Explanation of the segmentation parrameters

[dandanaus]

Hi team,

I am looking to train using my own data and wanted to get some basic explanation on the relevance of the parrameters

sample_num = 2048

batch_size = 12

num_epochs = 256

In my settings file i have :

num_class = 2

sample_num = 2048

batch_size = 12

num_epochs = 256

label_weights = []
for c in range(num_class):
label_weights.append(1.0)

learning_rate_base = 0.001
decay_steps = 20000
decay_rate = 0.7
learning_rate_min = 1e-6

step_val = 500

weight_decay = 0.0

Trying to train on aerial data with 8 points per sq meter density and i have a training dataset of 800 000 000 million points.

I am seing some over-fitting very early on with these values:

num_class = 2

sample_num = 12288

batch_size = 6

num_epochs = 8096

label_weights = []
for c in range(num_class):
label_weights.append(1.0)

learning_rate_base = 0.001
decay_steps = 20000
decay_rate = 0.7
learning_rate_min = 1e-6

step_val = 500

weight_decay = 0.0

jitter = 0.0
jitter_val = 0.0

rotation_range = [0, math.pi/32., 0, 'u']
rotation_range_val = [0, 0, 0, 'u']
rotation_order = 'rxyz'

scaling_range = [0.0, 0.0, 0.0, 'g']
scaling_range_val = [0, 0, 0, 'u']

sample_num_variance = 1 // 8
sample_num_clip = 1 // 4

Any advice?

[dandanaus]

Here is a snap of the loss for the validation set - very lumpy

[aniketpant83]

Any updates on this?