python-2.7 - 为什么 FCN32 输出是黑色图像（所有像素都为零值）？

Question

我正在尝试从 FCN 32 获取输出。我使用 pascalcontext-fcn32-heavy.caffemodel 预训练模型训练了 FCN32。我可以运行 5 个类别的灰度图像。但是，在推理过程中，输出全为零（黑色图像）。这是推理代码：

import numpy as np
from PIL import Image
import sys
import scipy.io as sio
from caffe.proto import caffe_pb2
import caffe
    caffe.set_device(0) 
    caffe.set_mode_gpu()

    # load image, subtract mean, and make dims C x H x W for Caffe

   img_name='/home/ss/caffe-pascalcontext-fcn32s/dataset/Test/PNG/image-061-023.png'    #+
    im = Image.open(img_name)

    in_ = np.array(im, dtype=np.float32)
    in_ = np.expand_dims(in_, axis=0)               #+
    print in_.shape
    #Read mean image
    '''####################'''
    mean_blob = caffe_pb2.BlobProto()
    with open('/home/ss/caffe-pascalcontext-fcn32s/input/FCN32_mean.binaryproto') as f:
        mean_blob.ParseFromString(f.read())
    mean_array = np.asarray(mean_blob.data, dtype=np.float32).reshape(
        (mean_blob.channels, mean_blob.height, mean_blob.width))
    in_ -= mean_array

    net_root = '/home/ss/caffe-pascalcontext-fcn32s'

    MODEL_DEF = net_root + '/deploy.prototxt'
    PRETRAINED = net_root + '/snapshot/FCN32s_train_iter_40000.caffemodel'
    # load net
    #net = caffe.Net('deploy.prototxt', 'snapshot/train_iter_640000.caffemodel', caffe.TEST)
    net = caffe.Net(MODEL_DEF,PRETRAINED, caffe.TEST)
    #net = caffe.Net('deploy.prototxt', 'snapshot_bak1/train_iter_400000.caffemodel', caffe.TEST)

    # shape for input (data blob is N x C x H x W), set data
    # put img to net
    net.blobs['data'].reshape(1, *in_.shape)  # 1: batch size, *in_.shape 3 channel ?
    net.blobs['data'].data[...] = in_

    # run net and take argmax for prediction
    output = net.forward()

    # print
    def print_param(output):
        # the blobs
        print '--------------------------'
        print 'the blobs'
        for k, v in net.blobs.items():
            print k, v.data.shape

        # the parameters
        print '--------------------------'
        print 'the paramsters'
        for k, v in net.params.items():
            print k, v[0].data.shape

        # the conv layer weights
        print '--------------------------'
        print 'the conv layer weights'
        print net.params['conv1_1'][0].data

        # the data blob 
        print '--------------------------'
        print 'the data blob'
        print net.blobs['data'].data

        # the conv1_1 blob
        print '--------------------------'
        print 'the conv1_1 blob'
        print net.blobs['conv1_1'].data

        # the pool1 blob
        print '--------------------------'
        print 'the pool1 blob'
        print net.blobs['pool1'].data

        weights = net.blobs['fc6'].data[0]
        print 'blobs fc6'
        print np.unique(weights)
        weights = net.blobs['fc7'].data[0]
        print 'blobs fc7'
        print np.unique(weights)
        weights = net.blobs['score_fr_sign'].data[0]
        print 'blobs score_fr_sign'
        print np.unique(weights)
        weights = net.blobs['upscore_sign'].data[0]
        print 'blobs upscore_sign'
        print np.unique(weights)
        weights = net.blobs['score'].data[0]
            print weights.shape             #+
            sio.savemat('scores.mat',{'weights':weights})   #+
        print 'blobs score'
        print np.unique(weights)

    print_param(output)

    out = net.blobs['score'].data[0].argmax(axis=0)
    print out           #+

    #np.savetxt("vote", out, fmt="%02d")
    np.savetxt("vote", out, fmt="%d")

    print im.height
    print im.width
    print out.shape, len(out.shape)

    def array2img(out):
        out1 = np.array(out, np.unit8)
        img = Image.fromarray(out1,'L')
        for x in range(img.size[0]):
            for y in range(img.size[1]):
                if not img.getpixel((x, y)) == 0:
                    print 'PLz', str(img.getpixel((x, y)))

        img.show()


    def show_pred_img(file_name):
        file = open(file_name, 'r')
        lines = file.read().split('\n')

