pixelNum = torch.sum(segBatch ).cpu().data.item()
for m in range(0, n + 1):
directErrs.append( torch.sum( (imPBatch - imP1Batch)
* (imPBatch - imP1Batch) * segBatch.expand_as(imPBatch) ) / pixelNum / 3.0 )
globalIllu2Errs.append( torch.sum( (globalIllu2s[m] - imP2Batch)
* (globalIllu2s[m] - imP2Batch) * segBatch.expand_as(imP2Batch) ) / pixelNum / 3.0 )
globalIllu3Errs.append(torch.sum( (globalIllu3s[m] - imP3Batch)
* (globalIllu3s[m] - imP3Batch) * segBatch.expand_as(imP3Batch) ) / pixelNum / 3.0 )
renderErrs.append( torch.sum( (renderImgs[m] - imPBatch)
* (renderImgs[m] - imPBatch) * segBatch.expand_as(imPBatch) ) / pixelNum / 3.0 )
renderGtErrs.append( torch.sum( (renderImgGts[m] - imPBatch)
* (renderImgGts[m] - imPBatch) * segBatch.expand_as(imPBatch) ) / pixelNum / 3.0 )
# Output testing error
utils.writeErrToScreen('directErrs', directErrs, epoch, j)
utils.writeErrToScreen('globalIllu2', globalIllu2Errs, epoch, j)
utils.writeErrToScreen('globalIllu3', globalIllu3Errs, epoch, j)
utils.writeErrToScreen('renderErrs', renderErrs, epoch, j)
utils.writeErrToScreen('renderGtErrs', renderGtErrs, epoch, j)
utils.writeErrToFile('directErrs', directErrs, testingLog, epoch, j)
utils.writeErrToFile('globalIllu2', globalIllu2Errs, testingLog, epoch, j)
utils.writeErrToFile('globalIllu3', globalIllu3Errs, testingLog, epoch, j)
utils.writeErrToFile('renderErrs', renderErrs, testingLog, epoch, j)
utils.writeErrToFile('renderGtErrs', renderGtErrs, testingLog, epoch, j)
directErrsNpList = np.concatenate( [directErrsNpList, utils.turnErrorIntoNumpy(directErrs) ], axis=0)
globalIllu2ErrsNpList = np.concatenate( [globalIllu2ErrsNpList, utils.turnErrorIntoNumpy(globalIllu2Errs)], axis=0)
globalIllu3ErrsNpList = np.concatenate( [globalIllu3ErrsNpList, utils.turnErrorIntoNumpy(globalIllu3Errs)], axis=0)
renderErrsNpList = np.concatenate( [renderErrsNpList, utils.turnErrorIntoNumpy(renderErrs)], axis=0 )
torch.sum(gtToPredNormal * gtNormal, dim=1), -1,
1)))
medianAngle_predToGt = 180.0 / np.pi * torch.median(
torch.acos(
torch.clamp(
torch.sum(normalPreds[m] * predToGtNormal, dim=1),
-1, 1)))
medianAngleErrs.append(
0.5 * (medianAngle_gtToPred + medianAngle_predToGt))
errorTotal = pointErrs[-1] * pw + normalErrs[-1] * nw
errorTotal.backward()
opPointNet.step()
utils.writeErrToScreen('point', pointErrs, epoch, j)
utils.writeErrToScreen('normal', normalErrs, epoch, j)
utils.writeErrToScreen('meanAngle', meanAngleErrs, epoch, j)
utils.writeErrToScreen('medianAngle', medianAngleErrs, epoch, j)
utils.writeErrToFile('point', pointErrs, trainingLog, epoch, j)
utils.writeErrToFile('normal', normalErrs, trainingLog, epoch, j)
utils.writeErrToFile('meanAngle', meanAngleErrs, trainingLog, epoch, j)
utils.writeErrToFile('medianAngle', medianAngleErrs, trainingLog,
epoch, j)
pointErrsNpList = np.concatenate(
[pointErrsNpList,
utils.turnErrorIntoNumpy(pointErrs)], axis=0)
normalErrsNpList = np.concatenate(
[normalErrsNpList,
depthErrSum = sum(depthErrs)
globalIllu1ErrSum = sum(globalIllu1Errs)
totalErr = albeW * albedoErrSum + normW * normalErrSum + rougW *roughErrSum \
+ deptW * depthErrSum + g1W * globalIllu1ErrSum
totalErr.backward()
# Update the network parameter
opEncoderInit.step()
opAlbedoInit.step()
opNormalInit.step()
opRoughInit.step()
opDepthInit.step()
# Output training error
utils.writeErrToScreen('albedo', albedoErrs, epoch, j)
utils.writeErrToScreen('normal', normalErrs, epoch, j)
utils.writeErrToScreen('rough', roughErrs, epoch, j)
utils.writeErrToScreen('depth', depthErrs, epoch, j)
utils.writeErrToScreen('globalIllu1', globalIllu1Errs, epoch, j)
utils.writeErrToFile('albedo', albedoErrs, trainingLog, epoch, j)
utils.writeErrToFile('normal', normalErrs, trainingLog, epoch, j)
