CS7319_project

This is the project code for CS 7319

Experiment Result

2 reconstruction performance

===================== AE ========================
weight analysis start:
weight analysis: fc[0].weight and dec_fc[0].weight:
transpose distance:  10.27124
inverse distance:  6.3389487
weight analysis: fc[1].weight and dec_fc[1].weight:
transpose distance:  2831.982
inverse distance:  4351.746
weight analysis: fc[2].weight and dec_fc[2].weight:
transpose distance:  1.0128667
inverse distance:  0.35673222

=====================Lmser with DPN ==========================
weight analysis start:
weight analysis: fc[0].weight and dec_fc[0].weight:
transpose distance:  0.0007783578
inverse distance:  0.0
weight analysis: fc[1].weight and dec_fc[1].weight:
transpose distance:  0.0009829665
inverse distance:  0.0
weight analysis: fc[2].weight and dec_fc[2].weight:
transpose distance:  4.3793516e-06
inverse distance:  0.0

=====================Lmser with DCW ========================
weight analysis start:
weight analysis: fc[0].weight[] and dec_fc[0].weight:
transpose distance:  0.0
inverse distance:  9.055498
weight analysis: fc[1].weight[] and dec_fc[1].weight:
transpose distance:  0.0
inverse distance:  565.1025
weight analysis: fc[2].weight[] and dec_fc[2].weight:
transpose distance:  0.0
inverse distance:  0.826585

3 reconstruction in few-shot learning scenarios

we sample a small train epoch from MNIST/F-MNIST, each epoch contain num_batch batches, n_per_batch samples per batch, way class per batch. We create the following classic few-shot learning scenario.

5-way, n-shot: num_batch=100, n_per_batch=n, way=5

other training parameters epoch=10

train loss

scenarios	AE	Lmser(DPN)	Lmser(DCW)
5-way 1-shot	0.60	0.44	0.41
5-way 2-shot	0.59	0.44	0.41
5-way 3-shot	0.58	0.42	0.41
5-way 4-shot	0.58	0.43	0.41
5-way 5-shot	0.57	0.41	0.41

test loss

scenarios	AE	Lmser(DPN)	Lmser(DCW)
5-way 1-shot	0.60	0.44	0.40
5-way 2-shot	0.59	0.44	0.40
5-way 3-shot	0.57	0.42	0.40
5-way 4-shot	0.58	0.42	0.40
5-way 5-shot	0.57	0.41	0.40