Merge pull request #25 from spaceml-org/vertexai

New Limb Masking + Nonlinear DL + F10.7 Experiment
spaceml-org · Oct 2, 2024 · e9dae8a · e9dae8a
2 parents 8c6ebd7 + 41be0b5
commit e9dae8a
Show file tree

Hide file tree

Showing 35 changed files with 49,749 additions and 547 deletions.
diff --git a/.gitignore b/.gitignore
@@ -2,6 +2,7 @@ wandb
 output
 outputs
 *.tar
+notebooks/imgs_for_google_emb/*
 notebooks/camera_ready/*/lightning_logs
 
 # aux directories

diff --git a/experiments/for_google_emb.yaml b/experiments/for_google_emb.yaml
@@ -0,0 +1,163 @@
+# default.yaml
+
+# MODEL SUMMARY
+#   | Name        | Type                   | Params
+# -------------------------------------------------------
+# 0 | autoencoder | MaskedAutoencoderViT3D | 333 M 
+# -------------------------------------------------------
+# 329 M     Trainable params
+# 4.7 M     Non-trainable params
+# 333 M     Total params
+# 1,335.838 Total estimated model params size (MB)
+
+# general
+log_level: 'DEBUG'
+experiment:
+  name: "mae-helioprojected-2011" # generate random name in wandb
+  project: "sdofm"
+  task: "pretrain" # options: train, evaluate (not implemented)
+  model: "samae"
+  backbone_checkpoint: null
+  seed: 0
+  disable_cuda: false
+  resuming: false
+  wandb:
+    enable: true
+    entity: "fdlx"
+    group: "sdofm-phase1"
+    job_type: "pretrain"
+    tags: []
+    notes: ""
+    output_directory: "wandb_output"
+    log_model: "all" # can be True (final checkpoint), False (no checkpointing), or "all" (for all epoches)
+  gcp_storage: # this will checkpoint all epoches, perhaps clean up this config
+    enabled: true
+    bucket: "sdofm-checkpoints"
+  fold: null
+  evaluate: false     # skip training and only evaluate (requires checkpoint to be set)
+  checkpoint: null    # this is the wandb run_id of the checkpoint to load
+  device: null        # this is set automatically using the disable_cuda flag and torch.cuda.is_available()
+  precision: '16' #-true'       # (32, 64) for cuda, ('32-true', '16-true', 'bf16-true') for tpu
+  log_n_batches: 1000 # log every n training batches
+  save_results: true  # save full results to file and wandb
+  accelerator: "auto"  # options are "auto", "gpu", "tpu", "ipu", or "cpu"
+  profiler: null #'XLAProfiler' # 'XLAProfiler' # options are XLAProfiler/PyTorchProfiler Warning: XLA for TPUs only works on single world size
+  distributed:
+    enabled: true
+    world_size: "auto" # The "auto" option recognizes the machine you are on, and selects the appropriate number of accelerators.
+  log_every_n_steps: 50
+
+# dataset configuration
+data:
+  min_date: '2015-02-01 00:00:00.00' # NOT IMPLEMENTED # minimum is '2010-09-09 00:00:11.08'
+  max_date: '2015-05-31 23:59:59.99' # NOT IMPLEMENTED # maximum is '2023-05-26 06:36:08.072'
+  month_splits: # non selected months will form training set 
+    train: [1] #,2,3,4,5,6,7,8,9,10]
+    val: [2]
+    test: [3,4]
+    holdout: []
+  num_workers: 16    # set appropriately for your machine
+  prefetch_factor: 3
+  num_frames: 1 # WARNING: This is only read for FINETUNING, model num_frames overrides in BACKBONE
+  drop_frame_dim: false
+  # output_directory: "wandb_output"
+  sdoml:
+    base_directory: "/mnt/sdoml"
+    sub_directory:
+      hmi: "HMI.zarr"
+      aia: "AIA.zarr"
+      eve: "EVE_legacy.zarr"
+      cache: "cache"
+    components: null # null for select all magnetic components ["Bx", "By", "Bz"]
+    wavelengths: null # null for select all wavelengths channels ["131A","1600A","1700A","171A","193A","211A","304A","335A","94A"]
+    ions: null # null to select all ion channels ["C III", "Fe IX", "Fe VIII", "Fe X", "Fe XI", "Fe XII", "Fe XIII", "Fe XIV", "Fe XIX", "Fe XV", "Fe XVI", "Fe XVIII", "Fe XVI_2", "Fe XX", "Fe XX_2", "Fe XX_3", "H I", "H I_2", "H I_3", "He I", "He II", "He II_2", "He I_2", "Mg IX", "Mg X", "Mg X_2", "Ne VII", "Ne VIII", "O II", "O III", "O III_2", "O II_2", "O IV", "O IV_2", "O V", "O VI", "S XIV", "Si XII", "Si XII_2"]
+    frequency: '12min' # smallest is 12min
