#!/usr/bin/env python3 # -*- coding: utf-8 -*- """ Created on Fri Apr 7 17:15:02 2017 @author: wowjoy """ #============================================================================== # Import module #============================================================================== import tensorflow as tf import model_tool from resnet_model import Resnet import shutil import time import os import pickle #============================================================================== # Global Flags and related preparation #============================================================================== batch_size =128 n_epoch = 200 train_iters = int(50000/batch_size) val_iters = int(10000/100) train_data = ['train.tfrecords'] test_data = ['test.tfrecords'] tf.set_random_seed(2017) #============================================================================== # Input and global variables #============================================================================== with tf.name_scope('Input'): phase_train = tf.placeholder(bool, name='Phase_train')#Train mode or validation mode global_step = tf.Variable(0, trainable=False, name='global_step') with tf.name_scope('Batch_input'): train_data_batch, train_label_batch = model_tool.make_train_batch(train_data, batch_size, flip=True, shuffle=True) test_data_batch, test_label_batch = model_tool.make_test_batch(test_data, 100, shuffle=False) image_batch, label_batch = tf.cond(phase_train, lambda:(train_data_batch, train_label_batch),#Return when Phase_train is True lambda:(test_data_batch, test_label_batch)#Return when Phase_train is False ) #============================================================================== # Network interface #============================================================================== config = tf.ConfigProto() config.gpu_options.allow_growth = True sess = tf.Session(config=config) logits = Resnet(image_batch, n = 5, train=phase_train, k=1, num_classes=100)# The number of layers is n*6+2, k is the multiply factor of filters saver = tf.train.Saver() #============================================================================== # Loss #============================================================================== with tf.name_scope('Loss'): loss_ = model_tool.softmax_sparse(logits=logits, labels=label_batch)#Loss of softmatx layer loss_re = tf.get_collection(tf.GraphKeys.REGULARIZATION_LOSSES)#Loss of regularization loss = tf.add_n([loss_]+loss_re)# Sum of two losses # tf.summary.scalar('loss', loss) #============================================================================== # Evaluation method #============================================================================== with tf.name_scope('Evaluation'): accuracy = model_tool.accuracy(logits=logits, labels=label_batch)#Common accuracy # tf.summary.scalar('accuracy', accuracy) #============================================================================== # Optimizer define #============================================================================== with tf.name_scope('Train_op'): lr = tf.train.exponential_decay(0.2, global_step, 50000/batch_size*60, 0.1, staircase=True) optimizer = tf.train.MomentumOptimizer(learning_rate=lr, momentum=0.9, use_nesterov=True) grads = optimizer.compute_gradients(loss) train_op = optimizer.apply_gradients(grads, global_step) grad_list = [] for grad, var in grads: if grad is not None and ('weights' in var.op.name): grad_list.append(tf.summary.histogram(var.op.name +'/gradients', grad)) #============================================================================== # Monitor #============================================================================== base_list = [tf.summary.scalar('loss', loss), tf.summary.scalar('accuracy', accuracy)] advanced_list = base_list + grad_list merge1 = tf.summary.merge(advanced_list) merge2 = tf.summary.merge(base_list) #============================================================================== # Save and summary #============================================================================== try: shutil.rmtree('logs/resnet6_2') except: pass train_writer = tf.summary.FileWriter(logdir='logs/resnet6_2/train', graph=sess.graph) val_writer = tf.summary.FileWriter(logdir='logs/resnet6_2/val') #merged = tf.summary.merge_all() init = tf.global_variables_initializer() sess.run(init) #saver.restore(sess,'logs/resnet6_2/model-78000') #============================================================================== # Train and test #============================================================================== coord = tf.train.Coordinator() threads = tf.train.start_queue_runners(sess=sess, coord=coord) def train_val(train_iters): for j in range(train_iters): _, train_summary = sess.run([train_op, merge1], feed_dict={phase_train:True}) train_writer.add_summary(train_summary, sess.run(global_step)) val_summary = sess.run(merge2, feed_dict={phase_train:False}) val_writer.add_summary(val_summary, sess.run(global_step)) print(sess.run(global_step)) print('finish') def time_cost(train_iters): for j in range(train_iters): a = time.time() _, train_summary = sess.run([train_op, merge1], feed_dict={phase_train:True}) b = time.time() train_writer.add_summary(train_summary, sess.run(global_step)) print('Train_op cost: %0.5f'%(time.time() - a)) print('Write summary cost: %0.5f'%(time.time() - b)) c = time.time() val_summary = sess.run(merged, feed_dict={phase_train:False}) print('validation cost: %0.5f'%(time.time() - c)) val_writer.add_summary(val_summary, sess.run(global_step)) print(sess.run(global_step)) def train(train_iters, monitor=None): sum_acc = 0 sum_loss = 0 for j in range(train_iters): _, train_acc, train_loss = sess.run([train_op, accuracy, loss], feed_dict={phase_train:True}) sum_acc += train_acc sum_loss += train_loss avg_acc = sum_acc/train_iters avg_loss = sum_loss/train_iters print('Training data\'s accuracy: %0.4f' %avg_acc) print('Training data\'s loss: %0.4f'%avg_loss) if monitor is not None: monitor['train_loss'].append(avg_loss) monitor['train_acc'].append(avg_acc) def val(iters, monitor=None): sum_acc = 0 sum_loss = 0 for i in range(iters): val_acc, val_loss = sess.run([accuracy, loss], feed_dict = {phase_train:False}) sum_acc += val_acc sum_loss += val_loss avg_acc = sum_acc/iters avg_loss = sum_loss/iters print('Validation data\'s accuracy: %0.4f' %avg_acc) print('Validation data\'s loss: %0.4f'%avg_loss) if monitor is not None: monitor['val_loss'].append(avg_loss) monitor['val_acc'].append(avg_acc) #============================================================================== # Main #============================================================================== monitor = {'train_loss':[], 'val_loss':[], 'train_acc':[], 'val_acc':[]} for i in range(n_epoch): a = time.time() # train_val(train_iters) train(train_iters, monitor) val(val_iters, monitor) print('epoch %d is finished'%(i+1)) print('This epoch cost: %f'%(time.time() - a)) # if i > 160: # val(val_iters) #val(val_iters) saver.save(sess=sess, save_path='logs/resnet6_2/model' , global_step=global_step) f = open('resnet6_2.pkl', 'wb') pickle.dump(monitor, f) f.close()