replace argparser in training with config file

Author: hangzhaomit

config/__init__.py

@@ -0,0 +1 @@
+from .defaults import _C as cfg

config/ade20k-resnet50dilated-ppm_deepsup.yaml

@@ -0,0 +1,36 @@
+DATASET:
+  root_dataset: "./data/"
+  list_train: "./data/training.odgt"
+  list_val: "./data/validation.odgt"
+  num_class: 150
+
+MODEL:
+  id: "baseline"
+  arch_encoder: "resnet50dilated"
+  arch_decoder: "ppm_deepsup"
+  weights_encoder: ""
+  weights_decoder: ""
+  fc_dim: 2048
+
+TRAIN:
+  batch_size_per_gpu: 2
+  num_epoch: 20
+  start_epoch: 1
+  epoch_iters: 5000
+  optim: "SGD"
+  lr_encoder: 0.02
+  lr_decoder: 0.02
+  lr_pow: 0.9
+  beta1: 0.9
+  weight_decay: 1e-4
+  deep_sup_scale: 0.4
+  fix_bn: False
+  workers: 16
+  imgSize: (300, 375, 450, 525, 600)
+  imgMaxSize: 1000
+  padding_constant: 8
+  segm_downsampling_rate: 8
+  random_flip: True
+  ckpt: "./ckpt"
+  disp_iter: 20
+  seed: 304

config/defaults.py

@@ -0,0 +1,84 @@
+from yacs.config import CfgNode as CN
+
+# -----------------------------------------------------------------------------
+# Config definition
+# -----------------------------------------------------------------------------
+
+_C = CN()
+
+# -----------------------------------------------------------------------------
+# Dataset
+# -----------------------------------------------------------------------------
+_C.DATASET = CN()
+_C.DATASET.root_dataset = "./data/"
+_C.DATASET.list_train = "./data/training.odgt"
+_C.DATASET.list_val = "./data/validation.odgt"
+_C.DATASET.num_class = 150
+
+# -----------------------------------------------------------------------------
+# Model
+# -----------------------------------------------------------------------------
+_C.MODEL = CN()
+# a name for identifying the model
+_C.MODEL.id = "baseline"
+# architecture of net_encoder
+_C.MODEL.arch_encoder = "resnet50dilated"
+# architecture of net_decoder
+_C.MODEL.arch_decoder = "ppm_deepsup"
+# weights to finetune net_encoder
+_C.MODEL.weights_encoder = ""
+# weights to finetune net_decoder
+_C.MODEL.weights_decoder = ""
+# number of feature channels between encoder and decoder
+_C.MODEL.fc_dim = 2048
+
+# -----------------------------------------------------------------------------
+# Training
+# -----------------------------------------------------------------------------
+_C.TRAIN = CN()
+_C.TRAIN.batch_size_per_gpu = 2
+# epochs to train for
+_C.TRAIN.num_epoch = 20
+# epoch to start training. useful if continue from a checkpoint
+_C.TRAIN.start_epoch = 1
+# iterations of each epoch (irrelevant to batch size)
+_C.TRAIN.epoch_iters = 5000
+
+_C.TRAIN.optim = "SGD"
+_C.TRAIN.lr_encoder = 0.02
+_C.TRAIN.lr_decoder = 0.02
+# power in poly to drop LR
+_C.TRAIN.lr_pow = 0.9
+# momentum for sgd, beta1 for adam
+_C.TRAIN.beta1 = 0.9
+# weights regularizer
+_C.TRAIN.weight_decay = 1e-4
+# the weighting of deep supervision loss
+_C.TRAIN.deep_sup_scale = 0.4
+# fix bn params, only under finetuning
+_C.TRAIN.fix_bn = False
+# number of data loading workers
+_C.TRAIN.workers = 16
+
+# input image size of short edge (int or tuple)
+_C.TRAIN.imgSize = (300, 375, 450, 525, 600)
+# maximum input image size of long edge
+_C.TRAIN.imgMaxSize = 1000
+# maxmimum downsampling rate of the network
+_C.TRAIN.padding_constant = 8
+# downsampling rate of the segmentation label
+_C.TRAIN.segm_downsampling_rate = 8
+# if horizontally flip images when training
+_C.TRAIN.random_flip = True
+
+# folder to output checkpoints
+_C.TRAIN.ckpt = "./ckpt"
+# frequency to display
+_C.TRAIN.disp_iter = 20
+# manual seed
+_C.TRAIN.seed = 304
+
+# -----------------------------------------------------------------------------
+# Testing
+# -----------------------------------------------------------------------------
+_C.TEST = CN()

dataset.py

@@ -51,9 +51,9 @@ def round2nearest_multiple(self, x, p):
 
 
 class TrainDataset(BaseDataset):
-    def __init__(self, odgt, opt, batch_per_gpu=1, **kwargs):
+    def __init__(self, root_dataset, odgt, opt, batch_per_gpu=1, **kwargs):
         super(TrainDataset, self).__init__(odgt, opt, **kwargs)
-        self.root_dataset = opt.root_dataset
+        self.root_dataset = root_dataset
         self.random_flip = opt.random_flip
         # down sampling rate of segm labe
         self.segm_downsampling_rate = opt.segm_downsampling_rate
@@ -101,7 +101,7 @@ def __getitem__(self, index):
         batch_records = self._get_sub_batch()
 
         # resize all images' short edges to the chosen size
-        if isinstance(self.imgSize, list):
+        if isinstance(self.imgSize, list) or isinstance(self.imgSize, tuple):
             this_short_size = np.random.choice(self.imgSize)
         else:
             this_short_size = self.imgSize
@@ -184,9 +184,9 @@ def __len__(self):
 
 
 class ValDataset(BaseDataset):
-    def __init__(self, odgt, opt, **kwargs):
+    def __init__(self, root_dataset, odgt, opt, **kwargs):
         super(ValDataset, self).__init__(odgt, opt, **kwargs)
-        self.root_dataset = opt.root_dataset
+        self.root_dataset = root_dataset
 
     def __getitem__(self, index):
         this_record = self.list_sample[index]