FCN-KITTI.py

# fine-tuning on KITTI

import tensorflow as tf
import numpy as np
import os
import scipy.io
import glob
import random
import BatchDatsetReader as dataset
import scipy.misc as misc
from six.moves import cPickle as pickle

# reset the graph
tf.reset_default_graph()

# reset tf.flags.FLAGS
import argparse
tf.reset_default_graph()
tf.flags.FLAGS = tf.python.platform.flags._FlagValues()
tf.flags._global_parser = argparse.ArgumentParser()

# set tf.flags.FLAGS
FLAGS = tf.flags.FLAGS
tf.flags.DEFINE_integer("batch_size","1","batch size for training")
tf.flags.DEFINE_string("logs_dir","logs/KITTI","path to logs directory")
tf.flags.DEFINE_string("data_dir","Data_zoo/KITTI/","path to dataset")
tf.flags.DEFINE_string("pickle_name","KITTI.pickle","pickle file of the data")
tf.flags.DEFINE_float("learning_rate","1e-4","learning rate for the optimizier")
tf.flags.DEFINE_string("model_dir","Model_zoo/","path to vgg model mat")
tf.flags.DEFINE_bool("debug","True","Debug model: True/False")
tf.flags.DEFINE_string("mode","train","Mode: train/ valid")
tf.flags.DEFINE_integer("max_iters","100001","max training iterations of batches")
tf.flags.DEFINE_integer("num_classes","11","mit_sceneparsing with (150+1) classes")
tf.flags.DEFINE_string("model_weights","http://www.vlfeat.org/matconvnet/models/beta16/imagenet-vgg-verydeep-19.mat","pretrained weights of the CNN in use")
tf.flags.DEFINE_string("full_model","full_model/","trained parameters of the whole network")
tf.flags.DEFINE_string("full_model_file","100000_model.ckpt","pretrained parameters of the whole network")
tf.flags.DEFINE_bool("load","True","load in pretrained parameters")
tf.flags.DEFINE_string("name","KITTI","dataset name")

# check if the CNN weights folder exist
if not os.path.exists(FLAGS.model_dir):
    os.makedirs(FLAGS.model_dir)
    
# check if the CNN weights file exist
weights_file = os.path.join(FLAGS.model_dir,FLAGS.model_weights.split('/')[-1])
if not os.path.exists(weights_file):
    print("\ndownloading "+weights_file+" ...")
    os.system("wget "+FLAGS.model_weights+" -P "+FLAGS.model_dir)
    print("download finished!\n")
else:
    print("\n"+weights_file+" has already been downloaded.\n")
    
# load the weights file
print("\nloading pretrained weights from: "+weights_file)
pretrain_weights = scipy.io.loadmat(weights_file)
print("loading finished!\n")

# the mean RGB
mean = pretrain_weights['normalization'][0][0][0] # shape(224,224,3)
mean_pixel = np.mean(mean,axis=(0,1)) # average on (height,width) to compute the mean RGB   

# the weights and biases
weights_biases = np.squeeze(pretrain_weights['layers'])

# network input data
dropout_prob = tf.placeholder(tf.float32,name="dropout_probability")
images = tf.placeholder(tf.float32,shape=[None,None,None,3],name="input_images")
annotations = tf.placeholder(tf.int32,shape=[None,None,None,1],name="input_annotations")

# subtract the mean image
processed_image = images - mean_pixel

# construct the semantic_seg network
with tf.variable_scope("semantic_seg"):
    # convs of the vgg net
    net = {}
    layers = [
        'conv1_1','relu1_1','conv1_2','relu1_2','pool1',
        'conv2_1','relu2_1','conv2_2','relu2_2','pool2',
        'conv3_1','relu3_1','conv3_2','relu3_2','conv3_3','relu3_3','conv3_4','relu3_4','pool3',
        'conv4_1','relu4_1','conv4_2','relu4_2','conv4_3','relu4_3','conv4_4','relu4_4','pool4',
        'conv5_1','relu5_1','conv5_2','relu5_2','conv5_3' #,'relu5_3','conv5_4','relu5_4','pool5'
    ]
    current = processed_image

