densenet tensorflow 中文汉字手写识别

densenet 中文汉字手写识别，代码如下：

100

101

102

103

104

105

106

107

108

109

110

111

112

113

114

115

116

117

118

119

120

121

122

123

124

125

126

127

128

129

130

131

132

133

134

135

136

137

138

139

140

141

142

143

144

145

146

147

148

149

150

151

152

153

154

155

156

157

158

159

160

161

162

163

164

165

166

167

168

169

170

171

172

173

174

175

176

177

178

179

180

181

182

183

184

185

186

187

188

189

190

191

192

193

194

195

196

197

198

199

200

201

202

203

204

205

206

207

208

209

210

211

212

213

214

215

216

217

218

219

220

221

222

223

224

225

226

227

228

229

230

231

232

233

234

235

236

237

238

239

240

241

242

243

244

245

246

247

248

249

250

251

252

253

254

255

256

257

258

259

260

261

262

263

264

265

266

267

268

269

270

271

272

273

274

275

276

277

278

279

280

281

282

283

284

285

286

287

288

289

290

291

292

293

294

295

296

297

298

299

300

301

302

303

304

305

306

307

308

309

310

311

312

313

314

315

316

317

318

319

320

321

322

323

324

325

326

327

328

329

330

331

332

333

334

335

336

337

338

339

340

341

342

343

344

345

346

347

348

349

350

351

352

353

354

355

356

357

358

359

360

361

362

363

364

365

366

367

368

369

370

371

372

373

374

375

376

377

378

379

380

381

382

383

384

385

386

387

388

389

390

391

392

393

394

395

396

397

398

399

400

401

402

403

404

405

406

407

408

409

410

411

412

413

414

415

416

417

418

419

420

421

422

423

424

425

426

427

428

429

430

431

432

433

434

435

436

437

438

439

440

441

442

443

444

445

446

447

448

449

450

451

452

453

454

455

456

457

458

459

460

461

462

463

464

465

466

467

468

469

470

471

472

473

474

475

476

477

478

479

480

481

482

483

484

485

486

487

488

489

import tensorflow as tf
import os
import random
import math
import tensorflow.contrib.slim as slim
import time
import logging
import numpy as np
import pickle
from PIL import Image
   
import tensorflow as tf
#from tflearn.layers.conv import global_avg_pool
from tensorflow.contrib.layers import batch_norm, flatten
from tensorflow.contrib.framework import arg_scope
import numpy as np
 
 
# Hyperparameter
growth_k = 12
nb_block = 2 # how many (dense block + Transition Layer) ?
init_learning_rate = 1e-4
epsilon = 1e-8 # AdamOptimizer epsilon
dropout_rate = 0.2
 
# Momentum Optimizer will use
nesterov_momentum = 0.9
weight_decay = 1e-4
 
# Label & batch_size
class_num = 3755
batch_size = 128
 
total_epochs = 50
 
 
def conv_layer(input, filter, kernel, stride=1, layer_name="conv"):
    with tf.name_scope(layer_name):
        network = tf.layers.conv2d(inputs=input, filters=filter, kernel_size=kernel, strides=stride, padding='SAME')
        return network
 
def Global_Average_Pooling(x, stride=1):
    #It is global average pooling without tflearn
    width = np.shape(x)[1]
    height = np.shape(x)[2]
    pool_size = [width, height]
    return tf.layers.average_pooling2d(inputs=x, pool_size=pool_size, strides=stride) # The stride value does not matter
    """
    return global_avg_pool(x, name='Global_avg_pooling')
    # But maybe you need to install h5py and curses or not
    """
 
 
def Batch_Normalization(x, training, scope):
    with arg_scope([batch_norm],
                   scope=scope,
                   updates_collections=None,
                   decay=0.9,
                   center=True,
                   scale=True,
                   zero_debias_moving_mean=True) :
        return tf.cond(training,
                       lambda : batch_norm(inputs=x, is_training=training, reuse=None),
                       lambda : batch_norm(inputs=x, is_training=training, reuse=True))
 
def Drop_out(x, rate, training) :
    return tf.layers.dropout(inputs=x, rate=rate, training=training)
 
def Relu(x):
    return tf.nn.relu(x)
 
def Average_pooling(x, pool_size=[2,2], stride=2, padding='VALID'):
    return tf.layers.average_pooling2d(inputs=x, pool_size=pool_size, strides=stride, padding=padding)
 
 
def Max_Pooling(x, pool_size=[3,3], stride=2, padding='VALID'):
    return tf.layers.max_pooling2d(inputs=x, pool_size=pool_size, strides=stride, padding=padding)
 
def Concatenation(layers) :
    return tf.concat(layers, axis=3)
 
def Linear(x) :
    return tf.layers.dense(inputs=x, units=class_num, name='linear')
 
