def train(aug_no, model_mode='mymodel', set_epochs=10, do_es=False):
train_dir = os.path.join(cwd, "da_concat_{}".format(aug_no))
train_data, train_label = inputDataCreator(train_dir,
224,
normalize=True,
one_hot=True)
validation_data, validation_label = inputDataCreator(validation_dir,
224,
normalize=True,
one_hot=True)
print("train data shape : ", train_data.shape)
print("train label shape : ", train_label.shape)
INPUT_SIZE = train_data.shape[1]
print("INPUT_SIZE: ", INPUT_SIZE)
CHANNEL = train_data.shape[3]
print("set channel : ", CHANNEL)
batch_size = 10
print("set batch_size : ", batch_size)
mh = ModelHandler(INPUT_SIZE, CHANNEL)
if model_mode == 'mymodel':
model = mh.buildMyModel()
elif model_mode == 'tlearn':
model = mh.buildTlearnModel(base='mnv1')
model.summary()
if do_es:
es = EarlyStopping(monitor='val_loss',
patience=5,
verbose=1,
mode='auto')
es = [es]
else:
es = None
history = model.fit(train_data,
train_label,
batch_size=batch_size,
epochs=set_epochs,
validation_data=(validation_data, validation_label),
callbacks=es,
verbose=1)
# make log dir -----
if do_es:
log_dir = os.path.join(cwd, 'log_with_es')
else:
log_dir = os.path.join(cwd, 'log')
os.makedirs(log_dir, exist_ok=True)
child_log_dir = os.path.join(log_dir, '{}_{}'.format(aug_no, model_mode))
os.makedirs(child_log_dir, exist_ok=True)
# save model & weights
model_file = os.path.join(child_log_dir,
'{}_{}_model.h5'.format(aug_no, model_mode))
model.save(model_file)
# save history
history_file = os.path.join(child_log_dir,
'{}_{}_history.pkl'.format(aug_no, model_mode))
with open(history_file, 'wb') as p:
pickle.dump(history.history, p)
print("export logs in ", child_log_dir)
def main():
cwd = os.getcwd()
sub_prj = os.path.dirname(cwd)
sub_prj_root = os.path.dirname(sub_prj)
prj_root = os.path.dirname(sub_prj_root)
data_dir = os.path.join(prj_root, "datasets")
data_src = os.path.join(data_dir, "small_721")
print("\ndata source: ", data_src)
use_da_data = False
if use_da_data:
train_dir = os.path.join(data_src, "train_with_aug")
else:
train_dir = os.path.join(data_src, "train")
validation_dir = os.path.join(data_src, "validation")
test_dir = os.path.join(data_src, "test")
print("train_dir: ", train_dir)
print("validation_dir: ", validation_dir)
print("test_dir: ", test_dir)
# data load ----------
data_gen = ImageDataGenerator(rescale=1. / 255)
train_generator = data_gen.flow_from_directory(train_dir,
target_size=target_size,
batch_size=batch_size,
shuffle=True,
class_mode='categorical')
validation_generator = data_gen.flow_from_directory(
validation_dir,
target_size=target_size,
batch_size=batch_size,
shuffle=True,
class_mode='categorical')
test_generator = data_gen.flow_from_directory(test_dir,
target_size=target_size,
batch_size=batch_size,
shuffle=False,
class_mode='categorical')
data_checker, label_checker = next(train_generator)
print("train data shape (in batch): ", data_checker.shape)
print("train label shape (in batch): ", label_checker.shape)
# print("validation data shape:", validation_data.shape)
# print("validation label shape:", validation_label.shape)
# print("test data shape:", test_data.shape)
# print("test label shape:", test_label.shape)
# build model ----------
mh = ModelHandler(input_size, channel)
model = mh.buildMyModel()
model.summary()
# instance EarlyStopping -----
es = EarlyStopping(
monitor='val_loss',
# monitor='val_accuracy',
patience=5,
verbose=1,
restore_best_weights=True)
print("\ntraining sequence start .....")
steps_per_epoch = train_generator.n // batch_size
validation_steps = validation_generator.n // batch_size
print(steps_per_epoch, " [steps / epoch]")
print(validation_steps, " (validation steps)")
start = time.time()
history = model.fit_generator(train_generator,
steps_per_epoch=steps_per_epoch,
epochs=set_epochs,
validation_data=validation_generator,
validation_steps=validation_steps,
callbacks=[es],
verbose=1)
elapsed_time = time.time() - start
print("elapsed time (for train): {} [sec]".format(elapsed_time))
# evaluate ----------
print("\nevaluate sequence...")
accs = history.history['accuracy']
losses = history.history['loss']
val_accs = history.history['val_accuracy']
val_losses = history.history['val_loss']
print("last val_acc: ", val_accs[len(val_accs) - 1])
test_steps = test_generator.n // batch_size
eval_res = model.evaluate_generator(test_generator,
steps=test_steps,
verbose=1)
print("result loss: ", eval_res[0])
print("result score: ", eval_res[1])
# logging and detail outputs -----
# make log_dirctory
log_dir = os.path.join(sub_prj, "outputs", "logs")
os.makedirs(log_dir, exist_ok=True)
model_log_dir = os.path.join(sub_prj, "outputs", "models")
os.makedirs(log_dir, exist_ok=True)
now = datetime.datetime.now()
child_log_dir = os.path.join(log_dir, "{0:%Y%m%d}".format(now))
os.makedirs(child_log_dir, exist_ok=True)
child_model_log_dir = os.path.join(model_log_dir, "{0:%Y%m%d}".format(now))
os.makedirs(child_model_log_dir, exist_ok=True)
# save model & weights
model_file = os.path.join(child_model_log_dir, "model.h5")
model.save(model_file)
print("\nexport model in ", child_model_log_dir)
print("\npredict sequence...")
