I did a small project to classify certain tropical fishes. The import of images worked well, but again, I faced a similar issue in handling large image archives.

(keeping three different classes of the image in separate folders)

SetDirectory[NotebookDirectory[]];
C01 = FileNames[All, "Fishes\\Guppy"];
C02 = FileNames[All, "Fishes\\Sword"];
C03 = FileNames[All, "Fishes\\Zebra"];

(define the association between the file path and the class)

Data01 = C01[[#]] -> "Guppy" & /@ Range[Length@C01];
Data02 = C02[[#]] -> "Sword" & /@ Range[Length@C02];
Data03 = C03[[#]] -> "Zebra" & /@ Range[Length@C03];

(union all defined associations)

Data = Union[Data01, Data02, Data03];

(Perform imports)

myData = Import@Data[[#]][[1]] -> Data[[#]][[2]] & /@Range[Length@Data];

(Train)

myClassify =  Classify[myData, TargetDevice -> "CPU"]

Thanks for your reply!

Even if the resolution of each image is 200x200 (which is a conservative estimate), then the total size of the imported images will be 30000x200x200x8=21.6 GB! I tried training on a smaller subset of the dataset, but the results were not that good and so I want to train on the entire dataset.

Each floating point number is 8 bytes right? So you would need to multiply by 8 as well. The computation that I did is just for grayscale images. The classes are approximately sampled, and there are around 1000-1300 images for each of the 30 classes. Even if the sampling approach worked, I still want to know if there is a way to feed a list of files to Classify, or alternatively how to write a sampler/data loader. It seems like a general and needed approach that may be needed if files are too large to fit into memory or if you want to train on a dataset that is online (i.e., each image is a URL reference and not saved on the disk). You can do both of these with NetTrain.