Diffusion development branch

NOTICE

@@ -1,3 +1,43 @@
+=======================================================================
+NVLABS/EDM (Elucidating the Design of Diffusion Models)
+
+This software incorporates code from the 'edm' repository.
+
+Original Copyright (c) 2022, NVIDIA CORPORATION & AFFILIATES. All rights reserved.
+
+The source code is available at:
+https://github.com/NVlabs/edm
+
+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
+
+=======================================================================
+google-deepmind/graphcast (several associated papers)
+
+This software incorporates code from the 'google-deepmind/graphcast' repository, with adaptations.
+
+Original Copyright 2024 DeepMind Technologies Limited.
+
+The source code is available at:
+https://github.com/google-deepmind/graphcast
+
+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
+
+=======================================================================
+facebookresearch/DiT (Scalable Diffusion Models with Transformers (DiT))
+
+This software incorporates code from the 'facebookresearch/DiT' repository, with adaptations.
+
+The source code is available at:
+https://github.com/facebookresearch/DiT
+
+The code and model weights are licensed under CC-BY-NC. 
+See https://raw.githubusercontent.com/facebookresearch/DiT/refs/heads/main/LICENSE.txt for details.
 This project includes code derived from project "DINOv2: Learning Robust Visual Features without Supervision",
 originally developed by Meta Platforms, Inc. and affiliates,
 licensed under the Apache License, Version 2.0.

