When i learn about deep learning, I found dataset with precomputed features form. Link (http://cs.stanford.edu/people/karpathy/deepimagesent/coco.zip). What's the different with usual dataset?
So generally you would compute say ResNet features on the images this is a time-consuming process so they've pre-computed those features and put them up there directly.
This works since in most applications you'd want to freeze those ResNet layers and just train the model past that part.
Pre-trained - You save your trained network with all its weight on a big and reliable dataset. Then use again for a similar requirement with further training on some additional data. This is a typical transfer-learning.
Precomputed - Pre-compute is the process of storing the output of all your layers before the last one re-use for some other purpose on the same dataset.
It makes the subsequent (beyond the first) training faster. Layer other than the last one will not change with new data.
Since only the last (classifier) layers change during training but not the earlier layers, we can convert the images to feature vectors just once and then train the classifier on these feature vectors instead of on the images.