    # sanity check
    print("processed_image: {}".format(processed_image.get_shape()))

    for i,name in enumerate(layers):
        type = name[:4]
        if type == 'conv':
            # matconvnet weights: (width, height, in_channels, out_channels)
            # tensorflow weights: (height, width, in_channels, out_channels)
            weights, biases = weights_biases[i][0][0][0][0]

            weights = np.transpose(weights,(1,0,2,3)) 
            biases = np.squeeze(biases)
            
            init = tf.constant_initializer(weights,dtype=tf.float32)
            weights = tf.get_variable(initializer=init,shape=weights.shape,name=name+"_w")
            
            init = tf.constant_initializer(biases,dtype=tf.float32)
            biases = tf.get_variable(initializer=init,shape=biases.shape,name=name+"_b")
            
            current = tf.nn.conv2d(current,weights,strides=[1,1,1,1],padding="SAME")
            current = tf.nn.bias_add(current,biases,name=name)

            # sanity check
            print("{}: {}".format(name,current.get_shape()))
        elif type == 'relu':
            current = tf.nn.relu(current,name=name)
            if FLAGS.debug:
                tf.histogram_summary(current.op.name+"/activation",current)
                tf.scalar_summary(current.op.name+"/sparsity",tf.nn.zero_fraction(current))
            # sanity check
            print("{}: {}".format(name,current.get_shape()))        
        elif type == 'pool':
            if name == 'pool5':
                current = tf.nn.max_pool(current,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME",name=name)
            else:
                current = tf.nn.avg_pool(current,ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME",name=name)
            # sanity check
            print("{}: {}".format(name,current.get_shape()))
        net[name] = current
             
    net['pool5'] = tf.nn.max_pool(net['conv5_3'],ksize=[1,2,2,1],strides=[1,2,2,1],padding="SAME",name=name)
    # sanity check
    print("pool5: {}".format(net['pool5'].get_shape()))
    
     # fcn6
    init = tf.truncated_normal(shape=[7,7,512,4096],stddev=0.02)
    fcn6_w = tf.get_variable(initializer=init,name="fcn6_w")

    init = tf.constant(0.0,shape=[4096])
    fcn6_b = tf.get_variable(initializer=init,name="fcn6_b")

    fcn6 = tf.nn.conv2d(net['pool5'],fcn6_w,strides=[1,1,1,1],padding="SAME")
    fcn6 = tf.nn.bias_add(fcn6,fcn6_b,name="fcn6")

    relu6 = tf.nn.relu(fcn6,name="relu6")
    if FLAGS.debug:
        tf.histogram_summary("relu6/activation", relu6, collections=None, name=None)
        tf.scalar_summary("relu6/sparsity", tf.nn.zero_fraction(relu6), collections=None, name=None)
    dropout6 = tf.nn.dropout(relu6, keep_prob=dropout_prob, noise_shape=None, seed=None, name="dropout6")
    # sanity check
    print("dropout6: {}".format(dropout6.get_shape()))

     # fcn7
    init = tf.truncated_normal(shape=[1,1,4096,4096],stddev=0.02)
    fcn7_w = tf.get_variable(initializer=init,name="fcn7_w")

    init = tf.constant(0.0,shape=[4096])
    fcn7_b = tf.get_variable(initializer=init,name="fcn7_b")

    fcn7 = tf.nn.conv2d(dropout6, fcn7_w, strides=[1,1,1,1], padding="SAME", use_cudnn_on_gpu=None, data_format=None, name=None)
    fcn7 = tf.nn.bias_add(fcn7, fcn7_b, data_format=None, name="fcn7")

    relu7 = tf.nn.relu(fcn7,name="relu7")
    if FLAGS.debug:
        tf.histogram_summary("relu7/activation", relu7, collections=None, name=None)
        tf.scalar_summary("relu7/sparsity", tf.nn.zero_fraction(relu7), collections=None, name=None)
    dropout7 = tf.nn.dropout(relu7, keep_prob=dropout_prob, noise_shape=None, seed=None, name="dropout7")
    # sanity check
    print("dropout7: {}".format(dropout7.get_shape()))

