param.MODE, w2v_step, param.USE_LSI, pse, le, targets_train,
pre_seq_train, post_seq_train, active_meds_train,
active_classes_train, depa_train, param.W2V_EMBEDDING_DIM,
param.SEQUENCE_LENGTH, param.N_ENCODER_LOSS_SAMPLES)
sample_test_generator = TransformedGenerator(
param.MODE, w2v_step, param.USE_LSI, pse, le, targets_test,
pre_seq_test, post_seq_test, active_meds_test, active_classes_test,
depa_test, param.W2V_EMBEDDING_DIM, param.SEQUENCE_LENGTH,
param.N_ENCODER_LOSS_SAMPLES)
batch_data_X_val, batch_data_y_val = sample_test_generator.__getitem__(
0)
samples_X_val = batch_data_X_val['pse_input']
samples_y_val = batch_data_y_val['main_output']
# Define the network and train
n = neural_network(param.MODE, param.L1L2RATIO)
if param.MODE == 'retrospective-gan':
custom_objects_dict = {
'autoencoder_accuracy': n.autoencoder_accuracy,
'autoencoder_false_neg_rate': n.autoencoder_false_neg_rate,
'Sampling': Sampling
}
val_monitor_losses = []
# TODO load previously saved models and resume
# The code to save and resume training the gan is buggy and doesn't proprely restore the model weights. Do not use for now.
# TODO fix model saving and resuming training with gan
'''
if i == initial_fold:
try:
le,
targets_val,
seq_val,
active_meds_val,
active_classes_val,
depa_val,
W2V_EMBEDDING_DIM,
SEQUENCE_LENGTH,
BATCH_SIZE,
shuffle=False)
#%%[markdown]
# #### Instantiate the model
#%%
n = neural_network()
callbacks = n.callbacks(SAVE_PATH)
model = n.define_model(LSTM_SIZE, N_LSTM, DENSE_PSE_SIZE, CONCAT_SIZE,
DENSE_SIZE, DROPOUT, L2_REG, SEQUENCE_LENGTH,
W2V_EMBEDDING_DIM, pse_shape, N_PSE_DENSE, N_DENSE,
output_n_classes)
print(model.summary())
tf.keras.utils.plot_model(model, to_file=os.path.join(SAVE_PATH, 'model.png'))
#%%[markdown]
# #### Train the model
#
# Use in_ipynb to check if running in Jupyter or not to print
# progress bars if in terminal and log only at epoch level if
# in Jupyter. This is a bug of Jupyter or Keras where progress
# bars will flood stdout slowing down and eventually crashing
filtered_targets,
filtered_pre_seqs,
filtered_post_seqs,
filtered_active_meds,
filtered_active_classes,
filtered_depas,
param.W2V_EMBEDDING_DIM,
param.SEQUENCE_LENGTH,
param.BATCH_SIZE,
shuffle=False)
#%%[markdown]
# ### Instantiate the model
#%%
n = neural_network(param.MODE)
if param.MODE in ['retrospective-autoenc', 'retrospective-gan']:
custom_objects_dict = {
'autoencoder_accuracy': n.autoencoder_accuracy,
'autoencoder_false_neg_rate': n.autoencoder_false_neg_rate,
'combined_l1l2loss': n.combined_l1l2loss
}
else:
custom_objects_dict = {
'sparse_top10_accuracy': n.sparse_top10_accuracy,
'sparse_top30_accuracy': n.sparse_top30_accuracy
}
model = tf.keras.models.load_model(os.path.join(save_path, 'model.h5'),
custom_objects=custom_objects_dict)