week1_image_classifier/README.md

Install required packages

!pip install -q datasets transformers accelerate evaluate pillow

Complete Food-101 Training - Single Cell

1. Setup

from google.colab import userdata import os import torch from datasets import load_dataset from transformers import ( AutoImageProcessor, AutoModelForImageClassification, TrainingArguments, Trainer ) import evaluate import numpy as np

Setup HF token

hf_token = userdata.get('HF_TOKEN') os.environ['HF_TOKEN'] = hf_token

Check GPU

print(f"GPU Available: {torch.cuda.is_available()}") print(f"GPU Name: {torch.cuda.get_device_name(0)}\n")

2. Load Dataset

print("Loading Food-101 dataset...") dataset = load_dataset("food101") print(f"Train samples: {len(dataset['train']):,}") print(f"Test samples: {len(dataset['validation']):,}\n")

Get class names

class_names = dataset['train'].features['label'].names print(f"Number of classes: {len(class_names)}\n")

3. Setup Preprocessing

model_name = "microsoft/resnet-50" processor = AutoImageProcessor.from_pretrained(model_name)

def preprocess_dataset(example): image = example['image'].convert("RGB") inputs = processor(image, return_tensors="pt") return { 'pixel_values': inputs['pixel_values'].squeeze(0), 'labels': example['label'] }

4. Split and Preprocess Data

print("Splitting dataset...") train_val = dataset['train'].train_test_split(test_size=0.1, seed=42)

print("Preprocessing train dataset (this takes 5-10 minutes)...") train_dataset = train_val['train'].map( preprocess_dataset, remove_columns=['image', 'label'], batched=False )

print("Preprocessing validation dataset...") val_dataset = train_val['test'].map( preprocess_dataset, remove_columns=['image', 'label'], batched=False )

print("Preprocessing test dataset...") test_dataset = dataset['validation'].map( preprocess_dataset, remove_columns=['image', 'label'], batched=False )

Set format

train_dataset.set_format(type='torch', columns=['pixel_values', 'labels']) val_dataset.set_format(type='torch', columns=['pixel_values', 'labels']) test_dataset.set_format(type='torch', columns=['pixel_values', 'labels'])

print(f"\nTrain: {len(train_dataset):,}") print(f"Validation: {len(val_dataset):,}") print(f"Test: {len(test_dataset):,}\n")

5. Load Model

print("Loading model...") id2label = {i: label for i, label in enumerate(class_names)} label2id = {label: i for i, label in enumerate(class_names)}

model = AutoModelForImageClassification.from_pretrained( model_name, num_labels=len(class_names), id2label=id2label, label2id=label2id, ignore_mismatched_sizes=True )

print(f"Model loaded: {model_name}") print(f"Total parameters: {sum(p.numel() for p in model.parameters()):,}\n")

6. Setup Metrics

accuracy_metric = evaluate.load("accuracy") f1_metric = evaluate.load("f1")

def compute_metrics(eval_pred): predictions, labels = eval_pred predictions = np.argmax(predictions, axis=1)

accuracy = accuracy_metric.compute(predictions=predictions, references=labels)
f1 = f1_metric.compute(predictions=predictions, references=labels, average='weighted')

return {
    'accuracy': accuracy['accuracy'],
    'f1': f1['f1']
}

7. Training Arguments

training_args = TrainingArguments( output_dir="./food101-resnet50", num_train_epochs=5, per_device_train_batch_size=64, per_device_eval_batch_size=64, learning_rate=2e-5, weight_decay=0.01, warmup_steps=500, logging_steps=100, eval_strategy="epoch", save_strategy="epoch", save_total_limit=2, load_best_model_at_end=True, metric_for_best_model="accuracy", greater_is_better=True, fp16=True, dataloader_num_workers=0, push_to_hub=True, hub_model_id="suchithnj12/food101-resnet50", hub_strategy="end", report_to="none" )

8. Create Trainer

trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=val_dataset, compute_metrics=compute_metrics )

9. Train

print("Starting training (this takes 2-3 hours)...") print("-" * 60) train_result = trainer.train()

print("\nTraining completed!") print(f"Final train loss: {train_result.metrics['train_loss']:.4f}") print(f"Training time: {train_result.metrics['train_runtime']/60:.2f} minutes\n")

10. Evaluate on Test Set

print("Evaluating on test set...") test_results = trainer.evaluate(test_dataset)

print("\nTest Results:") print(f"Accuracy: {test_results['eval_accuracy']:.4f}") print(f"F1 Score: {test_results['eval_f1']:.4f}\n")

11. Push to Hub

print("Pushing model to HuggingFace Hub...") trainer.push_to_hub(commit_message="Food-101 ResNet-50 trained model")

print(f"\nModel available at: https://huggingface.co/suchithnj12/food101-resnet50") print("Training pipeline completed successfully!")

ing builder script:   4.20k/? [00:00<00:00, 416kB/s] Downloading builder script:   6.79k/? [00:00<00:00, 645kB/s] Starting training (this takes 2-3 hours)...

[5330/5330 1:56:41, Epoch 5/5] Epoch Training Loss Validation Loss Accuracy F1 1 4.496600 4.458635 0.166337 0.154736 2 3.814000 3.742320 0.298614 0.261986 3 3.257200 3.191649 0.362508 0.330396 4 2.935900 2.900200 0.397096 0.368665 5 2.847600 2.818526 0.408317 0.380199

Training completed! Final train loss: 3.6109 Training time: 116.77 minutes

Evaluating on test set... [395/395 05:23]

Test Results: Accuracy: 0.4450 F1 Score: 0.4139

Pushing model to HuggingFace Hub... Processing Files (2 / 2)      : 100%  95.1MB / 95.1MB, 9.91MB/s   New Data Upload               : 100%  95.1MB / 95.1MB, 9.91MB/s     ...esnet50/model.safetensors: 100%  95.1MB / 95.1MB               ...esnet50/training_args.bin: 100%  5.84kB / 5.84kB            

Model available at: https://huggingface.co/suchithnj12/food101-resnet50 Training pipeline completed successfully!