我正在尝试在OpenCv或Tensorflow中实现分段项目，目前我在Tensorflow中的代码有一些问题。您遇到此错误了吗？怎么处理呢？我的训练图像和测试图像的宽度和高度相同，所以现在是问题所在。
我有以下代码：
import numpy as np
from scipy import misc
import tensorflow as tf
import shutil
import scipy.io as sio
import os
# --------------------------------------------------
# setup
# --------------------------------------------------
imsize = 360
nclasses = 3
nchannels = 1
ntrain = 60
nvalid = 10
ntest = 23
nsteps = 10#000
batchsize = 4
nextraconvs = 1
restore_variables = True#False
ks = 3
nout0 = 16
nout1 = 2*nout0
nout2 = 2*nout1
nout3 = 2*nout2
dsf0 = 2 # downsampling factor, first layer
dsf1 = 2 # downsampling factor, second layer
dsf2 = 2 # downsampling factor, third layer
# noutX = [nchannels, nout0, nout1]
# dsfX = [dsf0]
# noutX = [nchannels, nout0, nout1, nout2]
# dsfX = [dsf0, dsf1]
noutX = [nchannels, nout0, nout1, nout2, nout3]
dsfX = [dsf0, dsf1, dsf2]
nlayers = len(dsfX)
impath = '/Users/Rares/Workplace/DH_NBCB'
train_writer_path = '/Users/Rares/Workspace/TFLog/UNet/Train'
valid_writer_path = '/Users/Rares/Workspace/TFLog/UNet/Valid'
out_log_path = '/Users/Rares/Workspace/TFLog/UNet'
out_model_path = '/Users/Rares/Workspace/TFModel/UNet.ckpt'
out_pm_path = '/Users/Rares/Workspace/Scratch'
def concat3(lst):
return tf.concat(lst,3)
# --------------------------------------------------
# data
# --------------------------------------------------
Train = np.zeros((ntrain,imsize,imsize,nchannels))
Valid = np.zeros((nvalid,imsize,imsize,nchannels))
Test = np.zeros((ntest,imsize,imsize,nchannels))
LTrain = np.zeros((ntrain,imsize,imsize,nclasses))
LValid = np.zeros((nvalid,imsize,imsize,nclasses))
LTest = np.zeros((ntest,imsize,imsize,nclasses))
for isample in range(0, ntrain):
path = '%s/TrainImages/I%03d.tif' % (impath,isample+1)
im = misc.imread(path).astype(np.float32)/255
im = im-np.mean(im)
im = im/np.std(im)
Train[isample,:,:,0] = im
path = '%s/TrainLabels/L%03d.tif' % (impath,isample+1)
im = misc.imread(path).astype(np.float32)
for i in range(nclasses):
if nclasses == 2:
LTrain[isample,:,:,i] = (im == 2*i)
else:
LTrain[isample,:,:,i] = (im == i)
for isample in range(0, nvalid):
path = '%s/ValidImages/I%03d.tif' % (impath,isample+1)
im = misc.imread(path).astype(np.float32)/255
im = im-np.mean(im)
im = im/np.std(im)
Valid[isample,:,:,0] = im
path = '%s/ValidLabels/L%03d.tif' % (impath,isample+1)
im = misc.imread(path).astype(np.float32)
for i in range(nclasses):
if nclasses == 2:
LValid[isample,:,:,i] = (im == 2*i)
else:
LValid[isample,:,:,i] = (im == i)
for isample in range(0, ntest):
path = '%s/TestImages/I%03d.tif' % (impath,isample+1)
im = misc.imread(path).astype(np.float32)/255
im = im-np.mean(im)
im = im/np.std(im)
Test[isample,:,:,0] = im
path = '%s/TestLabels/L%03d.tif' % (impath,isample+1)
im = misc.imread(path).astype(np.float32)
for i in range(nclasses):
if nclasses == 2:
LTest[isample,:,:,i] = (im == 2*i)
else:
LTest[isample,:,:,i] = (im == i)
# --------------------------------------------------
# downsampling layer
# --------------------------------------------------
with tf.name_scope('placeholders'):
tf_data = tf.placeholder("float", shape=[None,imsize,imsize,nchannels],name='data')
tf_labels = tf.placeholder("float", shape=[None,imsize,imsize,nclasses],name='labels')
def down_samp_layer(data,index):
with tf.name_scope('ld%d' % index):
ldX_weights1 = tf.Variable(tf.truncated_normal([ks, ks, noutX[index], noutX[index+1]], stddev=0.1),name='kernel1')
ldX_weights_extra = []
for i in range(nextraconvs):
ldX_weights_extra.append(tf.Variable(tf.truncated_normal([ks, ks, noutX[index+1], noutX[index+1]], stddev=0.1),name='kernel_extra%d' % i))
c_00 = tf.nn.relu(tf.nn.conv2d(data, ldX_weights1, strides=[1, 1, 1, 1], padding='SAME'),name='conv')
for i in range(nextraconvs):
c_00 = tf.nn.relu(tf.nn.conv2d(c_00, ldX_weights_extra[i], strides=[1, 1, 1, 1], padding='SAME'),name='conv_extra%d' % i)
return tf.nn.max_pool(c_00, ksize=[1, dsfX[index], dsfX[index], 1], strides=[1, dsfX[index], dsfX[index], 1], padding='SAME',name='maxpool')
# --------------------------------------------------
# bottom layer
# --------------------------------------------------
with tf.name_scope('lb'):
lb_weights1 = tf.Variable(tf.truncated_normal([ks, ks, noutX[nlayers], noutX[nlayers+1]], stddev=0.1),name='kernel1')
