def train_autokeras(self):
#Load images
train_data, train_labels = load_image_dataset(csv_file_path=self.TRAIN_CSV_DIR, images_path=self.RESIZE_TRAIN_IMG_DIR)
test_data, test_labels = load_image_dataset(csv_file_path=self.TEST_CSV_DIR, images_path=self.RESIZE_TEST_IMG_DIR)
train_data = train_data.astype('float32') / 255
test_data = test_data.astype('float32') / 255
print("Train data shape:", train_data.shape)
clf = ImageClassifier(verbose=True, path=self.TEMP_DIR, resume=False)
clf.fit(train_data, train_labels, time_limit=self.TIME)
clf.final_fit(train_data, train_labels, test_data, test_labels, retrain=True)
evaluate_value = clf.evaluate(test_data, test_labels)
print("Evaluate:", evaluate_value)
# clf.load_searcher().load_best_model().produce_keras_model().save(MODEL_DIR)
# clf.export_keras_model(MODEL_DIR)
clf.export_autokeras_model(self.MODEL_DIR)
#统计训练信息
dic = {}
ishape = clf.cnn.searcher.input_shape
dic['n_train'] = train_data.shape[0] #训练总共用了多少图
dic['n_classes'] = clf.cnn.searcher.n_classes
dic['input_shape'] = str(ishape[0]) + 'x' + str(ishape[1]) + 'x' + str(ishape[2])
dic['history'] = clf.cnn.searcher.history
dic['model_count'] = clf.cnn.searcher.model_count
dic['best_model'] = clf.cnn.searcher.get_best_model_id()
best_model = [item for item in dic['history'] if item['model_id'] == dic['best_model']]
if len(best_model) > 0:
dic['loss'] = best_model[0]['loss']
dic['metric_value'] = best_model[0]['metric_value']
dic['evaluate_value'] = evaluate_value
self.traininfo = dic
def train_autokeras(RESIZE_TRAIN_IMG_DIR, RESIZE_TEST_IMG_DIR, TRAIN_CSV_DIR,
TEST_CSV_DIR, TIME):
# Load images
train_data, train_labels = load_image_dataset(
csv_file_path=TRAIN_CSV_DIR, images_path=RESIZE_TRAIN_IMG_DIR) # 加载数据
test_data, test_labels = load_image_dataset(
csv_file_path=TEST_CSV_DIR, images_path=RESIZE_TEST_IMG_DIR)
train_data = train_data.astype('float32') / 255
test_data = test_data.astype('float32') / 255
clf = ImageClassifier(verbose=True)
clf.fit(train_data, train_labels, time_limit=TIME) # 找最优模型
clf.final_fit(train_data,
train_labels,
test_data,
test_labels,
retrain=True) # 最优模型继续训练
y = clf.evaluate(test_data, test_labels)
print("测试集精确度:", y)
score = clf.evaluate(train_data, train_labels) # score: 0.8139240506329114
print("训练集精确度:", score)
clf.export_keras_model(MODEL_DIR) # 储存
def train_autokeras(RESIZE_TRAIN_IMG_DIR, TRAIN_CSV_DIR, RESIZE_TEST_IMG_DIR,
TEST_CSV_DIR, TIME):
#Load images
train_data, train_labels = load_image_dataset(
csv_file_path=TRAIN_CSV_DIR, images_path=RESIZE_TRAIN_IMG_DIR)
test_data, test_labels = load_image_dataset(
csv_file_path=TEST_CSV_DIR, images_path=RESIZE_TEST_IMG_DIR)
train_data = train_data.astype('float32') / 255
test_data = test_data.astype('float32') / 255
print("Train data shape:", train_data.shape)
clf = ImageClassifier(verbose=True)
clf.fit(train_data, train_labels, time_limit=TIME)
clf.final_fit(train_data,
train_labels,
test_data,
test_labels,
retrain=True)
y = clf.evaluate(test_data, test_labels)
print("Evaluate:", y)
#Predict the category of the test image
img = load_img(PREDICT_IMG_PATH)
x = img_to_array(img)
x = x.astype('float32') / 255
x = np.reshape(x, (1, RESIZE, RESIZE, 3))
print("x shape:", x.shape)
y = clf.predict(x)
print("predict:", y)
clf.load_searcher().load_best_model().produce_keras_model().save(MODEL_DIR)
#Save model architecture diagram
model = load_model(MODEL_DIR)
plot_model(model, to_file=MODEL_PNG)
def train_autokeras(self):
time_limit = self.projectinfo['parameter_time']
#Load images
train_data, train_labels = load_image_dataset(csv_file_path=self.project_train_labels_csv,
images_path=self.project_resize_train_dir)
test_data, test_labels = load_image_dataset(csv_file_path=self.project_test_labels_csv,
images_path=self.project_resize_test_dir)