utils.writeErrToFile('rough', roughErrs, trainingLog, epoch, j)
utils.writeErrToFile('depth', depthErrs, trainingLog, epoch, j)
utils.writeErrToFile('globalIllu1', globalIllu1Errs, trainingLog,
epoch, j)
albedoErrsNpList = np.concatenate(
[albedoErrsNpList,
utils.turnErrorIntoNumpy(albedoErrs)], axis=0)
normalErrsNpList = np.concatenate(
[normalErrsNpList,
torch.sum((globalIllu3s[m] - 0.5 * (imP3Batch + 1)) *
(globalIllu3s[m] - 0.5 *
(imP3Batch + 1)) * segBatch.expand_as(imBatch)) /
pixelNum / 3.0)
for m in range(0, len(renderedEnvs)):
imgEnvErrs.append(
torch.sum((renderedEnvs[m] - 0.5 * (imEBatch + 1)) *
(renderedEnvs[m] - 0.5 *
(imEBatch + 1)) * segBatch.expand_as(imBatch)) /
pixelNum / 3.0)
for m in range(0, len(SHPreds)):
envErrs.append(
torch.mean((SHPreds[m] - SHBatch) * (SHPreds[m] - SHBatch)))
# Output testing error
utils.writeErrToScreen('albedo', albedoErrs, epoch, j)
utils.writeErrToScreen('normal', normalErrs, epoch, j)
utils.writeErrToScreen('rough', roughErrs, epoch, j)
utils.writeErrToScreen('depth', depthErrs, epoch, j)
utils.writeErrToScreen('globalIllu1', globalIllu1Errs, epoch, j)
utils.writeErrToScreen('globalIllu2', globalIllu2Errs, epoch, j)
utils.writeErrToScreen('globalIllu3', globalIllu3Errs, epoch, j)
utils.writeErrToScreen('imgEnv', imgEnvErrs, epoch, j)
utils.writeErrToScreen('env', envErrs, epoch, j)
utils.writeErrToFile('albedo', albedoErrs, testingLog, epoch, j)
utils.writeErrToFile('normal', normalErrs, testingLog, epoch, j)
utils.writeErrToFile('rough', roughErrs, testingLog, epoch, j)
utils.writeErrToFile('depth', depthErrs, testingLog, epoch, j)
utils.writeErrToFile('globalIllu1', globalIllu1Errs, testingLog, epoch, j)
utils.writeErrToFile('globalIllu2', globalIllu2Errs, testingLog, epoch, j)
globalIllu3Errs.append(
torch.sum((globalIllu3s[m] - imP3Batch) *
(globalIllu3s[m] - imP3Batch) *
segBatch.expand_as(imP3Batch)) / pixelNum / 3.0)
globalIllu2ErrSum = sum(globalIllu2Errs)
globalIllu3ErrSum = sum(globalIllu3Errs)
totalErr = g2W * globalIllu2ErrSum + g3W * globalIllu3ErrSum
totalErr.backward()
opGlobalIllu1to2.step()
opGlobalIllu2to3.step()
# Output training error
utils.writeErrToScreen('globalIllu2', globalIllu2Errs, epoch, j)
utils.writeErrToScreen('globalIllu3', globalIllu3Errs, epoch, j)
utils.writeErrToFile('globalIllu2', globalIllu2Errs, trainingLog,
epoch, j)
utils.writeErrToFile('globalIllu3', globalIllu3Errs, trainingLog,
epoch, j)
globalIllu2ErrsNpList = np.concatenate(
[globalIllu2ErrsNpList,
utils.turnErrorIntoNumpy(globalIllu2Errs)],
axis=0)
globalIllu3ErrsNpList = np.concatenate(
[globalIllu3ErrsNpList,
utils.turnErrorIntoNumpy(globalIllu3Errs)],
axis=0)
if j < 1000:
depthPred, depthErr = depthPair[0], depthPair[1]
# Back propagate the gradients
totalErr = 4 * albeW * albedoErr + normW * normalErr \
+ rougW *roughErr + deptW * depthErr
totalErr.backward()
# Update the network parameter
opEncoder.step()
opAlbedo.step()
opNormal.step()
opRough.step()
opDepth.step()
# Output training error
utils.writeErrToScreen('albedo', [albedoErr], epoch, j )
utils.writeErrToScreen('normal', [normalErr], epoch, j )
utils.writeErrToScreen('rough', [roughErr], epoch, j)
utils.writeErrToScreen('depth', [depthErr], epoch, j)
utils.writeErrToFile('albedo', [albedoErr], trainingLog, epoch, j)
utils.writeErrToFile('normal', [normalErr], trainingLog, epoch, j)
utils.writeErrToFile('rough', [roughErr], trainingLog, epoch, j)
utils.writeErrToFile('depth', [depthErr], trainingLog, epoch, j)
albedoErrsNpList = np.concatenate( [albedoErrsNpList, utils.turnErrorIntoNumpy( [albedoErr] )], axis=0)
normalErrsNpList = np.concatenate( [normalErrsNpList, utils.turnErrorIntoNumpy( [normalErr] )], axis=0)
