Merging the latent diffusion model

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 +1,339 @@
+# (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: 512
+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: 512
+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: 6
+fe_num_heads: 16
+fe_dropout_rate: 0.1
+fe_with_qk_lnorm: True
+fe_diffusion_model: True
+fe_diffusion_model_conditioning: "forecast" # options: "date_time", "time", "forecast"
+fe_diffusion_model_conditioning_type: "cross_attn" # options: "cross_attn", "ada_ln"
+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)
+forecast_att_dense_rate: 1.0
+with_step_conditioning: True # False
+# Diffusion related parameters
+diffusion_conditioning_embed_dim: 32
+frequency_embedding_dim: 256
+embedding_dim: 512
+sigma_min: 0.002
+sigma_max: 80
+sigma_data: 1.0
+rho: 7
+p_mean: 1.5
+p_std: 1.2
+
+healpix_level: 5
+
+# Use 2D RoPE instead of traditional global positional encoding
+# When True: uses 2D RoPE based on healpix cell coordinates (lat/lon)
+# When False: uses traditional pe_global positional encoding
+rope_2D: True
+mlp_type: swiglu
+use_xsa: True
+# mlp_type: mlp
+# use_xsa: False
+
+with_mixed_precision: True
+with_flash_attention: True
+compile_model: False
+with_fsdp: False
+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.*|.*embed_engine.*|.*fe.*|.*ae_local_engine.*|.*ae_local_global_engine.*|.*ae_global_engine.*"
+# freeze_modules: ".*latent_pre_norm.*|.*latent_heads.*|.*encoder.*|.*StreamEmbedder_ERA5.*|.*embed_engine.*|.*embed_engine.*|.*ae_local_engine.*|.*ae_local_global_engine.*|.*ae_global_engine.*"
+# freeze_modules: ""
+# load_chkpt: {'run_id': 't0bdz7qn', 'epoch': -1}  # multi-var d2048 hl5, sigma_data=1.7
+# load_chkpt: {'run_id': 'dcl584vo', 'epoch': -1}  # z500 d2048 hl5, sigma_data=159.08
+# load_chkpt: {'run_id': 'tvkicam9', 'epoch': -1}  # z500 d2048 hl3 enc-lnorm, sigma_data=1.0
+# load_chkpt: {'run_id': 'q9grso75', 'epoch': -1}  # z500 d2048 hl3, sigma_data=39.2936
+# load_chkpt: {'run_id': 'qxivdyqz', 'epoch': -1}  # z500 d2048 hl5 enc-lnorm, sigma_data=1.0
+# load_chkpt: {'run_id': 'h8x1qgz3', 'epoch': -1}  # z500 d128 hl5, sigma_data=12.93
+# load_chkpt: {'run_id': '', 'epoch': -1}  # z500 d128 hl5 enc-lnorm, sigma_data=1.0
+# load_chkpt: {'run_id': 'wvpb76ai', 'epoch': -1}  # multi-var d2048 hl3 enc-lnorm, sigma_data=1.0
+# load_chkpt: {'run_id': 'ae4wlc5m', 'epoch': -1}  # multi-var d2048 hl3, sigma_data=2.7047
+# load_chkpt: {'run_id': 'r45iwyns', 'epoch': -1}  # multi-var d512 hl3, sigma_data=1.1785
+# load_chkpt: {'run_id': 'ydka6uql', 'epoch': -1}  # multi-var d512 hl4, sigma_data=0.827
+# load_chkpt: {'run_id': 'lwjkb3y4', 'epoch': -1}  # multi-var d512 hl5, sigma_data=0.5789
+# load_chkpt: {'run_id': 'v8kd6xc1', 'epoch': -1}  # multi-var d512 hl5 nopos, sigma_data=0.6481
+# load_chkpt: {'run_id': 'lwjkb3y4', 'epoch': -1}  # multi-var d512 hl5 enc-lnorm, sigma_data=1.0
+# load_chkpt: {'run_id': 'y1gu5md8', 'epoch': -1}  # multi-var d512 hl5, sigma_dqta=1.0, diffusion-full-pipeline
+# load_chkpt: {'run_id': 'mal6u4gc', 'epoch': -1}  # multi-var d512 hl5, sigma_dqta=1.0, geoinfos 64 epochs, diffusion-full-pipeline
+# load_chkpt: {'run_id': 'zrpncqb0', 'epoch': -1}  # multi-var d512 hl5, sigma_dqta=1.0, geoinfos 196 epochs, diffusion-full-pipeline
+# load_chkpt: {'run_id': 'm6fs8wvj', 'epoch': -1}  # multi-var d512 hl5, sigma_data=1.0, swiglu xsa geoinfos, diffusion-full-pipeline
