Upload train_vit_oxford_pets.py

train_vit_oxford_pets.py

@@ -0,0 +1,118 @@
+
+import os
+import numpy as np
+from datasets import load_dataset
+from transformers import (
+    ViTImageProcessor,
+    ViTForImageClassification,
+    TrainingArguments,
+    Trainer,
+)
+import evaluate
+from huggingface_hub import notebook_login
+
+# ── 1. Hugging Face Login ─────────────────────────────────
+# Erstelle einen Token auf: https://huggingface.co/settings/tokens
+# Typ: "write"
+notebook_login()   # gibt einen Login-Dialog aus
+
+# ── 2. Dataset laden ──────────────────────────────────────
+print("Lade Dataset...")
+dataset = load_dataset("pcuenq/oxford-pets")
+print(dataset)
+
+# Labels extrahieren
+label_names = dataset["train"].features["label"].names
+id2label = {i: label for i, label in enumerate(label_names)}
+label2id = {label: i for i, label in enumerate(label_names)}
+num_labels = len(label_names)
+print(f"Anzahl Klassen: {num_labels}")
+print("Labels:", label_names)
+
+# ── 3. Preprocessing ──────────────────────────────────────
+MODEL_NAME = "google/vit-base-patch16-224-in21k"
+processor = ViTImageProcessor.from_pretrained(MODEL_NAME)
+
+def preprocess(batch):
+    images = [img.convert("RGB") for img in batch["image"]]
+    inputs = processor(images=images, return_tensors="pt")
+    inputs["labels"] = batch["label"]
+    return inputs
+
+dataset = dataset.map(preprocess, batched=True, batch_size=32)
+dataset.set_format(type="torch", columns=["pixel_values", "labels"])
+
+# Train/Val Split (falls kein eigener Val-Split vorhanden)
+if "validation" not in dataset:
+    split = dataset["train"].train_test_split(test_size=0.15, seed=42)
+    train_ds = split["train"]
+    val_ds   = split["test"]
+else:
+    train_ds = dataset["train"]
+    val_ds   = dataset["validation"]
+
+print(f"Train: {len(train_ds)}, Val: {len(val_ds)}")
+
+# ── 4. Modell laden ───────────────────────────────────────
+model = ViTForImageClassification.from_pretrained(
+    MODEL_NAME,
+    num_labels=num_labels,
+    id2label=id2label,
+    label2id=label2id,
+    ignore_mismatched_sizes=True,
+)
+
+# ── 5. Metriken ───────────────────────────────────────────
+accuracy_metric = evaluate.load("accuracy")
+
+def compute_metrics(eval_pred):
+    logits, labels = eval_pred
+    predictions = np.argmax(logits, axis=-1)
+    return accuracy_metric.compute(predictions=predictions, references=labels)
+
+# ── 6. Training ───────────────────────────────────────────
+# WICHTIG: Ersetze "DEIN_HF_USERNAME" mit deinem Hugging Face Benutzernamen!
+HF_USERNAME = "DEIN_HF_USERNAME"
+MODEL_REPO  = f"{HF_USERNAME}/vit-oxford-pets"
+
+training_args = TrainingArguments(
+    output_dir="./vit-oxford-pets",
+    num_train_epochs=5,
+    per_device_train_batch_size=32,
+    per_device_eval_batch_size=32,
+    warmup_steps=200,
+    weight_decay=0.01,
+    logging_dir="./logs",
+    logging_steps=50,
+    evaluation_strategy="epoch",
+    save_strategy="epoch",
+    load_best_model_at_end=True,
+    metric_for_best_model="accuracy",
+    push_to_hub=True,
+    hub_model_id=MODEL_REPO,
+    report_to="none",
+)
+
+trainer = Trainer(
+    model=model,
+    args=training_args,
+    train_dataset=train_ds,
+    eval_dataset=val_ds,
+    compute_metrics=compute_metrics,
+)
+
+# ── 7. Training starten ───────────────────────────────────
+print("Starte Training...")
+train_result = trainer.train()
+print("Training abgeschlossen!")
+
+# Trainings-Log für README speichern
+log_history = trainer.state.log_history
+print("\nTrainings-Log:")
+for entry in log_history:
+    if "eval_accuracy" in entry:
+        print(entry)
+
+# ── 8. Modell auf Hugging Face hochladen ──────────────────
+trainer.push_to_hub()
+print(f"\nModell hochgeladen: https://huggingface.co/{MODEL_REPO}")