def get_cnn(dataprovider=None):
if dataprovider is None:
dataprovider = dp
en = CnnEmbedding(output_size=100,
cnn_layers_per_block=1,
block_feature_counts=[32, 64],
fc_layer_feature_counts=[100],
hidden_activation='relu',
final_activation='relu',
batch_norm_for_init_layer=True,
dropout_after_max_pooling=0.5,
dropout_after_fc=0.5)
c_nn = ClusterNNTry04_Ddbc(dp,
20,
en,
lstm_layers=0,
lstm_units=1,
cluster_count_dense_layers=0,
cluster_count_dense_units=1,
output_dense_layers=0,
output_dense_units=1,
cluster_count_lstm_layers=0,
cluster_count_lstm_units=1)
c_nn.minibatch_size = 35
c_nn.validate_every_nth_epoch = 10
c_nn.optimizer = Adadelta(clipnorm=0.5)
c_nn.use_cluster_count_loss = False
c_nn.use_similarities_loss = False
c_nn.fixed_cluster_count_output = dp.get_max_cluster_count()
return c_nn
def get_cnn(dataprovider=None):
if dataprovider is None:
dataprovider = dp
en = CnnEmbedding(
output_size=256, cnn_layers_per_block=1, block_feature_counts=[32, 64, 128],
fc_layer_feature_counts=[256], hidden_activation=LeakyReLU(), final_activation=LeakyReLU(),
batch_norm_for_init_layer=False, batch_norm_after_activation=True, batch_norm_for_final_layer=True
)
c_nn = ClusterNNTry00_V135(dataprovider, 20, en, lstm_layers=14, internal_embedding_size=96*3, cluster_count_dense_layers=1, cluster_count_dense_units=256,
output_dense_layers=0, output_dense_units=256, cluster_count_lstm_layers=1, cluster_count_lstm_units=128,
kl_embedding_size=128, kl_divergence_factor=0., simplified_center_loss_factor=0.,
cluster_assignment_regularization_factor=10.0, use_v02_cluster_assignment_loss=True)
c_nn.include_self_comparison = False
c_nn.weighted_classes = True
c_nn.class_weights_approximation = 'stochastic'
c_nn.minibatch_size = 25
c_nn.class_weights_post_processing_f = lambda x: np.sqrt(x)
c_nn.set_loss_weight('similarities_output', 5.0)
c_nn.optimizer = Adadelta(lr=5.0)
validation_factor = 10
c_nn.early_stopping_iterations = 15001
c_nn.validate_every_nth_epoch = 10 * validation_factor
c_nn.validation_data_count = c_nn.minibatch_size * validation_factor
# c_nn.prepend_base_name_to_layer_name = False
return c_nn
ds_dir + 'datasets/TIMIT/traininglist_100/testlist_20.txt')
dp = TIMITDataProvider(
# data_dir=top_dir + "/test/TIMIT_mini", cache_directory=top_dir + "/test/cache",
data_dir=top_dir + "/TIMIT",
cache_directory=top_dir + "/test/cache",
min_cluster_count=1,
max_cluster_count=5,
return_1d_audio_data=False,
test_classes=TIMIT_lst,
validate_classes=TIMIT_lst,
concat_audio_files_of_speaker=True)
en = CnnEmbedding(output_size=256,
cnn_layers_per_block=1,
block_feature_counts=[32, 64],
fc_layer_feature_counts=[],
hidden_activation=LeakyReLU(),
final_activation=LeakyReLU(),
batch_norm_for_init_layer=False,
batch_norm_after_activation=True,
batch_norm_for_final_layer=True)
c_nn = ClusterNNTry00_V23(dp,
20,
en,
lstm_layers=7,
internal_embedding_size=96,
cluster_count_dense_layers=1,
cluster_count_dense_units=256,
output_dense_layers=1,
output_dense_units=256,
cluster_count_lstm_layers=2,
dp = TIMITDataProvider(
# data_dir=top_dir + "/test/TIMIT_mini", cache_directory=top_dir + "/test/cache",
data_dir=top_dir + "/TIMIT", cache_directory=top_dir + "/test/cache",
min_cluster_count=1,
max_cluster_count=5,
return_1d_audio_data=False,
test_classes=TIMIT_lst[:100],
validate_classes=TIMIT_lst[100:],
concat_audio_files_of_speaker=True,
minimum_snippets_per_cluster=2,
window_width=200
)
en = CnnEmbedding(
