Check correcness - Githubissues

anagainaru / medical_DNN

Study patterns for medical AI applications and possible optimizations for deployment on HPC systems

0 stars 0 forks source link

Closed anagainaru closed 2 years ago

anagainaru commented 2 years ago

For ADIOS

Patch extraction results
- Total number of patches is the same?
- In the classification code read the ADIOS variable and the initial png file and compare each byte.

For Pytorch

Pytorch vs Tenserflow for both the initial code and ADOIS
Classification results should be similar (should give similar number of detection patches)
```
cat patch-level-lym.txt | grep -v "0.0" | wc -l
```
The patches identified inside patch-level-lym.txt should be around the same areas.

anagainaru commented 2 years ago

The acceptable threshold for each model are:

0.5 MENNDL
0.42 VGG16
0.1 Inception

awk '{if ($3>0.5) print $0}' patch-level-lym.txt | wc -l

The path-level file has location probability x y prob [necrosis]

Use the dice coefficient