        #img_name = str(sys.argv[1])
        im = Image.open(img_name)
        im_pixel = im.load()

        img = Image.new('RGB', im.size, "black")
        pixels = img.load()

        w, h = 0, 0
        for l in lines:
            w = 0
            if len(l) > 0:
                word = l.split(' ')
                for x in word:
                    if int(x) == 1:
                        pixels[w, h] = im_pixel[w, h]
                    w += 1
                h += 1
        print im.size
        #img.show()
        img.save(img_name+'_result.png')
    show_pred_img('vote')

这是推理的日志信息：

the blobs
data (1, 1, 256, 256)
data_input_0_split_0 (1, 1, 256, 256)
data_input_0_split_1 (1, 1, 256, 256)
conv1_1 (1, 64, 454, 454)
conv1_2 (1, 64, 454, 454)
pool1 (1, 64, 227, 227)
conv2_1 (1, 128, 227, 227)
conv2_2 (1, 128, 227, 227)
pool2 (1, 128, 114, 114)
conv3_1 (1, 256, 114, 114)
conv3_2 (1, 256, 114, 114)
conv3_3 (1, 256, 114, 114)
pool3 (1, 256, 57, 57)
conv4_1 (1, 512, 57, 57)
conv4_2 (1, 512, 57, 57)
conv4_3 (1, 512, 57, 57)
pool4 (1, 512, 29, 29)
conv5_1 (1, 512, 29, 29)
conv5_2 (1, 512, 29, 29)
conv5_3 (1, 512, 29, 29)
pool5 (1, 512, 15, 15)
fc6 (1, 4096, 9, 9)
fc7 (1, 4096, 9, 9)
score_fr_sign (1, 5, 9, 9)
upscore_sign (1, 5, 320, 320)
score (1, 5, 256, 256)
--------------------------
the paramsters
conv1_1 (64, 1, 3, 3)
conv1_2 (64, 64, 3, 3)
conv2_1 (128, 64, 3, 3)
conv2_2 (128, 128, 3, 3)
conv3_1 (256, 128, 3, 3)
conv3_2 (256, 256, 3, 3)
conv3_3 (256, 256, 3, 3)
conv4_1 (512, 256, 3, 3)
conv4_2 (512, 512, 3, 3)
conv4_3 (512, 512, 3, 3)
conv5_1 (512, 512, 3, 3)
conv5_2 (512, 512, 3, 3)
conv5_3 (512, 512, 3, 3)
fc6 (4096, 512, 7, 7)
fc7 (4096, 4096, 1, 1)
score_fr_sign (5, 4096, 1, 1)
upscore_sign (5, 1, 64, 64)
--------------------------
the conv layer weights
[[[[ 0.  0.  0.]
   [ 0.  0.  0.]
   [ 0.  0.  0.]]]

...
 .
 .
 .       

 [[[ 0.  0.  0.]
   [ 0.  0.  0.]
   [ 0.  0.  0.]]]]
--------------------------
the data blob
[[[[ 29.32040787  20.31391525  20.30148506 ...,  10.41113186  11.42486095
      6.42949915]
   [ 33.32374954  21.31280136  22.30037117 ...,   9.40779209  10.42189217
      8.43079758]
   [ 36.32300568  25.30816269  25.29183578 ...,  10.40148449  11.41818142
     10.42838573]
   ..., 
   [ 34.64990616  31.65658569  30.65714264 ...,   4.           2.99981451
      0.99962896]
   [ 39.65788651  33.65769958  29.65974045 ...,   5.99981451   4.99944353
      0.99888682]
   [ 41.6641922   34.66493607  30.66567802 ...,   5.99962902   2.99907231
      3.99833035]]]]
--------------------------
the conv1_1 blob
[[[[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  ..., 
  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]]]
--------------------------
the pool1 blob
[[[[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  ..., 
  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]

