Sophiex/kerem/pr/transformer head

config/config_jepa.yml

@@ -0,0 +1,305 @@
+# (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: 1024
+ae_local_num_blocks: 2
+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: 2
+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: 8
+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 # CrossAttentionAdaNormConditioning
+pred_adapter_kv: False
+pred_self_attention: True
+pred_dyadic_dims: False
+pred_mlp_adaln: True
+num_class_tokens: 1
+num_register_tokens: 7
+
+# 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_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
+
+healpix_level: 5
+
+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
+
+latent_heads_allowed: True
+
+
+freeze_modules: ""
+
+norm_type: "LayerNorm"
+
+
+#####################################
+
+streams_directory: "./config/streams/era5_1deg/"
+# streams_directory: "./config/streams/era5_nppatms_synop/"
+streams: ???
+
+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: 8
+  rng_seed: ???
+
+
+# config for training
+training_config:
+
+  # training_mode: "masking", "student_teacher", "latent_loss"
+  training_mode: ["student_teacher"]
+
+  num_mini_epochs: 32
+  samples_per_mini_epoch: 256
+  shuffle: True
+
+  start_date: 1979-01-01T00:00
+  end_date: 2022-12-31T00:00
+
+  time_window_step: 06:00:00
+  time_window_len: 06:00:00
+
+  window_offset_prediction : 0
+
+  learning_rate_scheduling :
+    lr_start: 1e-6
+    lr_max: 5e-5
+    lr_final_decay: 1e-6
+    lr_final: 0.0
+    num_steps_warmup: 512
+    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 : {
+    "student-teacher": {
+        enabled: True,
+        type: LossLatentSSLStudentTeacher,
+        weight: 1.0,
+        loss_fcts : {
+          "JEPA": {
+            'weight': 8, "loss_extra_args": {}, "out_dim": 2048, "head": transformer,
+            "pred_num_blocks": 24, "pred_num_heads": 12, "pred_with_qk_lnorm": True, "pred_intermediate_dim": 768, 
+            "pred_dropout_rate": 0.1,
+            target_source_correspondence: {0 : {0 : "complement"} },
+          }
+        },
+        target_and_aux_calc: { "EMATeacher" : 
+          { ema_ramp_up_ratio : 0.09,
+            ema_halflife_in_thousands: 1e-3,
+            model_param_overrides : { latent_heads_allowed: False },
+          }
+        }
+        }
+  }
+
+  model_input: {
+    "random_easy" : {
+      # masking strategy: "random", "forecast"
+      masking_strategy: "random",
+      num_samples: 1,
+      num_steps_input: 1,
+      masking_strategy_config : { 
+        diffusion_rn : True, 
+        rate : 0.6,
+        rate_sampling: False
+      },
+    },
+  }
+
+  target_input: {
+    "random_easy_target" : {
+      masking_strategy: "healpix",
+      num_samples: 1,
+      masking_strategy_config : { rate : 0.2, hl_mask: 0, rate_sampling: False },
+    },
+  }
+
+  forecast :
+      time_step: 00:00:00
+      num_steps: 0
+      policy: null
+
+
+# validation config; full validation config is merge of training and validation config
+validation_config: 
+
+  samples_per_mini_epoch: 256
+  shuffle: False
+
+  start_date: 2023-10-01T00:00
+  end_date: 2023-12-31T00:00
+
+  # whether to track the exponential moving average of weights for validation
+  validate_with_ema: 
+    enabled : False
+    ema_ramp_up_ratio: 0.09
+    ema_halflife_in_thousands: 1e-3
+
+  # number of validation samples that are written to disk
+  write_num_samples: 0
+  # output streams to write; default all
+  output_streams: null
+
+  # run validation before training starts (mainly for model development)
+  validate_before_training: False
+
+  # losses: {
+  #   "physical": {
+  #       type: LossPhysical,
+  #       weight: 1.0,
+  #       loss_fcts: {
+  #         "mse": {
+  #           weight: 1.0,
+  #         },
+  #       },
+  #     },
+  # }
+
+  # Requires enabled flags
+  # model_input: {
+  #   "random_easy" : {
+  #     enabled : False,
+  #   },
+  #   "random_hard" : {
+  #     enabled : False,
+  #   },
+  #   "strategy1" : {
+  #   # "random", "healpix". Masking strategy to use for model input for masking
+  #     masking_strategy: "forecast",
+  #     num_samples: 1,
+  #     masking_strategy_config : { diffusion_rn : True, rate : 0.4 },
+  #     # relationship: "independent", "subset", "disjoint".
+  #     relationship: "independent",
+  #     num_steps_input: 1,
+  #     }
+  # }
+
+# 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/config_physical_jepa.yml

@@ -64,6 +64,12 @@ 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
 
+sslpred_num_blocks: 12
+sslpred_num_heads: 12
+sslpred_dropout_rate: 0.1
+sslpred_with_qk_lnorm: True 
+sslpred_intermediate_dim: 384
+
 healpix_level: 5
 
 with_mixed_precision: True

integration_tests/jepa1.yaml

@@ -11,6 +11,12 @@ lr_steps_warmup: 2
 lr_steps_cooldown: 2
 loader_num_workers: 8
 
+sslpred_num_blocks: 12
+sslpred_num_heads: 12
+sslpred_dropout_rate: 0.1
+sslpred_with_qk_lnorm: True 
+sslpred_intermediate_dim: 384
+
 train_log:
   log_interval: 1
 ### Example validation and training config for student-teacher with JEPA