6 Data flows through Caffe as [Blobs](net_layer_blob.html#blob-storage-and-communication).
7 Data layers load input and save output by converting to and from Blob to other formats.
8 Common transformations like mean-subtraction and feature-scaling are done by data layer configuration.
9 New input types are supported by developing a new data layer -- the rest of the Net follows by the modularity of the Caffe layer catalogue.
11 This data layer definition
15 # DATA layer loads leveldb or lmdb storage DBs for high-throughput.
17 # the 1st top is the data itself: the name is only convention
19 # the 2nd top is the ground truth: the name is only convention
21 # the DATA layer configuration
24 source: "examples/mnist/mnist_train_lmdb"
25 # type of DB: LEVELDB or LMDB (LMDB supports concurrent reads)
27 # batch processing improves efficiency.
30 # common data transformations
32 # feature scaling coefficient: this maps the [0, 255] MNIST data to [0, 1]
37 loads the MNIST digits.
39 **Tops and Bottoms**: A data layer makes **top** blobs to output data to the model.
40 It does not have **bottom** blobs since it takes no input.
42 **Data and Label**: a data layer has at least one top canonically named **data**.
43 For ground truth a second top can be defined that is canonically named **label**.
44 Both tops simply produce blobs and there is nothing inherently special about these names.
45 The (data, label) pairing is a convenience for classification models.
47 **Transformations**: data preprocessing is parametrized by transformation messages within the data layer definition.
55 mean_file_size: mean.binaryproto
56 # for images in particular horizontal mirroring and random cropping
57 # can be done as simple data augmentations.
58 mirror: 1 # 1 = on, 0 = off
59 # crop a `crop_size` x `crop_size` patch:
60 # - at random during training
61 # - from the center during testing
66 **Prefetching**: for throughput data layers fetch the next batch of data and prepare it in the background while the Net computes the current batch.
68 **Multiple Inputs**: a Net can have multiple inputs of any number and type. Define as many data layers as needed giving each a unique name and top. Multiple inputs are useful for non-trivial ground truth: one data layer loads the actual data and the other data layer loads the ground truth in lock-step. In this arrangement both data and label can be any 4D array. Further applications of multiple inputs are found in multi-modal and sequence models. In these cases you may need to implement your own data preparation routines or a special data layer.
70 *Improvements to data processing to add formats, generality, or helper utilities are welcome!*
74 Refer to the layer catalogue of [data layers](layers.html#data-layers) for close-ups on each type of data Caffe understands.
78 For on-the-fly computation deployment Nets define their inputs by `input` fields: these Nets then accept direct assignment of data for online or interactive computation.