def lb(hidden):
return tf.nn.relu(tf.nn.conv2d(hidden, lb_weights1, strides=[1, 1, 1, 1], padding='SAME'),name='conv')
# --------------------------------------------------
# downsampling
# --------------------------------------------------
with tf.name_scope('downsampling'):
dsX = []
dsX.append(tf_data)
for i in range(nlayers):
dsX.append(down_samp_layer(dsX[i],i))
b = lb(dsX[nlayers])
# --------------------------------------------------
# upsampling layer
# --------------------------------------------------
def up_samp_layer(data,index):
with tf.name_scope('lu%d' % index):
luX_weights1 = tf.Variable(tf.truncated_normal([ks, ks, noutX[index+1], noutX[index+2]], stddev=0.1),name='kernel1')
luX_weights2 = tf.Variable(tf.truncated_normal([ks, ks, noutX[index]+noutX[index+1], noutX[index+1]], stddev=0.1),name='kernel2')
luX_weights_extra = []
for i in range(nextraconvs):
luX_weights_extra.append(tf.Variable(tf.truncated_normal([ks, ks, noutX[index+1], noutX[index+1]], stddev=0.1),name='kernel2_extra%d' % i))
out_size = imsize
for i in range(index):
out_size /= dsfX[i]
deconv_shape = tf.stack([batchsize, 90, 90, 32])
output_shape = [batchsize,out_size,out_size,noutX[index+1]]
us = tf.nn.relu(tf.nn.conv2d_transpose(data, luX_weights1, output_shape=deconv_shape, strides=[1, dsfX[index], dsfX[index], 1], padding='SAME'),name='conv1')
cc = concat3([dsX[index],us])
cv = tf.nn.relu(tf.nn.conv2d(cc, luX_weights2, strides=[1, 1, 1, 1], padding='SAME'),name='conv2')
for i in range(nextraconvs):
cv = tf.nn.relu(tf.nn.conv2d(cv, luX_weights_extra[i], strides=[1, 1, 1, 1], padding='SAME'),name='conv2_extra%d' % i)
return cv
# --------------------------------------------------
# final (top) layer
# --------------------------------------------------
with tf.name_scope('lt'):
lt_weights1 = tf.Variable(tf.truncated_normal([1, 1, noutX[1], nclasses], stddev=0.1),name='kernel')
def lt(hidden):
return tf.nn.conv2d(hidden, lt_weights1, strides=[1, 1, 1, 1], padding='SAME',name='conv')
# --------------------------------------------------
# upsampling
# --------------------------------------------------
with tf.name_scope('upsampling'):
usX = []
usX.append(b)
for i in range(nlayers):
usX.append(up_samp_layer(usX[i],nlayers-1-i))
t = lt(usX[nlayers])
# --------------------------------------------------
# optimization
# --------------------------------------------------
with tf.name_scope('optim'):
sm = tf.nn.softmax(t,-1)
loss = tf.reduce_mean(-tf.reduce_sum(tf.multiply(tf_labels,tf.log(sm)),3))
opt_op = tf.train.MomentumOptimizer(1e-3,0.9).minimize(loss)
# opt_op = tf.train.GradientDescentOptimizer(1e-2).minimize(loss)
# error
with tf.name_scope('eval'):
error = []
for iclass in range(nclasses):
labels0 = tf.reshape(tf.to_int32(tf.slice(tf_labels,[0,0,0,iclass],[-1,-1,-1,1])),[batchsize,imsize,imsize])
predict0 = tf.reshape(tf.to_int32(tf.equal(tf.argmax(sm,3),iclass)),[batchsize,imsize,imsize])
correct = tf.multiply(labels0,predict0)
ncorrect0 = tf.reduce_sum(correct)
nlabels0 = tf.reduce_sum(labels0)
error.append(1-tf.to_float(ncorrect0)/tf.to_float(nlabels0))
errors = tf.tuple(error)
# --------------------------------------------------
# inspection
# --------------------------------------------------
with tf.name_scope('loss'):
tf.summary.scalar('avg_cross_entropy', loss)
for iclass in range(nclasses):
tf.summary.scalar('avg_pixel_error_%d' % iclass, error[iclass])
with tf.name_scope('histograms'):
tf.summary.histogram('ds0',dsX[1])
with tf.name_scope('images'):
split0 = tf.slice(sm,[0,0,0,0],[-1,-1,-1,1])
split1 = tf.slice(sm,[0,0,0,1],[-1,-1,-1,1])
tf.summary.image('pm0',split0)
tf.summary.image('pm1',split1)
merged = tf.summary.merge_all()
# --------------------------------------------------
# session
# --------------------------------------------------
saver = tf.train.Saver()
sess = tf.Session(config=tf.ConfigProto(allow_soft_placement=True)) # config parameter needed to save variables when using GPU
if os.path.exists(out_log_path):
shutil.rmtree(out_log_path)
train_writer = tf.summary.FileWriter(train_writer_path, sess.graph)
valid_writer = tf.summary.FileWriter(valid_writer_path, sess.graph)
if restore_variables:
saver.restore(sess, out_model_path)
print("Model restored.")