roughErrsNpList = np.concatenate( [roughErrsNpList, utils.turnErrorIntoNumpy( [roughErr] )], axis=0)
depthErrsNpList = np.concatenate( [depthErrsNpList, utils.turnErrorIntoNumpy( [depthErr] )], axis=0)
if j < 1000:
torch.acos(
torch.clamp(
torch.sum(normal2Preds[m] * normal2Batch, dim=1), -1,
1)[seg1Batch.squeeze(1) != 0])))
else:
medianAngle2Errs.append(0 * torch.sum(normal2Preds[m]))
for m in range(0, len(renderedImgs)):
renderedErrs.append(
torch.sum(
(renderedImgs[m] - imBgBatch) *
(renderedImgs[m] - imBgBatch) * seg2Batch.expand_as(imBatch)) /
pixel2Num / 3.0)
# Output training error
utils.writeErrToScreen('normal1', normal1Errs, epoch, j)
utils.writeErrToScreen('medianAngle1', medianAngle1Errs, epoch, j)
utils.writeErrToScreen('meanAngle1', meanAngle1Errs, epoch, j)
utils.writeErrToScreen('normal2', normal2Errs, epoch, j)
utils.writeErrToScreen('medianAngle2', medianAngle2Errs, epoch, j)
utils.writeErrToScreen('meanAngle2', meanAngle2Errs, epoch, j)
utils.writeErrToScreen('rendered', renderedErrs, epoch, j)
utils.writeErrToFile('normal1', normal1Errs, trainingLog, epoch, j)
utils.writeErrToFile('medianAngle1', medianAngle1Errs, trainingLog, epoch,
j)
utils.writeErrToFile('meanAngle1', meanAngle1Errs, trainingLog, epoch, j)
utils.writeErrToFile('normal2', normal2Errs, trainingLog, epoch, j)
utils.writeErrToFile('medianAngle2', medianAngle2Errs, trainingLog, epoch,
j)
utils.writeErrToFile('meanAngle2', meanAngle2Errs, trainingLog, epoch, j)
depthPred, depthErr = depthPair[0], depthPair[1]
# Back propagate the gradients
totalErr = 4 * albeW * albedoErr + normW * normalErr \
+ rougW *roughErr + deptW * depthErr
totalErr.backward()
# Update the network parameter
opEncoder.step()
opAlbedo.step()
opNormal.step()
opRough.step()
opDepth.step()
# Output training error
utils.writeErrToScreen('albedo', [albedoErr], epoch, j)
utils.writeErrToScreen('normal', [normalErr], epoch, j)
utils.writeErrToScreen('rough', [roughErr], epoch, j)
utils.writeErrToScreen('depth', [depthErr], epoch, j)
utils.writeErrToFile('albedo', [albedoErr], trainingLog, epoch, j)
utils.writeErrToFile('normal', [normalErr], trainingLog, epoch, j)
utils.writeErrToFile('rough', [roughErr], trainingLog, epoch, j)
utils.writeErrToFile('depth', [depthErr], trainingLog, epoch, j)
albedoErrsNpList = np.concatenate( [albedoErrsNpList, utils.turnErrorIntoNumpy( [albedoErr] )], axis=0)
normalErrsNpList = np.concatenate( [normalErrsNpList, utils.turnErrorIntoNumpy( [normalErr] )], axis=0)
roughErrsNpList = np.concatenate( [roughErrsNpList, utils.turnErrorIntoNumpy( [roughErr] )], axis=0)
depthErrsNpList = np.concatenate( [depthErrsNpList, utils.turnErrorIntoNumpy( [depthErr] )], axis=0)
if j < 1000:
renderedImPred, renderErr = renderPair[0], renderPair[1]
imBatch = renderPair[2]
# Back propagate the gradients
totalErr = renW * renderErr + recW * reconstErr
totalErr.backward()
# Update the network parameter
opLightEncoder.step()
opAxisDecoder.step()
opLambDecoder.step()
opWeightDecoder.step()
# Output training error
utils.writeErrToScreen('albedo', [albedoErr], epoch, j)
utils.writeErrToScreen('normal', [normalErr], epoch, j)
utils.writeErrToScreen('rough', [roughErr], epoch, j)
utils.writeErrToScreen('depth', [depthErr], epoch, j)
utils.writeErrToScreen('reconstErrors', [reconstErr], epoch, j)
utils.writeErrToScreen('renderErrors', [renderErr], epoch, j)
utils.writeErrToFile('albedo', [albedoErr], trainingLog, epoch, j)
utils.writeErrToFile('normal', [normalErr], trainingLog, epoch, j)
utils.writeErrToFile('rough', [roughErr], trainingLog, epoch, j)
utils.writeErrToFile('depth', [depthErr], trainingLog, epoch, j)
utils.writeErrToFile('reconstErrors', [reconstErr], trainingLog, epoch, j)
utils.writeErrToFile('renderErrors', [renderErr], trainingLog, epoch, j)