def tongue_basic_gen_trainer(tongue_image_arrays, tongue_yaofangs, tongue_image_shape, nb_yao,
gen_model_path=None, train_on_batch=False,
use_data_augment=False):
'''
@param use_tfidf_tensor: flag of use tfidf tensor or not with different tensorization function
'''
total_tongue_x, total_y = tongue2text_gen.data_tensorization(
tongue_image_arrays, tongue_yaofangs, tongue_image_shape, nb_yao)
# train_x = total_tongue_x[500:]
# train_y = total_y[500:]
# train_x = np.concatenate((total_tongue_x[: 2000], total_tongue_x[2500 :]), axis=0)
# train_y = np.concatenate((total_y[: 2000], total_y[2500 :]), axis=0)
# train_x = np.concatenate((total_tongue_x[: 4000], total_tongue_x[4500 :]), axis=0)
# train_y = np.concatenate((total_y[: 4000], total_y[4500 :]), axis=0)
# train_x = np.concatenate((total_tongue_x[: 6000], total_tongue_x[6500 :]), axis=0)
# train_y = np.concatenate((total_y[: 6000], total_y[6500 :]), axis=0)
train_x = total_tongue_x[: len(total_tongue_x) - 500]
train_y = total_y[: len(total_y) - 500]
del(total_tongue_x)
del(total_y)
if use_data_augment == True:
# just can be use on service 225 with big memory
datagen = image_augment.image_augment_gen()
train_x, train_y = image_augment.data_tensoration_augment(
datagen, train_x, train_y)
scaling_act_type = 'binary'
print('training 2 * cnn + mlp tongue2text gen model------on_batch: %d------scaling_activation: %s...' %
(train_on_batch, scaling_act_type))
tongue_gen_model = tongue2text_gen.k_cnns_mlp(
yao_indices_dim=nb_yao, tongue_image_shape=tongue_image_shape, with_compile=True)
if train_on_batch == True:
trained_tongue_gen_model, history = tongue2text_gen.trainer_on_batch(
tongue_gen_model, train_x, train_y)
else:
record_path = None
if gen_model_path != None:
record_path = gen_model_path.replace('json', 'h5')
trained_tongue_gen_model, history = tongue2text_gen.trainer(
tongue_gen_model, train_x, train_y, best_record_path=record_path)
if gen_model_path != None:
tongue2text_gen.storageModel(model=trained_tongue_gen_model, frame_path=gen_model_path,
replace_record=False)
print('history: {0}'.format(history))
return trained_tongue_gen_model
def tongue_gen_deeper_2pipeline_2outputs_trainer(tongue_image_arrays,
tongue_yaofangs,
tongue_image_shape,
base_model_name,
nb_yao,
lda_model_path,
gen_model_path=None,
lda_replace=False,
use_tfidf_tensor=False,
use_data_augment=False):
'''
can not use train_on_batch
@param lda_replace: flag of lda need replace by a new one or not
@param base_model: the name of base model in {'vgg16', 'vgg19', 'resnet50'}
'''
if os.path.exists(lda_model_path) == False or lda_replace == True:
tongue_yaofangs_str = lda.list_int2str(tongue_yaofangs)
lda_model, dictionary = lda.lda_trainer(tongue_yaofangs_str,
lda_model_path)
else:
lda_model, dictionary = lda.loadModelfromFile(lda_model_path,
readOnly=True)
total_tongue_x, total_y, total_aux_y = tongue2text_gen.data_tensorization_lda(
tongue_image_arrays,
tongue_yaofangs,
tongue_image_shape,
nb_yao,
lda_model.num_topics,
lda_model,
dictionary,
use_tfidf_tensor=use_tfidf_tensor)
# train_tongue_x = total_tongue_x[500 :]
# train_y = total_y[500 :]
# train_aux_y = total_aux_y[500 :]
# train_tongue_x = np.concatenate((total_tongue_x[: 2000], total_tongue_x[2500 :]), axis=0)
# train_y = np.concatenate((total_y[: 2000], total_y[2500 :]), axis=0)
# train_aux_y = np.concatenate((total_aux_y[: 2000], total_aux_y[2500 :]), axis=0)
# train_tongue_x = np.concatenate((total_tongue_x[: 4000], total_tongue_x[4500 :]), axis=0)
# train_y = np.concatenate((total_y[: 4000], total_y[4500 :]), axis=0)
# train_aux_y = np.concatenate((total_aux_y[: 4000], total_aux_y[4500 :]), axis=0)
# train_tongue_x = np.concatenate((total_tongue_x[: 6000], total_tongue_x[6500 :]), axis=0)
# train_y = np.concatenate((total_y[: 6000], total_y[6500 :]), axis=0)
# train_aux_y = np.concatenate((total_aux_y[: 6000], total_aux_y[6500 :]), axis=0)
train_tongue_x = total_tongue_x[:len(total_tongue_x) - 500]
train_y = total_y[:len(total_y) - 500]
train_aux_y = total_aux_y[:len(total_aux_y) - 500]
del (total_tongue_x)
del (total_y)
del (total_aux_y)
if use_data_augment == True:
# just can be use on service 225 with big memory
datagen = image_augment.image_augment_gen()
train_tongue_x, train_y, train_aux_y = image_augment.data_tensoration_augment_withaux(
datagen, train_tongue_x, train_y, train_aux_y)
scaling_act_type = 'tfidf' if use_tfidf_tensor else 'binary'
print(
'training 2 * cnn + mlp with double output(lda) tongue2text gen model------scaling_activation: %s...'
