mlmech

Structure

1. Pre-trained Encoder simple Decoder

Encoder:

• VGG16 (Julian)
• MobileNet_v2 (Lena)
• ResNet50_v2 (Samuel)

Test Case

• EarlyStopping/Shift = True
  validation_split=0.01
  
  • Single Class (A)
  • Multiclass (B)
    with 50 epochs
• EarlyStopping/Shift = False
  validation_split=0.1
  
  • Single Class (C)
  • Multiclass (D)
    with 100 epochs

Results MobileNet_v2

• A: results/lena/mobilenetV2/oneLabel/standard_V2
• B: results/lena/mobilenetV2/multiLabel/standard
• C: results/lena/mobilenetV2/oneLabel/withoutShift/noEarlyStopping/epochs100_validationSplit0-1
• D: results/lena/mobilenetV2/multiLabel/withoutShift/noEarlyStopping_epochs100_validationSplit0-1

Results VGG16

• A: results/julian/vgg16_1
• B: results/julian/vgg16_3
• C: results/julian/vgg16_7
• D. results/julian/vgg16_6

Results REsNEt50V2

• A: results/samuel/ResNet50V2_wS_SC_ES_50epochs
• B: results/samuel/ResNet50V2_wS_MC_ES_50epochs
• C: results/samuel/ResNet50V2_nS_SC_100epochs
• D. results/samuel/ResNet50V2_nS_MC_100epochs

Decoder:

Very simple only UpSampling2D and Conv2D

Other Stuff:

VGG16:

• results/julian/vgg16_5 all trainable

2. Simple Segmentation Architectures

• U-Net (self build architecture) (Julian)
  no shifts, early stopping applied (?), epochs=?
  
  • Single Class:
    • results/julian/unet_4; 256x256; epochs: 50; complexity: 4; EarlyStopping: false
    • results/julian/unet_256x3072; 256x3072; epochs: 20; complexity: 4; EarlyStopping: false
    • results/julian/unet_256x3072_2; 256x3072; epochs: 20; complexity: 3; EarlyStopping: false
    • results/julian/unet_256x3072_3; 256x3072; epochs: 20; complexity: 5; EarlyStopping: false
  • Multiclass
    • results/julian/unet_5; 256x256; epochs: 50; complexity: 5; EarlyStopping: false
    • results/julian/unet_256x3072_4; 256x3072; epochs: 50; complexity: 5; EarlyStopping: false
• SegNet (self build architecture) (Lena)
  no shifts, early stopping applied (?), epochs=?
  
  • Single Class
  • Multiclass
• FCN (Best Encoder with stronger Decoder) (Samuel)
  no shifts, early stopping applied (?), epochs=?
  
  • Single Class
  • Multi class
    Find out about the effect of dropout, batchnormalization and General average pooling
    Analyse in use with a pretrained encoder, only decoder tuned.

3. Pre-trained Encoder and Advanced Decoder in advanced architecture

Implement the best (or the fastes?) encoder in the best architecture (Unet/Segnet/FCN) with a stronger decoder.

Compare

• Loss
• Benchmark
• Validation Loss
• Opitcal Result

On Plus

• Big sized picture used in a self build architecture
• Use a front-net in front of a pretrained encoder in order
  to use the real picture size and also adjust the decoder for
  the real sized pictures. --> More parameters to learn with!

Presentation

Requirements

GPL Ghostscript: https://www.ghostscript.com/download/gsdnld.html

pstoedit: http://www.calvina.de/pstoedit/

Worklfow

1. Open MyTeXPoint.exe you can fin it here (a small window will open).
2. Open Presentation
3. Click on a text box (The window will increase)
4. Edit text
5. Compile (Strg + Shift + Enter)
6. Maybe you have to configurate then browse for the asked packages They should be found normally here:
   • gwin32c.exe: "C:\Prpgramm Files (x86)\gs\gs9.xx\bin\gswin32c.exe"
   • pstoedit.exe: "C:\Prpgramm Files\pstoedit\pstoedit.exe"
7. Then just click test and afterwars save

Issues

• If there is an error during compiling you have to close the temrinal before you can compile again