img_width = X_train.shape[3]
#print X_train.shape
# Step1b: init variables
n_batch = 32
num_epochs = 100
# Step1c: init batch iterators
from utils import batchiterator
batchitertrain = batchiterator.BatchIterator(batch_indices=range(
X_train.shape[0]),
batchsize=n_batch,
data=(X_train, y_train))
num_data = batchitertrain.n
batchitertrain = batchiterator.threaded_generator(batchitertrain, 10)
batchiterval = batchiterator.BatchIterator(batch_indices=range(X_val.shape[0]),
batchsize=n_batch,
data=(X_val, y_val))
batchiterval = batchiterator.threaded_generator(batchiterval, 10)
# Step2: define model
l_in = lasagne.layers.InputLayer(shape=(n_batch, n_channels, img_height,
img_width))
labels = T.ivector('label')
l_ram = ram.layers.RAMLayer(
l_in, # input images (n_batch x n_channels x img_height x img_width)
k=1, # number of glimps scales
patch=8, # size of glimps patch
import time
from utils import batchiterator
import numpy as np
import theano
batchsize = 32
print "imagenet batchiterator"
batchitertrain = batchiterator.ImagenetBatchIterator(
batchsize=batchsize,
database='/media/data0/image/ilsvrc12/ilsvrc12_train_lmdb')
batchitertrain = batchiterator.threaded_generator(batchitertrain, 3)
time.sleep(1.) # pause 1. seconds
num_epoch = 30
for i in xrange(num_epoch):
start_time = time.time()
[X_batch, y_batch] = batchitertrain.next()
end_time = time.time()
time.sleep(0.4) # approximated processing time
print "time (", i, "): ", end_time - start_time, " sec"
print "imagenet batchiterator (crop)"
batchitertrain = batchiterator.ImagenetBatchCropIterator(
batchsize=batchsize,
crop='random',
crop_height=224,
crop_width=224,
import time
from utils import batchiterator
import numpy as np
import theano
batchsize = 32
print "imagenet batchiterator"
batchitertrain = batchiterator.ImagenetBatchIterator(batchsize=batchsize, database='/media/data0/image/ilsvrc12/ilsvrc12_train_lmdb')
batchitertrain = batchiterator.threaded_generator(batchitertrain,3)
time.sleep(1.) # pause 1. seconds
num_epoch = 30
for i in xrange(num_epoch):
start_time = time.time()
[X_batch, y_batch] = batchitertrain.next()
end_time = time.time()
time.sleep(0.4) # approximated processing time
print "time (", i, "): ", end_time-start_time, " sec"
print "imagenet batchiterator (crop)"
batchitertrain = batchiterator.ImagenetBatchCropIterator(batchsize=batchsize, crop='random', crop_height=224, crop_width=224, flip=True, database='/media/data0/image/ilsvrc12/ilsvrc12_train_lmdb')
batchitertrain = batchiterator.threaded_generator(batchitertrain,3)
time.sleep(1.) # pause 1. seconds
num_epoch = 30
X_train, y_train, X_val, y_val, X_test, y_test = load_dataset()
n_data = X_train.shape[0]
n_channels = X_train.shape[1]
img_height = X_train.shape[2]
img_width = X_train.shape[3]
#print X_train.shape
# Step1b: init variables
n_batch = 32
num_epochs = 100
# Step1c: init batch iterators
from utils import batchiterator
batchitertrain = batchiterator.BatchIterator(batch_indices=range(X_train.shape[0]), batchsize=n_batch, data=(X_train, y_train))
num_data = batchitertrain.n
batchitertrain = batchiterator.threaded_generator(batchitertrain,10)
batchiterval = batchiterator.BatchIterator(batch_indices=range(X_val.shape[0]), batchsize=n_batch, data=(X_val, y_val))
batchiterval = batchiterator.threaded_generator(batchiterval,10)
# Step2: define model
l_in = lasagne.layers.InputLayer(shape=(n_batch, n_channels, img_height, img_width))
labels = T.ivector('label')
l_ram=ram.layers.RAMLayer(l_in, # input images (n_batch x n_channels x img_height x img_width)
k=1, # number of glimps scales
patch=8, # size of glimps patch
n_steps=6, # number of glimps steps
lambda_=10.0, # mixing ratio between
n_h_g=128, # number of hidden units in h_g (in glimps network)
