Overview This model is a high-performance image classification model built using the InceptionResNetV2 architecture as a feature extractor, achieving an impressive 98% accuracy on the validation set. It is designed to classify fashion product images into their respective categories, such as Apparel, Footwear, and Accessories.
The base model is pre-trained on the ImageNet dataset, and custom dense layers are added to fine-tune the model for the specific dataset used.
Model Architecture Base Model: InceptionResNetV2 (pre-trained on ImageNet, frozen during initial training). Custom Layers: Global Average Pooling: Reduces spatial dimensions to a single vector. Dense Layers: 2048 units, ReLU activation, batch normalization, and 50% dropout. 1024 units, ReLU activation, batch normalization, and 50% dropout. 512 units, ReLU activation, batch normalization, and 30% dropout. Output Layer: Fully connected layer with num_classes units and softmax activation. Architecture Summary plaintext Copy code Layer (type) Output Shape Param #
input_1 (InputLayer) [(None, 224, 224, 3)] 0 inception_resnet_v2 (Functional) (None, 5, 5, 1536) 54336736 global_average_pooling2d (GlobalAv (None, 1536) 0 dense (Dense) (None, 2048) 3147776 batch_normalization (BatchNormaliz (None, 2048) 8192 dropout (Dropout) (None, 2048) 0 dense_1 (Dense) (None, 1024) 2098176 batch_normalization_1 (BatchNormal (None, 1024) 4096 dropout_1 (Dropout) (None, 1024) 0 dense_2 (Dense) (None, 512) 524800 batch_normalization_2 (BatchNormal (None, 512) 2048 dropout_2 (Dropout) (None, 512) 0 dense_3 (Dense) (None, num_classes) (varies)
Trainable params: ~6M+ (if base model frozen initially). Dataset The model was trained on the Fashion Product Images Small dataset, which contains product images labeled by masterCategory.
Source: Fashion Product Images Small on Kaggle
Dataset Size: Training Images: ~35,000. Validation Images: ~10,000. Classes: 6 (e.g., Apparel, Footwear, Accessories).
Preprocessing: Images resized to 224x224 pixels. Pixel values normalized to [0, 1]. Training Details Optimizer: Adam with a learning rate of 0.0001. Loss Function: Sparse Categorical Cross-Entropy. Metrics: Accuracy. Batch Size: 32. Data Augmentation: Rotation: ±20°. Width/Height Shifts: ±20%. Zoom, shear, and horizontal flips. Validation: Used a separate validation set with rescaled images (no augmentation). Performance Validation Accuracy: 98%. Validation Loss: ~0.10. The high accuracy indicates the model's robustness in classifying diverse fashion products with minimal overfitting.