+    mask_with_hmi_threshold: null # None/null for no mask, float for threshold
+    feature_engineering:
+      enabled: true
+      dclass: 'HelioProjected'
+
+# model configurations
+model:
+  # PRETRAINERS
+  mae:
+    img_size: 512
+    patch_size: 16
+    num_frames: 1
+    tubelet_size: 1
+    in_chans: 9
+    embed_dim: 128
+    depth: 24
+    num_heads: 16
+    decoder_embed_dim: 512
+    decoder_depth: 8
+    decoder_num_heads: 16
+    mlp_ratio: 4.0
+    norm_layer: 'LayerNorm'
+    norm_pix_loss: False
+    masking_ratio: 0.75
+  samae:
+    # uses all parameters as in mae plus these
+    masking_type: "solar_aware"  # 'random' or 'solar_aware'
+    active_region_mu_degs: 15.73
+    active_region_std_degs: 6.14
+    active_region_scale: 1.0
+    active_region_abs_lon_max_degs: 60
+    active_region_abs_lat_max_degs: 60
+  nvae:
+    use_se: true
+    res_dist: true
+    num_x_bits: 8
+    num_latent_scales: 3  # 5
+    num_groups_per_scale: 1  # 16
+    num_latent_per_group: 1  # 10
+    ada_groups: true
+    min_groups_per_scale: 1
+    num_channels_enc: 30
+    num_channels_dec: 30
+    num_preprocess_blocks: 2   # 1
+    num_preprocess_cells: 2
+    num_cell_per_cond_enc: 2
+    num_postprocess_blocks: 2  # 1
+    num_postprocess_cells: 2
+    num_cell_per_cond_dec: 2
+    num_mixture_dec: 1
+    num_nf: 2
+    kl_anneal_portion: 0.3
+    kl_const_portion: 0.0001
+    kl_const_coeff: 0.0001
+    # learning_rate: 1e-2
+    # weight_decay: 3e-4
+    weight_decay_norm_anneal: true
+    weight_decay_norm_init: 1.
+    weight_decay_norm: 1e-2        
+
+  # FINE-TUNERS
+  degragation:
+    num_neck_filters: 32
+    output_dim: 1 # not sure why this is implemented for autocorrelation, should be a scalar
+    loss: "mse" # options: "mse", "heteroscedastic"
+    freeze_encoder: true
+
+  # ML optimization arguments:
+  opt:
+    loss: "mse" # options: "mae", "mse", "mape"
+    scheduler: "constant" #other options: "cosine", "plateau", "exp"
+    scheduler_warmup: 0
+    batch_size: 1
+    learning_rate: 0.0001
+    weight_decay: 3e-4 # 0.0
+    optimiser: "adam"
+    epochs: 2
+    patience: 2
+
+# hydra configuration
+# hydra:
+#   sweeper:
+#     params:
+#       model.mae.embed_dim: 256, 512
+#       model.mae.masking_ratio: 0.5, 0.75
+#       model.samae.masking_type: "random", "solar_aware"
+
+hydra:
+  mode: RUN
diff --git a/experiments/pretrain_32.2M_mae_HP_r512_e128_p16.yaml b/experiments/pretrain_32.2M_mae_HP_r512_e128_p16.yaml
@@ -37,7 +37,7 @@ experiment:
   evaluate: false     # skip training and only evaluate (requires checkpoint to be set)
   checkpoint: null    # this is the wandb run_id of the checkpoint to load
   device: null        # this is set automatically using the disable_cuda flag and torch.cuda.is_available()
-  precision: '16' #-true'       # (32, 64) for cuda, ('32-true', '16-true', 'bf16-true') for tpu
+  precision: '16-mixed' #-true'       # (32, 64) for cuda, ('32-true', '16-true', 'bf16-true') for tpu
   log_n_batches: 1000 # log every n training batches
   save_results: true  # save full results to file and wandb
   accelerator: "auto"  # options are "auto", "gpu", "tpu", "ipu", or "cpu"
@@ -73,6 +73,7 @@ data:
     ions: null # null to select all ion channels ["C III", "Fe IX", "Fe VIII", "Fe X", "Fe XI", "Fe XII", "Fe XIII", "Fe XIV", "Fe XIX", "Fe XV", "Fe XVI", "Fe XVIII", "Fe XVI_2", "Fe XX", "Fe XX_2", "Fe XX_3", "H I", "H I_2", "H I_3", "He I", "He II", "He II_2", "He I_2", "Mg IX", "Mg X", "Mg X_2", "Ne VII", "Ne VIII", "O II", "O III", "O III_2", "O II_2", "O IV", "O IV_2", "O V", "O VI", "S XIV", "Si XII", "Si XII_2"]
     frequency: '12min' # smallest is 12min
     mask_with_hmi_threshold: null # None/null for no mask, float for threshold
+    apply_mask: false
     feature_engineering:
       enabled: true
       dclass: 'HelioProjected'
@@ -148,9 +149,12 @@ model:
     learning_rate: 0.0001
     weight_decay: 3e-4 # 0.0
     optimiser: "adam"
-    epochs: 2
+    epochs: 100
     patience: 2
 
