experiment_name = "experiment_{}_batch_size_{}_bn_{}_dr{}_nl_{}_ml_{}".format(
experiment_prefix, batch_size, batch_norm, dropout_rate, n_layers, max_len)
saved_models_filepath, logs_filepath = build_experiment_folder(
experiment_name, logs_path)
filepath = "{}/best_validation_{}".format(saved_models_filepath,
experiment_name)
checkpoint = ModelCheckpoint(filepath,
monitor='val_acc',
verbose=1,
save_best_only=True,
mode='max')
from multimodaldata import get_data
train_set_audio, valid_set_audio, test_set_audio, train_set_text, valid_set_text, test_set_text, \
train_set_visual, valid_set_visual, test_set_visual, \
y_train, y_valid, y_test = get_data(max_len_audio=20, max_len_text=15, max_len_visual=20)
filepath = "{}/best_validation_{}_".format(saved_models_filepath,
experiment_name)
weights = "{}_weights{}.h5"
# pdb.set_trace()
k = 3
m = 2
# AUDIO
model1_in = Input(name="Audio_Covarep",
shape=(train_set_audio.shape[1], train_set_audio.shape[2]))
model1_cnn = Conv1D(filters=64, kernel_size=k, activation='relu')(model1_in)
model1_mp = MaxPooling1D(m)(model1_cnn)
model1_fl = Flatten()(model1_mp)
model1_dense = Dense(128, activation="relu",
W_regularizer=l2(0.0001))(model1_fl)
experiment_name)
weights = "{}late_fusion_weights{}.h5"
# pdb.set_trace()
k = 3
m = 2
for max_len in [15, 20, 25, 30]:
for dropout_rate in [0.0, 0.1, 0.2]:
for n_layers in [1, 2, 3]:
logging.info("New experiment")
logging.info("*" * 30)
train_set_audio, valid_set_audio, test_set_audio, train_set_text, valid_set_text, test_set_text, \
train_set_visual, valid_set_visual, test_set_visual, \
y_train, y_valid, y_test = get_data(max_len_audio=max_len, max_len_text=max_len, max_len_visual=max_len)
# AUDIO
model1_in = Input(name="Audio_Covarep",
shape=(train_set_audio.shape[1],
train_set_audio.shape[2]))
model1_cnn = Conv1D(filters=64, kernel_size=k,
activation='relu')(model1_in)
model1_mp = MaxPooling1D(m)(model1_cnn)
model1_fl = Flatten()(model1_mp)
model1_dropout = Dropout(dropout_rate)(model1_fl)
model1_dense = Dense(128,
activation="relu",
W_regularizer=l2(0.0001))(model1_dropout)
for i in range(2, n_layers + 1):
model1_dropout = Dropout(dropout_rate)(model1_dense)