Closed yamins81 closed 10 years ago
yamins81
commented
10 years ago @daseibert Can you also review this?
@daseibert I noticed that you removed the img_size and num_colors params from the CroppedGeneralDataProvider and then hard-coded the values for those params. I've put them back, because I want this to be able to work in general. Notice that you can set them in the call to convnet.py (using the img-size and img-channels params).
yamins81
commented
10 years ago Some documentation is needed.
Here is a snipped from a script that uses it:
python script:
dp_params = {'dataset_name': ('dldata.stimulus_sets.synthetic.synthetic_datasets', 'TrainingDataset'), 'batch_size': 128, 'meta_attribute': 'obj', 'perm_type': 'random', 'perm_seed': 0, 'preproc': {'resize_to': (128, 128, 3), 'mode': 'RGB', 'dtype': 'float32', 'normalize': False} }
if name == "main": op = ConvNet.get_options_parser() op, load_dic = IGPUModel.parse_options(op) nr.seed(op.options['random_seed'].value) model = ConvNet(op, load_dic, dp_params=dp_params) model.start()
Command line: python run_synthetic_training.py --data-path=/export/imgnet_storage_full/yamins_skdata/sythetic_batches_0 --crop=7 --save-path=/export/imgnet_storage_full/ --test-range=950-999 --train-range=0-949 --layer-def=/home/yamins/archconvnets/archconvnets/convnet/ut_model_full/layer_nofc_0.cfg --layer-params=/home/darren/archconvnets/archconvnets/convnet/ut_model_full/layer-params.cfg --data-provider=general-cropped --test-freq=50 --conserve-mem=1 --max-filesize=99999999 --img-size=128
@ardila can you review this