Files used to process dataset

fo2hub.py

@@ -0,0 +1,9 @@
+import os
+
+import fiftyone as fo
+from fiftyone.utils.huggingface import push_to_hub
+import huggingface_hub
+huggingface_hub.login("XXXXXXXXXXXXXX")
+pokemon = fo.load_dataset("pokemon")
+
+push_to_hub(pokemon, "pokemon", chunk_size=500, exist_ok=True)

pokemon2fo.py

@@ -0,0 +1,46 @@
+import fiftyone as fo
+
+
+# Downloaded the images from hugging face using git clone
+# git clone https://huggingface.co/datasets/fcakyon/pokemon-classification
+POKEMON_DIR = "/home/spousty/data/pokemon-classification/data"
+
+# There is a subdir for test, train, valid
+train_pokemon = fo.Dataset.from_dir(
+    dataset_dir=POKEMON_DIR + "/train",
+    progress=True,
+    dataset_type=fo.types.ImageClassificationDirectoryTree,
+    name="train_pokemon",
+    overwrite=True,
+    persistent=True
+)
+
+test_pokemon = fo.Dataset.from_dir(
+    dataset_dir=POKEMON_DIR + "/test",
+    progress=True,
+    dataset_type=fo.types.ImageClassificationDirectoryTree,
+    name="test_pokemon",
+    overwrite=True,
+    persistent=True
+)
+
+valid_pokemon = fo.Dataset.from_dir(
+    dataset_dir=POKEMON_DIR + "/valid",
+    progress=True,
+    dataset_type=fo.types.ImageClassificationDirectoryTree,
+    name="valid_pokemon",
+    overwrite=True,
+    persistent=True
+)
+
+train_pokemon.tag_samples("train")
+test_pokemon.tag_samples("test")
+valid_pokemon.tag_samples("valid")
+
+pokemon = train_pokemon.clone(name="pokemon", persistent=True)
+pokemon.merge_samples(test_pokemon)
+pokemon.merge_samples(valid_pokemon)
+pokemon.save()
+
+session = fo.launch_app(pokemon)
+session.wait()

train_resnet_pokemon.py

@@ -0,0 +1,82 @@
+import torch
+import torchvision.transforms.v2 as T
+
+import fiftyone as fo
+from fiftyone import ViewField as F
+import fiftyone.train as fot
+
+def main():
+
+    dataset = fo.load_dataset("pokemon")
+
+    dataset = dataset.match_tags("train")
+
+    model_config = fot.model.classification.TorchvisionResnetModelConfig(
+        get_item_class_name='ImageClassificationGetItem',
+        get_item_inference_class_name='ImageClassificationInferenceGetItem',
+        train_transforms_list='standard',
+        inference_transforms_list='standard',
+        classes=dataset.distinct('ground_truth.label'),
+        weights=None
+    )
+    train_run_config = fot.run.TrainConfig(
+        model_config,
+        data={
+            'dataloader': {
+                'batch_size': 128,
+            }
+        },
+        train_regime = {
+            "num_steps": 3800,  # 100 epochs × 38 steps/epoch
+            "validate_every": 38,  # Validate every epoch
+            "early_stop_after": 456,  # 6 epochs without improvement
+            "grad_clip": 15,
+        },
+        optimizer = {
+            "name": "SGD",
+            "kwargs": {
+                "lr": 1e-2,  # Lower initial learning rate
+                "momentum": 0.9,
+                "weight_decay": 1e-3,
+            },
+            "scheduler": {
+                "name": "CosineAnnealingLR",
+                "kwargs": {
+                    "T_max": 3800,  # Total steps
+                    "eta_min": 1e-6,
+                }
+            }
+        }
+    )
+
+    incidentals={
+            "logging" : {
+                "wandb" : {
+                    'project':'resnet-pokemon'
+                },
+                "log_every" : 6,
+            },
+            # "gi_kwargs" : {
+            #     "field_mapping" : {
+            #         "labels_field" : "ground_truth.label"
+            #     }
+            # }
+        }
+
+    with fot.context(embeddings_model="resnet18-imagenet-torch",
+                     embeddings_field='embed-resnet18-imagenet',
+                     checkpoint_dir="./.checkpoints"):
+        print(fot.context)
+        print(fot.context.checkpoint_dir)
+        results = fot.run.train(
+            samples=dataset,
+            config=train_run_config,
+            key=f"resnet18_classifier",
+            incidentals=incidentals
+        )
+
+    return results
+
+
+if __name__ == "__main__":
+    main()