def debug(exp, bench=None, net=None):
if bench is None:
bench = pbench.PoseBenchmark(azimuthOnly=False, classes=['car'])
preds = get_predictions(exp, bench, net=net, debugMode=True)
gtPose = bench.giveTestPoses('car')
gtAz, pdAz = [], []
gtEl, pdEl = [], []
testPreds = []
for i in range(len(gtPose)):
az, el, _ = gtPose[i]
az,_ = pep.format_label(az, exp.dPrms_, bins=exp.dPrms_.azBins)
el,_ = pep.format_label(el, exp.dPrms_, bins=exp.dPrms_.elBins)
gtAz.append(az)
gtEl.append(el)
pdFloat, pdBins = preds[i]
testPreds.append(pdFloat)
paz, pel, _ = pdBins
pdAz.append(paz)
pdEl.append(pel)
gtAz = np.array(gtAz)
pdAz = np.array(pdAz)
gtEl = np.array(gtEl)
pdEl = np.array(pdEl)
errs = bench.evaluatePredictions('car', testPreds)
print (180*np.median(errs)/np.pi)
return np.array(gtPose), np.array(testPreds), gtAz, pdAz
def save_evaluation(exp, numIter, bench=None, forceWrite=False):
resFile = get_result_filename(exp, numIter)
#Check if result file exists
if osp.exists(resFile) and not forceWrite:
print ('%s exists' % resFile)
return
#Get the benchmark object
print ('Loading Benchmark Object')
if bench is None:
bench = pbench.PoseBenchmark(classes=PASCAL_CLS)
#Make the net
net = make_deploy_net(exp, numIter)
#Start evaluation
print ('Starting evaluation')
res = edict()
mds = []
for i, cls in enumerate(PASCAL_CLS):
res[cls] = edict()
res[cls]['pd'] = get_predictions(exp, bench, className=cls,
net=net)
res[cls]['gt'] = bench.giveTestPoses(cls)
res[cls]['err'] = bench.evaluatePredictions(cls, res[cls]['pd'])
mds.append(180 * (np.median(res[cls]['err'])/np.pi))
print ('Median accuracy on %s is %f' % (cls, mds[i]))
res[cls]['imn'], res[cls]['bbox'] = bench.giveTestInstances(cls)
mds = np.array(mds)
print ('MEAN ACCURACY %f' % np.mean(mds))
pickle.dump(res, open(resFile, 'w'))
def save_nn_results(cls='car', bench=None):
dat = get_data_dict('train')
fullKeys = dat.keys()
idKeys = [osp.split(k)[1] for k in dat.keys()]
trainFiles, trainLbs = get_cls_set_files('train', cls=cls)
if bench is None:
bench = pbench.PoseBenchmark(classes=[cls])
imNames, bbox = bench.giveTestInstances(cls)
dat
def stupid_debug(exp, bench=None):
bench = pbench.PoseBenchmark(azimuthOnly=True, classes=['car'])
gtPose = bench.giveTestPoses('car')
pdPose = []
for i in range(len(gtPose)):
a, e, _ = gtPose[i]
aBin,_ = pep.format_label(a, exp.dPrms_, bins=exp.dPrms_.azBins)
eBin,_ = pep.format_label(e, exp.dPrms_, bins=exp.dPrms_.elBins)
az = pep.unformat_label(aBin, None,
exp.dPrms_, bins=exp.dPrms_.azBins)
el = pep.unformat_label(eBin, None,
exp.dPrms_, bins=exp.dPrms_.elBins)
pdPose.append([az, el, 0])
errs = bench.evaluatePredictions('car', pdPose)
print (np.median(errs))
def find_test_keys(cls='car', bench=None, dat=None):
if dat is None:
dat = get_data_dict('test')
fullKeys = dat.keys()
idKeys = [osp.split(k)[1] for k in fullKeys]
testKeys = []
if bench is None:
bench = pbench.PoseBenchmark(classes=[cls])
imNames, bbox = bench.giveTestInstances(cls)
for nm in imNames:
_, pascalid = osp.split(nm)
pascalid = pascalid[0:-4]
idx = idKeys.index(pascalid)
if len(dat[fullKeys[idx]].coords) > 1:
pdb.set_trace()
testKeys.append(fullKeys[idx])
return testKeys
def save_imdata():
