AlexNet模型实现（4. 基于Tensorflow的python实现）

基于Tensorflow提供的深度学习架构，在阿里云的数加平台运行，效率很高，但是由于目前频繁获取数据导致请求消耗金钱太多。在本地四核PC运行效率比C++实现也要高。

贴码：

# -*- coding: utf-8 -*-import osos.environ["CUDA_VISIBLE_DEVICES"] = "0,1"import sysimport argparseimport numpy as npimport tensorflow as tfFLAGS = Noneclass slqAlexNet():def __init__(self):self.accuray = 0# tf.set_random_seed(2017)self.weights = {# kernel map: 11*11, input map: 3, output map: 96'c1conv':tf.get_variable('c1conv', [11,11,3,96], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# output map: 96's1pool':tf.get_variable('s1pool', [96,], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# kernel map: 5*5, input map: 48, output map: 96'c2conv':tf.get_variable('c2conv', [5,5,96,256], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# output map: 256's2pool':tf.get_variable('s2pool', [256,], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# kernel map: 3*3, input map: 256, output map: 384'c3conv':tf.get_variable('c3conv', [3,3,256,384], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# kernel map: 3*3, input map: 192, output map: 384'c4conv':tf.get_variable('c4conv', [3,3,384,384], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# kernel map: 3*3, input map: 192, output map: 256'c5conv':tf.get_variable('c5conv', [3,3,384,256], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# output map: 256's5pool':tf.get_variable('s5pool', [256,], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# intput map: 9216, output map: 4096'f1conn':tf.get_variable('f1conn', [9216,4096], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# input map: 4096, output map: 4096'f2conn':tf.get_variable('f2conn', [4096, 4096], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),# input map: 4096, output map: 1000'f3conn':tf.get_variable('f3conn', [4096, 1000], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01, seed=None, dtype=tf.float32)),}self.bias={'c1bias':tf.get_variable('c1bias', [96,], initializer=tf.zeros_initializer(dtype=tf.float32)),'s1bias':tf.get_variable('s1bias', [96,], initializer=tf.zeros_initializer(dtype=tf.float32)),'c2bias':tf.get_variable('c2bias', [256,], initializer=tf.ones_initializer(dtype=tf.float32)),'s2bias':tf.get_variable('s2bias', [256,], initializer=tf.zeros_initializer(dtype=tf.float32)),'c3bias':tf.get_variable('c3bias', [384,], initializer=tf.zeros_initializer(dtype=tf.float32)),'c4bias':tf.get_variable('c4bias', [384,], initializer=tf.ones_initializer(dtype=tf.float32)),'c5bias':tf.get_variable('c5bias', [256,], initializer=tf.ones_initializer(dtype=tf.float32)),'s5bias':tf.get_variable('s5bias', [256,], initializer=tf.zeros_initializer(dtype=tf.float32)),'f1bias':tf.get_variable('f1bias', [4096,], initializer=tf.ones_initializer(dtype=tf.float32)),'f2bias':tf.get_variable('f2bias', [4096,], initializer=tf.ones_initializer(dtype=tf.float32)),'f3bias':tf.get_variable('f3bias', [1000,], initializer=tf.zeros_initializer(dtype=tf.float32)),}def __mode__(self, data):c1map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(data, self.weights['c1conv'], strides=[1,4,4,1], padding='VALID'), self.bias['c1bias']))# s1map = tf.nn.tanh(tf.nn.bias_add(tf.nn.max_pool(c1map, ksize=[1,3,3,1], strides=[1,2,2,1], padding='VALID'), self.bias['s1bias']))s1map = tf.nn.bias_add(tf.nn.max_pool(c1map, ksize=[1,3,3,1], strides=[1,2,2,1], padding='VALID'), self.bias['s1bias'])drops1 = tf.nn.dropout(s1map, 0.5)c2map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(drops1, self.weights['c2conv'], strides=[1,1,1,1], padding='SAME'), self.bias['c2bias']))# c2map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(s1map, self.weights['c2conv'], strides=[1,1,1,1], padding='SAME'), self.bias['c2bias']))# s2map = tf.nn.tanh(tf.nn.bias_add(tf.nn.max_pool(c2map, ksize=[1,3,3,1], strides=[1,2,2,1], padding='VALID'), self.bias['s2bias']))s2map = tf.nn.bias_add(tf.nn.max_pool(c2map, ksize=[1,3,3,1], strides=[1,2,2,1], padding='VALID'), self.bias['s2bias'])c3map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(s2map, self.weights['c3conv'], strides=[1,1,1,1], padding='SAME'), self.bias['c3bias']))dropc3 = tf.nn.dropout(c3map, 0.5)c4map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(dropc3, self.weights['c4conv'], strides=[1,1,1,1], padding='SAME'), self.bias['c4bias']))# c4map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(c3map, self.weights['c4conv'], strides=[1,1,1,1], padding='SAME'), self.bias['c4bias']))dropc4 = tf.nn.dropout(c4map, 0.5)c5map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(dropc4, self.weights['c5conv'], strides=[1,1,1,1], padding='SAME'), self.bias['c5bias']))# c5map = tf.nn.relu(tf.nn.bias_add(tf.nn.conv2d(c4map, self.weights['c5conv'], strides=[1,1,1,1], padding='SAME'), self.bias['c5bias']))# s5map = tf.nn.tanh(tf.nn.bias_add(tf.nn.max_pool(c5map, ksize=[1,3,3,1], strides=[1,2,2,1], padding='VALID'), self.bias['s5bias']))s5map = tf.nn.bias_add(tf.nn.max_pool(c5map, ksize=[1,3,3,1], strides=[1,2,2,1], padding='VALID'), self.bias['s5bias'])flatten = tf.reshape(s5map, [-1, self.weights['f1conn'].get_shape().as_list()[0]])drop1 = tf.nn.dropout(flatten, 0.5)f1map = tf.nn.relu(tf.matmul(drop1, self.weights['f1conn'])+self.bias['f1bias'])drop2 = tf.nn.dropout(f1map, 0.5)f2map = tf.nn.relu(tf.matmul(drop2, self.weights['f2conn'])+self.bias['f2bias'])# f3map = tf.nn.tanh(tf.matmul(f2map, self.weights['f3conn'])+self.bias['f3bias'])f3map = tf.matmul(f2map, self.weights['f3conn'])+self.bias['f3bias']return f3mapdef train(self):witdthBytes = 227heightBytes = 227depthBytes = 3recBytes = 3*227*227print("create file dirctory list ...")trainbatch1 = os.path.join(FLAGS.buckets, "imgTrainBatch01")trainbatch2 = os.path.join(FLAGS.buckets, "imgTrainBatch02")trainbatch3 = os.path.join(FLAGS.buckets, "imgTrainBatch03")trainbatch4 = os.path.join(FLAGS.buckets, "imgTrainBatch04")trainbatch5 = os.path.join(FLAGS.buckets, "imgTrainBatch05")filename_queue = tf.train.string_input_producer([trainbatch1, trainbatch2, trainbatch3, trainbatch4, trainbatch5], shuffle=False)filename_queue_t = tf.train.string_input_producer([os.path.join(FLAGS.buckets, "imgTestArray")])filename_out_module = os.path.join(FLAGS.buckets, "alexnet.module")reader = tf.FixedLengthRecordReader(record_bytes=recBytes)print("create graph read train ...")key,value = reader.read(filename_queue)bytes = tf.decode_raw(value,out_type=tf.uint8)originalImg  = tf.reshape(bytes, [depthBytes,heightBytes,witdthBytes])print(originalImg)imgtrain = tf.transpose(originalImg,[1,2,0])print(imgtrain)imgtrain = tf.cast(imgtrain, tf.float32) * 2.0/255 - 1.0print(imgtrain)imgtrain = tf.reshape(imgtrain, shape=[-1, 227, 227, 3])print(imgtrain)print("create graph read test ...")key_t,value_t = reader.read(filename_queue_t)bytes_t = tf.decode_raw(value_t,out_type=tf.uint8)originalImg_t  = tf.reshape(tf.strided_slice(bytes_t,[0],[recBytes]),[depthBytes,heightBytes,witdthBytes])img_t = tf.transpose(originalImg,[1,2,0])img_t = tf.cast(img_t, tf.float32) * 2.0/255 - 1.0img_t = tf.reshape(img_t, shape=[-1, 227, 227, 3])label_reader = tf.FixedLengthRecordReader(record_bytes=1)label_handle = tf.train.string_input_producer([os.path.join(FLAGS.buckets, "imgTrainLabel")])label_t = tf.train.string_input_producer([os.path.join(FLAGS.buckets, "imgTestLabel")])print("create graph read train label ...")key_tr,value_tr = label_reader.read(label_handle)bytes_tr = tf.decode_raw(value_tr,out_type=tf.uint8)label_train  = tf.strided_slice(bytes_tr,[0],[1])one_hot_label_train = tf.one_hot(label_train, 1000, 1.0, 0.0, dtype=tf.float32)print("create graph read test label ...")key_tl,value_tl = label_reader.read(label_t)bytes_tl = tf.decode_raw(value_tl,out_type=tf.uint8)label_test  = tf.strided_slice(bytes_tl,[0],[1])one_hot_label_test = tf.one_hot(label_test, 1000, 1.0, 0.0, dtype=tf.float32)print("create step ...")train_tep = self.