else:
sess.run(tf.global_variables_initializer())
# --------------------------------------------------
# train
# --------------------------------------------------
batch_data = np.zeros((batchsize,imsize,imsize,nchannels))
batch_labels = np.zeros((batchsize,imsize,imsize,nclasses))
for i in range(nsteps):
# train
perm = np.arange(ntrain)
np.random.shuffle(perm)
for j in range(batchsize):
batch_data[j,:,:,:] = Train[perm[j],:,:,:]
batch_labels[j,:,:,:] = LTrain[perm[j],:,:,:]
summary,_ = sess.run([merged,opt_op],feed_dict={tf_data: batch_data, tf_labels: batch_labels})
train_writer.add_summary(summary, i)
# validation
perm = np.arange(nvalid)
np.random.shuffle(perm)
for j in range(batchsize):
batch_data[j,:,:,:] = Valid[perm[j],:,:,:]
batch_labels[j,:,:,:] = LValid[perm[j],:,:,:]
summary, es = sess.run([merged, errors],feed_dict={tf_data: batch_data, tf_labels: batch_labels})
valid_writer.add_summary(summary, i)
e = np.mean(es)
print('step %05d, e: %f' % (i,e))
if i == 0:
if restore_variables:
lowest_error = e
else:
lowest_error = np.inf
if np.mod(i,100) == 0 and e < lowest_error:
lowest_error = e
print("Model saved in file: %s" % saver.save(sess, out_model_path))
# --------------------------------------------------
# test
# --------------------------------------------------
if not os.path.exists(out_pm_path):
os.makedirs(out_pm_path)
for i in range(ntest):
j = np.mod(i,batchsize)
batch_data[j,:,:,:] = Test[i,:,:,:]
batch_labels[j,:,:,:] = LTest[i,:,:,:]
if j == batchsize-1 or i == ntest-1:
output = sess.run(sm,feed_dict={tf_data: batch_data, tf_labels: batch_labels})
for k in range(j+1):
sio.savemat('%s/PM_I%02d.mat' % (out_pm_path,i-j+k+1), {'classProbs':output[k,:,:,:]})
for l in range(nclasses):
misc.imsave('%s/I%d_PM%d.png' % (out_pm_path,i-j+k+1,l),output[k,:,:,l])
# --------------------------------------------------
# clean-up
# --------------------------------------------------
train_writer.close()
valid_writer.close()
sess.close()
运行代码时，出现以下错误，我不确定如何解决：
During handling of the above exception, another exception occurred:
Traceback (most recent call last):
File "segm.py", line 198, in <module>
usX.append(up_samp_layer(usX[i],nlayers-1-i))
File "segm.py", line 174, in up_samp_layer
cv = tf.nn.relu(tf.nn.conv2d(cc, luX_weights2, strides=[1, 1, 1, 1], padding='SAME'),name='conv2')
File "C:\Users\Rares\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\ops\gen_nn_ops.py", line 1042, in conv2d
data_format=data_format, dilations=dilations, name=name)
File "C:\Users\Rares\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\framework\op_def_library.py", line 787, in _apply_op_helper
op_def=op_def)
File "C:\Users\Rares\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\framework\ops.py", line 3414, in create_op
op_def=op_def)
File "C:\Users\Rares\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\framework\ops.py", line 1756, in __init__
control_input_ops)
File "C:\Users\Rares\AppData\Local\Programs\Python\Python36\lib\site-packages\tensorflow\python\framework\ops.py", line 1592, in _create_c_op
raise ValueError(str(e))
ValueError: Dimensions must be equal, but are 64 and 96 for 'upsampling/lu2/Conv2D' (op: 'Conv2D') with input shapes: [4,90,90,64], [3,3,96,64].