AN2DL Challenges

Artificial Neural Networks and Deep Learning @ PoliMi, a.y. 2020

• Filippo Colombo (@fcolombo7)
• Giovanni Del Vecchio (@giovannidv8)

---

Challenge 3

• https://github.com/chingyaoc/VQA-tensorflow
• https://medium.com/@harshareddykancharla/visual-question-answering-with-hierarchical-question-image-co-attention-c5836684a180

Steps and results history

---

Challenge 2

Steps and results history

U-Net like architecture
Bipbip dataset

Model Name	IoU Bipbip	Input size	Num of levels	Starting filter depth size	Conv. per block	Regularization	Note
U-Net_s32	0.6533	original	4	32	2	-	no normalization
U-Net_s16_norm-input	0.0669	original	4	16	2	-	with normalization --- NO! errore sul test set
U-Net_s16_adapt-lr+batch_norm	0.6543	original	4	16	2	-	batch norm and adaptive learning rate
U-Net_sXX_adapt-lr+batch_norm+l1	---	original	4	16/32	2	l1	batch norm and adaptive learning rate, with l1: the net does not learn (both 32 and 16)
U-Net_s32_adapt-lr+batch_norm+drop	0.6642	original	4	32	2	dropout=0.2	batch norm, adaptive learning rate [1e-3 - 1e-5]
U-Net_s32_adapt-lr_v2+batch_norm+drop	0.6877	original	4	32	2	dropout=0.2	batch norm, adaptive learning rate [1e-4 - 1e-5]
U-Net_s32_1convXblock_6depth	0.6971	original	6	32	1	-	batch norm, adaptive learning rate [1e-4 - 1e-5]

---

Challenge 1

Note

• Hyperparameter Optimization with Scikit-Learn, Scikit-Opt and Keras
• K-Fold in KerasTuner
• Hyperparameter Optimization with Keras

---

Results history

CNN from skratch

Filename	Result	Input size	Feature extractor depth	kernel size	stride	N. Dense Hidden Layer	N. Neurons (dense layers)	Dropout	L2	info
results_Nov11_09-10-27.csv	0.70666	512x512	6	3x3	(1,1)	1	256	-	-	-
results_Nov11_13-39-53.csv	0.57333	254x256	4	3x3	(2,2)	2	128	0.2	0.001	other test with a larger kernel size has been done, but there were no imporvements
results_Nov11_09-38-44.csv	0.58444	128x128	6	3x3	(1,1)	3	64	-	-	-
results_Nov12_21-27-13.csv	0.70666	256x256	5	3x3	(1,1)	1	512	0.2	0.001	-

CNN with Transfer Learning

Filename	Result	Input size	Base model	N. Dense Hidden Layer	N. Neurons (dense layers)	Dropout	L2	info
results_Nov12_21-36-14.csv	0.84222	256x256	VGG16	1	256	-	-	-
results_Nov12_21-36-14.csv	0.77111	256x256	VGG16	2	512+128	-	-	-
results_Nov14_13-40-16.csv	0.88444	400x400	VGG16	1	448	0.1	-	-
results_Nov14.csv	0.91555	400x400	VGG16	1	448	0.2	-	augmentation v2: rotation_range=20, width_shift_range=0.3, height_shift_range=0.3, zoom_range=0.4, horizontal_flip=True, shear_range=10, channel_shift_range=100, fill_mode='reflect', rescale=1./255
results_Nov15_10-37-55.csv	0.94000	400x400	VGG16	1	512	0.2	-	augmentation v2
results_Nov15_14-31-13.csv	0.93777	400x400	VGG16	2	512+448	0.1	-	augmentation v2
results_Nov15_23-11-35.csv	0.93333	400x400	VGG16	2	512+512	0.2	-	augmentation v2
results_Nov16_10-58-01.csv	0.92000	400x400	VGG16	2	512+512	-	0.001	augmentation v2
results_Nov15_10-37-55.csv	0.94000	448x448	VGG16	1	576	0.2	-	augmentation v2
results_Nov17_21-18-42.csv	0.94000	448x448	VGG16	1	600	0.2	-	augmentation v3: rotation_range=30, width_shift_range=0.2, height_shift_range=0.2, zoom_range=0.4, horizontal_flip=True, brightness_range = [0.6, 1.0], shear_range=30, channel_shift_range=50, fill_mode='reflect', rescale=1./255
results_Nov18_07-50-09.csv	0.94444	448x448	VGG16	1	512	0.2	-	augmentation v2