train_data = train_data.astype('float32') / 255
test_data = test_data.astype('float32') / 255
self.log.info("Train data shape: %d" % train_data.shape[0])
clf = ImageClassifier(verbose=True, path=self.project_tmp_dir, resume=False)
clf.fit(train_data, train_labels, time_limit=time_limit)
clf.final_fit(train_data, train_labels, test_data, test_labels, retrain=True)
evaluate_value = clf.evaluate(test_data, test_labels)
self.log.info("Evaluate: %f" % evaluate_value)
clf.export_autokeras_model(self.project_mod_path)
#统计训练信息
dic = {}
ishape = clf.cnn.searcher.input_shape
dic['n_train'] = train_data.shape[0] #训练总共用了多少图
dic['n_classes'] = clf.cnn.searcher.n_classes
dic['input_shape'] = str(ishape[0]) + 'x' + str(ishape[1]) + 'x' + str(ishape[2])
dic['history'] = clf.cnn.searcher.history
dic['model_count'] = clf.cnn.searcher.model_count
dic['best_model'] = clf.cnn.searcher.get_best_model_id()
best_model = [item for item in dic['history'] if item['model_id'] == dic['best_model']]
if len(best_model) > 0:
dic['loss'] = best_model[0]['loss']
dic['metric_value'] = best_model[0]['metric_value']
dic['evaluate_value'] = evaluate_value
return dic
get_ipython().system(u'mkdir models')
clf = ImageClassifier(path="models/", verbose=True)
clf.fit(x_train, y_train, time_limit=hours_for_training * 60 * 60)
# In[ ]:
x_val, y_val = load_image_dataset(csv_file_path="split_data/val/label.csv",
images_path="split_data/resized-val/")
print(x_val.shape)
print(y_val.shape)
# In[ ]:
clf.final_fit(x_train, y_train, x_val, y_val, retrain=True)
y = clf.evaluate(x_val, y_val)
print(y)
# In[ ]:
x_test, y_test = load_image_dataset(csv_file_path="split_data/test/label.csv",
images_path="split_data/resized-test/")
print(x_test.shape)
print(y_test.shape)
# In[ ]:
y = clf.evaluate(x_test, y_test)
print(y)
# In[ ]:
if __name__ == '__main__':
# 需要把数据放到 ~/.keras/dataset 中，不然下载会报错
(x_train, y_train), (x_test, y_test) = mnist.load_data()
print(x_train.shape)
# (60000, 28, 28)
print('增加一个维度，以符合格式要求')
x_train = x_train.reshape(x_train.shape + (1, ))
print(x_train.shape)
# (60000, 28, 28, 1)
x_test = x_test.reshape(x_test.shape + (1, ))
# 指定模型更新路径
clf = ImageClassifier(path="automodels/", verbose=True)
# 限制为 4 个小时
# 搜索部分
gap = 6
clf.fit(x_train[::gap], y_train[::gap], time_limit=4 * 60 * 60)
# 用表现最好的再训练一次
clf.final_fit(x_train[::gap],
y_train[::gap],
x_test,
y_train,
retrain=True)
y = clf.evaluate(x_test, y_test)
print(y)
print("导出训练好的模型")
clf.export_autokeras_model("automodels/auto_mnist_model")
print("可视化模型")
visualize("automodels/")
train_data = train_data.astype('float32') / 255
test_data = test_data.astype('float32') / 255
print("train data shape:", train_data.shape)
# 使用图片识别器
clf = ImageClassifier(verbose=True)
# 给其训练数据和标签，训练的最长时间可以设定，假设为1分钟，autokers会不断找寻最优的网络模型
clf.fit(train_data, train_labels, time_limit=1 * 60)
# 找到最优模型后，再最后进行一次训练和验证
clf.final_fit(train_data,
train_labels,
test_data,
test_labels,
retrain=True)
# 给出评估结果
y = clf.evaluate(test_data, test_labels)
print("evaluate:", y)
# 给一个图片试试预测是否准确
img = load_img(PREDICT_IMG_PATH)
x = img_to_array(img)
x = x.astype('float32') / 255
x = np.reshape(x, (1, IMAGE_SIZE, IMAGE_SIZE, 3))
print("x shape:", x.shape)
# 最后的结果是一个numpy数组，里面是预测值4，意味着是马，说明预测准确
y = clf.predict(x)
print("predict:", y)
# 导出我们生成的模型
clf.export_keras_model(MODEL_DIR)
from keras.datasets import mnist from autokeras.image.image_supervised import ImageClassifier # gather data (x_train, y_train), (x_test, y_test) = mnist.load_data() x_train = x_train.reshape(x_train.shape + (1, )) x_test = x_test.reshape(x_test.shape + (1, )) # train the model model = ImageClassifier(verbose=True) model.fit(x_train, y_train, time_limit=15 * 60) model.final_fit(x_train, y_train, x_test, y_test, retrain=False) y = model.evaluate(x_test, y_test) print(y)