__mode__(imgtrain)test_tep = self.__mode__(img_t)print("create train graph ...")train_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=one_hot_label_train, logits=train_tep))train_loop = tf.train.GradientDescentOptimizer(0.0001).minimize(train_loss)# print("create test graph ...")# test_loop_one = tf.equal(tf.argmax(test_tep, 1), tf.argmax(one_hot_label_test, 1))init=tf.global_variables_initializer()saver = tf.train.Saver()self.savePara()with tf.Session() as sess:print("start session ...")sess.run(init) coord = tf.train.Coordinator()threads = tf.train.start_queue_runners(coord=coord)max_iter = 100 iter=0 while iter < max_iter: self.accuray = np.array([0,0,0,0,0])for i in range(52500):if (0 == (i%5000)):print("train tep ", i)sess.run(train_loop)for i in range(5000):if (0 == (i%500)):print("test tep ", i)test_v = sess.run(test_tep)v_a = np.array([0.,1.,2.,3.,4.])i_a = np.array([test_v[0,0],test_v[0,1],test_v[0,2],test_v[0,3],test_v[0,4]])for j in range(5,1000):if (test_v[j] > np.min(i_a)):i_a[int(np.argmin(i_a))] = test_v[0,j]v_a[int(np.argmin(i_a))] = jtest_l = sess.run(label_test)# sortfor j in range(4):for k in range(j+1, 5):if (i_a[j] < i_a[k]):ta = i_a[j]i_a[j] = i_a[k]i_a[k] = tata = v_a[j]v_a[j] = v_a[k]v_a[k] = tafor j in range(5):if (v_a[i] == test_l):self.accuray[i] += 1# valid_v = sess.run(test_loop_one)# if (False != valid_v[0]):# self.accuray += 1.iter+=1print("epock loop ", iter, " acc ", self.accuray)if (np.sum(self.accuray)/5000.0 > 0.98):saver.save(sess, filename_out_module)breakif (iter == max_iter):saver.save(sess, filename_out_module)coord.request_stop()#queue need to be closedcoord.join(threads)def savePara(self):tf.add_to_collection('alexparam', self.weights['c1conv'])tf.add_to_collection('alexparam', self.weights['s1pool'])tf.add_to_collection('alexparam', self.weights['c2conv'])tf.add_to_collection('alexparam', self.weights['s2pool'])tf.add_to_collection('alexparam', self.weights['c3conv'])tf.add_to_collection('alexparam', self.weights['c4conv'])tf.add_to_collection('alexparam', self.weights['c5conv'])tf.add_to_collection('alexparam', self.weights['s5pool'])tf.add_to_collection('alexparam', self.weights['f1conn'])tf.add_to_collection('alexparam', self.weights['f2conn'])tf.add_to_collection('alexparam', self.weights['f3conn'])tf.add_to_collection('alexparam', self.bias['c1bias'])tf.add_to_collection('alexparam', self.bias['s1bias'])tf.add_to_collection('alexparam', self.bias['c2bias'])tf.add_to_collection('alexparam', self.bias['s2bias'])tf.add_to_collection('alexparam', self.bias['c3bias'])tf.add_to_collection('alexparam', self.bias['c4bias'])tf.add_to_collection('alexparam', self.bias['c5bias'])tf.add_to_collection('alexparam', self.bias['s5bias'])tf.add_to_collection('alexparam', self.bias['f1bias'])tf.add_to_collection('alexparam', self.bias['f2bias'])tf.add_to_collection('alexparam', self.bias['f3bias'])def main(_):print("create alexnet obj ...")alexNet = slqAlexNet()print("start train ...")alexNet.train()if __name__ == '__main__':parser = argparse.ArgumentParser()#bucketsparser.add_argument('--buckets', type=str, default='',help='input data path')#checkpointparser.add_argument('--checkpointDir', type=str, default='',help='output model path')FLAGS, _ = parser.parse_known_args()print(FLAGS.buckets)tf.app.run(main=main)