else:
# make dummy random data just for testing model inits
train = RandomEMDataIterator(name='train')
test = RandomEMDataIterator(name='test')
if not args.model_file:
# create the model based on the architecture specified via command line
arch = EMModelArchitecture.init_model_arch(
args.model_arch, train.parser.nclass,
not train.parser.independent_labels)
model = Model(layers=arch.layers)
# allocate saving counts for training labels here so they can be saved with convnet checkpoints
model.batch_meta = {
'prior_train_count':
np.zeros((train.parser.nclass, ), dtype=np.int64),
'prior_total_count':
np.zeros((1, ), dtype=np.int64)
}
if hasattr(model, 'batch_meta'):
train.parser.batch_meta['prior_train_count'] = model.batch_meta[
'prior_train_count']
train.parser.batch_meta['prior_total_count'] = model.batch_meta[
'prior_total_count']
assert (train.nmacrobatches > 0
) # no training batches specified and not in write_output mode
macro_epoch = model.epoch_index // train.nmacrobatches + 1
macro_batch = model.epoch_index % train.nmacrobatches + 1
if args.data_config and macro_batch > train.batch_range[0]:
print(
dim_ordering=args.dim_ordering, batch_range=args.test_range, name='test',
isTest=True, concatenate_batches=True, NBUF=args.nbebuf,
image_in_size=args.image_in_size) if args.callback_args['eval_freq'] else None
else:
# make dummy random data just for testing model inits
train = RandomEMDataIterator(name='train')
test = RandomEMDataIterator(name='test')
if not args.model_file:
# create the model based on the architecture specified via command line
arch = EMModelArchitecture.init_model_arch(args.model_arch, train.parser.nclass,
not train.parser.independent_labels)
model = Model(layers=arch.layers)
# allocate saving counts for training labels here so they can be saved with convnet checkpoints
model.batch_meta = {'prior_train_count':np.zeros((train.parser.nclass,),dtype=np.int64),
'prior_total_count':np.zeros((1,),dtype=np.int64)}
if hasattr(model,'batch_meta'):
train.parser.batch_meta['prior_train_count'] = model.batch_meta['prior_train_count']
train.parser.batch_meta['prior_total_count'] = model.batch_meta['prior_total_count']
assert( train.nmacrobatches > 0 ) # no training batches specified and not in write_output mode
macro_epoch = model.epoch_index//train.nmacrobatches+1
macro_batch = model.epoch_index%train.nmacrobatches+1
if args.data_config and macro_batch > train.batch_range[0]:
print('Model loaded at model epoch %d, setting to training batch %d' % (model.epoch_index,macro_batch,))
train.reset_batchnum(macro_batch)
# print out epoch and batch as they were in cuda-convnets2, starting at 1
print('Training from epoch %d to %d with %d/%d training/testing batches per epoch, %d examples/batch' \
% (macro_epoch, args.epochs, train.nmacrobatches, test.nmacrobatches if test else 0,