pred = model.predict_generator(test_generator, steps=test_steps, verbose=1)
test_label = []
for i in range(test_steps):
_, tmp_tl = next(test_generator)
if i == 0:
test_label = tmp_tl
else:
test_label = np.vstack((test_label, tmp_tl))
label_name_list = []
for i in range(len(test_label)):
if test_label[i][0] == 1:
label_name_list.append('cat')
elif test_label[i][1] == 1:
label_name_list.append('dog')
#print("result: ", pred)
df_pred = pd.DataFrame(pred, columns=['cat', 'dog'])
df_pred['class'] = df_pred.idxmax(axis=1)
df_pred['label'] = pd.DataFrame(label_name_list, columns=['label'])
df_pred['collect'] = (df_pred['class'] == df_pred['label'])
confuse = df_pred[df_pred['collect'] == False].index.tolist()
collect = df_pred[df_pred['collect'] == True].index.tolist()
print(df_pred)
print("\nwrong recognized indeices are ", confuse)
print(" wrong recognized amount is ", len(confuse))
print("\ncollect recognized indeices are ", collect)
print(" collect recognized amount is ", len(collect))
print("\nwrong rate: ", 100 * len(confuse) / len(test_label), " %")
# save history
save_dict = {}
save_dict['last_loss'] = losses[len(losses) - 1]
save_dict['last_acc'] = accs[len(accs) - 1]
save_dict['last_val_loss'] = val_losses[len(val_losses) - 1]
save_dict['last_val_acc'] = val_accs[len(val_accs) - 1]
save_dict['n_confuse'] = len(confuse)
save_dict['eval_loss'] = eval_res[0]
save_dict['eval_acc'] = eval_res[1]
save_dict['elapsed_time'] = elapsed_time
print(save_dict)
df_result = pd.DataFrame(save_dict.values(), index=save_dict.keys())
csv_file = os.path.join(child_log_dir, "result.csv")
df_result.to_csv(csv_file)
print("\nexport history in ", csv_file)
def main(N,
LEARN_PATH,
MODE,
BUILD_MODEL,
EPOCHS=60,
BATCH_SIZE=20,
FINE_TUNE_AT=81):
total_data, total_label = inputDataCreator(os.path.join(
LEARN_PATH, "natural"),
224,
normalize=True)
#one_hot=True
print("\ntotal_data shape: ", total_data.shape)
print("total_label shape: ", total_label.shape)
if MODE == 'auged':
auged_dir = os.path.join(LEARN_PATH, "auged")
EPOCHS = EPOCHS // 2
total_auged_data, total_auged_label = inputDataCreator(auged_dir,
224,
normalize=True,
one_hot=True)
print("\n total auged_data : ", total_auged_data.shape)
input_size = total_data.shape[1]
channel = total_data.shape[3]
mh = ModelHandler(input_size, channel)
skf = StratifiedKFold(n_splits=5)
k = 0
for traval_idx, test_idx in skf.split(total_data, total_label):
print("\nK-Fold Cross-Validation k:{} ==========".format(k))
print("\ntrain indices: \n", traval_idx)
print("\ntest indices: \n", test_idx)
test_data = total_data[test_idx]
test_label = total_label[test_idx]
print("-----*-----*-----")
traval_data = total_data[traval_idx]
traval_label = total_label[traval_idx]
# print(traval_data.shape)
# print(traval_label.shape)
traval_label = np.identity(2)[traval_label.astype(np.int8)]
test_label = np.identity(2)[test_label.astype(np.int8)]
train_data, train_label, validation_data, validation_label, _, _ = dataSplit(
traval_data,
traval_label,
train_rate=3 / 4,
validation_rate=1 / 4,
test_rate=0)
if MODE == 'auged':
print("\nadd auged data to train_data...")
auged_traval_data = total_auged_data[traval_idx]
auged_traval_label = total_auged_label[traval_idx]
auged_train_data, auged_train_label, _, _, _, _ = dataSplit(
auged_traval_data,
auged_traval_label,
train_rate=3 / 4,
validation_rate=1 / 4,
test_rate=0)
print(" append auged data: ", auged_train_data.shape)
print("\n concatnate auged data with native data...")
train_data = np.vstack((train_data, auged_train_data))
train_label = np.vstack((train_label, auged_train_label))
print(" Done.")
print("\ntrain data shape: ", train_data.shape)
print("train label shape: ", train_label.shape)
print("\nvalidation data shape: ", validation_data.shape)
print("validation label shape: ", validation_label.shape)
print("\ntest data shape: ", test_data.shape)
print("test label shape: ", test_label.shape)
es = EarlyStopping(monitor='val_loss',
patience=5,
verbose=1,
restore_best_weights=True,
mode='auto')
print("set epochs: ", EPOCHS)
if BUILD_MODEL == 'mymodel':
model = mh.buildMyModel()
# normal train ----------
print("\ntraining sequence start .....")
start = time.time()
history = model.fit(train_data,
train_label,
BATCH_SIZE,
epochs=EPOCHS,
vlidation_data=(validation_data,
validation_label),
callbacks=[es],
verbose=2)
elapsed_time = time.time() - start
elif BUILD_MODEL == 'tlearn':
# あとで重みの解凍をできるように base_model を定義
base_model = mh.buildMnv1Base()
base_model.trainable = False
model = mh.addChead(base_model)
print("\ntraining sequence start .....")
start = time.time()
# 準備体操 -----
print("\nwarm up sequence .....")
model.summary()
_history = model.fit(train_data,
train_label,
BATCH_SIZE,
epochs=10,
validation_data=(validation_data,
validation_label),
callbacks=[es],
verbose=2)
# fine tuning -----
print("\nfine tuning.....")