  [[ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   ..., 
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]
   [ 0.  0.  0. ...,  0.  0.  0.]]]]
blobs fc6
[ 0.]
blobs fc7
[ 0.]
blobs score_fr_sign
[-1.61920226 -1.34294271  0.07809996  0.60521388  2.2788291 ]
blobs upscore_sign
[-1.61920238 -1.61920226 -1.61920214 ...,  2.27882886  2.2788291
  2.27882934]
(5, 256, 256)
blobs score
[-1.61920238 -1.61920226 -1.61920214 -1.59390223 -1.59390211 -1.5689975
 -1.54330218 -1.54330206 -1.51918805 -1.49270213 -1.49270201 -1.4709599
 -1.46937859 -1.44210207 -1.44210196 -1.42273164 -1.41956913 -1.39150202
 -1.3915019  -1.37608469 -1.37450349 -1.36975968 -1.34294283 -1.34294271
 -1.3429426  -1.34090197 -1.34090185 -1.32943773 -1.32627523 -1.32195926
 -1.31995022 -1.30130363 -1.2903018  -1.28437209 -1.2827909  -1.27999234
 -1.27999222 -1.27804708 -1.27014089 -1.25999236 -1.23970175 -1.23930645
 -1.23802543 -1.23802531 -1.23614395 -1.22981894 -1.22033143 -1.21999264
 -1.21868122 -1.19605839 -1.19605827 -1.195822   -1.19424069 -1.18949699
 -1.1891017  -1.18910158 -1.18159068 -1.17999291 -1.17736995 -1.17052197
 -1.15409136 -1.15233755 -1.14917505 -1.14285004 -1.14130461 -1.13999307
 -1.13850164 -1.13850152 -1.13605869 -1.13336253 -1.12071252 -1.11212444
 -1.11043441 -1.1088531  -1.10410941 -1.10261631 -1.09999335 -1.09620309
 -1.09474754 -1.08790159 -1.08790147 -1.08513427 -1.07090306 -1.07015753
 -1.07015741 -1.06853116 -1.06536865 -1.06523943 -1.06392801 -1.05999362
 -1.05904365 -1.05343628 -1.04955614 -1.03730154 -1.03730142 -1.03690612
 -1.02820921 -1.02819049 -1.02786267 -1.02662802 -1.02523971 -1.0218842
 -1.02109361 -1.0199939  -1.013978   -1.01212502 -1.00290918 -0.99179727
 -0.99048585 -0.98867792 -0.98788732 -0.98670143 -0.98670137 -0.9865514
 -0.98622358 -0.98622352 -0.98472482 -0.97999406 -0.97839981 -0.97128415
 -0.97081381 -0.9689123  -0.95626229 -0.95573193 -0.95310903 -0.94914663
 -0.94786316 -0.94756538 -0.9442566  -0.94425654 -0.94282162 -0.94044977
 -0.93999434 -0.93491536 -0.92950261 -0.9238466  -0.92097807 -0.91966659
 -0.9157322  -0.91040593 -0.90961534 -0.90917486 -0.90724343 -0.90228963
 -0.90091842 -0.89999455 -0.89143091 -0.88819134 -0.88622415 -0.88360125
 -0.8787809  -0.87835538 -0.87324655 -0.8716653  -0.87048656 -0.86692154
 -0.86032271 -0.86032265 -0.85999483 -0.85901529 -0.85278171 -0.85147029
 -0.84794647 -0.84753585 -0.84688014 -0.8409785  -0.83608711 -0.8329246
 -0.83179826 -0.8265996  -0.81999505 -0.81933933 -0.81835574 -0.81835568
 -0.81711209 -0.81671637 -0.81147051 -0.80556893 -0.80360168 -0.80050892
 -0.79892766 -0.79418391 -0.79310995 -0.78720838 -0.78627765 -0.7858969
 -0.78196251 -0.77999532 -0.77540517 -0.76622486 -0.76493073 -0.76176822
 -0.75544322 -0.75507742 -0.75442165 -0.75245446 -0.7472086  -0.73933983
 -0.73093385 -0.72935259 -0.72884804 -0.72460884 -0.72425795 -0.72294647
 -0.71901208 -0.71245474 -0.70327443 -0.69693691 -0.6937744  -0.69343841
 -0.69081551 -0.68556964 -0.67770082 -0.66452122 -0.66393042 -0.66293997
 -0.66261894 -0.65868455 -0.65212721 -0.63442242 -0.63210559 -0.63179946
 -0.6265536  -0.60622585 -0.60491437 -0.60127115 -0.60097998 -0.57802927
 -0.57540637 -0.55114424 -0.54983276 -0.52425915 -0.49868551  0.02900147
  0.03048873  0.03197598  0.03205225  0.03346324  0.03361578  0.03495049
  0.0351793   0.03525557  0.03643775  0.03674283  0.03689536  0.037925
  0.03830635  0.03853516  0.03861143  0.03941226  0.03986987  0.04017495
  0.04032749  0.04089952  0.0414334   0.04181475  0.04204356  0.04211983
  0.04238677  0.04299692  0.04345454  0.04375962  0.04387403  0.04391216
  0.04456045  0.04509434  0.04536128  0.04547568  0.04570449  0.04578076
  0.04612397  0.04673413  0.04684854  0.04719175  0.04749683  0.04759216