config/config_diffusion.yml

@@ -0,0 +1,282 @@
+# (C) Copyright 2025 WeatherGenerator contributors.
+#
+# This software is licensed under the terms of the Apache Licence Version 2.0
+# which can be obtained at http://www.apache.org/licenses/LICENSE-2.0.
+#
+# In applying this licence, ECMWF does not waive the privileges and immunities
+# granted to it by virtue of its status as an intergovernmental organisation
+# nor does it submit to any jurisdiction.
+
+embed_orientation: "channels"
+embed_unembed_mode: "block"
+embed_dropout_rate: 0.1
+
+ae_local_dim_embed: 2048
+ae_local_num_blocks: 0
+ae_local_num_heads: 16
+ae_local_dropout_rate: 0.1
+ae_local_with_qk_lnorm: True
+
+ae_local_num_queries: 1
+ae_local_queries_per_cell: False
+ae_adapter_num_heads: 16
+ae_adapter_embed: 128
+ae_adapter_with_qk_lnorm: True
+ae_adapter_with_residual: True
+ae_adapter_dropout_rate: 0.1
+
+ae_global_dim_embed: 2048
+ae_global_num_blocks: 4
+ae_global_num_heads: 32
+ae_global_dropout_rate: 0.1
+ae_global_with_qk_lnorm: True
+# TODO: switching to < 1 triggers triton-related issues.
+# See https://github.com/ecmwf/WeatherGenerator/issues/1050
+ae_global_att_dense_rate: 1.0
+ae_global_block_factor: 64
+ae_global_mlp_hidden_factor: 2
+ae_global_trailing_layer_norm: False
+
+ae_aggregation_num_blocks: 0
+ae_aggregation_num_heads: 32
+ae_aggregation_dropout_rate: 0.1
+ae_aggregation_with_qk_lnorm: True
+ae_aggregation_att_dense_rate: 1.0
+ae_aggregation_block_factor: 64
+ae_aggregation_mlp_hidden_factor: 2
+
+decoder_type: PerceiverIOCoordConditioning #  Main options PerceiverIOCoordConditioning or Linear
+pred_adapter_kv: False
+pred_self_attention: True
+pred_dyadic_dims: False
+pred_mlp_adaln: True
+num_class_tokens: 0
+num_register_tokens: 0
+
+# number of steps offset applied to first target window; if set to zero and forecast_steps=0 then
+# one is training an auto-encoder
+fe_num_blocks: 2
+fe_num_heads: 16
+fe_dropout_rate: 0.1
+fe_with_qk_lnorm: True
+fe_diffusion_model: True
+fe_layer_norm_after_blocks: [7]  # Index starts at 0. Thus, [3] adds a LayerNorm after the fourth layer
+fe_impute_latent_noise_std: 0.0  # 1e-4
+# currently fixed to 1.0 (due to limitations with flex_attention and triton)
+forecast_att_dense_rate: 1.0
+with_step_conditioning: True # False
+# Diffusion related parameters
+frequency_embedding_dim: 256
+embedding_dim: 512
+sigma_min: 0.002
+sigma_max: 50000
+sigma_data: 0.5
+rho: 7
+p_mean: 0.0  # -1.2
+p_std: 1.2  # 1.2
+# Encoder weights (set to null to not load a pretrained encoder)
+# chkpt_encoder_weights: "./models/dhb9q2yo/dhb9q2yo_chkpt00126.chkpt"
+chkpt_encoder_weights: "dhb9q2yo"
+chkpt_encoder_mini_epoch: 126
+
+healpix_level: 5
+
+with_mixed_precision: True
+with_flash_attention: True
+compile_model: False
+with_fsdp: True
+attention_dtype: bf16
+mixed_precision_dtype: bf16
+mlp_norm_eps: 1e-5
+norm_eps: 1e-4
+
+latent_noise_kl_weight: 0.0 # 1e-5
+latent_noise_gamma: 2.0
+latent_noise_saturate_encodings: 5 
+latent_noise_use_additive_noise: False
+latent_noise_deterministic_latents: True
+
+
+freeze_modules: ".*latent_pre_norm.*|.*latent_heads.*|.*pred_heads.*|.*target_token_engines.*|.*embed_target_coords.*|.*encoder.*|.*StreamEmbedder_ERA5.*|.*embed_engine.*|.*embed_engine.*|.*ae_local_engine.*|.*ae_local_global_engine.*|.*ae_global_engine.*"
+# freeze_modules: ".*latent_pre_norm.*|.*latent_heads.*|.*encoder.*|.*StreamEmbedder_ERA5.*|.*embed_engine.*|.*ae_local_engine.*|.*ae_local_global_engine.*|.*ae_global_engine.*"
+load_chkpt: {}
+
+norm_type: "LayerNorm"
+
+#####################################
+
+streams_directory: "./config/streams/era5_1deg/"
+streams: ???
+
+# type of zarr_store
+zarr_store: "zip" # "zarr" for LocalStore, "zip" for ZipStore
+
+general:
+
+  # mutable parameters
+  istep: 0
+  rank: ???
+  world_size: ???
+
+  # local_rank, 
+  # with_ddp,
+  # data_path_*, 
+  # model_path, 
+  # run_path, 
+  # path_shared_
+
+  multiprocessing_method: "fork"
+
+  desc: ""
+  run_id: ???
+  run_history: []
+
+# logging frequency in the training loop (in number of batches)
+train_log_freq:
+  terminal: 10
+  metrics: 20
+  checkpoint: 250
+
+# parameters for data loading
+data_loading :
+
+  num_workers: 12
+  rng_seed: ???
+  repeat_data_in_mini_epoch : True
+
+  # pin GPU memory for faster transfer; it is possible that enabling memory_pinning with 
+  # FSDP2 + DINOv2 can cause the job to hang and trigger a PyTorch timeout error.
+  # If this happens, you can disable the flag, but performance will drop on GH200.
+  memory_pinning: True
+
+
+# config for training
+training_config:
+
+  # training_mode: "masking", "student_teacher", "latent_loss"
