This source code mimics the googlenet that won the ILSVRC14 challenge with slight modifications in the fully connected layers in all the 3 outputs to accomodate the use case described below.
Hardware Peak Utilisation:
- CPU: 32 core
- Virtual Memory: 25 GB
- RES Memory: 15 GB
- Runtime: ~35 mins
software:
- python: 3.5.2
- tensorflow:2.3.0
Data Set:
- source: Kaggle
- userid: puneet6060
- data-set: intel-image-classification
Normenclature:
- bell1: The bottom most fully connected FC NN emerging from the inception module 4a as described in the googlnet incarnation table.
- bell2: The middle fully connecneted FC NN emerging from the output of inception module 4d as described in the googlnet incarnation table.
- mainbell: The top most fully connecneted FC NN emerging from the last inception module 5b as described in the googlnet incarnation table.
Optimiser: Adam
- leraning rate: 0.001 decaying by 50% after each epoch
- beta1(first moment): 0.09
- beta2 (second moment):0.999
- loss function: SparseCategoricalCrossentropy with logits
- epoch: 3
Observations:
- epoch 1: Accuracy for both training and validation is in the sequence bell1 > bell2 > mainbell
- epoch 2: Accuracy for both training and validation is in the sequence bell2 >~ bell1 >~ mainbell
- epoch 3: Accuracy for both training and validation is in the sequence bell2 =~ bell1 =~ mainbell
"~" represents: with a very tiny margin.
Accuracy:
Loss:
Note:
- All the FC NN have been modified to accomodate this use case.
- The Conv and Pooling layers remain as described in the googlenet paper.