    # fcn8
    init = tf.truncated_normal(shape=[1,1,4096,FLAGS.num_classes],stddev=0.02)
    fcn8_w = tf.get_variable(initializer=init,name="fcn8_w")

    init = tf.constant(0.0,shape=[FLAGS.num_classes])
    fcn8_b = tf.get_variable(initializer=init,name="fcn8_b")

    fcn8 = tf.nn.conv2d(dropout7, fcn8_w, strides=[1,1,1,1], padding="SAME", use_cudnn_on_gpu=None, data_format=None, name=None)
    fcn8 = tf.nn.bias_add(fcn8, fcn8_b, data_format=None, name="fcn8")
    # sanity check
    print("fcn8: {}".format(fcn8.get_shape()))

    # deconv1 + net['pool4']: x32 -> x16
    s = 2
    k = 2*s
    in_channel = FLAGS.num_classes
    out_channel = net['pool4'].get_shape()[3].value
    out_shape = tf.shape(net['pool4'])

    init = tf.truncated_normal(shape=[k,k,out_channel,in_channel],stddev=0.02)
    deconv1_w = tf.get_variable(initializer=init,name="deconv1_w")

    init = tf.constant(0.0,shape=[out_channel])
    deconv1_b = tf.get_variable(initializer=init,name="deconv1_b")

    # sanity check
    print("deconv1 output_shape: {}".format(net['pool4'].get_shape()))

    deconv1 = tf.nn.conv2d_transpose(fcn8, deconv1_w, output_shape=out_shape, strides=[1,s,s,1], padding='SAME', name=None)
    deconv1 = tf.nn.bias_add(deconv1, deconv1_b, data_format=None, name="deconv1")

    fuse1 = tf.add(deconv1, net['pool4'], name="fuse1")
    
    # deconv2 + net['pool3']: x16 -> x8
    s = 2
    k = 2*s
    in_channel = out_channel
    out_channel = net['pool3'].get_shape()[3].value
    out_shape = tf.shape(net['pool3'])

    init = tf.truncated_normal(shape=[k,k,out_channel,in_channel],stddev=0.02)
    deconv2_w = tf.get_variable(initializer=init,name="deconv2_w")

    init = tf.constant(0.0,shape=[out_channel])
    deconv2_b = tf.get_variable(initializer=init,name="deconv2_b")

    deconv2 = tf.nn.conv2d_transpose(fuse1, deconv2_w, output_shape=out_shape, strides=[1,s,s,1], padding='SAME', name=None)
    deconv2 = tf.nn.bias_add(deconv2, deconv2_b, data_format=None, name="deconv2")

    fuse2 = tf.add(deconv2,net['pool3'],name="fuse2")

    # deconv3: x8 -> image_size
    s = 8
    k = 2*s
    in_channel = out_channel
    out_channel = FLAGS.num_classes
    out_shape = tf.pack([tf.shape(processed_image)[0],tf.shape(processed_image)[1],tf.shape(processed_image)[2],out_channel])
            
    init = tf.truncated_normal(shape=[k,k,out_channel,in_channel],stddev=0.02)
    deconv3_w = tf.get_variable(initializer=init,name="deconv3_w")

    init = tf.constant(0.0,shape=[out_channel])
    deconv3_b = tf.get_variable(initializer=init,name="deconv3_b")

    deconv3 = tf.nn.conv2d_transpose(fuse2, deconv3_w, output_shape=out_shape, strides=[1,s,s,1], padding='SAME', name=None)
    deconv3 = tf.nn.bias_add(deconv3, deconv3_b, data_format=None, name="deconv3")

    # per-pixel prediction
    annotations_pred = tf.argmax(deconv3, dimension=3, name=None)
    annotations_pred = tf.expand_dims(annotations_pred, dim=3, name="prediction")

# log images, annotations, annotations_pred
tf.image_summary("images", images, max_images=1, collections=None, name=None)
tf.image_summary("annotations", tf.cast(annotations,tf.uint8), max_images=1, collections=None, name=None)
tf.image_summary("annotations_pred", tf.cast(annotations_pred,tf.uint8), max_images=1, collections=None, name=None)

# construct the loss
loss = tf.nn.sparse_softmax_cross_entropy_with_logits(deconv3, tf.squeeze(annotations, squeeze_dims=[3]))
loss = tf.reduce_mean(loss, reduction_indices=None, keep_dims=False, name="pixel-wise_cross-entropy_loss")