  0.04764936  0.0476875   0.04837392  0.04890781  0.04925102  0.04928916
  0.04951797  0.04959423  0.05001372  0.05003278  0.05003279  0.05062388
  0.05108149  0.05138657  0.05153911  0.05165351  0.05233994  0.05247341
  0.05247341  0.05287382  0.05325517  0.05348398  0.05356025  0.054056
  0.05466616  0.05491403  0.05491403  0.05512378  0.05542885  0.05558139
  0.05645849  0.05699238  0.05735466  0.05735466  0.05737372  0.05760253
  0.0576788   0.05886098  0.05931859  0.05962367  0.05977621  0.05979528
  0.05979528  0.06126347  0.06164481  0.06187363  0.06194989  0.0622359
  0.06223591  0.06366596  0.06397104  0.06412357  0.06467653  0.06606845
  0.06629726  0.06637353  0.06711715  0.06847093  0.06862348  0.06955777
  0.06955778  0.07087342  0.0709497   0.0719984   0.0719984   0.07327592
  0.07443902  0.07443903  0.0756784   0.07687964  0.07687965  0.07809995
  0.07809996  0.07809997  0.22473885  0.23626392  0.24778898  0.24838002
  0.25931406  0.26049611  0.27083912  0.27261221  0.27320322  0.28236419
  0.28472832  0.28591037  0.29388925  0.29684439  0.29861748  0.29920852
  0.30541432  0.3089605   0.31132463  0.31250668  0.31693938  0.3210766
  0.32403174  0.32580483  0.32639587  0.32846448  0.33319271  0.33673888
  0.33910298  0.33998954  0.34028506  0.34530881  0.349446    0.35151461
  0.35240114  0.35417423  0.35476527  0.35742489  0.36215314  0.36303967
  0.36569929  0.36806342  0.36880219  0.36880222  0.36924547  0.36954099
  0.37486026  0.37899747  0.38165709  0.38195261  0.3837257   0.38431671
  0.38756737  0.38771513  0.38771516  0.39229563  0.39584181  0.39820591
  0.39938796  0.40027452  0.40559378  0.40662807  0.40973097  0.41268614
  0.4144592   0.41505024  0.41889194  0.42362016  0.42554098  0.42554101
  0.42716634  0.42953047  0.43071252  0.43750936  0.44164655  0.44445392
  0.44445395  0.44460171  0.44637477  0.44696581  0.45612678  0.45967296
  0.46203706  0.46321911  0.46336687  0.4633669   0.4747442   0.47769934
  0.47947243  0.48006344  0.48227981  0.48227984  0.49336162  0.49572572
  0.49690777  0.50119275  0.51197904  0.5137521   0.51434314  0.52010566
  0.52010572  0.53059644  0.53177851  0.53901857  0.53901863  0.54921389
  0.54980487  0.55793154  0.56783128  0.57684445  0.57684451  0.58644873
  0.59575737  0.59575742  0.60521382  0.60521388  0.60521394  0.84621561
  0.88961124  0.93300694  0.93523234  0.97640258  0.98085344  1.01979828
  1.02647448  1.02869999  1.06319392  1.07209563  1.07654643  1.10658967
  1.11771667  1.12439299  1.12661839  1.14998531  1.16333783  1.17223942
  1.17669034  1.19338095  1.20895886  1.22008598  1.22676229  1.22898769
  1.23677659  1.25458002  1.26793253  1.27683413  1.28017235  1.28128505
  1.30020106  1.31577897  1.32356799  1.32690609  1.3335824   1.3358078
  1.34582222  1.36362553  1.36696362  1.37697804  1.38587976  1.38866138
  1.3886615   1.39033055  1.39144325  1.41147208  1.42704999  1.43706429
  1.43817711  1.44485331  1.4470787   1.45931852  1.45987487  1.45987499
  1.47712183  1.49047434  1.49937606  1.50382698  1.50716507  1.52719378
  1.53108823  1.53108835  1.5427717   1.55389881  1.56057513  1.56280053
  1.57726574  1.59506905  1.6023016   1.60230172  1.60842156  1.61732328
  1.62177408  1.6473664   1.66294444  1.67351508  1.6735152   1.67407143
  1.68074775  1.68297315  1.71746719  1.7308197   1.7397213   1.74417222
  1.74472845  1.74472857  1.78756785  1.79869497  1.80537117  1.80759656
  1.81594181  1.81594193  1.81594205  1.85766852  1.86657023  1.87102103
  1.88715529  1.88715541  1.9277693   1.9344455   1.9366709   1.95836878
  1.99786997  2.00232077  2.02958202  2.02958226  2.06797075  2.07019615
  2.10079551  2.10079575  2.1380713   2.17200899  2.20817208  2.24322224
  2.24322248  2.27882886  2.2788291   2.27882934]
256
256
(256, 256) 2
(256, 256)

