

examples/generative/corrdiff/conf/config_generate_gefs_hrrr.yaml

@@ -0,0 +1,36 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+hydra:
+  job:
+    chdir: true
+    name: gefs_hrrr_generation
+  run:
+    dir: output/${hydra:job.name}
+
+# Get defaults
+defaults:
+
+  # Dataset
+  - dataset/gefs_hrrr
+
+  # Sampler
+  - sampler/stochastic
+  #- sampler/deterministic
+
+  # Generation
+  - generation/patched_based_gefs_hrrr
+    #- generation/patched_based

examples/generative/corrdiff/conf/config_training_gefs_diffusion.yaml

@@ -0,0 +1,34 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+hydra:
+  job:
+    chdir: true
+    name: gefs_hrrr_diffusion
+  run:
+    dir: ./outputs/${hydra:job.name}
+
+# Get defaults
+defaults:
+
+  # Dataset
+  - dataset/gefs_hrrr
+
+  # Model
+  - model/corrdiff_patched_diffusion_gefs_hrrr
+
+  # Training
+  - training/corrdiff_patched_diffusion_gefs_hrrr

examples/generative/corrdiff/conf/config_training_gefs_regression.yaml

@@ -0,0 +1,34 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+hydra:
+  job:
+    chdir: true
+    name: gefs_hrrr_regression
+  run:
+    dir: ./outputs/${hydra:job.name}
+
+# Get defaults
+defaults:
+
+  # Dataset
+  - dataset/gefs_hrrr
+
+  # Model
+  - model/corrdiff_regression_gefs_hrrr
+
+  # Training
+  - training/corrdiff_regression_gefs_hrrr

examples/generative/corrdiff/conf/dataset/cwb_train.yaml

@@ -26,4 +26,4 @@ ds_factor: 4
 min_path: null
 max_path: null
 global_means_path: null
-global_stds_path: null
+global_stds_path: null

examples/generative/corrdiff/conf/dataset/gefs_hrrr.yaml

@@ -0,0 +1,26 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+type: gefs_hrrr
+data_path: /data
+stats_path: modulus/examples/generative/corrdiff/stats.json
+output_variables: ["u10m", "v10m", "t2m", "precip", "cat_snow", "cat_ice", "cat_freez", "cat_rain", "cat_none"]
+prob_variables: ["cat_snow", "cat_ice", "cat_freez", "cat_rain"]
+input_surface_variables: ["u10m", "v10m", "t2m", "q2m", "sp", "msl", "precipitable_water"]
+input_isobaric_variables: ['u1000', 'u925', 'u850', 'u700', 'u500', 'u250', 'v1000', 'v925', 'v850', 'v700', 'v500', 'v250', 'z1000', 'z925', 'z850', 'z700', 'z500', 'z200', 't1000', 't925', 't850', 't700', 't500', 't100',  'r1000', 'r925', 'r850', 'r700', 'r500', 'r100']
+ds_factor: 4
+train: False
+hrrr_window: [[1,1057], [4,1796]] # need dims to be divisible by 16 [[0,1024], [0,1024]]

examples/generative/corrdiff/conf/dataset/hrrrmini.yaml

@@ -17,4 +17,4 @@
 type: hrrr_mini
 data_path: /data/corrdiff-mini/hrrr_mini_train.nc
 stats_path: /data/corrdiff-mini/stats.json
-output_variables: ['10u', '10v']
+output_variables: ['10u', '10v']

examples/generative/corrdiff/conf/generation/patched_based_gefs_hrrr.yaml

@@ -0,0 +1,68 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+num_ensembles: 1
+  # Number of ensembles to generate per input
+seed_batch_size: 1
+  # Size of the batched inference
+inference_mode: all
+  # Choose between "all" (regression + diffusion), "regression" or "diffusion"
+patch_size: 448
+patch_shape_x: 448
+patch_shape_y: 448
+  # Patch size. Patch-based sampling will be utilized if these dimensions differ from 
+  # img_shape_x and img_shape_y
+overlap_pixels: 4
+  # Number of overlapping pixels between adjacent patches
+boundary_pixels: 2
+  # Number of boundary pixels to be cropped out. 2 is recommanded to address the boundary
+  # artifact.
+hr_mean_conditioning: true
+gridtype: learnable
+N_grid_channels: 100
+sample_res: full
+  # Sampling resolution
+times_range: null
+times:
+  - "2024011212f00"
+  - "2024011212f03"
+  - "2024011212f06"
+  - "2024011212f09"
+  - "2024011212f12"
+  - "2024011212f15"
+  - "2024011212f18"
+  - "2024011212f21"
+  - "2024011212f24"
+
+has_lead_time: true
+
+perf:
+  force_fp16: false
+    # Whether to force fp16 precision for the model. If false, it'll use the precision
+    # specified upon training.
+  use_torch_compile: false
+    # whether to use torch.compile on the diffusion model
+    # this will make the first time stamp generation very slow due to compilation overheads
+    # but will significantly speed up subsequent inference runs
+  num_writer_workers: 1
+    # number of workers to use for writing file
+    # To support multiple workers a threadsafe version of the netCDF library must be used
+
+io:
+  res_ckpt_filename: EDMPrecondSRV2_updated.0.5821440.mdlus
+    # Checkpoint filename for the diffusion model  
+  reg_ckpt_filename: UNet_updated.0.1960960.mdlus
+    # Checkpoint filename for the mean predictor model

examples/generative/corrdiff/conf/model/corrdiff_patched_diffusion_gefs_hrrr.yaml

@@ -0,0 +1,24 @@
+# SPDX-FileCopyrightText: Copyright (c) 2023 - 2024 NVIDIA CORPORATION & AFFILIATES.
+# SPDX-FileCopyrightText: All rights reserved.
+# SPDX-License-Identifier: Apache-2.0
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+name: lt_aware_patched_diffusion
+  # Name of the preconditioner
+hr_mean_conditioning: True
+  # High-res mean (regression's output) as additional condition
+scale_cond_input: True
+  # If true, also scales the input conditioning
+  # For backward compatibility, this is true by default
+  # We recommend setting this to false for new training runs