def output(batchImages, batchPclA, batchPclB, batchtMatT, batchtMatP,
batchTFrecFileIDs, **kwargs):
"""
TODO: SIMILAR TO DATA INPUT -> WE NEED A QUEUE RUNNER TO WRITE THIS OFF TO BE FASTER
Everything evaluated
Warp second image based on predicted HAB and write to the new address
Args:
Returns:
Raises:
ValueError: If no dataDir
"""
for i in range(kwargs.get('activeBatchSize')):
# split for depth dimension
depthA, depthB = np.asarray(np.split(batchImages[i], 2, axis=2))
pclATransformed, tMatRes, depthATransformed = _apply_prediction(
batchPclA[i], batchtMatT[i], batchtMatP[i], **kwargs)
# Write each Tensorflow record
filename = str(batchTFrecFileIDs[i][0]) + "_" + str(
batchTFrecFileIDs[i][1]) + "_" + str(batchTFrecFileIDs[i][2])
tfrecord_io.tfrecord_writer(batchTFrecFileIDs[i], pclATransformed,
batchPclB[i], depthATransformed, depthB,
tMatRes,
kwargs.get('warpedOutputFolder') + '/',
filename)
return
def output_loop(batchImages, batchPclA, batchPclB, bargetT, targetP,
batchTFrecFileIDs, i, **kwargs):
"""
TODO: SIMILAR TO DATA INPUT -> WE NEED A QUEUE RUNNER TO WRITE THIS OFF TO BE FASTER
Everything evaluated
Warp second image based on predicted HAB and write to the new address
Args:
Returns:
Raises:
ValueError: If no dataDir
"""
# split for depth dimension
depthA, depthB = np.asarray(np.split(batchImages[i], 2, axis=2))
depthB = depthB.reshape(kwargs.get('imageDepthRows'),
kwargs.get('imageDepthCols'))
pclATransformed, targetRes, depthATransformed = _apply_prediction(
batchPclA[i], bargetT[i], targetP[i], **kwargs)
# Write each Tensorflow record
filename = str(batchTFrecFileIDs[i][0]) + "_" + str(
batchTFrecFileIDs[i][1]) + "_" + str(batchTFrecFileIDs[i][2])
tfrecord_io.tfrecord_writer(batchTFrecFileIDs[i], pclATransformed,
batchPclB[i], depthATransformed, depthB,
targetRes,
kwargs.get('warpedOutputFolder') + '/',
filename)
if kwargs.get('phase') == 'train':
folderTmat = kwargs.get('tMatTrainDir')
else:
folderTmat = kwargs.get('tMatTestDir')
write_predictions(batchTFrecFileIDs[i], targetP[i], folderTmat)
return
def output_with_test_image_files(batchImageOrig, batchImage, batchPOrig, batchTHAB, batchPHAB, batchTFrecFileIDs, **kwargs):
"""
TODO: SIMILAR TO DATA INPUT -> WE NEED A QUEUE RUNNER TO WRITE THIS OFF TO BE FASTER
Everything evaluated
Warp second image based on predicted HAB and write to the new address
Args:
Returns:
Raises:
ValueError: If no dataDir
"""
imagesOutputFolder = kwargs.get('testLogDir')+'/images/'
_set_folders(imagesOutputFolder)
## for each value call the record writer function
corrupt_patchOrig = 0
corrupt_patchPert = 0
corrupt_imageOrig = 0
dataJson = {'pOrig':[], 'tHAB':[], 'pHAB':[]}
for i in range(kwargs.get('activeBatchSize')):
dataJson['pOrig'] = batchPOrig[i].tolist()
dataJson['tHAB'] = batchTHAB[i].tolist()
dataJson['pHAB'] = batchPHAB[i].tolist()
write_json_file(imagesOutputFolder+str(batchTFrecFileIDs[i][0])+'_'+str(batchTFrecFileIDs[i][1]), dataJson)
cv2.imwrite(imagesOutputFolder+str(batchTFrecFileIDs[i][0])+'_'+ str(batchTFrecFileIDs[i][1])+'_fullOrig.jpg', batchImageOrig[i]*255)
orig, pert = np.asarray(np.split(batchImage[i], 2, axis=2))
cv2.imwrite(imagesOutputFolder+str(batchTFrecFileIDs[i][0])+'_'+str(batchTFrecFileIDs[i][1])+'_ob.jpg', orig*255)
cv2.imwrite(imagesOutputFolder+str(batchTFrecFileIDs[i][0])+'_'+str(batchTFrecFileIDs[i][1])+'_pb.jpg', pert*255)