mh.setFineTune(base_model, model, FINE_TUNE_AT)
model.summary()
history = model.fit(train_data,
train_label,
BATCH_SIZE,
epochs=EPOCHS,
validation_data=(validation_data,
validation_label),
callbacks=[es],
verbose=2)
elapsed_time = time.time() - start
# training end
accs = history.history['accuracy']
losses = history.history['loss']
val_accs = history.history['val_accuracy']
val_losses = history.history['val_loss']
log_dir = os.path.join(os.path.dirname(cwd), "flog")
os.makedirs(log_dir, exist_ok=True)
"""
child_log_dir = os.path.join(log_dir, "{}_{}_{}".format(MODE, BUILD_MODEL, no))
os.makedirs(child_log_dir, exist_ok=True)
# save model & weights
model_file = os.path.join(child_log_dir, "{}_{}_{}_model.h5".format(MODE, BUILD_MODEL, no))
model.save(model_file)
# save history
history_file = os.path.join(child_log_dir, "{}_{}_{}_history.pkl".format(MODE, BUILD_MODEL, no))
with open(history_file, 'wb') as p:
pickle.dump(history.history, p)
print("\nexport logs in ", child_log_dir)
"""
print("\npredict sequence...")
pred = model.predict(test_data, batch_size=10, verbose=2)
label_name_list = []
for i in range(len(test_label)):
if test_label[i][0] == 1:
label_name_list.append('cat')
elif test_label[i][1] == 1:
label_name_list.append('dog')
df_pred = pd.DataFrame(pred, columns=['cat', 'dog'])
df_pred['class'] = df_pred.idxmax(axis=1)
df_pred['label'] = pd.DataFrame(label_name_list, columns=['label'])
df_pred['collect'] = (df_pred['class'] == df_pred['label'])
confuse = df_pred[df_pred['collect'] == False].index.tolist()
collect = df_pred[df_pred['collect'] == True].index.tolist()
print(df_pred)
print("\nwrong recognized indeices are ", confuse)
print(" wrong recognized amount is ", len(confuse))
print("\ncollect recognized indeices are ", collect)
print(" collect recognized amount is ", len(collect))
print("\nwrong rate: ", 100 * len(confuse) / len(test_label), " %")
print("\nevaluate sequence...")
eval_res = model.evaluate(test_data,
test_label,
batch_size=10,
verbose=2)
print("result loss: ", eval_res[0])
print("result score: ", eval_res[1])
# ----------
save_dict = {}
save_dict['last_loss'] = losses[len(losses) - 1]
save_dict['last_acc'] = accs[len(accs) - 1]
save_dict['last_val_loss'] = val_losses[len(val_losses) - 1]
save_dict['last_val_acc'] = val_accs[len(val_accs) - 1]
save_dict['n_confuse'] = len(confuse)
save_dict['eval_loss'] = eval_res[0]
save_dict['eval_acc'] = eval_res[1]
save_dict['elapsed_time'] = elapsed_time
print(save_dict)
if k == 0:
df_result = pd.DataFrame(save_dict.values(),
index=save_dict.keys())
else:
series = pd.Series(save_dict)
df_result[k] = series
print(df_result)
# undefine ----------
# del total_data, total_label
del traval_data, traval_label
if MODE == 'auged':
# del total_auged_data, total_auged_label
del auged_traval_data, auged_traval_label
del auged_train_data, auged_train_label
del train_data, train_label
del validation_data, validation_label
del test_data, test_label
del model
del _history, history
# clear session against OOM Error
keras.backend.clear_session()
gc.collect()
k += 1
csv_file = os.path.join(
log_dir, "sample_{}_{}_{}_result.csv".format(N, MODE, BUILD_MODEL))
df_result.to_csv(csv_file)
print("\nexport {} as CSV.".format(csv_file))
def main(data_mode, model_mode, no, set_epochs=60, do_es=False):
batch_size = 10
if data_mode == 'native':
data_dir = os.path.join(cwd, "experiment_{}".format(no))
total_data, total_label = inputDataCreator(data_dir,
224,
normalize=True,
one_hot=True)
train_data, train_label, validation_data, validation_label, test_data, test_label = dataSplit(
total_data, total_label)
"""
print("\ntrain data shape: ", train_data.shape)
print("train label shape: ", train_label.shape)
print("\nvalidation data shape: ", validation_data.shape)
print("validation label shape: ", validation_label.shape)
print("\ntest data shape: ", test_data.shape)
print("test label shape: ", test_label.shape)
"""
datagen = ImageDataGenerator()
train_generator = datagen.flow(
train_data,
train_label,
# target_size=(224, 224),
batch_size=batch_size,
shuffle=True)
validation_generator = datagen.flow(
validation_data,
validation_label,
# target_size=(224, 224),
batch_size=batch_size,
shuffle=True)
test_generator = datagen.flow(
test_data,
test_label,
# target_size=(224, 224),
batch_size=batch_size,
shuffle=True)
elif data_mode == 'auged':
set_epochs = int(set_epochs / 2)
data_dir = os.path.join(cwd, "concat_experiment_{}".format(no))
train_dir = os.path.join(data_dir, "train")
validation_dir = os.path.join(data_dir, "validation")
test_dir = os.path.join(data_dir, "test")
datagen = ImageDataGenerator(rescale=1 / 255.0)
train_generator = datagen.flow_from_directory(train_dir,
target_size=(224, 224),
batch_size=batch_size,
shuffle=True,
class_mode='categorical')
validation_generator = datagen.flow_from_directory(
validation_dir,
target_size=(224, 224),
batch_size=batch_size,
shuffle=True,
class_mode='categorical')
test_generator = datagen.flow_from_directory(test_dir,
target_size=(224, 224),
batch_size=batch_size,
shuffle=True,
class_mode='categorical')
data_sample, label_sample = next(train_generator)
input_size = data_sample.shape[1]
channel = data_sample.shape[3]
print("input_size: {} | channel: {}".format(input_size, channel))
print("set epochs: ", set_epochs)
mh = ModelHandler(input_size, channel)
if model_mode == 'mymodel':
model = mh.buildMyModel()
elif model_mode == 'tlearn':
model = mh.buildTlearnModel(base='mnv1')
model.summary()
if do_es:
es = EarlyStopping(monitor='val_loss',
patience=5,
verbose=1,
mode='auto')
es = [es]
else:
es = None
print("\ntraining sequence start .....")
