from gluoncv import model_zoo as mzoo from mxnet import autograd, gluon, init, nd from mxnet.gluon import nn, loss as gloss from indian_dataset import IndianDataset num_epochs = 5 batch_size = 100 out_put_num = 16 dropout_rate = 0.8 ctx = mx.gpu() #modelname = 'indian_try' modelname = 'indian_conv_msra' para_filepath = os.path.join(this_dir, '..', 'symbols', 'para', '%s.params' % (modelname)) # dataset train_data = gluon.data.DataLoader(dataset=IndianDataset(train=True), batch_size=batch_size, shuffle=True, last_batch='rollover') val_data = gluon.data.DataLoader(dataset=IndianDataset(train=False), batch_size=batch_size, shuffle=False) # model net = nn.Sequential() #net.add( # nn.Dense(500,activation='relu'), # nn.Dense(256,activation='relu'), # nn.Dropout(dropout_rate), # nn.Dense(out_put_num,activation='sigmoid') # )
from mxnet import autograd, gluon, init, nd from mxnet.gluon import nn, loss as gloss from indian_dataset import IndianDataset num_epochs = 5 batch_size = 100 out_put_num = 16 dropout_rate=0.8 stochastic_ratio = 0.01 val_acc_bk = 0 ctx = mx.gpu() #modelname = 'indian_try' modelname = 'indian_simple_ema' para_filepath = os.path.join(this_dir,'..','symbols','para','%s.params'%(modelname)) # dataset train_data = gluon.data.DataLoader(dataset=IndianDataset(train=True), batch_size=batch_size ,shuffle=True,last_batch='rollover') val_data = gluon.data.DataLoader(dataset=IndianDataset(train=False), batch_size=batch_size ,shuffle=False) # g(x) : stochastic input augmentation function #def g(x): # return x + nd.random.normal(0,stochastic_ratio,shape=x.shape) # model basemodel_zoo = 'simple2' net = symbols.get_model(basemodel_zoo) net_t = symbols.get_model(basemodel_zoo) #net.initialize(mx.init.Xavier(magnitude=2.24)) #net.initialize(mx.init.MSRAPrelu()) #net_t.initialize(mx.init.MSRAPrelu()) net_t.load_parameters(os.path.join(this_dir,'..','symbols','para','%s.params'%("indian_simple20.870362")) )