# Get the difference of tHAB and pHAB, and make new perturbed image based on that
cHAB = batchTHAB[i]-batchPHAB[i]
# put them in correct form
HAB = np.asarray([[cHAB[0], cHAB[1], cHAB[2], cHAB[3]],
[cHAB[4], cHAB[5], cHAB[6], cHAB[7]]], np.float32)
pOrig = np.asarray([[batchPOrig[i][0], batchPOrig[i][1], batchPOrig[i][2], batchPOrig[i][3]],
[batchPOrig[i][4], batchPOrig[i][5], batchPOrig[i][6], batchPOrig[i][7]]])
if kwargs.get('warpOriginalImage'):
patchOrig, patchPert = _warp_w_orig_newTarget(batchImageOrig[i], batchImage[i], pOrig, HAB, **kwargs)
# NOT DEVELOPED YET
#imageOrig, imagePert = _warp_w_orig_newOrig(batchImageOrig[i], batchImage[i], pOrig, batchPHAB[i], **kwargs)
else:
patchOrig, patchPert = _warp_wOut_orig_newTarget(batchImage[i], batchPHAB[i])
cv2.imwrite(imagesOutputFolder+str(batchTFrecFileIDs[i][0])+'_'+str(batchTFrecFileIDs[i][1])+'_op.jpg', patchOrig*255)
cv2.imwrite(imagesOutputFolder+str(batchTFrecFileIDs[i][0])+'_'+str(batchTFrecFileIDs[i][1])+'_pp.jpg', patchPert*255)
# Write each Tensorflow record
fileIDs = str(batchTFrecFileIDs[i][0]) + '_' + str(batchTFrecFileIDs[i][1])
tfrecord_io.tfrecord_writer(batchImageOrig[i], patchOrig, patchPert, pOrig, HAB,
kwargs.get('warpedOutputFolder')+'/',
fileIDs, batchTFrecFileIDs[i])
if batchImageOrig[i].shape[0] != 240:
corrupt_imageOrig+=1
if patchOrig.shape[0]!=128:
corrupt_patchOrig+=1
if patchPert.shape[0]!=128:
corrupt_patchPert+=1
return corrupt_imageOrig, corrupt_patchOrig, corrupt_patchPert
def output(batchImageOrig, batchImage, batchPOrig, batchTHAB, batchPHAB, batchPrevPredHAB, batchTFrecFileIDs, **kwargs):
"""
TODO: SIMILAR TO DATA INPUT -> WE NEED A QUEUE RUNNER TO WRITE THIS OFF TO BE FASTER
Everything evaluated
Warp second image based on predicted HAB and write to the new address
Args:
Returns:
Raises:
ValueError: If no dataDir
"""
## for each value call the record writer function
corrupt_patchOrig=0
corrupt_patchPert=0
corrupt_imageOrig=0
for i in range(kwargs.get('activeBatchSize')):
##### THIS IS THE ADDED PREDICTION SUM
cHAB = batchPrevPredHAB[i]+batchPHAB[i]
# put them in correct form
additivePredHAB = np.asarray([[cHAB[0], cHAB[1], cHAB[2], cHAB[3]],
[cHAB[4], cHAB[5], cHAB[6], cHAB[7]]], np.float32)
pOrig = np.asarray([[batchPOrig[i][0], batchPOrig[i][1], batchPOrig[i][2], batchPOrig[i][3]],
[batchPOrig[i][4], batchPOrig[i][5], batchPOrig[i][6], batchPOrig[i][7]]])
##### THIS IS THE CORRECT FUNCTION USING ADDITIVE PREDICTION ON ORIGINAL NON CHANGED IMGORIG
### PATCH ORIG MOVES TOWARDS PATCH PERT
### patchOrig is only changing, patchPert is never changing
patchOrig, patchPert = _warp_w_orig_newOrig(batchImageOrig[i], batchImage[i], pOrig, additivePredHAB, **kwargs)
#### THIS HAB IS THE NEW TARGET HAB
# Get the difference of tHAB and pHAB, and make new perturbed image based on that
cHAB = batchTHAB[i]-batchPHAB[i]
# put them in correct form
newTargetHAB = np.asarray([[cHAB[0], cHAB[1], cHAB[2], cHAB[3]],
[cHAB[4], cHAB[5], cHAB[6], cHAB[7]]], np.float32)
# Write each Tensorflow record
fileIDs = str(batchTFrecFileIDs[i][0]) + '_' + str(batchTFrecFileIDs[i][1])
tfrecord_io.tfrecord_writer(batchImageOrig[i], patchOrig, patchPert, pOrig, newTargetHAB, additivePredHAB,
kwargs.get('warpedOutputFolder')+'/',
fileIDs, batchTFrecFileIDs[i])
if batchImageOrig[i].shape[0] != 240:
corrupt_imageOrig+=1
if patchOrig.shape[0]!=128:
corrupt_patchOrig+=1
if patchPert.shape[0]!=128:
corrupt_patchPert+=1
return corrupt_imageOrig, corrupt_patchOrig, corrupt_patchPert
def output(batchImageOrig, batchImage, batchPOrig, batchTHAB, batchPHAB, batchTFrecFileIDs, **kwargs):
"""
TODO: SIMILAR TO DATA INPUT -> WE NEED A QUEUE RUNNER TO WRITE THIS OFF TO BE FASTER
Everything evaluated
Warp second image based on predicted HAB and write to the new address
Args:
Returns:
Raises:
ValueError: If no dataDir
"""
warpedImageFolder = kwargs.get('warpedOutputFolder')+'/'
## for each value call the record writer function
corrupt_patchOrig=0
corrupt_patchPert=0
corrupt_imageOrig=0
for i in range(kwargs.get('activeBatchSize')):
# Get the difference of tHAB and pHAB, and make new perturbed image based on that
cHAB = batchTHAB[i]-batchPHAB[i]
# put them in correct form
HAB = np.asarray([[cHAB[0], cHAB[1], cHAB[2], cHAB[3]],
[cHAB[4], cHAB[5], cHAB[6], cHAB[7]]], np.float32)
pOrig = np.asarray([[batchPOrig[i][0], batchPOrig[i][1], batchPOrig[i][2], batchPOrig[i][3]],
[batchPOrig[i][4], batchPOrig[i][5], batchPOrig[i][6], batchPOrig[i][7]]])
if kwargs.get('warpOriginalImage'):
patchOrig, patchPert = _warp_w_orig_newTarget(batchImageOrig[i], batchImage[i], pOrig, HAB, **kwargs)
# NOT DEVELOPED YET
#imageOrig, imagePert = _warp_w_orig_newOrig(batchImageOrig[i], batchImage[i], pOrig, batchPHAB[i], **kwargs)
else:
patchOrig, patchPert = _warp_wOut_orig_newTarget(batchImage[i], batchPHAB[i])
# Write each Tensorflow record
fileIDs = str(batchTFrecFileIDs[i][0]) + '_' + str(batchTFrecFileIDs[i][1])
tfrecord_io.tfrecord_writer(batchImageOrig[i], patchOrig, patchPert, pOrig, HAB,
warpedImageFolder,
fileIDs, batchTFrecFileIDs[i])
if batchImageOrig[i].shape[0] != 240:
corrupt_imageOrig+=1
if patchOrig.shape[0]!=128:
corrupt_patchOrig+=1
if patchPert.shape[0]!=128:
corrupt_patchPert+=1
return corrupt_imageOrig, corrupt_patchOrig, corrupt_patchPert
def perturb_writer(ID, idx, imgOrig, imgPatchOrig, imgPatchPert, HAB, pOrig,
tfRecFolder):
##### original patch
#filename = filenameWrite.replace(".jpg", "_"+ str(idx) +"_orig.jpg")
#cv2.imwrite(patchFolder+filename, imgPatchOrig)
##### perturbed patch
#filename = filenameWrite.replace(".jpg", "_"+ str(idx) +"_pert.jpg")
#cv2.imwrite(patchFolder+filename, imgPatchPert)
##### HAB
#filename = filenameWrite.replace(".jpg", "_"+ str(idx) +"_HAB.csv")
#with open(HABFolder+filename, 'w', newline='') as f:
# writer = csv.writer(f)
# if (HAB.ndim > 1):
# writer.writerows(HAB)
# else:
# writer.writerow(HAB)
##### Original square
#filename = filenameWrite.replace(".jpg", "_"+ str(idx) +".csv")
#with open(squareFolder+filename, 'w', newline='') as f:
# writer = csv.writer(f)
# if (pOrig.ndim > 1):
# writer.writerows(pOrig)
# else:
# writer.writerow(pOrig)
# Tensorflow record
filename = str(ID) + "_" + str(idx)
fileID = [ID, idx]
tfrecord_io.tfrecord_writer(imgOrig, imgPatchOrig, imgPatchPert, pOrig,
HAB, HAB * 0, tfRecFolder, filename, fileID)
#imgOp = image_process_subMean_divStd(imgPatchOrig)
#imgPp = image_process_subMean_divStd(imgPatchPert)
#tfrecord_writer(imgOp, imgPp, HAB, pOrig, tfRecFolder+filename)
# Tensorflow record in range -1 and 1
#filename = filenameWrite.replace(".jpg", "_"+ str(idx))
#imgOp = image_process_subMean_divStd_n1p1(imgPatchOrig)
#imgPp = image_process_subMean_divStd_n1p1(imgPatchPert)
#tfrecord_writer(imgOp, imgPp, HAB, pOrig, tfRecFolderN1P1+filename)
return