import tflearn
from tflearn.layers.core import  input_data,dropout,fully_connected
from tflearn.layers.conv import conv_2d,max_pool_2d
from tflearn.layers.normalization import local_response_normalization
from tflearn.layers.estimator import regression
from tflearn.datasets import oxflower17

# 加载数据
X,Y = oxflower17.load_data(one_hot=True,resize_pics=(227,227))

network = input_data(shape=[None,227,227,3])
network = conv_2d(network,nb_filter=96,filter_size=11,strides=4,activation="relu")
network = max_pool_2d(network,kernel_size=3,strides=2)
network = local_response_normalization(network)


network = conv_2d(network,nb_filter=256,filter_size=5,strides=4,activation="relu")
network = max_pool_2d(network,kernel_size=3,strides=2)
network = local_response_normalization(network)

network = conv_2d(network,nb_filter=384,filter_size=3,activation="relu")
network = conv_2d(network,nb_filter=384,filter_size=3,activation="relu")
network = conv_2d(network,nb_filter=256,filter_size=3,activation="relu")
network = max_pool_2d(network,kernel_size=3,strides=2)
network = local_response_normalization(network)
network = fully_connected(network,4096,activation="tanh")
network = dropout(network,0.5)
network = fully_connected(network,4096,activation="tanh")
network = dropout(network,0.5)
network = fully_connected(network,17,activation="softmax")
network = regression(network,optimizer="momentum",loss="categorical_crossentropy",
                     learning_rate=0.001) # 回归学习率

model = tflearn.DNN(network,checkpoint_path="model_alexnet",max_checkpoints=1,tensorboard_verbose=2)

model.fit(X,Y,n_epoch=2000,validation_set=0.1,shuffle=True,show_metric=True,
          batch_size=64,snapshot_step=200,
          snapshot_epoch=False,run_id='alexnet_oxflowers17')