The current pipeline offers a limited selection of models, which restricts users from easily experimenting with and benchmarking different architectures. This makes it difficult to find the optimal model for a specific dataset, especially when balancing trade-offs like speed versus accuracy.
The core requirement is to add support for popular families like ResNetV2, DenseNet, and Inception, ensuring the existing transfer learning logic—freezing base layers and adding a new classification head—is fully compatible with each new model. The implementation must be seamless for the user, allowing them to select any new architecture simply by changing the model_name string in their configuration.
The current pipeline offers a limited selection of models, which restricts users from easily experimenting with and benchmarking different architectures. This makes it difficult to find the optimal model for a specific dataset, especially when balancing trade-offs like speed versus accuracy.
The core requirement is to add support for popular families like ResNetV2, DenseNet, and Inception, ensuring the existing transfer learning logic—freezing base layers and adding a new classification head—is fully compatible with each new model. The implementation must be seamless for the user, allowing them to select any new architecture simply by changing the model_name string in their configuration.