def test_metrics(self):
Y = np.random.randint(0, 2, size=(2, 5, 5))
Yhat = np.random.randint(0, 2, size=(2, 5, 5))
C, acc, prec, recall, f1 = emlib.metrics(Y, Yhat, display=False)
prec2, recall2, f12, supp = smetrics(np.reshape(Y, (Y.size, )),
np.reshape(Yhat, (Yhat.size, )))
self.assertAlmostEqual(prec, prec2[1])
self.assertAlmostEqual(recall, recall2[1])
self.assertAlmostEqual(f1, f12[1])
def test_metrics(self):
Y = np.random.randint(0,2,size=(2,5,5))
Yhat = np.random.randint(0,2,size=(2,5,5))
C,acc,prec,recall,f1 = emlib.metrics(Y, Yhat, display=False)
prec2, recall2, f12, supp = smetrics(np.reshape(Y, (Y.size,)),
np.reshape(Yhat, (Yhat.size,)))
self.assertAlmostEqual(prec, prec2[1])
self.assertAlmostEqual(recall, recall2[1])
self.assertAlmostEqual(f1, f12[1])
def _train_network(args):
""" Main CNN training loop.
Creates PyCaffe objects and calls train_one_epoch until done.
"""
#----------------------------------------
# parse information from the prototxt files
#----------------------------------------
solverParam = caffe_pb2.SolverParameter()
text_format.Merge(open(args.solver).read(), solverParam)
netFn = solverParam.net
netParam = caffe_pb2.NetParameter()
text_format.Merge(open(netFn).read(), netParam)
batchDim = emlib.infer_data_dimensions(netFn)
assert(batchDim[2] == batchDim[3]) # tiles must be square
print('[emCNN]: batch shape: %s' % str(batchDim))
if args.outDir:
outDir = args.outDir # overrides snapshot prefix
else:
outDir = str(solverParam.snapshot_prefix) # unicode -> str
if not os.path.isdir(outDir):
os.mkdir(outDir)
# choose a synthetic data generating function
if args.rotateData:
syn_func = lambda V: _xform_minibatch(V, arbitraryRotation=True)
print('[emCNN]: WARNING: applying arbitrary rotations to data. This may degrade performance in some cases...\n')
else:
syn_func = lambda V: _xform_minibatch(V, arbitraryRotation=False)
#----------------------------------------
# Create the Caffe solver
# Note this assumes a relatively recent PyCaffe
#----------------------------------------
solver = caffe.SGDSolver(args.solver)
solverMD = SGDSolverMemoryData(solver, solverParam)
solverMD.print_network()
#----------------------------------------
# Load data
#----------------------------------------
bs = border_size(batchDim)
print "[emCNN]: tile radius is: %d" % bs
print "[emCNN]: loading training data..."
Xtrain, Ytrain = _load_data(args.emTrainFile,
args.labelsTrainFile,
tileRadius=bs,
onlySlices=args.trainSlices,
omitLabels=args.omitLabels)
print "[emCNN]: loading validation data..."
Xvalid, Yvalid = _load_data(args.emValidFile,
args.labelsValidFile,
tileRadius=bs,
onlySlices=args.validSlices,
omitLabels=args.omitLabels)
#----------------------------------------
# Do training; save results
#----------------------------------------
omitLabels = set(args.omitLabels).union([-1,]) # always omit -1
currEpoch = 1
sys.stdout.flush()
while not solverMD.is_training_complete():
print "[emCNN]: Starting epoch %d" % currEpoch
train_one_epoch(solverMD, Xtrain, Ytrain,
batchDim, outDir,
omitLabels=omitLabels,
data_augment=syn_func)
currEpoch += 1
print "[emCNN]: Making predictions on validation data..."
Mask = np.ones(Xvalid.shape, dtype=np.bool)
Mask[Yvalid<0] = False
Prob = predict(solver.net, Xvalid, Mask, batchDim)
# discard mirrored edges and form class estimates
Yhat = np.argmax(Prob, 0)
Yhat[Mask==False] = -1;
Prob = prune_border_4d(Prob, bs)
Yhat = prune_border_3d(Yhat, bs)
# compute some metrics
print('[emCNN]: Validation set performance:')
emlib.metrics(prune_border_3d(Yvalid, bs), Yhat, display=True)
solver.net.save(str(os.path.join(outDir, 'final.caffemodel')))
np.save(os.path.join(outDir, 'Yhat.npz'), Prob)
scipy.io.savemat(os.path.join(outDir, 'Yhat.mat'), {'Yhat' : Prob})
print('[emCNN]: training complete.')
