model1 = emb_cnn_pre(INPUT_SHAPE, NUM_OUTPUTS)
model1.compile(optimizer='rmsprop', loss='mse')
start = time.time()
earlyStopping=EarlyStopping(monitor='loss', patience=3, verbose=0, mode='auto', restore_best_weights=True)
history = model1.fit([x_pre], [y_pre],
batch_size=36,
epochs=EPOCHS,
callbacks=[earlyStopping],
shuffle=False,
verbose=0)
end = time.time()
pre_epochs[i,k] = len(history.history['loss'])
print('Pre-training stopped at epoch ',pre_epochs[i,k], "after", end-start)
model = emb_cnn_full(model1, INPUT_SHAPE, NUM_CLASSES)
custom_callback = CustomCallback(k,i,x_test,y_test,folder,True)
csv_logger = CSVLogger(folder+str(k)+'/log_'+"{:03d}".format(i)+'.csv')
folder_chkpts = 'checkpoints/'+folder+str(k)+'/'+str(i)
if not os.path.exists(folder_chkpts):
os.makedirs(folder_chkpts)
filepath=folder_chkpts+"/model-{epoch:02d}"+"{:03d}".format(i)+".hdf5"
checkpoints = ModelCheckpoint(filepath, verbose=0, save_best_only=False, mode='auto', period=1)
callbacks = [custom_callback,csv_logger,checkpoints]
model.compile(loss={"class_output": 'categorical_crossentropy', "emb_inout": 'binary_crossentropy'},
loss_weights=[1, 1],
optimizer='adam',
metrics={"class_output": ['accuracy',acc_likelihood, acc_threshold], "emb_inout": ['mse']})
batch_size=BATCH,
epochs=EPOCHS,
callbacks=[earlyStopping],
shuffle=True,
verbose=0)
pre_epochs[i, k] = len(history.history['loss'])
print('Pre-training stopped at epoch ', pre_epochs[i, k])
batch_norm = True
if (i < 5):
batch_norm = False
model = emb_cnn_full(model1,
INPUT_SHAPE,
NUM_CLASSES,
embodied=False,
batch_norm=batch_norm)
custom_callback = CustomCallback(k, i, x_test, y_test, folder)
csv_logger = CSVLogger(folder + str(k) + '/log_' + "{:03d}".format(i) +
'.csv')
folder_chkpts = 'checkpoints/' + folder + str(k) + '/' + str(i)
if not os.path.exists(folder_chkpts):
os.makedirs(folder_chkpts)
filepath = folder_chkpts + "/model-{epoch:02d}" + "{:03d}".format(
i) + ".hdf5"
checkpoints = ModelCheckpoint(filepath,
verbose=0,
[emb_split[k % div::div]],
batch_size=BATCH,
epochs=EPOCHS,
callbacks=[earlyStopping],
shuffle=True,
verbose=0)
pre_epochs[i, k] = len(history.history['loss'])
print('Pre-training stopped at epoch ', pre_epochs[i, k])
batch_norm = True
if (i < 5):
batch_norm = False
model = emb_cnn_full(model1,
INPUT_SHAPE,
NUM_CLASSES,
batch_norm=batch_norm)
custom_callback = CustomCallback(k, i, x_test, y_test, folder, True)
csv_logger = CSVLogger(folder + str(k) + '/log_' + "{:03d}".format(i) +
'.csv')
folder_chkpts = 'checkpoints/' + folder + str(k) + '/' + str(i)
if not os.path.exists(folder_chkpts):
os.makedirs(folder_chkpts)
filepath = folder_chkpts + "/model-{epoch:02d}" + "{:03d}".format(
i) + ".hdf5"
checkpoints = ModelCheckpoint(filepath,
verbose=0,
save_best_only=False,
mode='auto',
model1 = emb_cnn_pre(INPUT_SHAPE, NUM_OUTPUTS)
model1.compile(optimizer='rmsprop', loss='mse')
earlyStopping=EarlyStopping(monitor='loss', patience=3, verbose=0, mode='auto', restore_best_weights=True)
history = model1.fit([x_pre], [y_pre],
batch_size=36,
epochs=EPOCHS,
callbacks=[earlyStopping],
shuffle=False,
verbose=0)
pre_epochs[i,k] = len(history.history['loss'])
print('Pre-training stopped at epoch ',pre_epochs[i,k])
model = emb_cnn_full(model1, INPUT_SHAPE, NUM_CLASSES, embodied=False)
plot_model(model,to_file=folder+'one-hot_pretraining.png')
custom_callback = CustomCallback(k,i,x_test,y_test,folder)
csv_logger = CSVLogger(folder+str(k)+'/log_'+"{:03d}".format(i)+'.csv')
folder_chkpts = 'checkpoints/'+folder+str(k)+'/'+str(i)
if not os.path.exists(folder_chkpts):
os.makedirs(folder_chkpts)
filepath=folder_chkpts+"/model-{epoch:02d}"+"{:03d}".format(i)+".hdf5"
checkpoints = ModelCheckpoint(filepath, verbose=0, save_best_only=False, mode='auto', period=1)
callbacks = [custom_callback,csv_logger,checkpoints]
model.compile(loss={"class_output": 'categorical_crossentropy'},