# Changes in v2.3.0
## License authentication
Along with the MANUFACIA pricing change, license authentication was implemented to limit usage after a certain period. After installation, it is necessary to input license information sent by MANUFACIA support in the initial page. See also [Input license information](../installation_manual_gpu/index_3.html#input-license-information) for further information.
## Support Ubuntu 20.04 LTS
Support of Ubuntu 18.04 LTS will be finished in April 2023, therefore, the newer version Ubuntu 20.04 LTS is now a recommended OS for MANUFACIA. It is still possible to use MANUFACIA on Ubuntu 18.04 LTS, but it is recommended to update to Ubuntu 20.04 LTS for this MANUFACIA release.
## Bug fix: Link to project overview of a lab disappears once going inside of a project
Next to MANUFACIA icon, there should be a link to top page of a lab, and also a project. Once going inside a project to see details, the link to top page of a lab disappears and users have to once click on MANUFACIA icon, then again choose the lab where they were. This bug has been fixed in this version.
# Changes in v2.2.0
## Introduced Matthews correlation coefficient for model performance index.
To evaluate AI models, it is not easy for non AI experts to evaluate several indices at the same time with care, such as datasets balance. In combination of accuracy and F-measure (precision/recall), there might be cases that they might get high score for a useless network.
Matthews correlation coefficient (MCC) which was introduced in this version takes the value from -1.0 to 1.0 for binary classification (anomaly detection). 1.0 represents the perfect prediction of the network, -1.0 means AI does 100% opposite, 0.0 something like random selection. For multi-class classification more than two classes, the lowest value can vary from -1.0 to 0.0 while the highest value to represent the perfect network is unchanged as 1.0.
Performance index of model reevaluated and "accuracy predict count" and "anomaly rate" were removed from the right of confusion matrix in model detail page. For further information about the meaning of each performance index, see [Users manual -> Performance index of neural network](../usermanual/index_4.html#performance-index-neural-network).
## Confusion matrix.
Improved how to display confusion matrix, especially for multi-class classification it will be automatically scaled to have better appearance. Before, it was hard to tell which matrix element is false positive or negative, the color of the title was changed to brighter one and the wrong user manual description was corrected.
See also [Users Manual -> Confusion matrix](../usermanual/index_4.html#confusion-matrix).
## Training loss curve, ROC curve and AUC.
Some charts were implemented in this version for a better understanding of the training. See also [Users manual -> Training loss curve](../usermanual/index_4.html#training-loss-curve) and [Users manual -> ROC curve and AUC](../usermanual/index_4.html#roc-curve)
## Model detail report function.
Most of the users are interested in sample detail to know how AI evaluated data correctly or for wrong. However, it took time and effort to click and open each interesting sample detail page to see how the visualized samples look like.
To reduce this effort, the report function was implemented.
In the model list page, now a new button "Print report of favorite models" is available to create a report of models marked as favorite in the table above in a new browser tab.
The report consists of the following items.
- Lab name and overview
- Project name and description
- Model ID
- Preprocessing setting detail
- Model performance indices
- Confusion matrix
- Visualized sample images
- Training loss curve
- ROC curve and AUC (in the area Model performance indices)
It can be printed out, or saved in PDF format from a browser.
## ZIP file upload for assets.
Assets can now be uploaded with a zipped file. It will make the process faster by about 20%.
Directory structure in the zip file will be fully ignored, its behavior is not guaranteed if there are files with the same name in different directories in the zip file.
Tag will not be automatically given by the system, users have to add tag name as same as for the case to select files to create an asset.
Moreover, it cannot be unzipped if zip file entry names are not encoded in UTF-8. For more information, see [User manual -> Register assets](../usermanual/index_2.html#register-assets).
## Improved threshold / distance calculation for binary classification.
For binary classification, the default threshold or distance was not always properly calculated and users had to take efforts to adjust it by hand. In this version, the default value will be calculated as shown below with the help of ROC curve which was also implemented in this version.