# log the loss
tf.scalar_summary("pixel-wise_cross-entropy_loss", loss, collections=None, name=None)

# log all the trainable variables
trainabel_vars = tf.trainable_variables()
if FLAGS.debug:
    for var in trainabel_vars:
        tf.histogram_summary(var.op.name+"/values", var, collections=None, name=None)
        tf.add_to_collection("sum(t ** 2) / 2 of all trainable_vars", tf.nn.l2_loss(var))
        
# construct the optimizier
optimizer = tf.train.AdamOptimizer(FLAGS.learning_rate)
gradients = optimizer.compute_gradients(loss,trainabel_vars)
if FLAGS.debug:
    # log the gradients
    for grad, var in gradients:
        tf.histogram_summary(var.op.name+"/gradients", grad, collections=None, name=None)
train_op = optimizer.apply_gradients(gradients)

# initialize the variables
print("\nInitializing the variables ...\n")
sess = tf.InteractiveSession()
tf.initialize_all_variables().run()

# set up the saver
print("\nSetting up the Saver ...\n")
saver = tf.train.Saver()
all_vars = tf.trainable_variables()
saver_KITTI = tf.train.Saver({"semantic_seg/conv1_1_w":all_vars[0],"semantic_seg/conv1_1_b":all_vars[1],"semantic_seg/conv1_2_w":all_vars[2],"semantic_seg/conv1_2_b":all_vars[3],
                              "semantic_seg/conv2_1_w":all_vars[4],"semantic_seg/conv2_1_b":all_vars[5],"semantic_seg/conv2_2_w":all_vars[6],"semantic_seg/conv2_2_b":all_vars[7],
                              "semantic_seg/conv3_1_w":all_vars[8],"semantic_seg/conv3_1_b":all_vars[9],"semantic_seg/conv3_2_w":all_vars[10],"semantic_seg/conv3_2_b":all_vars[11],"semantic_seg/conv3_3_w":all_vars[12],"semantic_seg/conv3_3_b":all_vars[13],"semantic_seg/conv3_4_w":all_vars[14],"semantic_seg/conv3_4_b":all_vars[15],
                              "semantic_seg/conv4_1_w":all_vars[16],"semantic_seg/conv4_1_b":all_vars[17],"semantic_seg/conv4_2_w":all_vars[18],"semantic_seg/conv4_2_b":all_vars[19],"semantic_seg/conv4_3_w":all_vars[20],"semantic_seg/conv4_3_b":all_vars[21],"semantic_seg/conv4_4_w":all_vars[22],"semantic_seg/conv4_4_b":all_vars[23],
                              "semantic_seg/conv5_1_w":all_vars[24],"semantic_seg/conv5_1_b":all_vars[25],"semantic_seg/conv5_2_w":all_vars[26],"semantic_seg/conv5_2_b":all_vars[27],"semantic_seg/conv5_3_w":all_vars[28],"semantic_seg/conv5_3_b":all_vars[29],
                              "semantic_seg/fcn6_w":all_vars[30],"semantic_seg/fcn6_b":all_vars[31],"semantic_seg/fcn7_w":all_vars[32],"semantic_seg/fcn7_b":all_vars[33],
                              "semantic_seg/deconv2_w":all_vars[38],"semantic_seg/deconv2_b":all_vars[39]})
if FLAGS.load:
    print("\nLoading pretrain parameters of the whole network ...\n")
    saver_KITTI.restore(sess, os.path.join(FLAGS.full_model,FLAGS.full_model_file))
    
# set the summary writer
print("\nSetting the summary writers ...\n")
summary_op = tf.merge_all_summaries()
if not os.path.exists(FLAGS.logs_dir):
    os.system("mkdir "+FLAGS.logs_dir)
if FLAGS.mode == 'train':
    if os.path.exists(FLAGS.logs_dir+"/train"):
        os.system("rm -r "+FLAGS.logs_dir+"/train")
    if os.path.exists(FLAGS.logs_dir+"/valid"):
        os.system("rm -r "+FLAGS.logs_dir+"/valid")
    train_writer = tf.train.SummaryWriter(FLAGS.logs_dir+"/train",sess.graph)
    valid_writer = tf.train.SummaryWriter(FLAGS.logs_dir+"/valid")
elif FLAGS.mode == 'valid':
    if os.path.exists(FLAGS.logs_dir+"/complete_valid"):
        os.system("rm -r "+FLAGS.logs_dir+"/complete_valid")
    valid_writer = tf.train.SummaryWriter(FLAGS.logs_dir+"/complete_valid")
    