+  training_mode: ["masking","student_teacher"] # ["student_teacher", "physical_loss"]  
+
+  num_mini_epochs: 150
+  samples_per_mini_epoch: 66
+  shuffle: True
+
+  start_date: 2012-06-01T00:00
+  end_date: 2012-06-01T18:00
+
+  time_window_step: 06:00:00
+  time_window_len: 06:00:00
+
+  learning_rate_scheduling :
+    lr_start: 5e-5 # 1e-6?
+    lr_max: 1e-4 # 5e-5?
+    lr_final_decay: 1e-6
+    lr_final: 0.0
+    num_steps_warmup: 64
+    num_steps_cooldown: 512
+    policy_warmup: "cosine"
+    policy_decay: "constant"
+    policy_cooldown: "linear"
+    parallel_scaling_policy: "sqrt"
+
+  optimizer:
+    grad_clip: 1.0
+    weight_decay: 0.1
+    log_grad_norms: False
+    adamw :
+      # parameters are scaled by number of DDP workers
+      beta1 : 0.975
+      beta2 : 0.9875
+      eps : 2e-08
+
+  losses : {
+    "physical": {
+        type: LossPhysical,
+        weight: 0.1,
+        loss_fcts: {
+          "mse": {},
+        },
+        target_and_aux_calc: "Physical",
+    },
+    "latent_diff": {
+        type: LossLatentDiffusion,
+        weight: 0.9,
+        target_and_aux_calc: DiffusionLatentTargetEncoder,
+        loss_fcts: { "mse": { }, },
+        }
+  }
+
+  model_input: {
+    "forecasting" : {
+      # masking strategy: "random", "healpix", "forecast"
+      masking_strategy: "forecast",
+      masking_strategy_config: {diffusion_rn: True},
+      num_samples: 3
+      }
+    }
+
+  forecast :
+      time_step: 06:00:00
+      num_steps: 1
+      offset: 0
+      policy: "fixed"
+
+
+# validation config; full validation config is merge of training and validation config
+validation_config: 
+
+  samples_per_mini_epoch: 16
+  shuffle: False
+
+  start_date: 2012-06-01T00:00
+  end_date: 2012-06-01T18:00
+
+  # whether to track the exponential moving average of weights for validation
+  validate_with_ema: 
+    enabled : True
+    ema_ramp_up_ratio: 0.09
+    ema_halflife_in_thousands: 1e-3
+
+  # parameters for validation samples that are written to disk
+  output : {
+    # number of samples that are written
+    num_samples: 1,
+    # write samples in normalized model space
+    normalized_samples: False,
+    # output streams to write; default all
+    streams: null,
+    }
+
+  # run validation before training starts (mainly for model development)
+  validate_before_training: True
+
+
+# test config; full test config is merge of validation and test config
+# test config is used by default when running inference
+
+# Tags for experiment tracking
+# These tags will be logged in MLFlow along with completed runs for train, eval, val
+# The tags are free-form, with the following rules:
+# - tags should be primitive types (strings, numbers, booleans). NO lists or dictionaries
+# - tags should not duplicate existing config entries.
+# - try to reuse existing tags where possible. MLFlow does not like having too many unique tags
+# - do not use long strings in values (less than 20 characters is a good rule of thumb, we may enforce this in the future)
+wgtags:
+  # The name of the organization of the person running the experiment.
+  # This may be autofilled in the future. Expected values are lowercase strings 
+  # e.g. "ecmwf", "cmcc", "metnor", "jsc", "escience"
+  org: null
+  # The Github issue corresponding to this run (number such as 1234)
+  # Github issues are the central point when running experiment and contain 
+  # links to hedgedocs, code branches, pull requests etc.
+  # It is recommended to associate a run with a Github issue.
+  issue: null
+  # The name of the experiment. This is a distinctive codename for the experiment campaign being run.
+  # This is expected to be the primary tag for comparing experiments in MLFlow, along with the
+  # issue number.
+  # Expected values are lowercase strings with no spaces, just underscores:
+  # Examples: "rollout_ablation_grid"  
+  exp: null
+  # *** Experiment-specific tags ***
+  # All extra tags (including lists, dictionaries, etc.) are treated 
+  # as strings by mlflow, so treat all extra tags as simple string key: value pairs.
+  grid: null

config/default_config.yml

@@ -59,6 +59,7 @@ fe_num_blocks: 6
 fe_num_heads: 16
 fe_dropout_rate: 0.1
 fe_with_qk_lnorm: True
+fe_diffusion_model: False
 fe_layer_norm_after_blocks: []  # Index starts at 0. Thus, [3] adds a LayerNorm after the fourth layer
 fe_impute_latent_noise_std: 0.0  # 1e-4
 # currently fixed to 1.0 (due to limitations with flex_attention and triton)
@@ -255,4 +256,4 @@ wgtags:
   # *** Experiment-specific tags ***
   # All extra tags (including lists, dictionaries, etc.) are treated 
   # as strings by mlflow, so treat all extra tags as simple string key: value pairs.
-  grid: null
+  grid: null