output_size=256, cnn_layers_per_block=1, block_feature_counts=[32, 64, 128], fc_layer_feature_counts=[512],
hidden_activation=LeakyReLU(), final_activation=LeakyReLU(),
batch_norm_for_init_layer=False, batch_norm_after_activation=True, batch_norm_for_final_layer=True,
dropout_init=0.25, dropout_after_max_pooling=[0.25, 0.25, 0.25], dropout_after_fc=0.25
)
def get_cnn():
c_nn = ClusterNNTry00_V98(dp, 20, en, lstm_layers=14, internal_embedding_size=96*3, cluster_count_dense_layers=1, cluster_count_dense_units=256,
output_dense_layers=0, output_dense_units=256, cluster_count_lstm_layers=1, cluster_count_lstm_units=128,
kl_embedding_size=128, kl_divergence_factor=0.1, simplified_center_loss_factor=0.5)
c_nn.include_self_comparison = False
c_nn.weighted_classes = True
c_nn.class_weights_approximation = 'stochastic'
c_nn.minibatch_size = 15
c_nn.class_weights_post_processing_f = lambda x: np.sqrt(x)
c_nn.set_loss_weight('similarities_output', 5.0)
c_nn.optimizer = Adadelta(lr=5.0)
concat_audio_files_of_speaker=True,
test_classes=TIMIT_lst[:200],
validate_classes=TIMIT_lst[:200],
# # Create at least two 1s snippets per speaker and create also some hints
# window_width=[(100, 100), (200, 200), (300, 300), (400, 400)],
# minimum_snippets_per_cluster=[(100, 100), (100, 100)],
# split_audio_pieces_longer_than_and_create_hints=100
window_width=256,
minimum_snippets_per_cluster=2)
en = CnnEmbedding(output_size=256,
cnn_layers_per_block=1,
block_feature_counts=[32, 64, 128],
fc_layer_feature_counts=[256],
hidden_activation=LeakyReLU(),
final_activation=LeakyReLU(),
batch_norm_for_init_layer=False,
batch_norm_after_activation=True,
batch_norm_for_final_layer=True,
max_pooling_size=[(4, 2), (2, 2), (2, 2)],
max_pooling_stride=[(4, 2), (2, 2), (2, 2)])
def get_cnn():
c_nn = ClusterNNTry00_V93(dp,
40,
en,
lstm_layers=14,
internal_embedding_size=96 * 3,
cluster_count_dense_layers=1,
cluster_count_dense_units=512,
output_dense_layers=1,
# data_dir=top_dir + "/test/TIMIT_mini", cache_directory=top_dir + "/test/cache",
data_dir=top_dir + "/TIMIT", cache_directory=top_dir + "/test/cache",
min_cluster_count=10,
max_cluster_count=10,
return_1d_audio_data=False,
train_classes=speaker_list,
test_classes=speaker_list,
validate_classes=speaker_list,
concat_audio_files_of_speaker=True
)
en = CnnEmbedding(
output_size=100, cnn_layers_per_block=1, block_feature_counts=[32, 64],
fc_layer_feature_counts=[100], hidden_activation='relu', final_activation='relu',
batch_norm_for_init_layer=True, cnn_filter_size=5
)
c_nn = ClusterNNTry04_Ddbc(dp, 20, en, lstm_layers=0, lstm_units=1, cluster_count_dense_layers=0, cluster_count_dense_units=1,
output_dense_layers=0, output_dense_units=1, cluster_count_lstm_layers=0, cluster_count_lstm_units=1)
c_nn.minibatch_size = 40
c_nn.validate_every_nth_epoch = 10
c_nn.optimizer = Adadelta(clipnorm=0.5)
c_nn.use_cluster_count_loss = False
c_nn.use_similarities_loss = False
c_nn.fixed_cluster_count_output = dp.get_max_cluster_count()
# c_nn.f_cluster_count = lambda: 10
# c_nn.minibatch_size = 200
'MPRT0', 'MCTT0', 'FEME0', 'MCRE0'
]
dp = TIMITDataProvider(
# data_dir=top_dir + "/test/TIMIT_mini", cache_directory=top_dir + "/test/cache",
data_dir=top_dir + "/TIMIT",
cache_directory=top_dir + "/test/cache",
min_cluster_count=1,
max_cluster_count=5,
return_1d_audio_data=False,
test_classes=TIMIT20_lst,
validate_classes=TIMIT20_lst,
concat_audio_files_of_speaker=True)
en = CnnEmbedding(output_size=128,
