Wm/1270/cropping

config/default_config.yml

@@ -1,5 +1,5 @@
-# streams_directory: "./config/streams/era5_1deg/"
-streams_directory: "./config/streams/era5_nppatms_synop/"
+streams_directory: "./config/streams/era5_1deg/"
+# streams_directory: "./config/streams/era5_nppatms_synop/"
 
 embed_orientation: "channels"
 embed_unembed_mode: "block"
@@ -119,28 +119,28 @@ ema_halflife_in_thousands: 1e-3
 
 # Student-teacher configuration (only used when training_mode == "student_teacher")
 # TODO: adapt so that the masking or forecast config entry also sits here
-training_config:
-  # when this is "masking", we are basically only using the model_input subconfig
-  training_mode: "masking"  # "masking", "student_teacher", "forecast"
-
-  model_input:
-    - masking_strategy: "random" # "random", "healpix". Masking strategy to use for model input for masking, and local (student) views when doing student-teacher
-      num_samples: 1  # if student-teacher, the number of local (student) views to generate
-      masking_strategy_config : { diffusion_rn : True, rate : 0.4 }
-      # masking_strategy_config: {"strategies": ["random", "healpix", "channel"], 
-      #                       "probabilities": [0.34, 0.33, 0.33],
-      #                       "hl_mask": 3, "mode": "per_cell",
-      #                       "same_strategy_per_batch": false
-      #                       }
-      # relationship: "independent" #, "subset", "disjoint". Relationship of student views to teacher view.
-      relationship: "complement"  # "independent", "subset", "disjoint". Relationship of student views to teacher view.
-      num_steps_input: 2
-      # loss : ibot
-      loss :
-        training :
-          - LossPhysical: {weight: 0.7, loss_fcts: [['mse', 0.8], ['mae', 0.2]]}
-        validation :
-          - LossPhysical: {weight: 1.0, loss_fcts: [['mse', 0.8]]}
+# training_config:
+#   # when this is "masking", we are basically only using the model_input subconfig
+#   training_mode: "masking"  # "masking", "student_teacher", "forecast"
+
+#   model_input:
+#     - masking_strategy: "random" # "random", "healpix". Masking strategy to use for model input for masking, and local (student) views when doing student-teacher
+#       num_samples: 1  # if student-teacher, the number of local (student) views to generate
+#       masking_strategy_config : { diffusion_rn : True, rate : 0.4 }
+#       # masking_strategy_config: {"strategies": ["random", "healpix", "channel"], 
+#       #                       "probabilities": [0.34, 0.33, 0.33],
+#       #                       "hl_mask": 3, "mode": "per_cell",
+#       #                       "same_strategy_per_batch": false
+#       #                       }
+#       # relationship: "independent" #, "subset", "disjoint". Relationship of student views to teacher view.
+#       relationship: "complement"  # "independent", "subset", "disjoint". Relationship of student views to teacher view.
+#       num_steps_input: 2
+#       # loss : ibot
+#       loss :
+#         training :
+#           - LossPhysical: {weight: 0.7, loss_fcts: [['mse', 0.8], ['mae', 0.2]]}
+#         validation :
+#           - LossPhysical: {weight: 1.0, loss_fcts: [['mse', 0.8]]}
     # - masking_strategy: "random"
     #   num_samples: 2  # if student-teacher, the number of local (student) views to generate
     #   masking_strategy_config : { diffusion_rn : True, rate : 0.4 }
@@ -187,12 +187,73 @@ training_config:
   #     num_samples: 1  # if student-teacher, the number of local (student) views to generate
   #     masking_strategy_config : { rate : 0.4 }
 
-  #   # masking_strategy: "cropping"  # Strategy for teacher (global) view: "random", "healpix"
-  #   # rate: 0.1  # Fraction of data to keep in global view (alternative: use "keep_m" for absolute count)
-  #   # num_samples: 2 # number of teacher views to generate
-  #   # hl_mask : 0 # healpix level to use for healpix masking strategy
-  #   # # keep_m: 100  # Alternative to rate: keep exactly this many parent cells
-  #   # rate_sampling: true  # randomly sample the rate per batch
+# ========================================
+# CROPPING EXAMPLES (cropping_healpix strategy)
+# ========================================
+# Spatial cropping: select and KEEP a spatially contiguous region, masking everything else
+# This is the inverse of healpix masking - perfect for SSL methods like DINO/JEPA/IBOT
+
+# Example 1: Basic spatial cropping (single global crop - 75% of sphere)
+training_config:
+  training_mode: "masking"
+  model_input:
+    - masking_strategy: "cropping_healpix"
+      num_samples: 1
+      masking_strategy_config:
+        hl_mask: 0          # HEALPix level for cropping (must be < healpix_level)
+        rate: 0.75          # Fraction of cells to KEEP (0.75 = keep 75%, mask 25%)
+        method: "geodesic_disk"  # Options: "disk", "random_walk", "geodesic_disk" (recommended)
+      relationship: "independent"
+      num_steps_input: 1
+      loss:
+        training:
+          - LossPhysical: {weight: 1.0, loss_fcts: [['mse', 1.0]]}
+
+# Example 2: DINO-style training (1 global + 3 local crops)
+# training_config:
+#   training_mode: "masking"
+#   target_input:
+#     - masking_strategy: "cropping_healpix"
+#       num_samples: 1
+#       masking_strategy_config: {hl_mask: 0, rate: 0.75, method: "geodesic_disk"}
+#   model_input:
+#     - masking_strategy: "cropping_healpix"
+#       num_samples: 3  # Creates 3 independent local crops
+#       masking_strategy_config: {hl_mask: 0, rate: 0.4, method: "geodesic_disk"}
+#       relationship: "independent"
+#       num_steps_input: 1
+#       loss:
+#         training:
+#           - LossPhysical: {weight: 1.0, loss_fcts: [['mse', 1.0]]}
+
+# Example 3: IBOT-style training (1 global teacher + 1 local student with controlled overlap)
+# training_config:
+#   training_mode: "masking"
+#   target_input:
+#     - masking_strategy: "cropping_healpix"
+#       num_samples: 1
+#       masking_strategy_config:
+#         hl_mask: 2          # HEALPix level 2 = 192 parent cells
+#         rate: 0.5           # Teacher crop: 50% of sphere
+#         method: "geodesic_disk"
+#   model_input:
+#     - masking_strategy: "cropping_healpix"
+#       num_samples: 1
+#       masking_strategy_config:
+#         hl_mask: 2          # Same level for consistent cropping
+#         rate: 0.3           # Student crop: 30% of sphere
+#         method: "geodesic_disk"
+#         overlap_ratio: 0.5  # 50% of student overlaps with teacher (15% overlap, 15% independent)
+#       relationship: "independent"
+#       num_steps_input: 1
+#       loss:
+#         training:
+#           - LossPhysical: {weight: 1.0, loss_fcts: [['mse', 1.0]]}
+#
+# Note: overlap_ratio controls what fraction of the STUDENT crop comes from the teacher:
+#   - 0.0 = completely disjoint (no overlap)
+#   - 0.5 = half of student overlaps with teacher
+#   - 1.0 = student is complete subset of teacher
 
 num_mini_epochs: 32
 samples_per_mini_epoch: 64 #4096