The default threshold / distance is now calculated as the point on the ROC curve that minimizes the function $FPR^2 + FNR^2$, where FPR stands for **F**alse **P**ositive **R**ate, FNR for **F**alse **N**egative **R**ate.
## Histogram improvement
Previous histogram just shows the validation results as distance changes. In the new version, histogram was improved to show more information including distribution of actual label. Its sample color now matches the one of confusion matrix: green for correct prediction, red for incorrect prediction. See details in [User manual -> Histogram](../usermanual/index_4.html#histogram)
## Histogram for supervised learning
Histogram was available only for unsupervised learning before. After implementing ROC curve,threshold and distance are now calculated with the use of ROC curve. As same as unsupervised, histogram is now available for supervised learning.
## Added a button to reset threshold / distance to the default setting
Threshold or distance can be modified after a training by slider, a button was newly added to reset to the default value.
## Speed up SmoothGrad calculation.
Achieved speed-up to process SmoothGrad image by enabling GPU usage.
## Greenia SDK limited for inference and acquiring results.
The SDK (Python library) were realized to infer or get results that is supposed to run on the server in which MANUFACIA is installed. It is different from Greenia Embedded SDK and is **NOT supposed to use on the edge device**.
## Updated Nvidia-Docker2/docker-compose versions.
Updated Nvidia-Docker2, docker-compose versions for maintenance reason. For further detail, see [Install manual -> Recommended system requirements](installation_manual_gpu/index_1.html#recommended-system-requirements)
## Renamed text "test" to "validation".
In MANUFACIA UI, it was written as "test" in place of using validation datasets. All those texts have been changed to "validation" except for reserved tag name "test", which may cause compatibility problem.
## MANUFACIA version number.
MANUFACIA version number plus build ID will now be displayed in the pop-up window on MANUFACIA logo top left.
## Bug fix: Training with vibration datasets fails in unsupervised learning.
A case was found that the training with vibration datasets failed in unsupervised learning, although the training with the same datasets in supervised learning finishes without problem. This problem was fixed in this version.
## Bug fix: Inappropriate handling of models with threshold/distance of NaN or Inf
Even if a model's threshold/distance during the training becomes NaN (Not a Number) or Inf (Infinity), it was evaluated as successful and was displayed in the scatter diagram or in model list. Trying to see the model detail, it was displayed "Failed to update". The behavior was corrected so that these models will be treated as training-failed models.
## Bug fix: Wrong sample detail was displayed
To select a sample in the model detail UI and show its detail, sometimes a wrong result was displayed. It happened when browsing models using <> keys and the internal variable to hold the current model ID was not properly overwritten and the incorrect model ID was used to show the sample detail. This bug was fixed in this version.
## Bug fix: Wrong percentage was displayed in labeling assets UI
The percentage of using datasets to the label was wrong and non-zero number was displayed, even if there is no datasets that will fulfill the condition of multiple tags at the same time that were set to the label. This problem was fixed in this version.
## Bug fix: Favorite model flag handling.
Favorite model flag was not properly handled while browsing models with <> keys or sometimes its state will not be saved. This bug was fixed in this version.
## Bug fix: Missing on-mouse binding behavior in the confusion matrix. (multi-class classification)
Having 20 labels in multi-class classification, the feature to highlight row and column of the confusion matrix at the position of mouse pointer, will not work at the right bottom corner. This problem was solved in this version.
## Bug fix: Percentage of assets displayed for each label in label assets UI.
In the label assets UI, the assets percentage that are linked to a label will be displayed. In the case that more then one tags are linked to a label and the overlapping assets of all tags is zero, the percentage non-zero value will be displayed, which is a major cause of training error.
In this version, the problem has been fixed to show the correct percentage of assets that will be actually used for training, so that users can see if the percentage linked to a label is not zero.
# Changes in v2.1.3
## Updated CUDA to version 11
This allows MANUFACIA to be used with recent graphics cards and recent Nvidia driver versions.
