Banana Image Recognition Using Deep Learning
Budget: $10 – $30 USD
TASK 1
Create an object detection model using R-CNN, Yolo, or SSD. Since there's no small dataset such as MNIST and Fashion-MNIST in object detection, you can use the dataset from the d2l/notebooks/mxnet/chapter_computer-vision/object-detection-dataset.ipynb notebook.
The dataset has 1000 banana images with different rotations and sizes placed randomly on some background images.
1.A Obtain the dataset.
Working on a small dataset such as the one from the object-detection-dataset.ipynb would allow you to train and test different models without taking hours/days to finish. Go to the notebook and obtain the dataset.
1.B Build an object detector.
• using YOLO, and SSD. (need two separate notebooks)
• Using the dataset, train, or fine-tune (a pre-trained model) to create an object detector.
• Test on the test split of the dataset and report the results.
• Show image examples of the detection performance.
• Note, all platforms, i.e., mxnet, pytorch, tensorflow (and keras), have pre-trained models. You can use them, but you'll need to fine-tune and show your edits to obtain your own results.
TASK 2
• Machine translation and seq2seq modeling
• Build a seq2seq model for machine translation (English to French)
• Using the example in seq2seq.ipynb, build your own model.
• The notebook shows how to load the dataset.
• Check the machine-translation-and-dataset.ipynb for more info.
• Use LSTM or GRU, With more than 2 layers and more than 128 hidden units.
• Report the loss after the training.
• Use the model to predict multiple sentences.
• Extra (optional): Use BLEU to evaluate the predicted translations.
Create an object detection model using R-CNN, Yolo, or SSD. Since there's no small dataset such as MNIST and Fashion-MNIST in object detection, you can use the dataset from the d2l/notebooks/mxnet/chapter_computer-vision/object-detection-dataset.ipynb notebook.
The dataset has 1000 banana images with different rotations and sizes placed randomly on some background images.
1.A Obtain the dataset.
Working on a small dataset such as the one from the object-detection-dataset.ipynb would allow you to train and test different models without taking hours/days to finish. Go to the notebook and obtain the dataset.
1.B Build an object detector.
• using YOLO, and SSD. (need two separate notebooks)
• Using the dataset, train, or fine-tune (a pre-trained model) to create an object detector.
• Test on the test split of the dataset and report the results.
• Show image examples of the detection performance.
• Note, all platforms, i.e., mxnet, pytorch, tensorflow (and keras), have pre-trained models. You can use them, but you'll need to fine-tune and show your edits to obtain your own results.
TASK 2
• Machine translation and seq2seq modeling
• Build a seq2seq model for machine translation (English to French)
• Using the example in seq2seq.ipynb, build your own model.
• The notebook shows how to load the dataset.
• Check the machine-translation-and-dataset.ipynb for more info.
• Use LSTM or GRU, With more than 2 layers and more than 128 hidden units.
• Report the loss after the training.
• Use the model to predict multiple sentences.
• Extra (optional): Use BLEU to evaluate the predicted translations.