cnn_layers_per_block=1,
block_feature_counts=[32, 64],
fc_layer_feature_counts=[],
hidden_activation='relu',
final_activation='tanh',
batch_norm_for_init_layer=True)
en.regularizer = l2(0.001)
c_nn = ClusterNNTry00_V17(dp,
20,
en,
lstm_layers=7,
lstm_units=96,
cluster_count_dense_layers=1,
cluster_count_dense_units=256,
output_dense_layers=1,
output_dense_units=256,
cluster_count_lstm_layers=2,
required_input_count = dp.get_required_input_count_for_full_test('test')
print("Required input count: {}".format(required_input_count))
data = dp.get_data(required_input_count, 1)
# dp = Birds200DataProvider(
# min_cluster_count=1,
# max_cluster_count=2,
# )
# dp = Simple2DPointDataProvider(min_cluster_count=1, max_cluster_count=2, allow_less_clusters=False)
# en = SimpleFCEmbedding(hidden_layers=[3])
# en = BDLSTMEmbedding()
# en = None
en = CnnEmbedding(block_feature_counts=[1, 2, 3], fc_layer_feature_counts=[4], output_size=3, dimensionality='auto')
# cnn = MinimalClusterNN(dp, 5, en, weighted_classes=True)
input_count = required_input_count
cnn = ClusterNNHint(dp, input_count, en, weighted_classes=True, lstm_layers=2, cluster_count_lstm_units=2, cluster_count_lstm_layers=1, cluster_count_dense_layers=1,
cluster_count_dense_units=1, output_dense_layers=1, output_dense_units=1)
cnn = ClusterNNMergedInputs(dp, input_count, en, weighted_classes=True)
# cnn.class_weights_approximation = 'stochastic'
# cnn.build_networks(print_summaries=True)
t_start = time()
cnn.build_networks(print_summaries=False, build_training_model=False)
t_end = time()
print("Required time to build the networks: {} s".format(t_end - t_start))
cnn.minibatch_size = 2
cnn.validate_every_nth_epoch = 1
from impl.nn.base.embedding_nn.cnn_embedding import CnnEmbedding
is_linux = platform == "linux" or platform == "linux2"
top_dir = "/cluster/home/meierbe8/data/MT/" if is_linux else "G:/tmp/"
ds_dir = "./" if is_linux else "../"
dp = FashionMNISTDataProvider(
min_cluster_count=1,
max_cluster_count=5,
)
en = CnnEmbedding(output_size=32,
cnn_layers_per_block=2,
block_feature_counts=[64, 128, 256],
fc_layer_feature_counts=[512],
hidden_activation=LeakyReLU(),
final_activation=LeakyReLU(),
batch_norm_for_init_layer=False,
batch_norm_after_activation=True,
batch_norm_for_final_layer=True,
dropout_init=.5,
dropout_after_max_pooling=[.5, .5, .5])
c_nn = ClusterNNTry00_V30(dp,
20,
en,
lstm_layers=7,
internal_embedding_size=96,
cluster_count_dense_layers=1,
cluster_count_dense_units=256,
output_dense_layers=1,
output_dense_units=256,
# data_dir=top_dir + "/test/TIMIT_mini", cache_directory=top_dir + "/test/cache",
data_dir=top_dir + "/TIMIT",
cache_directory=top_dir + "/test/cache",
min_cluster_count=1,
max_cluster_count=5,
return_1d_audio_data=True,
test_classes=TIMIT_lst,
validate_classes=TIMIT_lst,
concat_audio_files_of_speaker=True,
minimum_snippets_per_cluster=[(200, 200), (100, 100)],
window_width=[(100, 200)])
en = CnnEmbedding(output_size=256,
cnn_layers_per_block=1,
block_feature_counts=[64, 128, 256],
fc_layer_feature_counts=[256],
hidden_activation=LeakyReLU(),
final_activation=LeakyReLU(),
batch_norm_for_init_layer=False,
batch_norm_after_activation=True,
batch_norm_for_final_layer=True,
dimensionality='1d')
c_nn = ClusterNNTry00_V51(dp,
20,
en,
lstm_layers=7,
internal_embedding_size=96,
cluster_count_dense_layers=1,
cluster_count_dense_units=256,
output_dense_layers=1,
output_dense_units=256,
cluster_count_lstm_layers=1,