+  misc:
+    limb_mask: false
+
 # hydra configuration
 # hydra:
 #   sweeper:

diff --git a/experiments/pretrain_32.2M_mae_log_r512_e128_p16.yaml b/experiments/pretrain_32.2M_mae_log_r512_e128_p16.yaml
@@ -0,0 +1,167 @@
+# default.yaml
+
+# MODEL SUMMARY
+#   | Name        | Type                   | Params
+# -------------------------------------------------------
+# 0 | autoencoder | MaskedAutoencoderViT3D | 333 M 
+# -------------------------------------------------------
+# 329 M     Trainable params
+# 4.7 M     Non-trainable params
+# 333 M     Total params
+# 1,335.838 Total estimated model params size (MB)
+
+# general
+log_level: 'DEBUG'
+experiment:
+  name: "mae-log-helioprojected-limbmasked-2011subset-r512-e128-p16" # generate random name in wandb
+  project: "sdofm"
+  task: "pretrain" # options: train, evaluate (not implemented)
+  model: "mae"
+  backbone_checkpoint: null
+  seed: 0
+  disable_cuda: false
+  resuming: false
+  wandb:
+    enable: true
+    entity: "fdlx"
+    group: "sdofm-phase1"
+    job_type: "pretrain"
+    tags: []
+    notes: ""
+    output_directory: "wandb_output"
+    log_model: "all" # can be True (final checkpoint), False (no checkpointing), or "all" (for all epoches)
+  gcp_storage: # this will checkpoint all epoches, perhaps clean up this config
+    enabled: true
+    bucket: "sdofm-checkpoints"
+  fold: null
+  evaluate: false     # skip training and only evaluate (requires checkpoint to be set)
+  checkpoint: null    # this is the wandb run_id of the checkpoint to load
+  device: null        # this is set automatically using the disable_cuda flag and torch.cuda.is_available()
+  precision: '16-mixed' #-true'       # (32, 64) for cuda, ('32-true', '16-true', 'bf16-true') for tpu
+  log_n_batches: 1000 # log every n training batches
+  save_results: true  # save full results to file and wandb
+  accelerator: "auto"  # options are "auto", "gpu", "tpu", "ipu", or "cpu"
+  profiler: null #'XLAProfiler' # 'XLAProfiler' # options are XLAProfiler/PyTorchProfiler Warning: XLA for TPUs only works on single world size
+  distributed:
+    enabled: true
+    world_size: "auto" # The "auto" option recognizes the machine you are on, and selects the appropriate number of accelerators.
+  log_every_n_steps: 50
+
+# dataset configuration
+data:
+  min_date: '2011-01-01 00:00:00.00' # NOT IMPLEMENTED # minimum is '2010-09-09 00:00:11.08'
+  max_date: '2011-03-31 23:59:59.99' # NOT IMPLEMENTED # maximum is '2023-05-26 06:36:08.072'
+  month_splits: # non selected months will form training set 
+    train: [1] #,2,3,4,5,6,7,8,9,10]
+    val: [2]
+    test: [3]
+    holdout: []
+  num_workers: 16    # set appropriately for your machine
+  prefetch_factor: 3
+  num_frames: 1 # WARNING: This is only read for FINETUNING, model num_frames overrides in BACKBONE
+  drop_frame_dim: false
+  # output_directory: "wandb_output"
+  sdoml:
+    base_directory: "/mnt/sdoml"
+    sub_directory:
+      hmi: "HMI.zarr"
+      aia: "AIA.zarr"
+      eve: "EVE_legacy.zarr"
+      cache: "cache"
