My Report

Created on December 1|Last edited on December 1

Comment

﻿
Section 1My findings.
/
﻿
 best traffic acc (50% data))
best car acc (50% data)
Expand 38 lines ...
   
  # log semantic segmentation in the new style
  def on_epoch_end(self, **kwargs):
-    num_log = 5
+    num_log = 6
    input_batch = self.learn.data.valid_ds[:num_log]
    mask_list = []
    for i, img in enumerate(input_batch):
Expand 75 lines ...
config.img_size = (360, 640)    # dimensions of resized image - can be 1 dim or tuple
-config.batch_size = 8           # Batch size during training
+config.batch_size = 6           # Batch size during training
config.epochs = 10            # Number of epochs for training
-config.encoder = "resnet18" #resnet34"
+config.encoder = "resnet34" #resnet34"
if config.encoder == "resnet18":
  encoder = models.resnet18     # encoder of unet (contracting path)
elif config.encoder == "resnet34":
  encoder = models.resnet34
elif config.encoder == "squeezenet1_0":
Expand 13 lines ...
-config.weight_decay = 0.097 #0.097     # weight decay applied on layers
+config.weight_decay = 0.08173 #0.097     # weight decay applied on layers
config.bn_weight_decay = True # whether weight decay is applied on batch norm layers
config.one_cycle = True         # use the "1cycle" policy -> https://arxiv.org/abs/1803.09820
# SWEEPS UNCOMMENT
-config.learning_rate = 0.001 #09084 #0.001     # learning rate
+config.learning_rate = 0.001311 #09084 #0.001     # learning rate
save_model = False
# Custom values to filter runs
# SWEEPS UNCOMMENT
-config.training_stages = 2
+config.training_stages = 3
# Data paths
path_data = Path('../../../../BigData/bdd100K/bdd100k/seg')
path_lbl = path_data / 'labels'
path_img = path_data / 'images'
# Associate a label to an input
get_y_fn = lambda x: path_lbl / x.parts[-2] / f'{x.stem}_train_id.png'
Expand 146 lines ...
Overall accuracy of top 100 runs
Overall accuracy of top 100 runs
Showing first 50 runs
051015Step0.50.60.70.8
2.02.53.03.54.04.55.05.56.06.57.07.58.0bat...0.0000.0010.0020.0030.0040.0050.0060.0070.0080.0090.0100.0110.012lea...0.000.010.020.030.040.050.060.070.080.090.10wei...alexnetresnet18resnet34enc...null0.300.350.400.450.500.550.600.650.700.750.800.850.90acc
Run set328
﻿
﻿
﻿
Run set328
﻿
﻿

Add a comment

			Expand 38 lines ...
39	39
40	40		# log semantic segmentation in the new style
41	41		def on_epoch_end(self, **kwargs):
42		-	num_log = 5
	42	+	num_log = 6
43	43		input_batch = self.learn.data.valid_ds[:num_log]
44	44		mask_list = []
45	45		for i, img in enumerate(input_batch):
			Expand 75 lines ...
121	121		config.img_size = (360, 640) # dimensions of resized image - can be 1 dim or tuple
122	122
123		-	config.batch_size = 8 # Batch size during training
	123	+	config.batch_size = 6 # Batch size during training
124	124		config.epochs = 10 # Number of epochs for training
125	125
126		-	config.encoder = "resnet 18 " #resnet34"
	126	+	config.encoder = "resnet 34 " #resnet34"
127	127		if config.encoder == "resnet18":
128	128		encoder = models.resnet18 # encoder of unet (contracting path)
129	129		elif config.encoder == "resnet34":
130	130		encoder = models.resnet34
131	131		elif config.encoder == "squeezenet1_0":
			Expand 13 lines ...
145		-	config.weight_decay = 0.0 9 7 #0.097 # weight decay applied on layers
	145	+	config.weight_decay = 0.0 81 7 3 #0.097 # weight decay applied on layers
146	146		config.bn_weight_decay = True # whether weight decay is applied on batch norm layers
147	147		config.one_cycle = True # use the "1cycle" policy -> https://arxiv.org/abs/1803.09820
148	148		# SWEEPS UNCOMMENT
149		-	config.learning_rate = 0.001 #09084 #0.001 # learning rate
	149	+	config.learning_rate = 0.001 311 #09084 #0.001 # learning rate
150	150		save_model = False
151	151
152	152		# Custom values to filter runs
153	153		# SWEEPS UNCOMMENT
154		-	config.training_stages = 2
	154	+	config.training_stages = 3
155	155
156	156		# Data paths
157	157		path_data = Path('../../../../BigData/bdd100K/bdd100k/seg')
158	158		path_lbl = path_data / 'labels'
159	159		path_img = path_data / 'images'
160	160
161	161		# Associate a label to an input
162	162		get_y_fn = lambda x: path_lbl / x.parts[-2] / f'{x.stem}_train_id.png'
			Expand 146 lines ...