 
 
class DenseNet():
    def __init__(self, x, nb_blocks, filters, training):
        self.nb_blocks = nb_blocks
        self.filters = filters
        self.training = training
        self.model = self.Dense_net(x)
 
 
    def bottleneck_layer(self, x, scope):
        # print(x)
        with tf.name_scope(scope):
            x = Batch_Normalization(x, training=self.training, scope=scope+'_batch1')
            x = Relu(x)
            x = conv_layer(x, filter=4 * self.filters, kernel=[1,1], layer_name=scope+'_conv1')
            x = Drop_out(x, rate=dropout_rate, training=self.training)
 
            x = Batch_Normalization(x, training=self.training, scope=scope+'_batch2')
            x = Relu(x)
            x = conv_layer(x, filter=self.filters, kernel=[3,3], layer_name=scope+'_conv2')
            x = Drop_out(x, rate=dropout_rate, training=self.training)
 
            # print(x)
 
            return x
 
    def transition_layer(self, x, scope):
        with tf.name_scope(scope):
            x = Batch_Normalization(x, training=self.training, scope=scope+'_batch1')
            x = Relu(x)
            x = conv_layer(x, filter=self.filters, kernel=[1,1], layer_name=scope+'_conv1')
            x = Drop_out(x, rate=dropout_rate, training=self.training)
            x = Average_pooling(x, pool_size=[2,2], stride=2)
 
            return x
 
    def dense_block(self, input_x, nb_layers, layer_name):
        with tf.name_scope(layer_name):
            layers_concat = list()
            layers_concat.append(input_x)
 
            x = self.bottleneck_layer(input_x, scope=layer_name + '_bottleN_' + str(0))
 
            layers_concat.append(x)
 
            for i in range(nb_layers - 1):
                x = Concatenation(layers_concat)
                x = self.bottleneck_layer(x, scope=layer_name + '_bottleN_' + str(i + 1))
                layers_concat.append(x)
 
            x = Concatenation(layers_concat)
 
            return x
 
    def Dense_net(self, input_x):
        x = conv_layer(input_x, filter=2 * self.filters, kernel=[7,7], stride=2, layer_name='conv0')
        x = Max_Pooling(x, pool_size=[3,3], stride=2)
 
 
 
        for i in range(self.nb_blocks) :
            # 6 -> 12 -> 48
            x = self.dense_block(input_x=x, nb_layers=4, layer_name='dense_'+str(i))
            x = self.transition_layer(x, scope='trans_'+str(i))
 
 
        """
        x = self.dense_block(input_x=x, nb_layers=6, layer_name='dense_1')
        x = self.transition_layer(x, scope='trans_1')
 
        x = self.dense_block(input_x=x, nb_layers=12, layer_name='dense_2')
        x = self.transition_layer(x, scope='trans_2')
 
        x = self.dense_block(input_x=x, nb_layers=48, layer_name='dense_3')
        x = self.transition_layer(x, scope='trans_3')
        """
 
        x = self.dense_block(input_x=x, nb_layers=32, layer_name='dense_final')
 
        # 100 Layer
        x = Batch_Normalization(x, training=self.training, scope='linear_batch')
        x = Relu(x)
        x = Global_Average_Pooling(x)
        x = flatten(x)
        x = Linear(x)
 
 
        # x = tf.reshape(x, [-1, 10])
        return x
 
   
def build_graph(top_k):
    # with tf.device('/cpu:0'):
    # keep_prob = tf.placeholder(dtype=tf.float32, shape=[], name='keep_prob')
    images = tf.placeholder(dtype=tf.float32, shape=[None, 64, 64, 1], name='image_batch')
    # label = tf.placeholder(tf.float32, shape=[None, 10])
    labels = tf.placeholder(dtype=tf.int64, shape=[None], name='label_batch')
    training_flag = tf.placeholder(tf.bool)
    logits = DenseNet(x=images, nb_blocks=nb_block, filters=growth_k, training=training_flag).model
    loss = tf.reduce_mean(tf.nn.sparse_softmax_cross_entropy_with_logits(logits=logits, labels=labels))
    # loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits))
 
