Branch: master
Find file Copy path
Fetching contributors…
Cannot retrieve contributors at this time
105 lines (77 sloc) 3.83 KB
import time
import sys
import numpy as np
import argparse
caffe_root = '/home/axj232/caffe/'
sys.path.insert(0, caffe_root + 'python')
import caffe
#parse the command line arguments
parser = argparse.ArgumentParser(description='Generate lymphoma outputs from Caffe DL models.')
parser.add_argument('fold',type=int,help="Which fold to generate outputfor")
args= parser.parse_args()
MODEL_FILE = './models/deploy_train32.prototxt'
PRETRAINED = './models/%d_caffenet_train_w32_iter_600000.caffemodel' % (FOLD)
#this window size needs to be exactly the same size as that used to extract the patches from the matlab version
wsize = 36
hwsize= int(wsize/2)
stride = 32 #incase we don't want to compute every pixel, we can skip
#load our mean file and reshape it accordingly
file=open('./DB_train_w32_%d.binaryproto' % FOLD ,'rb')
data =
#make sure we use teh GPU otherwise things will take a very long time
#load the model
net = caffe.Classifier(MODEL_FILE, PRETRAINED, mean=means,
image_dims=(36, 36))
conf_matrix = np.matrix('0 0 0; 0 0 0; 0 0 0') #initialize confusion matrix
wrong_files = [] # and a list to store any files which might be classified incorrectly
start_time = time.time()
with open('./test_w32_parent_%d.txt' % FOLD,'rb') as test_file:
total_correct = 0
totals = 0
for main_image in test_file: #for each of the files in the test file, load them and analyze them
print main_image,
c = main_image[0] #get the first letter of the file...from this we can tell what the class is supposed to be
if (c == 'C'):
aclass = 0
elif (c == 'F'):
aclass = 1
elif (c == 'M'):
aclass = 2
else: #if its not a C F or M, then we don't know what class it belongs to
print 'UNKNOWN!', main_image
main_image = main_image.strip()
main_image = "./images/"+main_image+".tif"
counts = [0, 0, 0]
image = #load the image....
for rowi in xrange(hwsize+1,image.shape[0]-hwsize,stride): #on a per image basis, compute some patches and agglomerate their predicted patches
print "%s\t (%.3f,%.3f)\t %d of %d" % (main_image,time.time()-start_time,time.time()-start_time_iter,rowi,image.shape[0]-hwsize)
start_time_iter = time.time()
for coli in xrange(hwsize+1,image.shape[1]-hwsize,stride):
patch = image[rowi-hwsize:rowi+hwsize, coli-hwsize:coli+hwsize,:] #extract the patch
prediction = net.predict([patch]) #get its prediction
#print prediction
pclass = prediction[0].argmax() #figure out which class was chosen, where 0 = CLL, 1= FL and 2= MCL
counts[pclass] = counts[pclass] + 1 #add this to the overall predictions (or votes) for this image
pfinal = np.argmax(counts) #now compute which class received the most "votes"
conf_matrix[pfinal, aclass] = conf_matrix[pfinal, aclass] + 1 #pfinal has our prediction and aclass has the actual, so we can update the confusion matrix appropriately
print aclass, pfinal, counts
if (pfinal == aclass):
total_correct = total_correct + 1 #if its correct, add it to the correct counts
wrong_files.append(main_image) #otherwise add the file to the wrong list
totals = totals + 1
print total_correct, totals, total_correct / (1.0 * totals)
print (conf_matrix) #finally print the confusion matrix
print "wrong files:"
for wrong_file in wrong_files: #and print the wrong images
print (wrong_file)