steps_per_epoch = train_generator.n // batch_size
validation_steps = validation_generator.n // batch_size
print(steps_per_epoch, " [steps / epoch]")
print(validation_steps, " (validation steps)")
start = time.time()
history = model.fit_generator(train_generator,
steps_per_epoch=steps_per_epoch,
epochs=set_epochs,
validation_data=validation_generator,
validation_steps=validation_steps,
verbose=1)
elapsed_time = time.time() - start
accs = history.history['accuracy']
losses = history.history['loss']
val_accs = history.history['val_accuracy']
val_losses = history.history['val_loss']
if do_es:
log_dir = os.path.join(cwd, "log_with_es")
else:
log_dir = os.path.join(cwd, "log")
os.makedirs(log_dir, exist_ok=True)
"""
child_log_dir = os.path.join(log_dir, "{}_{}_{}".format(data_mode, model_mode, no))
os.makedirs(child_log_dir, exist_ok=True)
# save model & weights
model_file = os.path.join(child_log_dir, "{}_{}_{}_model.h5".format(data_mode, model_mode, no))
model.save(model_file)
# save history
history_file = os.path.join(child_log_dir, "{}_{}_{}_history.pkl".format(data_mode, model_mode, no))
with open(history_file, 'wb') as p:
pickle.dump(history.history, p)
print("\nexport logs in ", child_log_dir)
"""
print("\npredict sequence...")
pred = model.predict_generator(test_generator, verbose=1)
label_name_list = []
if data_mode == 'auged':
test_steps = test_generator.n // batch_size
for i in range(test_steps):
_, tmp = next(test_generator)
if i == 0:
test_label = tmp
else:
test_label = np.vstack((test_label, tmp))
for i in range(len(test_label)):
if test_label[i][0] == 1:
label_name_list.append('cat')
elif test_label[i][1] == 1:
label_name_list.append('dog')
df_pred = pd.DataFrame(pred, columns=['cat', 'dog'])
df_pred['class'] = df_pred.idxmax(axis=1)
df_pred['label'] = pd.DataFrame(label_name_list, columns=['label'])
df_pred['collect'] = (df_pred['class'] == df_pred['label'])
confuse = df_pred[df_pred['collect'] == False].index.tolist()
collect = df_pred[df_pred['collect'] == True].index.tolist()
print(df_pred)
print("\nwrong recognized indeices are ", confuse)
print(" wrong recognized amount is ", len(confuse))
print("\ncollect recognized indeices are ", collect)
print(" collect recognized amount is ", len(collect))
print("\nwrong rate: ", 100 * len(confuse) / len(test_label), " %")
print("\nevaluate sequence...")
eval_res = model.evaluate(test_generator, verbose=1)
print("result loss: ", eval_res[0])
print("result score: ", eval_res[1])
# ----------
save_dict = {}
save_dict['last_loss'] = losses[len(losses) - 1]
save_dict['last_acc'] = accs[len(accs) - 1]
save_dict['last_val_loss'] = val_losses[len(val_losses) - 1]
save_dict['last_val_acc'] = val_accs[len(val_accs) - 1]
save_dict['n_confuse'] = len(confuse)
save_dict['eval_loss'] = eval_res[0]
save_dict['eval_acc'] = eval_res[1]
save_dict['elapsed_time'] = elapsed_time
print(save_dict)
# undefine validable ----------
del datagen
del train_generator, validation_generator, test_generator
# due to CPU memory ---------
# del train_data, train_label
# del validation_data, validation_label
# del test_data, test_label
# del set_epochs
# due to GPU memory ---------
# del mh,
del model
del history
# del accs, losses, val_accs, val_losses
# del pred, df_pred, label_name_list
# del confuse, collect
# del eval_res
keras.backend.clear_session()
gc.collect()
return save_dict
def main(data_mode, model_mode, no, set_epochs=60, do_es=False):
cwd = os.getcwd()
data_dir = os.path.join(cwd, "experiment_{}".format(no))
total_data, total_label = inputDataCreator(
data_dir,
224,
normalize=True,
#one_hot=True
)
print("\ntotal_data shape: ", total_data.shape)
print("total_label shape: ", total_label.shape)
if data_mode == 'auged':
base_dir, data_dir_name = os.path.split(data_dir)
data_dir_name = "auged_" + data_dir_name
auged_dir = os.path.join(base_dir, data_dir_name)
set_epochs = int(set_epochs / 2)
total_auged_data, total_auged_label = inputDataCreator(auged_dir,
224,
normalize=True,
one_hot=True)
print("\n total auged_data : ", total_auged_data.shape)
input_size = total_data.shape[1]
channel = total_data.shape[3]
mh = ModelHandler(input_size, channel)
skf = StratifiedKFold(n_splits=10)
k = 0
for traval_idx, test_idx in skf.split(total_data, total_label):
print("\nK-Fold Cross-Validation k:{} ==========".format(k))
print("\ntrain indices: \n", traval_idx)
print("\ntest indices: \n", test_idx)
test_data = total_data[test_idx]
test_label = total_label[test_idx]
print("-----*-----*-----")
traval_data = total_data[traval_idx]
traval_label = total_label[traval_idx]
# print(traval_data.shape)
# print(traval_label.shape)
traval_label = np.identity(2)[traval_label.astype(np.int8)]
test_label = np.identity(2)[test_label.astype(np.int8)]
train_data, train_label, validation_data, validation_label, _, _ = dataSplit(
traval_data,
traval_label,
train_rate=2 / 3,
validation_rate=1 / 3,
test_rate=0)
if data_mode == 'auged':
print("\nadd auged data to train_data...")