# Changes in v2.1.2
## Bug fix: Project training result will not be displayed.
If there is any model whose training failed, MANUFACIA was not able to display the result and the screen would be totally white. This problem which appeared in v2.1.1 was fixed in v2.1.2.
# Changes in v2.1.1
## Accelerated training.
In v2.1, new network creation algorithm was implemented to optimize network size, but it took longer time to create compared to the previous version v2.0. In v2.1.1, internal communication between modules was optimized and the training has been accelerated. Beside these, with the help of this change, GPU will be used more efficiently.
## Improved SmoothGrad display.
By changing how to display SmoothGrad, unnecessary grid will not be shown anymore.
## Improved behavior to upload assets.
If many files are selected to upload assets, sometimes the browser got stuck. This problem has been solved by handling thumbnails asynchronously.
## Changed initial axis parameter setting in model evaluation graph.
Initial axis parameter setting in model evaluation graph was changed to “network size / accuracy” which seems to be more important to users.
## Changed axis scale for network size.
Axis scale for network size was changed to logarithmic, so that scattering of CNN models can be well seen with models from ResNet or MobileNet that are much bigger than them.
## Added model creation time for each in model evaluation table.
Model creation time was added in the model evaluation table.
## Changed the preprocessing sequence of vibration datasets.
Standardization/normalization was moved to in front of FFT process in preprocessing chain.
## Changed preprocessing default setting.
Interpolate option was switched off for default for time-series and vibration datasets.
## MANUFACIA icon.
Updated MANUFACIA icon on the top left.
## Bug fix: Although install-manufacia command fails, it will be displayed in console that it was successful.
Even if docker is not properly installed, install-manufacia command shows in console that the installation was successful. In v2.1.1 this bug was fixed.
## Bug fix: Browsing models with “<>” keys in the model result page from unsupervised learning, histogram will be switched to confusion matrix.
While browsing model result after an unsupervised training. Histogram was switched to confusion matrix, whenever user tries to show the result of previous/next model by pressing “<” or “>” button. This behavior was fixed in v2.1.1.
# Changes in v2.1
## ResNet50 and MobileNetV2 models were implemented for MANUFACIA 2D image recognition.
To improve the accuracy of image recognition in supervised learning, pre-trained ResNet50 (50 layers) and MobileNetV2 (53 layers) were additionally implemented to the current specification. To use the new neural networks to generate AI models, image dataset size should be larger than 224x224.
Considering the balance of processing time and the performance, maximum size is now limited to 512 x 512. And then abundant preprocessing scaling is now not allowed to activate.
These two neural networks will be used in priority to the existing neural network. By setting the number of AI models to two, only the models by ResNet50 and MobileNetV2 will be created.
With this change, the cartridge format has been changed and the V2.1 does not keep backward compatibility to any older versions.
## License authentication has been changed.
To be able to install MANUFACIA into a server without the Internet connection, licensing system has been changed.
## Network size was reduced.
The generated neural network model size by previous version was unnecessarily big. In this version, the network generation module has been improved to put out the model in appropriate size for all time-series, vibration, and image datasets.
## Added support 8-bit depth image file.
On the previous version, it was not possible to use 8-bit depth image file. The support for this image file was added in Ver.2.1. It is confirmed that inference on edge device works without any problem with this image format.
## CUDNN flag was enabled to accelerate training.
Enabled CUDNN flag in Caffe2 to efficiently use GPU(CUDA) and speed up the process.
## Bug Fix: manufacia-ctrl update will not always work.
manufacia-ctrl update command was removed from the command list. In place of this update command, the system should be first uninstalled and then be newly installed with “manufacia-ctrl start” command. The new version 2.1 does not have backward compatibility and should be clean installed.
## Bug Fix: "manufacia-ctrl start" command creates an empty file called "1" every time it is run.
manufacia-ctrl start command created empty files called "1" in the current directory every time it is run. This bug has been resolved in the version 2.1.