    """
    l2_loss = tf.add_n([tf.nn.l2_loss(var) for var in tf.trainable_variables()])
    optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate, momentum=nesterov_momentum, use_nesterov=True)
    train = optimizer.minimize(cost + l2_loss * weight_decay)
    In paper, use MomentumOptimizer
    init_learning_rate = 0.1
    but, I'll use AdamOptimizer
    """
 
    global_step = tf.get_variable("step", [], initializer=tf.constant_initializer(0.0), trainable=False)
    rate = tf.train.exponential_decay(2e-4, global_step, decay_steps=2000, decay_rate=0.97, staircase=True)
    optimizer = tf.train.AdamOptimizer(learning_rate=rate, epsilon=epsilon)
    train_op = optimizer.minimize(loss, global_step=global_step)
 
    accuracy = tf.reduce_mean(tf.cast(tf.equal(tf.argmax(logits, 1), labels), tf.float32))
 
    probabilities = logits
    tf.summary.scalar('loss', loss)
    tf.summary.scalar('accuracy', accuracy)
    merged_summary_op = tf.summary.merge_all()
    predicted_val_top_k, predicted_index_top_k = tf.nn.top_k(probabilities, k=top_k)
    accuracy_in_top_k = tf.reduce_mean(tf.cast(tf.nn.in_top_k(probabilities, labels, top_k), tf.float32))
 
    return {'images': images,
            'labels': labels,
            'training_flag': training_flag,
            'top_k': top_k,
            'global_step': global_step,
            'train_op': train_op,
            'loss': loss,
            'accuracy': accuracy,
            'accuracy_top_k': accuracy_in_top_k,
            'merged_summary_op': merged_summary_op,
            'predicted_distribution': probabilities,
            'predicted_index_top_k': predicted_index_top_k,
            'predicted_val_top_k': predicted_val_top_k} 
 
logger = logging.getLogger('Training a chinese write char recognition')
logger.setLevel(logging.INFO)
# formatter = logging.Formatter('%(asctime)s - %(name)s - %(levelname)s - %(message)s')
ch = logging.StreamHandler()
ch.setLevel(logging.INFO)
logger.addHandler(ch)
  
run_mode = "train"
charset_size = class_num
max_steps = 122002
save_steps = 1000
cur_test_acc = 0
   
"""
# for online 3755 words training
checkpoint_dir = '/aiml/dfs/checkpoint_888/'
train_data_dir = '/aiml/data/train/'
test_data_dir = '/aiml/data/test/'
log_dir = '/aiml/dfs/'
"""
  
  
checkpoint_dir = './checkpoint_densenet/'
train_data_dir = './data/train/'
test_data_dir = './data/test/'
log_dir = './'
  
  
tf.app.flags.DEFINE_string('mode', run_mode, 'Running mode. One of {"train", "valid", "test"}')
tf.app.flags.DEFINE_boolean('random_flip_up_down', True, "Whether to random flip up down")
tf.app.flags.DEFINE_boolean('random_brightness', True, "whether to adjust brightness")
tf.app.flags.DEFINE_boolean('random_contrast', True, "whether to random constrast")
   
tf.app.flags.DEFINE_integer('charset_size', charset_size, "Choose the first `charset_size` character to conduct our experiment.")
tf.app.flags.DEFINE_integer('image_size', 64, "Needs to provide same value as in training.")
tf.app.flags.DEFINE_boolean('gray', True, "whether to change the rbg to gray")
tf.app.flags.DEFINE_integer('max_steps', max_steps, 'the max training steps ')
tf.app.flags.DEFINE_integer('eval_steps', 50, "the step num to eval")
tf.app.flags.DEFINE_integer('save_steps', save_steps, "the steps to save")
   
tf.app.flags.DEFINE_string('checkpoint_dir', checkpoint_dir, 'the checkpoint dir')
tf.app.flags.DEFINE_string('train_data_dir', train_data_dir, 'the train dataset dir')
tf.app.flags.DEFINE_string('test_data_dir', test_data_dir, 'the test dataset dir')
tf.app.flags.DEFINE_string('log_dir', log_dir, 'the logging dir')
   
##############################
# resume training
tf.app.flags.DEFINE_boolean('restore', True, 'whether to restore from checkpoint')
##############################
  
tf.app.flags.DEFINE_boolean('epoch', 10, 'Number of epoches')
tf.app.flags.DEFINE_boolean('batch_size', 128, 'Validation batch size')
FLAGS = tf.app.flags.FLAGS
   
   
class DataIterator:
    def __init__(self, data_dir):
        # Set FLAGS.charset_size to a small value if available computation power is limited.
        truncate_path = data_dir + ('%05d' % FLAGS.charset_size)
        print(truncate_path)
        self.image_names = []
        for root, sub_folder, file_list in os.walk(data_dir):
            if root < truncate_path:
                self.image_names += [os.path.join(root, file_path) for file_path in file_list]
        random.shuffle(self.image_names)
        self.labels = [int(file_name[len(data_dir):].split(os.sep)[0]) for file_name in self.image_names]
   
