僅做記錄,后面慢慢整理
訓練函數:
from skimage import io, transform # skimage模塊下的io transform(圖像的形變與縮放)模塊 import glob # glob 文件通配符模塊 import os # os 處理文件和目錄的模塊 import tensorflow as tf import numpy as np # 多維數據處理模塊 import time # 數據集地址 path = 'E:/tensor_data/powerpoint/test_database/' # 模型保存地址 model_path = 'E:/tensor_data/powerpoint/model/fc_model.ckpt' # 將所有的圖片resize成100*100 w = 100 h = 100 c = 3 print("開始執行讀取圖片和數據處理") # 讀取圖片+數據處理 def read_img(path): # os.listdir(path) 返回path指定的文件夾包含的文件或文件夾的名字的列表 # os.path.isdir(path)判斷path是否是目錄 # b = [x+x for x in list1 if x+x<15 ] 列表生成式,循環list1,當if為真時,將x+x加入列表b print(os.listdir(path)) '''for x in os.listdir(path): if os.path.isdir(path+x): print(x)''' cate = [path + x for x in os.listdir(path) if os.path.isdir(path + x)] print("數據集地址:"+path) imgs = [] labels = [] for idx, folder in enumerate(cate): # glob.glob(s+'*.py') 從目錄通配符搜索中生成文件列表 for im in glob.glob(folder + '/*.jpg'): # 輸出讀取的圖片的名稱 print('reading the images:%s' % (im)) # io.imread(im)讀取單張RGB圖片 skimage.io.imread(fname,as_grey=True)讀取單張灰度圖片 # 讀取的圖片 img = io.imread(im) # skimage.transform.resize(image, output_shape)改變圖片的尺寸 img = transform.resize(img, (w, h)) # 將讀取的圖片數據加載到imgs[]列表中 imgs.append(img) # 將圖片的label加載到labels[]中,與上方的imgs索引對應 labels.append(idx) # 將讀取的圖片和labels信息,轉化為numpy結構的ndarr(N維數組對象(矩陣))數據信息 return np.asarray(imgs, np.float32), np.asarray(labels, np.int32) # 調用讀取圖片的函數,得到圖片和labels的數據集 data, label = read_img(path) # 打亂順序 # 讀取data矩陣的第一維數(圖片的個數) num_example = data.shape[0] # 產生一個num_example范圍,步長為1的序列 arr = np.arange(num_example) # 調用函數,打亂順序 np.random.shuffle(arr) # 按照打亂的順序,重新排序 data = data[arr] label = label[arr] # 將所有數據分為訓練集和驗證集 ratio = 0.8 s = np.int(num_example * ratio) x_train = data[:s] y_train = label[:s] x_val = data[s:] y_val = label[s:] # -----------------構建網絡---------------------- # 本程序cnn網絡模型,共有7層,前三層為卷積層,后三層為全連接層,前三層中,每層包含卷積、激活、池化層 # 占位符設置輸入參數的大小和格式 x = tf.placeholder(tf.float32, shape=[None, w, h, c], name='x') y_ = tf.placeholder(tf.int32, shape=[None, ], name='y_') def inference(input_tensor, train, regularizer): # -----------------------第一層---------------------------- with tf.variable_scope('layer1-conv1'): # 初始化權重conv1_weights為可保存變量,大小為5x5,3個通道(RGB),數量為32個 conv1_weights = tf.get_variable("weight", [5, 5, 3, 32], initializer=tf.truncated_normal_initializer(stddev=0.1)) # 初始化偏置conv1_biases,數量為32個 conv1_biases = tf.get_variable("bias", [32], initializer=tf.constant_initializer(0.0)) # 卷積計算,tf.nn.conv2d為tensorflow自帶2維卷積函數,input_tensor為輸入數據, # conv1_weights為權重,strides=[1, 1, 1, 1]表示左右上下滑動步長為1,padding='SAME'表示輸入和輸出大小一樣,即補0 conv1 = tf.nn.conv2d(input_tensor, conv1_weights, strides=[1, 1, 1, 1], padding='SAME') # 激勵計算,調用tensorflow的relu函數 relu1 = tf.nn.relu(tf.nn.bias_add(conv1, conv1_biases)) with tf.name_scope("layer2-pool1"): # 池化計算,調用tensorflow的max_pool函數,strides=[1,2,2,1],表示池化邊界,2個對一個生成,padding="VALID"表示不操作。 