ZER-0-NE

// The model overfits around 20 epochs. The saved checkpoint model works well on the test set by giving just 18 false positives
// as compared to 36 in other cases. The model is shallow as compared to other researchers. The batch size=128 and small lr
// were significant factor in getting low losses. Next try would be to increase lr and dropout.

// This was trained on the new dataset


train_data_path = 'detection_data_1/train/'
validation_data_path = 'detection_data_1/val/'
test_data_path = 'detection_data_1/test/'

ZER-0-NE

// The val acc. does not increase beyond 95% and the training acc. somewhat seems to get lower gradually. Without reducing 
// the lr there is a possibility to converge.


train_data_path = 'detection_data_1/train/'
validation_data_path = 'detection_data_1/val/'
test_data_path = 'detection_data_1/test/'

#Parametres
img_width, img_height = 150, 250

ZER-0-NE

img_width, img_height = 300, 500
*********************************************************

if K.image_data_format() == 'channels_first':
    input_shape = (3, img_width, img_height)
else:
    input_shape = (img_width, img_height, 3)


model = Sequential()

ZER-0-NE

img_width, img_height = 200, 300

*********************************************************

if K.image_data_format() == 'channels_first':
    input_shape = (3, img_width, img_height)
else:
    input_shape = (img_width, img_height, 3)

ZER-0-NE

img_width, img_height = 150, 250

*********************************************************

if K.image_data_format() == 'channels_first':
    input_shape = (3, img_width, img_height)
else:
    input_shape = (img_width, img_height, 3)