def create_and_save_data_images(path, filename):
data_generator = gen_data.DataGen(
data_path=
'/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-corpus-images',
corpus_path=
'/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-20120323.corpus'
)
dict = data_generator.create_dictionary()
np.save(path + '/' + '' + filename + '_' + 'dict.npy', dict)
e_input, d_input, d_output = data_generator.read_path_images()
np.save(path + '/' + '' + filename + '_' + 'e_input.npy', e_input)
np.save(path + '/' + '' + filename + '_' + 'd_input.npy', d_input)
np.save(path + '/' + '' + filename + '_' + 'd_output.npy', d_output)
def evaluate_image_model(model, arch, op, latent_dim, num):
data_generator = gen_data.DataGen(
data_path=
"/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-corpus-images/",
corpus_path=
"/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-20120323.corpus"
)
data_generator.load_from_file(sys.argv[2], sys.argv[3])
net_eval = eval_net.NetEval(model,
data_generator.dict,
load_model=True,
latent_dim=latent_dim,
mix=("merge" in arch))
acc = 0
for i in range(num):
try:
if arch == "std":
encoder_input_data, decoder_input_data, decoder_target_data = data_generator.get_random_sample(
)
encoder_input_data = np.expand_dims(encoder_input_data, 0)
elif arch == "std_conv" or arch == "deep_conv" or arch == "vgg_19_retrain_False" or arch == "vgg_19_retrain_True":
if not op:
encoder_input_data, decoder_input_data, decoder_target_data = data_generator.get_random_image_sample(
)
else:
encoder_input_data, decoder_input_data, decoder_target_data = data_generator.get_random_image_op_sample(
)
encoder_input_data = np.expand_dims(encoder_input_data, 0)
elif arch == "std_conv_merge" or arch == "deep_conv_merge":
if not op:
encoder_input_data, decoder_input_data, decoder_target_data, encoder_input_op = data_generator.get_random_mix_sample(
)
else:
encoder_input_data, decoder_input_data, decoder_target_data, encoder_input_op = data_generator.get_random_mix_op_sample(
)
if not is_same_input_size(encoder_input_data,
encoder_input_op):
continue
encoder_input_data = [
np.expand_dims(encoder_input_data, 0),
np.expand_dims(encoder_input_op, 0)
]
acc += net_eval.test_edit_distance(
encoder_input_data, np.expand_dims(decoder_target_data, 0))
except ValueError or OSError:
continue
return acc / num
def __init__(self,
model_path,
data_location,
data_name,
latent_dim=1024,
arch="std"):
self.net_generator = gen_net.NetGen()
if arch == "std":
self.model, self.model_name = self.net_generator.get_std_net(
latent_dim)
elif arch == "std_conv":
self.model, self.model_name = self.net_generator.get_std_conv_net(
latent_dim)
elif arch == "deep_conv":
self.model, self.model_name = self.net_generator.get_deep_conv_net(
latent_dim)
elif arch == "std_conv_merge":
self.model, self.model_name = self.net_generator.get_std_conv_merge_net(
latent_dim)
elif arch == "deep_conv_merge":
self.model, self.model_name = self.net_generator.get_deep_conv_merge_net(
latent_dim)
elif arch == "vgg_19_true":
self.model, self.model_name = self.net_generator.get_vgg_19_image_net(
latent_dim, trainable=True)
elif arch == "vgg_19_false":
self.model, self.model_name = self.net_generator.get_vgg_19_image_net(
latent_dim, trainable=False)
self.data_generator = gen_data.DataGen(
data_path=
"/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-corpus-images/",
corpus_path=
"/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-20120323.corpus"
)
self.data_generator.load_from_file(data_location, data_name)
#self.data_generator.compute_pca_of_image_set()
#self.data_generator.load_from_file(data_location, data_name)
#self.data_generator.split_testset(0.1)
self.path = model_path
def test_data_generation():
data_generator = gen_data.DataGen(
data_path=
'/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-corpus-images',
corpus_path=
'/home/johannes/Documents/master_data/jkummert_master_thesis/rwth/rwth-phoenix-full-20120323.corpus'
)
e_input, d_input, d_output = data_generator.read_path()
print("Encoder Input")
print(e_input[0].shape)
print(e_input[0])
print("Decoder Input")
print(d_input[0].shape)
print(d_input[0])
print("Decoder Output")
print(d_output[0].shape)
print(d_output[0])
e_input_b, d_input_b, d_output_b = data_generator.create_batch(50)
print("Encoder Input Batch")
print(e_input_b.shape)
print("Decoder Input")
print(d_input_b.shape)
print("Decoder Output")
print(d_output_b.shape)