In [22]:
%pylab inline
import warnings
warnings.filterwarnings("ignore")
import nolearn
from nolearn.lasagne import NeuralNet
import readdata
import lasagne
from lasagne import layers
from sklearn import metrics
import detectobjects as det
import os.path
from scipy import misc
import cv2
from progress_bar import ProgressBar
import shapefeatures
from sklearn import ensemble
In [23]:
opts = {'img_dir': '../data/plasmodium/',
'annotation_dir': '../data/plasmodium/',
'detection_probability_threshold': 0.5,
'detection_overlap_threshold': 0.3,
'gauss': 1,
'patch_size': (50,50),
'image_downsample' : 2,
'detection_step': 5,
'patch_creation_step': 40,
'object_class': None,
'negative_training_discard_rate': .9
}
opts['patch_stride_training'] = int(opts['patch_size'][0]*.25)
In [24]:
trainfiles, valfiles, testfiles = readdata.create_sets(opts['img_dir'], train_set_proportion=.5,
test_set_proportion=.5,
val_set_proportion=0)
train_y, train_X = readdata.create_patches(trainfiles, opts['annotation_dir'], opts['img_dir'], opts['patch_size'][0], opts['patch_stride_training'], grayscale=False, progressbar=True, downsample=opts['image_downsample'], objectclass=opts['object_class'], negative_discard_rate=opts['negative_training_discard_rate'])
test_y, test_X = readdata.create_patches(testfiles, opts['annotation_dir'], opts['img_dir'], opts['patch_size'][0], opts['patch_stride_training'], grayscale=False, progressbar=True, downsample=opts['image_downsample'], objectclass=opts['object_class'], negative_discard_rate=opts['negative_training_discard_rate'])
# Cut down on disproportionately large numbers of negative patches
train_X, train_y = readdata.balance(train_X, train_y, mult_neg=100)
#test_X, test_y = readdata.balance(test_X, test_y, mult_neg=100)
# Create rotated and flipped versions of the positive patches
train_X, train_y = readdata.augment_positives(train_X, train_y)
test_X, test_y = readdata.augment_positives(test_X, test_y)
print '\n'
print '%d positive training examples, %d negative training examples' % (sum(train_y), len(train_y)-sum(train_y))
print '%d positive testing examples, %d negative testing examples' % (sum(test_y), len(test_y)-sum(test_y))
print '%d patches (%.1f%% positive)' % (len(train_y)+len(test_y), 100.*((sum(train_y)+sum(test_y))/(len(train_y)+len(test_y))))
View a random selection of positive and negative patches to see if they look right
In [25]:
N_samples_to_display = 10
pos_indices = np.where(test_y)[0]
pos_indices = pos_indices[np.random.permutation(len(pos_indices))]
for i in range(N_samples_to_display):
plt.subplot(2,N_samples_to_display,i+1)
example_pos = test_X[pos_indices[i],:,:,:]
example_pos = np.swapaxes(example_pos,0,2)
plt.imshow(example_pos)
plt.tick_params(axis='both', which='both', bottom='off', top='off', labelbottom='off', right='off', left='off', labelleft='off')
neg_indices = np.where(test_y==0)[0]
neg_indices = neg_indices[np.random.permutation(len(neg_indices))]
for i in range(N_samples_to_display,2*N_samples_to_display):
plt.subplot(2,N_samples_to_display,i+1)
example_neg = test_X[neg_indices[i],:,:,:]
example_neg = np.swapaxes(example_neg,0,2)
plt.imshow(example_neg)
plt.tick_params(axis='both', which='both', bottom='off', top='off', labelbottom='off', right='off', left='off', labelleft='off')
plt.gcf().set_size_inches(1.5*N_samples_to_display,3)
In [ ]:
def CNN(n_epochs):
net1 = NeuralNet(
layers=[
('input', layers.InputLayer),
('conv1', layers.Conv2DLayer), #Convolutional layer. Params defined below
('pool1', layers.MaxPool2DLayer), # Like downsampling, for execution speed
('conv2', layers.Conv2DLayer),
('hidden3', layers.DenseLayer),
('output', layers.DenseLayer),
],
input_shape=(None, 3, opts['patch_size'][0]/opts['image_downsample'],
opts['patch_size'][0]/opts['image_downsample']),
conv1_num_filters=7,
conv1_filter_size=(5, 5),
conv1_nonlinearity=lasagne.nonlinearities.rectify,
