- images are now loaded for each batch instead
of all at the start in lambda_coco.py
- checkpoint location is now checked for existing
weights and loaded if available
- evaluation now pauses via pdb and correctly divides
the image sizes
- coco image requires a 'coco' folder with the COCO
'val2014' and 'annotations' folder downloaded from the
cocodataset website
- the script splits up the dataset by snipping the bounding-
box from the images and saving it into a seperate folder
for later parsing by the siamese network
- worse performance than with initial design
- vgg16 pretrained weights are used for the base
network, which is then piped into a custom head
model, which
- flattens the layer (previously done in the base model)
+ Dense Layer
+ Normalization
+ Activation
- training split with 360 fruits used, same as previous mode
- maximum prediction level around 0.95 after ca 60 epochs
