How josemiuoc/tfm-uoc works
Project Overview and Main Components
• This project is focused on improving breast cancer image diagnosis using deep neural networks. It encompasses data preprocessing, training with various CNN architectures (including standard CNNs, deep convolutional networks with transfer learning, and vision transformers), comparative performance evaluation (e.g., using Keras and GPU-accelerated models), and ensemble predictions (inteligencia colectiva).
a. Data Ingestion and Preprocessing
b. Dataset Fusion
c. Model Training Modules (including separate blocks for CNN training on each dataset and experiments with transfer learning, vision transformers, etc.)
d. Comparative Analysis Module (e.g., Keras models versus other CNN implementations)
e. Ensemble Prediction Module (inteligencia colectiva)
f. (Implicitly) The cloud-based environment (Google Colab) with GPU dependency.
a. And Create Blocks for Each Notebook:
• Data Loading and Preparation: Create blocks for "CBIS-DDSM-CargaDatos.ipynb", "CMMD_CargaDatos.ipynb" and the fusion block "DATASETS_UNIDOS.ipynb". These blocks represent the ingestion and harmonization of two separate datasets (CBIS-DDSM and CMMD) and their combination into a unified dataset.
• CNN Training Modules: Create individual blocks for "CBIS-DDSM_CNN.ipynb" and "CMMD_CNN.ipynb". These blocks represent the training and validation of CNN models on each respective dataset.
• Advanced Architectures: Include blocks for "DenseArchitectures_TransferLearning.ipynb" (for training with deep CNNs and transfer learning) and "Vision_Transformers.ipynb" (for experiments with transformer architectures).
• Comparative Evaluation: Add a block for "Keras_CNNs.ipynb" to highlight performance comparisons between Keras implementations and the original CNN experiments.
• Ensemble/Inteligencia Colectiva: Add a block for "Predicciones_InteligenciaColectiva.ipynb" to denote the module that combines the predictions of several models.
b. Relationships and Data Flow:
• Outputs of all training modules (CNNs, transfer learning models, vision transformers, etc.) to a centralized “Evaluation / Predictions” block.
• From there, indicate that these results feed into the ensemble predictions block.
• Optionally, include the external dependency on Google Colab (with GPU nodes noted) to indicate the runtime environment.
c. Architectural Patterns and Technologies:
• Experimental and notebook-style workflow typical in research settings.
• That the project separates data preparation, model training, and evaluation, which is a common layered architecture for deep learning pipelines.
• Mention the importance of GPU-accelerated computing (Colab PRO, T4/L4 GPUs) as a key infrastructure dependency.
• That transfer learning and ensemble learning (inteligencia colectiva) indicate modular, reuse-oriented design.