% scaling_act_type)
print('use base_model: ' + base_model_name)
image_input, base_model = tongue2text_deeper_gen.k_base_model(
tongue_image_shape, model_name=base_model_name)
tongue_gen_model = tongue2text_deeper_gen.k_2pipeline_mlp_2outputs(
yao_indices_dim=nb_yao,
topics_dim=lda_model.num_topics,
image_input=image_input,
base_model=base_model,
with_compile=True)
# trained_tongue_gen_model, history = tongue2text_gen.trainer(
# tongue_gen_model, train_tongue_x, train_y, train_aux_y)
record_path = None
if gen_model_path != None:
record_path = gen_model_path.replace('json', 'h5')
trained_tongue_gen_model, history = tongue2text_gen.trainer(
tongue_gen_model,
train_tongue_x,
train_y,
train_aux_y=train_aux_y,
batch_size=16,
epochs=_new_training_epochs,
best_record_path=record_path)
if gen_model_path != None:
tongue2text_gen.storageModel(model=trained_tongue_gen_model,
frame_path=gen_model_path,
replace_record=False)
print('history:')
for item in history:
print('\n{0}'.format(item))
return trained_tongue_gen_model
def tongue_gen_deeper_1pipeline_trainer(tongue_image_arrays,
tongue_yaofangs,
tongue_image_shape,
base_model_name,
nb_yao,
gen_model_path=None,
train_on_batch=False,
use_tfidf_tensor=False,
use_data_augment=True):
'''
@param base_model: the name of base model in {'vgg16', 'vgg19', 'resnet50'}
'''
total_tongue_x, total_y = tongue2text_gen.data_tensorization(
tongue_image_arrays, tongue_yaofangs, tongue_image_shape, nb_yao)
# train_x = total_tongue_x[500:]
# train_y = total_y[500:]
# train_x = np.concatenate((total_tongue_x[: 2000], total_tongue_x[2500 :]), axis=0)
# train_y = np.concatenate((total_y[: 2000], total_y[2500 :]), axis=0)
# train_x = np.concatenate((total_tongue_x[: 4000], total_tongue_x[4500 :]), axis=0)
# train_y = np.concatenate((total_y[: 4000], total_y[4500 :]), axis=0)
# train_x = np.concatenate((total_tongue_x[: 6000], total_tongue_x[6500 :]), axis=0)
# train_y = np.concatenate((total_y[: 6000], total_y[6500 :]), axis=0)
train_x = total_tongue_x[:len(total_tongue_x) - 500]
train_y = total_y[:len(total_y) - 500]
del (total_tongue_x)
del (total_y)
if use_data_augment == True:
# just can be use on service 225 with big memory
datagen = image_augment.image_augment_gen()
train_x, train_y = image_augment.data_tensoration_augment(
datagen, train_x, train_y)
scaling_act_type = 'tfidf' if use_tfidf_tensor else 'binary'
print(
'training 1pipeline + mlp tongue2text gen model------on_batch: %d------scaling_activation: %s...'