auged_traval_data = total_auged_data[traval_idx]
auged_traval_label = total_auged_label[traval_idx]
auged_train_data, auged_train_label, _, _, _, _ = dataSplit(
auged_traval_data,
auged_traval_label,
train_rate=2 / 3,
validation_rate=1 / 3,
test_rate=0)
print(" append auged data: ", auged_train_data.shape)
print("\n concatnate auged data with native data...")
train_data = np.vstack((train_data, auged_train_data))
train_label = np.vstack((train_label, auged_train_label))
print(" Done.")
print("\ntrain data shape: ", train_data.shape)
print("train label shape: ", train_label.shape)
print("\nvalidation data shape: ", validation_data.shape)
print("validation label shape: ", validation_label.shape)
print("\ntest data shape: ", test_data.shape)
print("test label shape: ", test_label.shape)
if model_mode == 'mymodel':
model = mh.buildMyModel()
elif model_mode == 'tlearn':
model = mh.buildTlearnModel(base='mnv1')
model.summary()
if do_es:
es = EarlyStopping(monitor='val_loss',
patience=5,
verbose=1,
mode='auto')
es = [es]
else:
es = None
batch_size = 10
print("set epochs: ", set_epochs)
print("\ntraining sequence start .....")
start = time.time()
history = model.fit(train_data,
train_label,
batch_size,
epochs=set_epochs,
validation_data=(validation_data,
validation_label),
callbacks=es,
verbose=1)
elapsed_time = time.time() - start
accs = history.history['accuracy']
losses = history.history['loss']
val_accs = history.history['val_accuracy']
val_losses = history.history['val_loss']
if do_es:
log_dir = os.path.join(cwd, "rlog_with_es")
else:
log_dir = os.path.join(cwd, "rlog")
os.makedirs(log_dir, exist_ok=True)
child_log_dir = os.path.join(
log_dir, "{}_{}_{}".format(data_mode, model_mode, no))
os.makedirs(child_log_dir, exist_ok=True)
"""
# save model & weights
model_file = os.path.join(child_log_dir, "{}_{}_{}_model.h5".format(data_mode, model_mode, no))
model.save(model_file)
# save history
history_file = os.path.join(child_log_dir, "{}_{}_{}_history.pkl".format(data_mode, model_mode, no))
with open(history_file, 'wb') as p:
pickle.dump(history.history, p)
print("\nexport logs in ", child_log_dir)
"""
print("\npredict sequence...")
pred = model.predict(test_data, batch_size=10, verbose=1)
label_name_list = []
for i in range(len(test_label)):
if test_label[i][0] == 1:
label_name_list.append('cat')
elif test_label[i][1] == 1:
label_name_list.append('dog')
df_pred = pd.DataFrame(pred, columns=['cat', 'dog'])
df_pred['class'] = df_pred.idxmax(axis=1)
df_pred['label'] = pd.DataFrame(label_name_list, columns=['label'])
df_pred['collect'] = (df_pred['class'] == df_pred['label'])
confuse = df_pred[df_pred['collect'] == False].index.tolist()
collect = df_pred[df_pred['collect'] == True].index.tolist()
print(df_pred)
print("\nwrong recognized indeices are ", confuse)
print(" wrong recognized amount is ", len(confuse))
print("\ncollect recognized indeices are ", collect)
print(" collect recognized amount is ", len(collect))
print("\nwrong rate: ", 100 * len(confuse) / len(test_label), " %")
print("\nevaluate sequence...")
eval_res = model.evaluate(test_data,
test_label,
batch_size=10,
verbose=1)
print("result loss: ", eval_res[0])
print("result score: ", eval_res[1])
# ----------
save_dict = {}
save_dict['last_loss'] = losses[len(losses) - 1]
save_dict['last_acc'] = accs[len(accs) - 1]
save_dict['last_val_loss'] = val_losses[len(val_losses) - 1]
save_dict['last_val_acc'] = val_accs[len(val_accs) - 1]
save_dict['n_confuse'] = len(confuse)
save_dict['eval_loss'] = eval_res[0]
save_dict['eval_acc'] = eval_res[1]
save_dict['elapsed_time'] = elapsed_time
print(save_dict)
if k == 0:
df_result = pd.DataFrame(save_dict.values(),
index=save_dict.keys())
else:
series = pd.Series(save_dict)
df_result[k] = series
print(df_result)
# undefine ----------
# del total_data, total_label
del traval_data, traval_label
if data_mode == 'auged':
# del total_auged_data, total_auged_label
del auged_traval_data, auged_traval_label
del auged_train_data, auged_train_label
del train_data, train_label
del validation_data, validation_label
del test_data, test_label
del model
del history
del pred
del df_pred, label_name_list, confuse, collect
del eval_res
del accs, losses, val_accs, val_losses
# clear session against OOM Error
keras.backend.clear_session()
gc.collect()
k += 1
csv_file = os.path.join(child_log_dir,
"{}_{}_result.csv".format(data_mode, model_mode))
df_result.to_csv(csv_file)
print("\nexport {} as CSV.".format(csv_file))
# delete valables at all in end of this program -----
del cwd, data_dir
del total_data, total_label
del save_dict, df_result
if data_mode == 'auged':
del total_auged_data, total_auged_label
del base_dir, data_dir_name, auged_dir
gc.collect()
def main():
cwd = os.getcwd()
sub_prj = os.path.dirname(cwd)
"""
sub_prj_root = os.path.dirname(sub_prj)
prj_root = os.path.dirname(sub_prj_root)
"""
data_dir = os.path.join(sub_prj, "datasets")
data_src = os.path.join(data_dir, "medium_721")
print("\ndata source: ", data_src)
"""
use_da_data = False
increase_val = False
print( "\nmode: Use Augmented data: {} | increase validation data: {}".format(use_da_data, increase_val) )
# First define original train_data only as train_dir
train_dir = os.path.join(data_dir, "train")
if (use_da_data == True) and (increase_val == False):