bench = pbench.PoseBenchmark(classes=PASCAL_CLS)
count = 0
dName = '/data0/pulkitag/data_sets/pascal3d/imCrop/test/im%d.jpg'
testList = []
count = 0
for cls in PASCAL_CLS:
print (cls)
imNames, bbox = bench.giveTestInstances(cls)
ims = get_imdata(imNames, bbox, svMode=True)
for i in range(ims.shape[0]):
svName = dName % count
scm.imsave(svName, ims[i])
testList.append([imNames[i], bbox[i], svName])
count += 1
outFile = 'pose-files/pascal_test_data.pkl'
pickle.dump({'testList': testList}, open(outFile, 'w'))
def debug_evaluate_data(exp, classes=['car'], isPlot=False): bench = pbench.PoseBenchmark(classes=classes) imNames, bbox = bench.giveTestInstances(classes[0]) ims = get_imdata(imNames, bbox, exp) if isPlot: plt.ion() fig = plt.figure() ax = fig.add_subplot(111) for i in range(ims.shape[0]): im = ims[i].transpose((1,2,0)) im = im[:,:,(2,1,0)] + 128 ax.imshow(im.astype(np.uint8)) plt.show() plt.draw() ip = raw_input() if ip == 'q': return plt.cla() return ims
def eval_accuracy_nn(bench=None, netName='caffe_pose_fc5', classes=['car'], visMatches=False):
modErr = []
if bench is None:
bench = pbench.PoseBenchmark(classes=classes)
#Train data
trainDat = transform_dict('train')
keyList = trainDat.keys()
#result data
resDat = pickle.load(open('pascal_results/pascal_results_%s.pkl' % netName,'r'))
resImList = [l[0] for l in resDat['testInfo']]
resBBox = [l[1] for l in resDat['testInfo']]
resKeys = resDat['nnKeys']
imNames, bbox = bench.giveTestInstances(classes[0])
gtPoses = bench.giveTestPoses(classes[0])
preds = []
if visMatches:
plt.ion()
fig = plt.figure()
ax = []
count = 1
for i in range(3):
for j in range(2):
ax.append(fig.add_subplot(2,3, count))
count += 1
exampleCount = 0
for nm, bb in zip(imNames, bbox):
#nm is the name of the image for which we want to find the pose
idx = [i for i,l in enumerate(resImList) if l == nm]
if len(idx) > 1:
for dd in idx:
resBox = resBBox[dd]
isFound = match_bbox(resBox, bb)
if isFound:
idx = [dd]
break
if not isFound:
pdb.set_trace()
assert len(idx)==1
idx = idx[0]
#The 1-NN
if visMatches:
dirName = osp.join(cfg.pths.pascal.dataDr, 'imCrop',
'imSz256_pad36_hash', 'imSz256_pad36_hash')
nnImNames = resKeys[idx]
ax[0].imshow(get_imdata([nm], [bb], svMode=True)[0])
for vv, visname in enumerate(nnImNames[0:5]):
im = scm.imread(osp.join(dirName, visname))
ax[vv+1].imshow(im)
plt.show()
plt.draw()
ip = raw_input()
if ip =='q':
return
key = resKeys[idx][0]
_, pred = trainDat[key]
#print (gtPoses[exampleCount])
modErr.append(find_theta_diff(pred[0], gtPoses[exampleCount][0], 'mod180'))
exampleCount += 1
pred = pred + (0.,)
preds.append(pred)
errs = bench.evaluatePredictions(classes[0], preds)
modErr = np.array(modErr)
mdModErr = 180 * np.median(modErr)/np.pi
mdErr = 180*np.median(errs)/np.pi
return mdModErr, mdErr
def get_car_bench_obj():
bench = pbench.PoseBenchmark('car')
return bench
def save_evaluation_multiple_iters(exp): bench = pbench.PoseBenchmark(classes=PASCAL_CLS) numIter = range(8000,60000,4000) for n in numIter: save_evaluation(exp, n, bench=bench)
def save_evaluation_multiple(): bench = pbench.PoseBenchmark(classes=PASCAL_CLS) for num in range(7): exp, numIter = get_exp(num) save_evaluation(exp, numIter, bench=bench)
