Fully trained model with large Imagenet dataset using ResNet50
Data Preparation: I load the ImageNet dataset, specifically the training images, and preprocess them. The dataset is organized into folders, each representing a class, and I load a few images from each class.
Model Building (Initial Attempt): I instantiate a ResNet-50 model with pre-trained weights from ImageNet and include the top layers for classification. Then, I make predictions on the loaded dataset and evaluate the model's accuracy.
102967424/102967424 [==============================] - 1s 0us/step
Model: "resnet50"
Layer (type) Output Shape Param # Connected to
==================================================================================================
input_1 (InputLayer) [(None, 224, 224, 3 0 [] )]
.....
predictions (Dense) (None, 1000) 2049000 ['avg_pool[0][0]']
==================================================================================================
_______________________________________________________________________________________
Visualization: I visualize some of the original images and the corresponding predictions made by the initial ResNet-50 model.
Transfer Learning: Next, I attempt transfer learning by using the ResNet-50 architecture as a feature extractor. I freeze the pre-trained layers and add new fully connected layers on top for fine-tuning to my specific classification task.
Model Evaluation (Transfer Learning): After fine-tuning, I make predictions with the transfer learning model and evaluate its accuracy.
Visualization of Transfer Learning Results: I visualize some of the images again, this time showing the predictions made by the transfer learning model.
Accuracy Comparison: Finally, I compare the accuracies of both the initial ResNet-50 model and the transfer learning model to see which performs better.
Overall, my project involves exploring the effectiveness of ResNet-50 for image classification on the ImageNet dataset, both with and without transfer learning, and comparing the performance of the models I build. This type of project is common in the field of deep learning, particularly for tasks like image classification.