    @property
    def size(self):
        return len(self.labels)
   
    @staticmethod
    def data_augmentation(images):
        if FLAGS.random_flip_up_down:
            # images = tf.image.random_flip_up_down(images)
            images = tf.contrib.image.rotate(images, random.randint(0, 15) * math.pi / 180, interpolation='BILINEAR')
        if FLAGS.random_brightness:
            images = tf.image.random_brightness(images, max_delta=0.3)
        if FLAGS.random_contrast:
            images = tf.image.random_contrast(images, 0.8, 1.2)
        return images
   
    def input_pipeline(self, batch_size, num_epochs=None, aug=False):
        images_tensor = tf.convert_to_tensor(self.image_names, dtype=tf.string)
        labels_tensor = tf.convert_to_tensor(self.labels, dtype=tf.int64)
        input_queue = tf.train.slice_input_producer([images_tensor, labels_tensor], num_epochs=num_epochs)
   
        labels = input_queue[1]
        images_content = tf.read_file(input_queue[0])
        images = tf.image.convert_image_dtype(tf.image.decode_png(images_content, channels=1), tf.float32)
        if aug:
            images = self.data_augmentation(images)
        new_size = tf.constant([FLAGS.image_size, FLAGS.image_size], dtype=tf.int32)
        images = tf.image.resize_images(images, new_size)
        image_batch, label_batch = tf.train.shuffle_batch([images, labels], batch_size=batch_size, capacity=50000,
                                                          min_after_dequeue=10000)
        return image_batch, label_batch
   
   
def train():
    print('Begin training')
    train_feeder = DataIterator(FLAGS.train_data_dir)
    test_feeder = DataIterator(FLAGS.test_data_dir)
    with tf.Session() as sess:
        train_images, train_labels = train_feeder.input_pipeline(batch_size=FLAGS.batch_size, aug=True)
        test_images, test_labels = test_feeder.input_pipeline(batch_size=FLAGS.batch_size)
        graph = build_graph(top_k=1)
        sess.run(tf.global_variables_initializer())
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(sess=sess, coord=coord)
        saver = tf.train.Saver()
   
        train_writer = tf.summary.FileWriter(FLAGS.log_dir + '/train', sess.graph)
        test_writer = tf.summary.FileWriter(FLAGS.log_dir + '/val')
        start_step = 0
        if FLAGS.restore:
            ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
            if ckpt:
                saver.restore(sess, ckpt)
                print("restore from the checkpoint {0}".format(ckpt))
                start_step += int(ckpt.split('-')[-1])
   
        logger.info(':::Training Start:::')
        try:
            while not coord.should_stop():
                start_time = time.time()
                train_images_batch, train_labels_batch = sess.run([train_images, train_labels])
                feed_dict = {graph['images']: train_images_batch,
                             graph['labels']: train_labels_batch,
                             graph['training_flag']: True}
                _, loss_val, train_summary, step = sess.run(
                    [graph['train_op'], graph['loss'], graph['merged_summary_op'], graph['global_step']],
                    feed_dict=feed_dict)
                train_writer.add_summary(train_summary, step)
                end_time = time.time()
                logger.info("the step {0} takes {1} loss {2}".format(step, end_time - start_time, loss_val))
                if step > FLAGS.max_steps:
                    break
                accuracy_test = 0
                if step % FLAGS.eval_steps == 1:
                    test_images_batch, test_labels_batch = sess.run([test_images, test_labels])
                    feed_dict = {graph['images']: test_images_batch,
                                 graph['labels']: test_labels_batch,
                                 graph['training_flag']: False}
                    accuracy_test, test_summary = sess.run(
                        [graph['accuracy'], graph['merged_summary_op']],
                        feed_dict=feed_dict)
                    test_writer.add_summary(test_summary, step)
                    logger.info('===============Eval a batch=======================')
                    logger.info('the step {0} test accuracy: {1}'
                                .format(step, accuracy_test))
                    logger.info('===============Eval a batch=======================')
                if step % FLAGS.save_steps == 1: 
                    logger.info('Save the ckpt of {0}'.format(step))
                    saver.save(sess, os.path.join(FLAGS.checkpoint_dir, 'my-model'),
                               global_step=graph['global_step'])
                global cur_test_acc
                cur_test_acc = accuracy_test
        except tf.errors.OutOfRangeError:
            logger.info('==================Train Finished================')
            saver.save(sess, os.path.join(FLAGS.checkpoint_dir, 'my-model'), global_step=graph['global_step'])
        finally:
            coord.request_stop()
        coord.join(threads)
   