我有两个主要问题：

我想知道为什么输出是黑色的？和
我如何知道何时停止运行算法（即迭代次数）？我真的不知道在那个阶段我可以停止微调的最佳迭代次数和损失值是多少。我停止了训练40,000 iterations，我对此一无所知。
分割的结果是否也必须是灰度图像（如输入），或者创建 RGB 结果图像对输出没有任何影响？

我真的不知道我做对了多少。很困惑:(有人有什么建议吗？我真的很感谢你的帮助。

score 0 · Accepted Answer

确保标签的数据类型是 uint8！我有同样的问题！

在训练之前，还要确保你的 prototxt 中有如下重量填充物！

layer {
name: "myupscore2"
type: "Deconvolution"
bottom: "myscore_fr"
top: "myupscore2"
param {
lr_mult: 5
}
convolution_param {
group :2
num_output: 2
weight_filler: { type: "bilinear" }
bias_term: false
kernel_size: 4
stride: 16
 }
 }

祝你好运！

score 0 · Accepted Answer

是的，这通常取决于您的图像大小！你检查过你的数据类型吗？你的图像和groundtruths都应该是uint8！

您是否还向您的 Deconv 图层添加了“组”行？

最好的

python-2.7 - 为什么 FCN32 输出是黑色图像（所有像素都为零值）？

2 回答 2

Related

Reference