Adding strategy config support for deepspeed

experiments/pretrain_32.2M_samae_deepspeed.yaml

@@ -0,0 +1,153 @@
+# 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
+experiment:
+  name: "pretrain_32.2M_samae_deepspeed"
+  project: "sdofm"
+  model: "samae"
+  task: "pretrain" # options: train, evaluate (not implemented)
+  seed: 0
+  disable_cuda: false
+  disable_wandb: false
+  wandb:
+    entity: "fdlx"
+    group: "sdofm-phase1"
+    job_type: "pretrain"
+    tags: []
+    notes: ""
+  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       # 16, 32, 64 'bfloat-16'
+  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"
+  strategy: "deepspeed_stage_2" # "auto"
+  distributed:
+    enabled: true
+    backend: "ddp"
+    # nproc_per_node: 1
+    # nnodes: 1
+    world_size: "auto" # The "auto" option recognizes the machine you are on, and selects the appropriate Accelerator.
+    # node_rank: 0
+    # local_rank: 0
+    # master_addr: "localhost"
+    # master_port: 12345
+
+# dataset configuration
+data:
+  min_date: '2010-09-09 00:00:11.08'
+  max_date: '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: [11]
+    test: [12]
+    holdout: []
+  num_workers: 16    # set appropriately for your machine
+  output_directory: "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
+
+# 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: defaults to nn.LayerNorm
+    norm_pix_loss: False
+  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
+  dimming:
+    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: 3
+    learning_rate: 0.0001
+    weight_decay: 3e-4 # 0.0
+    optimiser: "adam"
+    epochs: 4
+    patience: 2
+
+# hydra configuration
+hydra:
+  mode: MULTIRUN
+  run:
+    dir: ${data.output_directory}/${now:%Y-%m-%d-%H-%M-%S}
+  sweep:
+    dir: ${hydra.run.dir}
+    subdir: ${hydra.job.num}

requirements.txt

@@ -12,4 +12,7 @@ lightning
 scipy
 astropy
 sunpy
-timm
+timm
+matplotlib
+overrides
+deepspeed

scripts/main.py

@@ -5,6 +5,7 @@
 import os
 import random
 import time
+import warnings
 from pathlib import Path
 
 import hydra
@@ -38,16 +39,19 @@ def main(cfg: DictConfig) -> None:
     )
 
     # set precision of torch tensors
-    if cfg.experiment.device == "cuda":
-        match cfg.experiment.precision:
-            case 64:
-                torch.set_default_tensor_type(torch.DoubleTensor)
-            case 32:
-                torch.set_default_tensor_type(torch.FloatTensor)
-            case _:
-                raise NotImplementedError(
-                    f"Precision {cfg.experiment.precision} not implemented"
-                )
+    # if cfg.experiment.device == "cuda":
+    match cfg.experiment.precision:
+        case 64:
+            torch.set_default_tensor_type(torch.DoubleTensor)
+        case 32:
+            torch.set_default_tensor_type(torch.FloatTensor)
+        case _:
+            warnings.warn(
+                f"Setting precision {cfg.experiment.precision} will pass through to the trainer but not other operations."
+            )
+            # raise NotImplementedError(
+            #     f"Precision {cfg.experiment.precision} not implemented"
+            # )
 
     # run experiment
     print(f"\nRunning with config:")

scripts/pretrain.py

@@ -137,6 +137,7 @@ def run(self):
                 max_epochs=self.cfg.model.opt.epochs,
                 precision=self.cfg.experiment.precision,
                 logger=self.logger,
+                strategy=self.cfg.experiment.strategy,
             )
         else:
             trainer = pl.Trainer(