+    components: null # null for select all magnetic components ["Bx", "By", "Bz"]
+    wavelengths: null # null for select all wavelengths channels ["131A","1600A","1700A","171A","193A","211A","304A","335A","94A"]
+    ions: null # null to select all ion channels ["C III", "Fe IX", "Fe VIII", "Fe X", "Fe XI", "Fe XII", "Fe XIII", "Fe XIV", "Fe XIX", "Fe XV", "Fe XVI", "Fe XVIII", "Fe XVI_2", "Fe XX", "Fe XX_2", "Fe XX_3", "H I", "H I_2", "H I_3", "He I", "He II", "He II_2", "He I_2", "Mg IX", "Mg X", "Mg X_2", "Ne VII", "Ne VIII", "O II", "O III", "O III_2", "O II_2", "O IV", "O IV_2", "O V", "O VI", "S XIV", "Si XII", "Si XII_2"]
+    frequency: '12min' # smallest is 12min
+    mask_with_hmi_threshold: null # None/null for no mask, float for threshold
+    apply_mask: true
+    feature_engineering:
+      enabled: true
+      dclass: 'Log'
+
+# model configurations
+model:
+  # PRETRAINERS
+  mae:
+    img_size: 512
+    patch_size: 16
+    num_frames: 1
+    tubelet_size: 1
+    in_chans: 9
+    embed_dim: 128
+    depth: 24
+    num_heads: 16
+    decoder_embed_dim: 512
+    decoder_depth: 8
+    decoder_num_heads: 16
+    mlp_ratio: 4.0
+    norm_layer: 'LayerNorm'
+    norm_pix_loss: False
+    masking_ratio: 0.75
+  samae:
+    # uses all parameters as in mae plus these
+    masking_type: "solar_aware"  # 'random' or 'solar_aware'
+    active_region_mu_degs: 15.73
+    active_region_std_degs: 6.14
+    active_region_scale: 1.0
+    active_region_abs_lon_max_degs: 60
+    active_region_abs_lat_max_degs: 60
+  nvae:
+    use_se: true
+    res_dist: true
+    num_x_bits: 8
+    num_latent_scales: 3  # 5
+    num_groups_per_scale: 1  # 16
+    num_latent_per_group: 1  # 10
+    ada_groups: true
+    min_groups_per_scale: 1
+    num_channels_enc: 30
+    num_channels_dec: 30
+    num_preprocess_blocks: 2   # 1
+    num_preprocess_cells: 2
+    num_cell_per_cond_enc: 2
+    num_postprocess_blocks: 2  # 1
+    num_postprocess_cells: 2
+    num_cell_per_cond_dec: 2
+    num_mixture_dec: 1
+    num_nf: 2
+    kl_anneal_portion: 0.3
+    kl_const_portion: 0.0001
+    kl_const_coeff: 0.0001
+    # learning_rate: 1e-2
+    # weight_decay: 3e-4
+    weight_decay_norm_anneal: true
+    weight_decay_norm_init: 1.
+    weight_decay_norm: 1e-2        
+
+  # FINE-TUNERS
+  degragation:
+    num_neck_filters: 32
+    output_dim: 1 # not sure why this is implemented for autocorrelation, should be a scalar
+    loss: "mse" # options: "mse", "heteroscedastic"
+    freeze_encoder: true
+
+  # ML optimization arguments:
+  opt:
+    loss: "mse" # options: "mae", "mse", "mape"
+    scheduler: "constant" #other options: "cosine", "plateau", "exp"
+    scheduler_warmup: 0
+    batch_size: 1
+    learning_rate: 0.0001
+    weight_decay: 3e-4 # 0.0
+    optimiser: "adam"
+    epochs: 2
+    patience: 2
+
+  misc:
+    limb_mask: true
+
+# hydra configuration
+# hydra:
+#   sweeper:
+#     params:
+#       model.mae.embed_dim: 256, 512
+#       model.mae.masking_ratio: 0.5, 0.75
+#       model.samae.masking_type: "random", "solar_aware"
+
+hydra:
+  mode: RUN