pool1_pool_size=(2, 2),
conv2_num_filters=12,
conv2_filter_size=(2, 2),
conv2_nonlinearity=lasagne.nonlinearities.rectify,
hidden3_num_units=500,
output_num_units=2,
output_nonlinearity=lasagne.nonlinearities.softmax,
update_learning_rate=0.0001,
update_momentum=0.9,
max_epochs=n_epochs,
verbose=1,
)
return net1
cnn = CNN(500).fit(train_X, train_y)
In [31]:
y_pred = cnn.predict_proba(test_X)
In [33]:
false_positive_rate, true_positive_rate, thresholds = metrics.roc_curve(test_y, y_pred[:,1])
roc_auc = metrics.auc(false_positive_rate, true_positive_rate)
precision, recall, thresholds = metrics.precision_recall_curve(test_y, y_pred[:,1])
average_precision = metrics.average_precision_score(test_y, y_pred[:, 1])
subplot(121)
plt.title('ROC: AUC = %0.2f'% roc_auc)
plt.plot(false_positive_rate, true_positive_rate, 'b')
plt.legend(loc='lower right')
plt.plot([0,1],[0,1],'r--')
plt.ylim([-.05, 1.05])
plt.xlim([-.05, 1.0])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
subplot(122)
plt.plot(recall, precision)
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.ylim([0.0, 1.05])
plt.xlim([0.0, 1.0])
plt.title('Precision-Recall: AP={0:0.2f}'.format(average_precision))
plt.legend(loc="lower left")
plt.gcf().set_size_inches(10,4)
#plt.savefig('figs/plasmodium-patchevaluation.png', bbox_inches='tight')
In [34]:
false_positive_rate, true_positive_rate, thresholds = metrics.roc_curve(test_y, y_pred[:,1])
true_positive_rate.shape, thresholds.shape
plt.plot(true_positive_rate, thresholds,label='True positive rate')
plt.plot(false_positive_rate, thresholds, label='False positive rate')
plt.xlabel('Threshold')
plt.legend(loc='upper left')
Out[34]:
Examine mistakes to understand network performance: false positives.
Find the negative-labelled patches with highest prediction score
In [35]:
neg_indices = np.where(test_y==0)[0]
neg_scores = y_pred[neg_indices,1]
neg_indices = neg_indices[neg_scores.argsort()]
neg_indices = neg_indices[::-1]
neg_scores = y_pred[neg_indices,1]
N_samples_to_display = 12
offset = 55
for i in range(N_samples_to_display,2*N_samples_to_display):
plt.subplot(2,N_samples_to_display,i+1)
example_neg = test_X[neg_indices[i+offset],:,:,:]
example_neg = np.swapaxes(example_neg,0,2)
plt.imshow(example_neg)
plt.title('%.3f' % neg_scores[i+offset])
plt.tick_params(axis='both', which='both', bottom='off', top='off', labelbottom='off', right='off', left='off', labelleft='off')
plt.gcf().set_size_inches(1.5*N_samples_to_display,3)
plt.savefig('figs/plasmodium-falsedetections.png', bbox_inches='tight')
See highest-scored test patches
In [36]:
prob_range = [.95,1.]
tmp_scores = y_pred.copy()[:,1]
tmp_scores[tmp_scores<prob_range[0]] = -1
tmp_scores[tmp_scores>prob_range[1]] = -1
pos_indices = tmp_scores.argsort()
pos_indices = pos_indices[::-1]
N_samples_to_display = 12
offset = 0
for i in range(N_samples_to_display,2*N_samples_to_display):
plt.subplot(2,N_samples_to_display,i+1)
example_neg = test_X[pos_indices[i+offset],:,:,:]
example_neg = np.swapaxes(example_neg,0,2)
plt.imshow(example_neg)
plt.title('%.3f' % (tmp_scores[pos_indices[i+offset]]))
plt.tick_params(axis='both', which='both', bottom='off', top='off', labelbottom='off', right='off', left='off', labelleft='off')
plt.gcf().set_size_inches(1.5*N_samples_to_display,3)
plt.savefig('figs/plasmodium-detectedpatches.png', bbox_inches='tight')
See lowest scored test patches
In [37]:
pos_indices = y_pred[:,1].argsort()
N_samples_to_display = 12
for i in range(N_samples_to_display,2*N_samples_to_display):
plt.subplot(2,N_samples_to_display,i+1)
example_neg = test_X[pos_indices[i],:,:,:]
example_neg = np.swapaxes(example_neg,0,2)
plt.imshow(example_neg)
plt.title('%.3f' % (y_pred[pos_indices[i],1]))
plt.tick_params(axis='both', which='both', bottom='off', top='off', labelbottom='off', right='off', left='off', labelleft='off')