pool1 = tf.nn.max_pool(relu1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="VALID") # -----------------------第二層---------------------------- with tf.variable_scope("layer3-conv2"): # 同上,不過參數的有變化,根據卷積計算和通道數量的變化,設置對應的參數 conv2_weights = tf.get_variable("weight", [5, 5, 32, 64], initializer=tf.truncated_normal_initializer(stddev=0.1)) conv2_biases = tf.get_variable("bias", [64], initializer=tf.constant_initializer(0.0)) conv2 = tf.nn.conv2d(pool1, conv2_weights, strides=[1, 1, 1, 1], padding='SAME') relu2 = tf.nn.relu(tf.nn.bias_add(conv2, conv2_biases)) with tf.name_scope("layer4-pool2"): pool2 = tf.nn.max_pool(relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID') # -----------------------第三層---------------------------- # 同上,不過參數的有變化,根據卷積計算和通道數量的變化,設置對應的參數 with tf.variable_scope("layer5-conv3"): conv3_weights = tf.get_variable("weight", [3, 3, 64, 128], initializer=tf.truncated_normal_initializer(stddev=0.1)) conv3_biases = tf.get_variable("bias", [128], initializer=tf.constant_initializer(0.0)) conv3 = tf.nn.conv2d(pool2, conv3_weights, strides=[1, 1, 1, 1], padding='SAME') relu3 = tf.nn.relu(tf.nn.bias_add(conv3, conv3_biases)) with tf.name_scope("layer6-pool3"): pool3 = tf.nn.max_pool(relu3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID') # -----------------------第四層---------------------------- # 同上,不過參數的有變化,根據卷積計算和通道數量的變化,設置對應的參數 with tf.variable_scope("layer7-conv4"): conv4_weights = tf.get_variable("weight", [3, 3, 128, 128], initializer=tf.truncated_normal_initializer(stddev=0.1)) conv4_biases = tf.get_variable("bias", [128], initializer=tf.constant_initializer(0.0)) conv4 = tf.nn.conv2d(pool3, conv4_weights, strides=[1, 1, 1, 1], padding='SAME') relu4 = tf.nn.relu(tf.nn.bias_add(conv4, conv4_biases)) with tf.name_scope("layer8-pool4"): pool4 = tf.nn.max_pool(relu4, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='VALID') nodes = 6 * 6 * 128 reshaped = tf.reshape(pool4, [-1, nodes]) # 使用變形函數轉化結構 # -----------------------第五層--------------------------- with tf.variable_scope('layer9-fc1'): # 初始化全連接層的參數,隱含節點為1024個 fc1_weights = tf.get_variable("weight", [nodes, 1024], initializer=tf.truncated_normal_initializer(stddev=0.1)) if regularizer != None: tf.add_to_collection('losses', regularizer(fc1_weights)) # 正則化矩陣 fc1_biases = tf.get_variable("bias", [1024], initializer=tf.constant_initializer(0.1)) # 使用relu函數作為激活函數 fc1 = tf.nn.relu(tf.matmul(reshaped, fc1_weights) + fc1_biases) # 采用dropout層,減少過擬合和欠擬合的程度,保存模型最好的預測效率 if train: fc1 = tf.nn.dropout(fc1, 0.5) # -----------------------第六層---------------------------- with tf.variable_scope('layer10-fc2'): # 同上,不過參數的有變化,根據卷積計算和通道數量的變化,設置對應的參數 fc2_weights = tf.get_variable("weight", [1024, 512], initializer=tf.truncated_normal_initializer(stddev=0.1)) if regularizer != None: tf.add_to_collection('losses', regularizer(fc2_weights)) fc2_biases = tf.get_variable("bias", [512], initializer=tf.constant_initializer(0.1)) fc2 = tf.nn.relu(tf.matmul(fc1, fc2_weights) + fc2_biases) if train: fc2 = tf.nn.dropout(fc2, 0.5) # -----------------------第七層---------------------------- with tf.variable_scope('layer11-fc3'): # 同上,不過參數的有變化,根據卷積計算和通道數量的變化,設置對應的參數 fc3_weights = tf.get_variable("weight", [512, 5], initializer=tf.truncated_normal_initializer(stddev=0.1)) if regularizer != None: tf.add_to_collection('losses', regularizer(fc3_weights)) fc3_biases = tf.get_variable("bias", [5], initializer=tf.constant_initializer(0.1)) logit = tf.add(tf.matmul(fc2, fc3_weights), fc3_biases, name="output") # matmul矩陣相乘 # 返回最后的計算結果 return logit # ---------------------------網絡結束--------------------------- # 設置正則化參數為0.0001 regularizer = tf.contrib.layers.l2_regularizer(0.0001) # 將上述構建網絡結構引入 logits = inference(x, False, regularizer) # (小處理)將logits乘以1賦值給logits_eval,定義name,方便在后續調用模型時通過tensor名字調用輸出tensor b = tf.constant(value=1, dtype=tf.float32) logits_eval = tf.multiply(logits, b, name='logits_eval') # b為1 # 設置損失函數,作為模型訓練優化的參考標准,loss越小,模型越優 loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=y_) # 設置整體學習率為α為0.001 train_op = tf.train.AdamOptimizer(learning_rate=0.001).minimize(loss) # 設置預測精度 correct_prediction = tf.equal(tf.cast(tf.argmax(logits, 1), tf.int32), y_) acc = tf.reduce_mean(tf.cast(correct_prediction, tf.float32)) # 定義一個函數,按批次取數據 def minibatches(inputs=None, targets=None, batch_size=None, shuffle=False): assert len(inputs) == len(targets) if shuffle: indices = np.arange(len(inputs)) np.random.shuffle(indices) for start_idx in range(0, len(inputs) - batch_size + 1, batch_size): if shuffle: excerpt = indices[start_idx:start_idx + batch_size] else: excerpt = slice(start_idx, start_idx + batch_size) yield inputs[excerpt], targets[excerpt] # 訓練和測試數據,可將n_epoch設置更大一些 # 迭代次數 n_epoch = 20#10 # 每次迭代輸入的圖片數據 batch_size = 64 saver = tf.train.Saver(max_to_keep=4) # 可以指定保存的模型個數,利用max_to_keep=4,則最終會保存4個模型( with tf.Session() as sess: # 初始化全局參數 sess.run(tf.global_variables_initializer()) # 開始迭代訓練,調用的都是前面設置好的函數或變量 for epoch in range(n_epoch): start_time = time.time() # training#訓練集 train_loss, train_acc, n_batch = 0, 0, 0 for x_train_a, y_train_a in minibatches(x_train, y_train, batch_size, shuffle=True): _, err, ac = sess.run([train_op, loss, acc], feed_dict={x: x_train_a, y_: y_train_a}) train_loss += err; train_acc += ac; n_batch += 1 print(" train loss: %f" % (np.sum(train_loss) / n_batch)) print(" train acc: %f" % (np.sum(train_acc) / n_batch)) # validation#驗證集 val_loss, val_acc, n_batch = 0, 0, 0 for x_val_a, y_val_a in minibatches(x_val, y_val, batch_size, shuffle=False): err, ac = sess.run([loss, acc], feed_dict={x: x_val_a, y_: y_val_a}) val_loss += err; val_acc += ac; n_batch += 1 print(" validation loss: %f" % (np.sum(val_loss) / n_batch)) print(" validation acc: %f" % (np.sum(val_acc) / n_batch)) # 保存模型及模型參數 if epoch % 2 == 0: saver.save(sess, model_path, global_step=epoch) print(sess.graph.name_scope)
測試代碼:
from skimage import io, transform import tensorflow as tf import numpy as np import os # os 處理文件和目錄的模塊 import glob # glob 文件通配符模塊 # 此程序作用於進行簡單的預測,取5個圖片來進行預測,如果有多數據預測,按照cnn.py中,讀取數據的方式即可 path = 'E:/tensor_data/powerpoint/test_powerpoint/' # 類別代表字典 flower_dict = {0: '其他', 1: '文檔', 2: '幻燈片', 3: '黑板', 4: '不可能出現的類別'} w = 100 h = 100 c = 3 # 讀取圖片+數據處理 def read_img(path): # os.listdir(path) 返回path指定的文件夾包含的文件或文件夾的名字的列表 # os.path.isdir(path)判斷path是否是目錄 # b = [x+x for x in list1 if x+x<15 ] 列表生成式,循環list1,當if為真時,將x+x加入列表b cate = [path + x for x in os.listdir(path) if os.path.isdir(path + x)] imgs = [] for idx, folder in enumerate(cate): # glob.glob(s+'*.py') 從目錄通配符搜索中生成文件列表 for im in glob.glob(folder + '/*.jpg'): # 輸出讀取的圖片的名稱 print('reading the images:%s' % (im)) # io.imread(im)讀取單張RGB圖片 skimage.io.imread(fname,as_grey=True)讀取單張灰度圖片 # 讀取的圖片 img = io.imread(im) # skimage.transform.resize(image, output_shape)改變圖片的尺寸 img = transform.resize(img, (w, h)) # 將讀取的圖片數據加載到imgs[]列表中 imgs.append(img) # 將圖片的label加載到labels[]中,與上方的imgs索引對應 # labels.append(idx) # 將讀取的圖片和labels信息,轉化為numpy結構的ndarr(N維數組對象(矩陣))數據信息 return np.asarray(imgs, np.float32) # 調用讀取圖片的函數,得到圖片和labels的數據集 data = read_img(path) with tf.Session() as sess: saver = tf.train.import_meta_graph('E:/tensor_data/powerpoint/model/fc_model.ckpt-18.meta') saver.restore(sess, tf.train.latest_checkpoint('E:/tensor_data/powerpoint/model/')) # sess:表示當前會話,之前保存的結果將被加載入這個會話 # 設置每次預測的個數 graph = tf.get_default_graph() x = graph.get_tensor_by_name("x:0") feed_dict = {x: data} logits = graph.get_tensor_by_name("logits_eval:0") # eval功能等同於sess(run) classification_result = sess.run(logits, feed_dict) # 打印出預測矩陣 print(classification_result) # 打印出預測矩陣每一行最大值的索引 print(tf.argmax(classification_result, 1).eval()) # 根據索引通過字典對應的分類 output = [] output = tf.argmax(classification_result, 1).eval() for i in range(len(output)): print("第", i + 1, "張圖片預測:" + flower_dict[output[i]])
這里生成的模型是ckpt,參考代碼CNN中是沒有指定輸入輸出結點名稱的,這里直接在源碼第11層修改即可。
使用Netron可以快速查看模型結構,找到輸入輸出結點名稱。
也可以使用代碼打印全部結點名稱:
import os import tensorflow as tf checkpoint_path=os.path.join('E:/tensor_data/powerpoint/model/fc_model.ckpt-18') reader=pywrap_tensorflow.NewCheckpointReader(checkpoint_path) var_to_shape_map=reader.get_variable_to_shape_map() for key in var_to_shape_map: print ('tensor_name: ',key)
拿到輸出結點名稱后,就可以使用腳本對ckpt模型轉換了,轉成pb格式
第一個參數是 ckpt模型地址,第二個是pb模型輸出地址,第三個是輸出結點
import tensorflow as tf def read_graph_from_ckpt(ckpt_path, out_pb_path, output_name): # 從meta文件加載網絡結構 saver = tf.train.import_meta_graph(ckpt_path + '.meta', clear_devices=True) graph = tf.get_default_graph() with tf.Session(graph=graph) as sess: sess.run(tf.global_variables_initializer()) # 從ckpt加載參數 saver.restore(sess, ckpt_path) output_tf = graph.get_tensor_by_name(output_name) # 固化 pb_graph = tf.graph_util.convert_variables_to_constants(sess, graph.as_graph_def(), [output_tf.op.name]) # 保存 with tf.gfile.FastGFile(out_pb_path, mode='wb') as f: f.write(pb_graph.SerializeToString()) read_graph_from_ckpt('E:/tensor_data/powerpoint/model/fc_model.ckpt-18', 'E:/tensor_data/powerpoint/model/idcard_seg.pb', 'layer11-fc3/output:0')
拿到pb模型后,再使用Netron查看就清晰了很多~~~~
由於我訓練模型是為了手機使用的,因此還需要將pb模型轉成tflite格式
查看官方文檔發現已經提供了轉換的py接口,直接使用就好啦~
input是輸入結點,output是輸出結點,使用Netron看一下就好了
生成的tflite在你的工程根目錄下
import tensorflow as tf graph_def_file = "E:/tensor_data/powerpoint/model/idcard_seg.pb" input_arrays = ["x"] output_arrays = ["layer11-fc3/output"] converter = tf.lite.TFLiteConverter.from_frozen_graph( graph_def_file, input_arrays, output_arrays) tflite_model = converter.convert() open("converted_model.tflite", "wb").write(tflite_model)
那個啥,完全沒有測試模型的准確率emmm先試試看吧!