# with_augmented data (no validation increase)
train_dir = os.path.join(data_dir, "train_with_aug")
validation_dir = os.path.join(data_dir, "val") # original validation data
# pair of decreaced train_data and increased validation data
if (increase_val == True):
train_dir = os.path.join(data_dir, "red_train")
if (use_da_data == True):
train_dir = os.path.join(data_dir, "red_train_with_aug")
validation_dir = os.path.join(data_dir, "validation")
"""
use_da_data = True
if use_da_data:
train_dir = os.path.join(data_src, "train_with_aug")
else:
train_dir = os.path.join(data_src, "train")
validation_dir = os.path.join(data_src, "validation")
test_dir = os.path.join(data_src, "test")
print("train_dir: ", train_dir)
print("validation_dir: ", validation_dir)
# print("test_dir: ", test_dir)
# data load ----------
data_gen = ImageDataGenerator(rescale=1. / 255)
train_generator = data_gen.flow_from_directory(train_dir,
target_size=target_size,
batch_size=batch_size,
shuffle=True,
class_mode='categorical')
validation_generator = data_gen.flow_from_directory(
validation_dir,
target_size=target_size,
batch_size=batch_size,
shuffle=True,
class_mode='categorical')
test_generator = data_gen.flow_from_directory(test_dir,
target_size=target_size,
batch_size=batch_size,
shuffle=False,
class_mode='categorical')
data_checker, label_checker = next(train_generator)
print("train data shape (in batch): ", data_checker.shape)
print("train label shape (in batch): ", label_checker.shape)
# print("validation data shape:", validation_data.shape)
# print("validation label shape:", validation_label.shape)
# print("test data shape:", test_data.shape)
# print("test label shape:", test_label.shape)
# build model ----------
mh = ModelHandler(input_size, channel)
model = mh.buildMyModel()
model.summary()
# instance EarlyStopping -----
es = EarlyStopping(
monitor='val_loss',
# monitor='val_accuracy',
patience=5,
verbose=1,
restore_best_weights=True)
print("\ntraining sequence start .....")
steps_per_epoch = train_generator.n // batch_size
validation_steps = validation_generator.n // batch_size
print(steps_per_epoch, " [steps / epoch]")
print(validation_steps, " (validation steps)")
start = time.time()
history = model.fit_generator(train_generator,
steps_per_epoch=steps_per_epoch,
epochs=set_epochs,
validation_data=validation_generator,
validation_steps=validation_steps,
callbacks=[es],
verbose=1)
print("elapsed time (for train): {} [sec]".format(time.time() - start))
print("\nevaluate sequence...")
test_steps = test_generator.n // batch_size
eval_res = model.evaluate_generator(test_generator,
steps=test_steps,
verbose=1)
print("result loss: ", eval_res[0])
print("result score: ", eval_res[1])
# confusion matrix -----
print("\nconfusion matrix")
pred = model.predict_generator(test_generator, steps=test_steps, verbose=3)
test_label = []
for i in range(test_steps):
_, tmp_tl = next(test_generator)
if i == 0:
test_label = tmp_tl
else:
test_label = np.vstack((test_label, tmp_tl))
idx_label = np.argmax(test_label, axis=-1) # one_hot => normal
idx_pred = np.argmax(pred, axis=-1) # 各 class の確率 => 最も高い値を持つ class
cm = confusion_matrix(idx_label, idx_pred)
# Calculate Precision and Recall
tn, fp, fn, tp = cm.ravel()
print(" | T | F ")
print("--+----+---")
print("N | {} | {}".format(tn, fn))
print("--+----+---")
print("P | {} | {}".format(tp, fp))
# 適合率 (precision):
# precision = tp/(tp+fp)
# print("Precision of the model is {}".format(precision))
# 再現率 (recall):
# recall = tp/(tp+fn)
# print("Recall of the model is {}".format(recall))
# save model -----
save_location = os.path.join(sub_prj, "outputs", "models")
if use_da_data:
save_file = os.path.join(save_location, "mymodel_auged.h5")
else:
save_file = os.path.join(save_location, "mymodel.h5")
model.save(save_file)
print("\nmodel has saved in", save_file)
def main(N, LEARN_PATH, DATA_MODE, BUILD_MODEL, EPOCHS=60):
sample_dir = os.path.join(LEARN_PATH, "sample_{}".format(N))
use_da_data = False
if use_da_data:
train_dir = os.path.join(sample_dir, "train_with_aug")
else:
train_dir = os.path.join(sample_dir, "train")
validation_dir = os.path.join(sample_dir, "validation")
test_dir = os.path.join(sample_dir, "test")
print("train_dir: ", train_dir)
print("validation_dir: ", validation_dir)
print("test_dir: ", test_dir)
# data load ----------
train_data, train_label = inputDataCreator(train_dir,
224,
normalize=True,
one_hot=True)
validation_data, validation_label = inputDataCreator(validation_dir,
224,
normalize=True,
one_hot=True)
test_data, test_label = inputDataCreator(test_dir,
224,
normalize=True,
one_hot=True)
print("\ntrain data shape: ", train_data.shape)
print("train label shape: ", train_label.shape)
print("\nvalidation data shape: ", validation_data.shape)
print("validation label shape: ", validation_label.shape)
input_size = train_data.shape[1]
channel = train_data.shape[3]
batch_size = 10
print("set epochs: ", EPOCHS)
mh = ModelHandler(input_size, channel)
if BUILD_MODEL == 'mymodel':
model = mh.buildMyModel()
elif BUILD_MODEL == 'tlearn':
model = mh.buildTlearnModel(base='mnv1')
model.summary()
"""
es = EarlyStopping(monitor='val_loss',
patience=5,
verbose=1,
mode='auto',
restore)
"""
# early stopping
es = keras.callbacks.EarlyStopping(monitor='val_loss',
patience=5,
restore_best_weights=True)
print("\ntraining sequence start .....")