% (train_on_batch, scaling_act_type))
print('use base_model: ' + base_model_name)
image_input, base_model = tongue2text_deeper_gen.k_base_model(
tongue_image_shape=tongue_image_shape, model_name=base_model_name)
tongue_gen_model = tongue2text_deeper_gen.k_1pipeline_mlp(
yao_indices_dim=nb_yao,
image_input=image_input,
base_model=base_model,
with_compile=True)
if train_on_batch == True:
trained_tongue_gen_model, history = tongue2text_gen.trainer_on_batch(
tongue_gen_model, train_x, train_y)
else:
record_path = None
if gen_model_path != None:
record_path = gen_model_path.replace('json', 'h5')
trained_tongue_gen_model, history = tongue2text_gen.trainer(
tongue_gen_model,
train_x,
train_y,
batch_size=16,
epochs=_new_training_epochs,
best_record_path=record_path)
if gen_model_path != None:
tongue2text_gen.storageModel(model=trained_tongue_gen_model,
frame_path=gen_model_path,
replace_record=False)
print('history: {0}'.format(history))
return trained_tongue_gen_model
def tongue_gen_biggertimes_cnn_withlda_trainer(tongue_image_arrays, tongue_yaofangs, tongue_image_shape, nb_yao, channel_times,
lda_model_path, gen_model_path=None, lda_replace=False, use_tfidf_tensor=False,
use_data_augment=False):
'''
can not use train_on_batch
@param lda_replace: flag of lda need replace by a new one or not
@param use_tfidf_tensor: flag of use tfidf tensor or not with different tensorization function
'''
if os.path.exists(lda_model_path) == False or lda_replace == True:
tongue_yaofangs_str = lda.list_int2str(tongue_yaofangs)
lda_model, dictionary = lda.lda_trainer(
tongue_yaofangs_str, lda_model_path)
else:
lda_model, dictionary = lda.loadModelfromFile(
lda_model_path, readOnly=True)
total_tongue_x, total_y, total_aux_y = tongue2text_gen.data_tensorization_lda(
tongue_image_arrays, tongue_yaofangs, tongue_image_shape, nb_yao, lda_model.num_topics,
lda_model, dictionary, use_tfidf_tensor=use_tfidf_tensor)
# train_tongue_x = total_tongue_x[500 :]
# train_y = total_y[500 :]
# train_aux_y = total_aux_y[500 :]
# train_tongue_x = np.concatenate((total_tongue_x[: 2000], total_tongue_x[2500 :]), axis=0)
# train_y = np.concatenate((total_y[: 2000], total_y[2500 :]), axis=0)
# train_aux_y = np.concatenate((total_aux_y[: 2000], total_aux_y[2500 :]), axis=0)
# train_tongue_x = np.concatenate((total_tongue_x[: 4000], total_tongue_x[4500 :]), axis=0)
# train_y = np.concatenate((total_y[: 4000], total_y[4500 :]), axis=0)
# train_aux_y = np.concatenate((total_aux_y[: 4000], total_aux_y[4500 :]), axis=0)
# train_tongue_x = np.concatenate((total_tongue_x[: 6000], total_tongue_x[6500 :]), axis=0)
# train_y = np.concatenate((total_y[: 6000], total_y[6500 :]), axis=0)
# train_aux_y = np.concatenate((total_aux_y[: 6000], total_aux_y[6500 :]), axis=0)
train_tongue_x = total_tongue_x[: len(total_tongue_x) - 500]
train_y = total_y[: len(total_y) - 500]
train_aux_y = total_aux_y[: len(total_aux_y) - 500]
del(total_tongue_x)
del(total_y)
del(total_aux_y)
if use_data_augment == True:
# just can be use on service 225 with big memory
datagen = image_augment.image_augment_gen()
train_tongue_x, train_y, train_aux_y = image_augment.data_tensoration_augment_withaux(
datagen, train_tongue_x, train_y, train_aux_y)
scaling_act_type = 'tfidf' if use_tfidf_tensor else 'binary'
print('training 2 * cnn + mlp with double output(lda) tongue2text gen model------scaling_activation: %s...' %
scaling_act_type)
if use_tfidf_tensor == True:
tongue_gen_model = tongue2text_gen.k_cnns_biggertimes_channels_mlp_2output(yao_indices_dim=nb_yao,
tongue_image_shape=tongue_image_shape,
topics_dim=lda_model.num_topics,
times_nb_filters=channel_times,
with_compile=True, scaling_activation='tfidf')
else:
tongue_gen_model = tongue2text_gen.k_cnns_biggertimes_channels_mlp_2output(yao_indices_dim=nb_yao,
tongue_image_shape=tongue_image_shape,
topics_dim=lda_model.num_topics,
times_nb_filters=channel_times,
with_compile=True, scaling_activation='binary')
# trained_tongue_gen_model, history = tongue2text_gen.trainer(
# tongue_gen_model, train_tongue_x, train_y, train_aux_y)
record_path = None
if gen_model_path != None:
record_path = gen_model_path.replace('json', 'h5')
trained_tongue_gen_model, history = tongue2text_gen.trainer(
tongue_gen_model, train_tongue_x, train_y, train_aux_y=train_aux_y,
batch_size=8,
best_record_path=record_path)
if gen_model_path != None:
tongue2text_gen.storageModel(model=trained_tongue_gen_model, frame_path=gen_model_path,
replace_record=False)
print('history:')
for item in history:
print('\n{0}'.format(item))
return trained_tongue_gen_model