Source: https://www.cse.wustl.edu/~z.cui/projects/mcnn/ ######################################################### # Python packages ## ######################################################### 0. numpy (version I use: 1.10.4) 1. Theano (version I use: 0.7.0) Useful link for installing theano: http://deeplearning.net/software/theano/install.html ######################################################### # How To Run ## ######################################################### mcnn.py trains on one of the ucr dataset, Trace, with multiscale convolutional neural network (MSCNN). It can run on both CPU and GPU. 1. run 'python mcnn.py' ## This command trains a MSCNN with default parameters. ## You should be able to get zero error on train set, ## validation set as well as test set. 2. run "THEANO_FLAGS='blas.ldflags=-lblas -lgfortran,mode=FAST_RUN, cuda.root=/usr/local/cuda,device=gpu,floatX=float32, lib.cnmem=1' python mcnn.py" ## This command trains exactly the same MSCNN as the first one. ## However, You can enjoy 10x speedup if you have GPU with ## CUDA installed using this command. 3. run 'python mcnn.py -h' for more information. ######################################################### # Standard Output ## ######################################################### increase factor is 29 , ori len 275 train size 2320 ,valid size 580 test size 2900 batch size 232 n_train_batches is 10 data dim is 371 --------------------------- building the model... training... ...epoch 1, valid err: 0.75000 | test err: 0.43000 | train err 0.73621, cost 2.2027 ...epoch 2, valid err: 0.75000 | test err: 0.43000 | train err 0.65603, cost 1.4417 ...epoch 3, valid err: 0.75000 | test err: 0.43000 | train err 0.54526, cost 1.0825 ...epoch 4, valid err: 0.75000 | test err: 0.43000 | train err 0.48319, cost 0.9583 ######################################################### Please contact me if you have any question.