start = time.time()
history = model.fit(train_data,
train_label,
batch_size,
epochs=EPOCHS,
validation_data=(validation_data, validation_label),
callbacks=[es],
verbose=2)
elapsed_time = time.time() - start
print("elapsed time (for train): {} [sec]".format(time.time() - start))
accs = history.history['accuracy']
losses = history.history['loss']
val_accs = history.history['val_accuracy']
val_losses = history.history['val_loss']
"""
# logging and detail outputs -----
# make log_dirctory
log_dir = os.path.join(sub_prj, "outputs", "logs")
os.makedirs(log_dir, exist_ok=True)
model_log_dir = os.path.join(sub_prj, "outputs", "models")
os.makedirs(log_dir, exist_ok=True)
now = datetime.datetime.now()
child_log_dir = os.path.join(log_dir, "{0:%Y%m%d}".format(now))
os.makedirs(child_log_dir, exist_ok=True)
child_model_log_dir = os.path.join(model_log_dir, "{0:%Y%m%d}".format(now))
os.makedirs(child_model_log_dir, exist_ok=True)
"""
"""
if do_es:
log_dir = os.path.join(cwd, "log_with_es")
else:
log_dir = os.path.join(cwd, "log")
os.makedirs(log_dir, exist_ok=True)
"""
"""
child_log_dir = os.path.join(log_dir, "{}_{}_{}".format(DATA_MODE, BUILD_MODEL, no))
os.makedirs(child_log_dir, exist_ok=True)
# save model & weights
model_file = os.path.join(child_log_dir, "{}_{}_{}_model.h5".format(DATA_MODE, BUILD_MODEL, no))
model.save(model_file)
# save history
history_file = os.path.join(child_log_dir, "{}_{}_{}_history.pkl".format(DATA_MODE, BUILD_MODEL, no))
with open(history_file, 'wb') as p:
pickle.dump(history.history, p)
print("\nexport logs in ", child_log_dir)
"""
print("\npredict sequence...")
pred = model.predict(test_data, batch_size=10, verbose=1)
label_name_list = []
for i in range(len(test_label)):
if test_label[i][0] == 1:
label_name_list.append('cat')
elif test_label[i][1] == 1:
label_name_list.append('dog')
df_pred = pd.DataFrame(pred, columns=['cat', 'dog'])
df_pred['class'] = df_pred.idxmax(axis=1)
df_pred['label'] = pd.DataFrame(label_name_list, columns=['label'])
df_pred['collect'] = (df_pred['class'] == df_pred['label'])
confuse = df_pred[df_pred['collect'] == False].index.tolist()
collect = df_pred[df_pred['collect'] == True].index.tolist()
print(df_pred)
print("\nwrong recognized indeices are ", confuse)
print(" wrong recognized amount is ", len(confuse))
print("\ncollect recognized indeices are ", collect)
print(" collect recognized amount is ", len(collect))
print("\nwrong rate: ", 100 * len(confuse) / len(test_label), " %")
print("\nevaluate sequence...")
eval_res = model.evaluate(test_data, test_label, batch_size=10, verbose=1)
print("result loss: ", eval_res[0])
print("result score: ", eval_res[1])
# ----------
save_dict = {}
save_dict['last_loss'] = losses[len(losses) - 1]
save_dict['last_acc'] = accs[len(accs) - 1]
save_dict['last_val_loss'] = val_losses[len(val_losses) - 1]
save_dict['last_val_acc'] = val_accs[len(val_accs) - 1]
save_dict['n_confuse'] = len(confuse)
save_dict['eval_loss'] = eval_res[0]
save_dict['eval_acc'] = eval_res[1]
save_dict['elapsed_time'] = elapsed_time
print(save_dict)
# 重そうなものは undefine してみる
#del train_data, train_label, validation_data, validation_label, test_data, test_label
del model
del history
#del pred
keras.backend.clear_session()
gc.collect()
return save_dict
def main(LEARN_PATH, INPUT_SIZE, CHANNEL, BATCH_SIZE, EPOCHS):
target_size = (INPUT_SIZE, INPUT_SIZE)
input_shape = (INPUT_SIZE, INPUT_SIZE, CHANNEL)
data_src = LEARN_PATH
print("\ndata source: ", data_src)
use_da_data = False
if use_da_data:
train_dir = os.path.join(data_src, "train_with_aug")
else:
train_dir = os.path.join(data_src, "train")
validation_dir = os.path.join(data_src, "validation")
test_dir = os.path.join(data_src, "test")
print("train_dir: ", train_dir)
print("validation_dir: ", validation_dir)
print("test_dir: ", test_dir)
# data load ----------
train_data, train_label = inputDataCreator(train_dir,
INPUT_SIZE,
normalize=True,
one_hot=True)
validation_data, validation_label = inputDataCreator(validation_dir,
INPUT_SIZE,
normalize=True,
one_hot=True)
test_data, test_label = inputDataCreator(test_dir,
INPUT_SIZE,
normalize=True,
one_hot=True)
"""
total_data, total_label = inputDataCreator(data_dir,
224,
normalize=True,
one_hot=True)
train_data, train_label, validation_data, validation_label, test_data, test_label = dataSplit(total_data, total_label)
"""
print("train data shape (in batch): ", train_data.shape)
print("train label shape (in batch): ", train_label.shape)
# print("validation data shape:", validation_data.shape)
# print("validation label shape:", validation_label.shape)
# print("test data shape:", test_data.shape)
# print("test label shape:", test_label.shape)
# build model ----------
mh = ModelHandler(INPUT_SIZE, CHANNEL)
model = mh.buildMyModel()
model.summary()
# instance EarlyStopping -----
es = EarlyStopping(monitor='val_loss',
# monitor='val_accuracy',
patience=5,
verbose=1,
restore_best_weights=True)
print("\ntraining sequence start .....")
