multilabel classification

Author: rramosp

Diff for: configs/multilabel.yaml

@@ -0,0 +1,14 @@
+dataloader:
+  _target_: earthtext.datamodules.chipmultilabel.ChipMultilabelModule
+  metadata_file:     /opt/data/california-worldcover-chips/california-worldcover-chips-osm-multilabels.parquet
+  embeddings_folder: /opt/data/california-worldcover-chips/embeddings_v0.2
+  get_strlabels: True
+  #num_workers: 10
+  min_ohe_count: 1
+
+model:
+  _target_: earthtext.models.multilabel.MultilabelModel
+  input_dim: 768
+  output_dim: 99
+  layers_spec: [512, 256, 128]
+  activation_fn: 'elu'

Diff for: notebooks/models/01 - clay model v0.2 test.ipynb

@@ -257,7 +257,7 @@
    "name": "python",
    "nbconvert_exporter": "python",
    "pygments_lexer": "ipython3",
-   "version": "3.10.9"
+   "version": "3.12.1"
   }
  },
  "nbformat": 4,

Diff for: notebooks/models/02 - random forest to esawc.ipynb renamed to notebooks/models/03 - random forest to esawc.ipynb

@@ -38,7 +38,8 @@ def __init__(
         get_strlabels = False,
         get_esawc_proportions = False,
         get_chip_id = False,
-        min_ohe_count = 1
+        min_ohe_count = 1,
+        cache_size = -1
     ):
 
         """
@@ -67,6 +68,9 @@ def __init__(
         self.metadata = self.metadata[chips_exists]
         logger.info(f"read {split} split with {len(self.metadata)} chip files (out of {nitems})")
 
+        logger.info(f"max cache size is {cache_size}")
+        self.cache_size = cache_size
+        self.cache = {}
 
     def prepare_data(self):
         """This is an optional preprocessing step to be defined in each dataloader.
@@ -80,11 +84,14 @@ def __len__(self):
 
 
     def __repr__(self):
-        return f"{self.__class__.__name__} {self.split} split with {len(self)} items"
+        return f"{self.__class__.__name__} {self.split} split with {len(self)} items, in cache {len(self.cache)} items"
 
 
     def __getitem__(self, idx):
 
+        if idx in self.cache.keys():
+            return self.cache[idx]
+
         r = {}        
 
         item = self.metadata.iloc[idx]
@@ -115,6 +122,9 @@ def __getitem__(self, idx):
         if self.get_esawc_proportions:
             r['esawc_proportions'] = str(item.esawc_proportions)
 
+        # store in cache
+        if self.cache_size == -1 or len(self.cache) < self.cache_size:
+            self.cache[idx] = r
 
         return r
 

Diff for: notebooks/models/04 - multilabel classification.ipynb

@@ -0,0 +1,48 @@
+import torch
+from torch import nn
+import torch.nn.functional as F
+import numpy as np
+
+def activation_from_str(activation_str):
+    if activation_str == 'relu':
+        return nn.ReLU()
+
+    if activation_str == 'elu':
+        return nn.ELU()
+
+    if activation_str == 'sigmoid':
+        return nn.Sigmoid()
+    
+    raise ValueError(f"unknown activation function string '{activation_str}'")
+
+class MultilabelModel(nn.Module):
+    """
+    assumes an input of shape [batch_size, h, w, 2, 2]
+    """
+    def __init__(self, input_dim, output_dim, layers_spec = [10], activation_fn='relu'):
+        super().__init__()
+
+        self.input_dim = input_dim
+        self.output_dim = output_dim
+        self.layers_spec = layers_spec
+        self.activation_fn = activation_fn
+
+        layers = [
+                nn.Linear(input_dim, layers_spec[0]),
+                activation_from_str(activation_fn)
+        ]
+        
+        for i in range(len(layers_spec)-1):
+            layers.append(nn.Linear(layers_spec[i], layers_spec[i+1]))
+            layers.append(activation_from_str(activation_fn))
+
+        layers.append(nn.Linear(layers_spec[-1], output_dim))
+        layers.append(activation_from_str('sigmoid'))
+        
+        self.layers = nn.Sequential(*layers)
+        
+    
+    def forward(self, x):
+        x = self.layers(x)        
+        return x
+        

Diff for: src/earthtext/datamodules/components/chipmultilabel.py

@@ -0,0 +1,29 @@
+import os
+import random
+
+import dotenv
+import hydra
+import lightning.pytorch as pl
+import numpy as np
+import omegaconf
+from loguru import logger
+from omegaconf import DictConfig
+
+
+
+@hydra.main(
+    version_base="1.1",
+    config_path="configs",
+    config_name="train.yaml",
+)
+
+def main(config: DictConfig):
+    print (config)
+
+    dataloader = hydra.utils.instantiate(config.dataloader)
+
+    print ("XX", dataloader)
+
+
+if __name__ == "__main__":
+    main()