+# load_chkpt: {'run_id': 'cgxt9imf', 'epoch': -1}  # diffusion model to fine-tune decoder, p_mean=0.5, SwiGLU+XSA+geoinfos, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'wo5mf2z4', 'epoch': -1}  # diffusion model to fine-tune decoder, p_mean=1.5, SwiGLU+XSA+geoinfos, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'zf6wnmpe', 'epoch': -1}  # multi-var d2048 hl5, sigma_data=1.832
+# load_chkpt: {'run_id': 'mivw6jda', 'epoch': -1}  # multi-var d2048 hl5 enc-lnorm, sigma_data=1.0
+# load_chkpt: {'run_id': 'j74tn8le', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=-1.5, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'j7lr0jws', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=-1.2, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'cbras2el', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=-0.5, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'kn3124hp', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=0.0, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'qqbu9852', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=0.5, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'vqsh3yrl', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=1.0, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'xl8h7vbt', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=1.5, based on m6fs8wvj backbone
+# load_chkpt: {'run_id': 'p9m2jwvc', 'epoch': -1}  # forecasting d512 hl5, diffusion-full-pipeline, p_mean=2.0, based on m6fs8wvj backbone
+
+
+norm_type: "LayerNorm"
+
+#####################################
+
+streams_directory: "./config/streams/era5_1deg_forecasting/"
+# streams_directory: "./config/streams/era5_1deg_forecasting_z500/"
+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_logging:
+  terminal: 10
+  metrics: 20
+  checkpoint: 250
+  log_grad_norms: False
+
+# parameters for data loading
+data_loading :
+
+  num_workers: 12
+  rng_seed: ???
+  repeat_data_in_mini_epoch : False
+
+  # 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"]
+
+  num_mini_epochs: 128
+  samples_per_mini_epoch: 4096
+  shuffle: True
+
+  start_date: 1979-01-01T00:00
+  end_date: 2022-12-31T18:00
+
+  time_window_step: 06:00:00
+  time_window_len: 06:00:00
+
+  learning_rate_scheduling :
+    lr_start: 1e-6 #5e-5
+    lr_max: 1e-5 #1e-4
+    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: 0.0,
+        loss_fcts: {
+          "mse": {},
+        },
+        target_and_aux_calc: "Physical",
+    },
+    "latent_diff": {
+        type: LossLatentDiffusion,
+        weight: 1.0,
+        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_steps_input: 2,
+      num_samples: 1,
+      }
+    }
+
+  target_input: {
+    "forecasting" : {
+      masking_strategy: "forecast",
+      masking_strategy_config: {diffusion_rn: True},
+      num_steps_input: 1,
+      num_samples: 1,
+      }
+    }
+
+  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: 
+
+  # Noise levels (eta values in standard normal space) at which to evaluate the
+  # diffusion model during validation. sigma = exp(eta * p_std + p_mean).
+  # Each value produces a separate validation pass with independently logged metrics.
+  validation_noise_levels: [1.0, 2.0, 3.0, 4.0]
+
+  samples_per_mini_epoch: 256
+  shuffle: True
+
+  start_date: 2023-10-01T00:00
+  end_date: 2023-12-31T18:00
+
+  # 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: 0,
+    # 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