start = time.time()
history = model.fit(train_data,
train_label,
batch_size=BATCH_SIZE,
epochs=EPOCHS,
validation_data=(validation_data, validation_label),
callbacks=[es],
verbose=2)
elapsed_time = time.time() - start
print( "elapsed time (for train): {} [sec]".format(elapsed_time) )
accs = history.history['accuracy']
losses = history.history['loss']
val_accs = history.history['val_accuracy']
val_losses = history.history['val_loss']
print("last val_acc: ", val_accs[len(val_accs)-1])
# evaluate ----------
print("\nevaluate sequence...")
eval_res = model.evaluate(test_data,
test_label,
batch_size=10,
verbose=2)
print("result loss: ", eval_res[0])
print("result score: ", eval_res[1])
# logging and detail outputs -----
# make log_dirctory
cwd = os.getcwd()
sub_prj = os.path.dirname(cwd)
log_dir = os.path.join(sub_prj, "outputs", "logs")
os.makedirs(log_dir, exist_ok=True)
model_log_dir = os.path.join(sub_prj, "outputs", "models")
os.makedirs(log_dir, exist_ok=True)
now = datetime.datetime.now()
child_log_dir = os.path.join(log_dir, "{0:%Y%m%d}".format(now))
os.makedirs(child_log_dir, exist_ok=True)
child_model_log_dir = os.path.join(model_log_dir, "{0:%Y%m%d}".format(now))
os.makedirs(child_model_log_dir, exist_ok=True)
# save model & weights
model_file = os.path.join(child_model_log_dir, "model.h5")
model.save(model_file)
print("\nexport model in ", child_model_log_dir)
# predict -> confusion matrix ----------
print("\npredict sequence...")
pred = model.predict(test_data,
batch_size=BATCH_SIZE,
verbose=2)
label_name_list = []
for i in range(len(test_label)):
if test_label[i][0] == 1:
label_name_list.append('cat')
elif test_label[i][1] == 1:
label_name_list.append('dog')
#print("result: ", pred)
df_pred = pd.DataFrame(pred, columns=['cat', 'dog'])
df_pred['class'] = df_pred.idxmax(axis=1)
df_pred['label'] = pd.DataFrame(label_name_list, columns=['label'])
df_pred['collect'] = (df_pred['class'] == df_pred['label'])
confuse = df_pred[df_pred['collect'] == False].index.tolist()
collect = df_pred[df_pred['collect'] == True].index.tolist()
print(df_pred)
print("\nwrong recognized indeices are ", confuse)
print(" wrong recognized amount is ", len(confuse))
print("\ncollect recognized indeices are ", collect)
print(" collect recognized amount is ", len(collect))
print("\nwrong rate: ", 100*len(confuse)/len(test_label), " %")
# save history
save_dict = {}
save_dict['last_loss'] = losses[len(losses)-1]
save_dict['last_acc'] = accs[len(accs)-1]
save_dict['last_val_loss'] = val_losses[len(val_losses)-1]
save_dict['last_val_acc'] = val_accs[len(val_accs)-1]
save_dict['n_confuse'] = len(confuse)
save_dict['eval_loss'] = eval_res[0]
save_dict['eval_acc'] = eval_res[1]
save_dict['elapsed_time'] = elapsed_time
print("save result dict:", save_dict)
df_result = pd.DataFrame(save_dict.values(), index=save_dict.keys())
csv_file = os.path.join( child_log_dir, "result.csv" )
df_result.to_csv(csv_file)
print("\nexport result in ", csv_file)
# confusion matrix -----
# Predict
# 0 | 1
# --+----+-----
# 0 | TN | FP
# label -------+-----
# 1 | FN | TP
print("\nconfusion matrix")
idx_label = np.argmax(test_label, axis=-1) # one_hot => normal
idx_pred = np.argmax(pred, axis=-1) # 各 class の確率 => 最も高い値を持つ class
cm = confusion_matrix(idx_label, idx_pred)
# Calculate Precision and Recall
tn, fp, fn, tp = cm.ravel()
print(" | T | F ")
print("--+----+---")
print("N | {} | {}".format(tn, fn))
print("--+----+---")
print("P | {} | {}".format(tp, fp))
# 適合率 (precision):
# sklearn.metrics => precision_score() にも libaray がある。
# # 入力は (idx_label, idx_pred)
# 「陽性と予測されたサンプルのうち, 正解したサンプルの割合」
# PPV (positive predictive value) とも呼ばれる。
precision = tp/(tp+fp)
print("Precision of the model is {}".format(precision))
# 再現率 (recall:
# sklearn.metrics => recall_score() にも library がある
# # 入力は (idx_label, idx_pred)
# 「実際に陽性のサンプルのうち, 正解したサンプルの割合」。
# sensitivity や hit rate,
# TPR (true positive rate, 真陽性率) などとも呼ばれる。
# # すごく大雑把にいえば
# # class 1 だけに対して考えた正解率の様な指標だと言える
# # (Negative 側がどれだけ正解/不正解かは don't care)
# # 逆に TN / (TN + FP) とすると
# # class 0 だけに対する正解率となる。
recall = tp/(tp+fn)
print("Recall of the model is {}".format(recall))