# read data_records from *.pickle
print("\nReading in and reprocessing all images ...\n")
# check if FLAGS.data_dir folder exist
if not os.path.exists(FLAGS.data_dir):
    os.makedirs(FLAGS.data_dir)
# check if the *.pickle file exist
pickle_file = os.path.join(FLAGS.data_dir,FLAGS.pickle_name)
# load data_records from *.pickle
with open(pickle_file,'rb') as f:
    pickle_records = pickle.load(f)
    train_records = pickle_records['trn']
    valid_records = pickle_records['test']
    del pickle_records

# initialize the data reader
print("Initializing the data reader...")
reader_options = {'different_size':True}
if FLAGS.mode == 'train':
    train_reader = dataset.BatchDatset(train_records,reader_options)
valid_reader = dataset.BatchDatset(valid_records,reader_options)

# check if FLAGS.full_model exist
if not os.path.exists(os.path.join(FLAGS.full_model,FLAGS.name)):
    os.makedirs(os.path.join(FLAGS.full_model,FLAGS.name))

# start training/ validation
print("\nStarting training/ validation...\n")
if FLAGS.mode == 'train':
    for itr in xrange(FLAGS.max_iters):
        # read next batch
        train_images, train_annotations = train_reader.next_image(FLAGS.batch_size)
        feed_dict = {images:train_images,annotations:train_annotations,dropout_prob:0.85}
        # training
        sess.run(train_op,feed_dict=feed_dict)
        # log training info
        if itr % 10 == 0:
            train_loss, train_summary = sess.run([loss,summary_op],feed_dict=feed_dict)
            train_writer.add_summary(train_summary,itr)
            print("Step: %d, train_loss: %f"%(itr,train_loss))
        # log valid info
        if itr % 100 == 0:
            valid_images, valid_annotations = valid_reader.next_image(FLAGS.batch_size)
            feed_dict = {images:valid_images,annotations:valid_annotations,dropout_prob:1.0}
            valid_loss, valid_summary = sess.run([loss,summary_op],feed_dict=feed_dict)
            valid_writer.add_summary(valid_summary,itr)
            print("==============================")
            print("Step: %d, valid_loss: %f"%(itr,valid_loss))
            print("==============================")
        # save snapshot
        if itr % 500 == 0:
            snapshot_name = os.path.join(os.path.join(FLAGS.full_model,FLAGS.name),str(itr)+"_model.ckpt")
            saver.save(sess,snapshot_name)
elif FLAGS.mode == 'valid':
    # quantitative results
    valid_images,valid_annotations=valid_reader.next_image(FLAGS.batch_size)
    feed_dict = {images:valid_images[:20],annotations:valid_annotations[:20],dropout_prob:1.0}
    valid_loss,valid_summary = sess.run([loss,summary_op],feed_dict=feed_dict)
    valid_writer.add_summary(valid_summary,FLAGS.max_iters)
    print("==============================")
    print("Step: %d, valid_loss: %f"%(FLAGS.max_iters,valid_loss))
    print("==============================")
    # qualitative results
    valid_images,valid_annotations=valid_reader.next_image(FLAGS.batch_size)
    feed_dict = {images:valid_images,annotations:valid_annotations,dropout_prob:1.0}
    annotations_pred_results = sess.run(annotations_pred,feed_dict=feed_dict)
    
    valid_annotations = np.squeeze(valid_annotations,axis=3)
    annotations_pred_results = np.squeeze(annotations_pred_results,axis=3)
    
    for n in xrange(FLAGS.batch_size):
        print("Saving %d-th valid tuples for qualitative comparisons..."%(n))
        misc.imsave(FLAGS.logs_dir+"/complete_valid/"+str(n)+"_image.png",valid_images[n].astype(np.uint8))
        misc.imsave(FLAGS.logs_dir+"/complete_valid/"+str(n)+"_annotation.png",valid_annotations[n].astype(np.uint8))
        misc.imsave(FLAGS.logs_dir+"/complete_valid/"+str(n)+"_prediction.png",annotations_pred_results[n].astype(np.uint8))
    print("saving finished!!!")