def run_model(
hdf5_file,
num_authors,
num_forms_per_author,
shingle_dim,
num_iters=10,
batch_size=32,
use_form=False,
lr=0.03,
loadweights=None,
):
# Create Keras model
model = create_model(num_authors, shingle_dim, lr)
if loadweights:
model.load_weights(loadweights)
# Create a mini_batcher for
iam_m = IAM_MiniBatcher(
hdf5_file,
num_authors,
num_forms_per_author,
shingle_dim=shingle_dim,
use_form=use_form,
default_mode=MiniBatcher.TRAIN,
batch_size=batch_size,
)
# Get validation dataset
[X_test, Y_test] = iam_m.get_test_batch(batch_size * 20)
X_test = np.expand_dims(X_test, 1)
Y_test = to_categorical(Y_test, num_authors)
print "test_size:", X_test.shape, Y_test.shape
for i in range(num_iters):
print "Starting Epoch: ", i
start_time = time.time()
# Get training batch
(X_train, Y_train) = iam_m.get_train_batch(batch_size * 100)
X_train = np.expand_dims(X_train, 1)
Y_train = to_categorical(Y_train, num_authors)
# TODO: Maybe we should only validate every N iters since right now we are doing 20% extra work every iter
model.fit(
X_train,
Y_train,
batch_size=batch_size,
nb_epoch=1,
show_accuracy=True,
verbose=1,
validation_data=(X_test, Y_test),
) # verbose=1
print "Elapsed Time: ", time.time() - start_time
# Set checkpoint every 500 iterations
if i % 500 == 0 and i != 0:
fname = "authors_%d_forms_per_author_%d_epoch_%d.hdf5" % (num_authors, num_forms_per_author, i)
model.save_weights("%s.hdf5" % fname, overwrite=True)
fname = "authors_%d_forms_per_author_%d_final.hdf5" % (num_authors, num_forms_per_author)
model.save_weights("%s.hdf5" % fname, overwrite=True)
def run_model(hdf5_file,
num_authors,
num_forms_per_author,
shingle_dim,
num_iters=10,
batch_size=32,
use_form=False,
lr=0.03,
loadweights=None):
# Create Keras model
model = create_model(num_authors, shingle_dim, lr)
if loadweights:
model.load_weights(loadweights)
# Create a mini_batcher for
iam_m = IAM_MiniBatcher(hdf5_file,
num_authors,
num_forms_per_author,
shingle_dim=shingle_dim,
use_form=use_form,
default_mode=MiniBatcher.TRAIN,
batch_size=batch_size)
# Get validation dataset
[X_test, Y_test] = iam_m.get_test_batch(batch_size * 20)
X_test = np.expand_dims(X_test, 1)
Y_test = to_categorical(Y_test, num_authors)
print 'test_size:', X_test.shape, Y_test.shape
for i in range(num_iters):
print 'Starting Epoch: ', i
start_time = time.time()
# Get training batch
(X_train, Y_train) = iam_m.get_train_batch(batch_size * 100)
X_train = np.expand_dims(X_train, 1)
Y_train = to_categorical(Y_train, num_authors)
# TODO: Maybe we should only validate every N iters since right now we are doing 20% extra work every iter
model.fit(X_train,
Y_train,
batch_size=batch_size,
nb_epoch=1,
show_accuracy=True,
verbose=1,
validation_data=(X_test, Y_test)) # verbose=1
print 'Elapsed Time: ', time.time() - start_time
# Set checkpoint every 500 iterations
if i % 500 == 0 and i != 0:
fname = 'authors_%d_forms_per_author_%d_epoch_%d.hdf5' % (
num_authors, num_forms_per_author, i)
model.save_weights('%s.hdf5' % fname, overwrite=True)
fname = 'authors_%d_forms_per_author_%d_final.hdf5' % (
num_authors, num_forms_per_author)
model.save_weights('%s.hdf5' % fname, overwrite=True)
from keras.optimizers import SGD
from keras.utils.np_utils import to_categorical
from keras.layers.normalization import BatchNormalization as BN
import matplotlib.pylab as plt
import sys
sys.path.append('/work/code/repo/d-script/')
# d-script imports
from data_iters.minibatcher import MiniBatcher
from data_iters.iam_hdf5_iterator import IAM_MiniBatcher
from fielutil import *
hdf5_file = '/memory/author_lines.hdf5'
num_forms_per_author=50; batch_size=32; num_authors=47; shingle_dim=(120,120); use_form=True
iam_m = IAM_MiniBatcher(hdf5_file, num_authors, num_forms_per_author, shingle_dim=shingle_dim, use_form=use_form, default_mode=MiniBatcher.TRAIN, batch_size=batch_size)
[X_test, Y_test] = iam_m.get_test_batch(batch_size*20)
X_test = np.expand_dims(X_test, 1)
X_test = randangle(X_test)
Y_test = to_categorical(Y_test, num_authors)
im = smi.imread('/fileserver/iam/forms/h07-025a.png')
im = 1.0-im/256.0
maxx, maxy = im.shape
maxx = maxx/3
maxy = maxy/3
halfim = im[ :maxx, :maxy ]
halfim = np.expand_dims( np.expand_dims( halfim, 0 ), 0 )
if True:
print "Loading original weights into GPU memory"