   
def validation():
    print('validation')
    test_feeder = DataIterator(FLAGS.test_data_dir)
   
    final_predict_val = []
    final_predict_index = []
    groundtruth = []
   
    with tf.Session() as sess:
        test_images, test_labels = test_feeder.input_pipeline(batch_size=FLAGS.batch_size, num_epochs=1)
        graph = build_graph(top_k=3)
   
        sess.run(tf.global_variables_initializer())
        sess.run(tf.local_variables_initializer())  # initialize test_feeder's inside state
   
        coord = tf.train.Coordinator()
        threads = tf.train.start_queue_runners(sess=sess, coord=coord)
   
        saver = tf.train.Saver()
        ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
        if ckpt:
            saver.restore(sess, ckpt)
            print("restore from the checkpoint {0}".format(ckpt))
   
        print(':::Start validation:::')
        try:
            i = 0
            acc_top_1, acc_top_k = 0.0, 0.0
            while not coord.should_stop():
                i += 1
                start_time = time.time()
                test_images_batch, test_labels_batch = sess.run([test_images, test_labels])
                feed_dict = {graph['images']: test_images_batch,
                             graph['labels']: test_labels_batch,
                             graph['training_flag']: False}
                batch_labels, probs, indices, acc_1, acc_k = sess.run([graph['labels'],
                                                                       graph['predicted_val_top_k'],
                                                                       graph['predicted_index_top_k'],
                                                                       graph['accuracy'],
                                                                       graph['accuracy_top_k']], feed_dict=feed_dict)
                final_predict_val += probs.tolist()
                final_predict_index += indices.tolist()
                groundtruth += batch_labels.tolist()
                acc_top_1 += acc_1
                acc_top_k += acc_k
                end_time = time.time()
                logger.info("the batch {0} takes {1} seconds, accuracy = {2}(top_1) {3}(top_k)"
                            .format(i, end_time - start_time, acc_1, acc_k))
   
        except tf.errors.OutOfRangeError:
            logger.info('==================Validation Finished================')
            acc_top_1 = acc_top_1 * FLAGS.batch_size / test_feeder.size
            acc_top_k = acc_top_k * FLAGS.batch_size / test_feeder.size
            logger.info('top 1 accuracy {0} top k accuracy {1}'.format(acc_top_1, acc_top_k))
        finally:
            coord.request_stop()
        coord.join(threads)
    return {'prob': final_predict_val, 'indices': final_predict_index, 'groundtruth': groundtruth}
   
   
def inference(image):
    print('inference')
    temp_image = Image.open(image).convert('L')
    temp_image = temp_image.resize((FLAGS.image_size, FLAGS.image_size), Image.ANTIALIAS)
    temp_image = np.asarray(temp_image) / 255.0
    temp_image = temp_image.reshape([-1, 64, 64, 1])
    with tf.Session() as sess:
        logger.info('========start inference============')
        # images = tf.placeholder(dtype=tf.float32, shape=[None, 64, 64, 1])
        # Pass a shadow label 0. This label will not affect the computation graph.
        graph = build_graph(top_k=3)
        saver = tf.train.Saver()
        ckpt = tf.train.latest_checkpoint(FLAGS.checkpoint_dir)
        if ckpt:
            saver.restore(sess, ckpt)
        predict_val, predict_index = sess.run([graph['predicted_val_top_k'], graph['predicted_index_top_k']],
                                              feed_dict={graph['images']: temp_image, graph['training_flag']: False})
    return predict_val, predict_index
   
   
def main(_):
    print(FLAGS.mode)
    if FLAGS.mode == "train":
        train()
    elif FLAGS.mode == 'validation':
        dct = validation()
        result_file = 'result.dict'
        logger.info('Write result into {0}'.format(result_file))
        with open(result_file, 'wb') as f:
            pickle.dump(dct, f)
        logger.info('Write file ends')
    elif FLAGS.mode == 'inference':
        image_path = './data/00098/102544.png'
        final_predict_val, final_predict_index = inference(image_path)
        logger.info('the result info label {0} predict index {1} predict_val {2}'.format(190, final_predict_index,
                                                                                         final_predict_val))
   
if __name__ == "__main__":
    tf.app.run()