Cuda-convnet

(24th-October-2020)

Code of the Supervision team of ILSVRC 2012 champion • nVidia GTX 580 GPUs • Linux and CUDA

• Data file creation

• - Make a Python list (samples are randomly ordered), divide it into multiple files (called batch) with cPickle and save - multiple training batches and one or more test batch

• 2. Write data reading function - data size and number of planes, with or without normalization

• 3. Think about the architecture 4. Execution of learning - One of training batches for validation - Iterative training (tens to hundreds of epochs) - Iterative training using all batches until test error continues to decrease - Reduce learning coefficient, repeat multiple times (on the order of 10 epochs) - End

Gray level 28 × 28 60000 training

· 10000 test sample → for training 6 batch for testing 1 batch