plt.gcf().set_size_inches(1.5*N_samples_to_display,3)
plt.savefig('figs/plasmodium-testpatches-lowprob.png', bbox_inches='tight')
In [ ]:
reload(det)
fname = testfiles[5]
imfile = opts['img_dir'] + fname
opts['detection_probability_threshold'] = 0.5
opts['detection_overlap_threshold'] = 0.1
found = det.detect(imfile, cnn, opts)
im = misc.imread(imfile)
plt.box(False)
plt.xticks([])
plt.yticks([])
annofile = opts['annotation_dir'] + fname[:-3] + 'xml'
bboxes = readdata.get_bounding_boxes_for_single_image(annofile)
for bb in bboxes:
bb = bb.astype(int)
cv2.rectangle(im, (bb[0],bb[2]), (bb[1],bb[3]), (255,255,255), 2)
for f in found:
f = f.astype(int)
cv2.rectangle(im, (f[0],f[1]), (f[2],f[3]), (255,0,0), 2)
plt.gcf().set_size_inches(10,10)
plt.title('Detected objects in %s' % (imfile))
plt.imshow(im)
#cv2.imwrite('detectionimages/detected-' + os.path.basename(imfile),im)
In [ ]:
featureset = [3,7,11,12,15,17]
centiles = [0,25,50,75,100]
pb = ProgressBar(train_X.shape[0])
train_X_f = []
for i in range(train_X.shape[0]):
if i % 100 == 0:
pb.step(i)
graypatch = cv2.cvtColor(np.swapaxes(train_X[i,:,:,:],0,2).astype('uint8'), cv2.COLOR_BGR2GRAY)
train_X_f.append(shapefeatures.extract(graypatch,attributes=featureset,centiles=centiles, momentfeatures=True))
train_X_f = np.vstack(train_X_f)
test_X_f = []
for i in range(test_X.shape[0]):
if i % 100 == 0:
pb.step(i)
graypatch = cv2.cvtColor(np.swapaxes(test_X[i,:,:,:],0,2).astype('uint8'), cv2.COLOR_BGR2GRAY)
test_X_f.append(shapefeatures.extract(graypatch,attributes=featureset,centiles=centiles, momentfeatures=True))
test_X_f = np.vstack(test_X_f)
In [ ]:
clf = ensemble.ExtraTreesClassifier(n_estimators=100, max_depth=5, n_jobs=-1)
clf.fit(train_X_f, train_y)
y_pred_CLF = clf.predict_proba(test_X_f)
In [ ]:
false_positive_rate_CNN, true_positive_rate_CNN, thresholds_CNN = metrics.roc_curve(test_y, y_pred[:,1])
roc_auc_CNN = metrics.auc(false_positive_rate_CNN, true_positive_rate_CNN)
precision_CNN, recall_CNN, thresholds_CNN = metrics.precision_recall_curve(test_y, y_pred[:,1])
average_precision_CNN = metrics.average_precision_score(test_y, y_pred[:, 1])
false_positive_rate_CLF, true_positive_rate_CLF, thresholds_CLF = metrics.roc_curve(test_y, y_pred_CLF[:,1])
roc_auc_CLF = metrics.auc(false_positive_rate_CLF, true_positive_rate_CLF)
precision_CLF, recall_CLF, thresholds_CLF = metrics.precision_recall_curve(test_y, y_pred_CLF[:,1])
average_precision_CLF = metrics.average_precision_score(test_y, y_pred_CLF[:, 1])
subplot(211)
plt.title('ROC' )
plt.plot(false_positive_rate_CNN, true_positive_rate_CNN, 'b', label='CNN: AUC=%.2f' % (roc_auc_CNN))
plt.plot(false_positive_rate_CLF, true_positive_rate_CLF, 'k--', label='ERT: AUC=%.2f' % (roc_auc_CLF))
plt.legend(loc='lower right')
plt.ylim([-.05, 1.05])
plt.xlim([-.05, 1.0])
plt.ylabel('True Positive Rate')
plt.xlabel('False Positive Rate')
subplot(212)
plt.plot(recall_CNN, precision_CNN, label='CNN: AP=%.2f' % (average_precision_CNN))
plt.plot(recall_CLF, precision_CLF,'k--', label='ERT: AP=%.2f' % (average_precision_CLF))
plt.xlabel('Recall')
plt.ylabel('Precision')
plt.ylim([0.0, 1.05])
plt.xlim([0.0, 1.0])
plt.title('Precision-Recall')
plt.legend(loc="lower left")
plt.gcf().set_size_inches(4,10)
plt.savefig('figs/plasmodium-patchevaluation.png', bbox_inches='tight')
In [ ]:
results = {
"false_positive_rate_CNN": false_positive_rate_CNN,
"true_positive_rate_CNN": true_positive_rate_CNN,
"false_positive_rate_CLF": false_positive_rate_CLF,
"true_positive_rate_CLF": true_positive_rate_CLF,
"roc_auc_CNN": roc_auc_CNN,
"roc_auc_CLF": roc_auc_CLF,
"recall_CNN": recall_CNN,
"precision_CNN": precision_CNN,
"average_precision_CNN": average_precision_CNN,
"recall_CLF": recall_CLF,
"precision_CLF": precision_CLF,
"average_precision_CLF": average_precision_CLF,
"opts": opts
}
import pickle
pickle.dump(results,open('plasmodium-results.pkl','w'))
In [ ]: