update apgcc framework and inference code

apgcc/config.py

@@ -0,0 +1,180 @@
+import os
+from easydict import EasyDict as edict
+import time
+import numpy as np
+
+# -----------------------------------------------------------------------------
+# Config definition
+# -----------------------------------------------------------------------------
+_C = edict()
+cfg = _C  # Alias for easy usage
+
+_C.TAG = 'APGCC'
+_C.SEED = 1229 # seed
+_C.GPU_ID = 0 # gpu_id, the gpu used for training
+_C.OUTPUT_DIR = './output/temp/' # output_dir, path where to save, empty for no saving
+_C.VIS = False # vis the predict sample
+
+# -----------------------------------------------------------------------------
+# MODEL
+# -----------------------------------------------------------------------------
+_C.MODEL = edict()
+_C.MODEL.ENCODER = 'vgg16_bn' # ['vgg16', 'vgg16_bn'] # backbone, Name of the convolutional backbone to use.
+_C.MODEL.ENCODER_kwargs = {"last_pool": False, # last layer downsample, False:feat4(H/16,W/16), True:feat4(H/32,W/32)  
+						  } 
+						#   "layers": 4,  # select number of bodies.
+						#   "fpn": True,  
+						#   "multi_grid": True,
+						#   "zero_init_residual": True,
+						#   "replace_stride_with_dilation": [False, False, False],
+						#   "feat_dims":64} 
+
+_C.MODEL.DECODER = 'basic' # ['basic', 'IFA'] # decoder 
+_C.MODEL.DECODER_kwargs = { "num_classes": 2,        # output num_classes, default:2 means confindence.
+							"inner_planes": 256,     # basic: 256, IFA: 64
+							"feat_layers":[3,4],     # control the number of decoder features. [1,2,3,4]
+							"pos_dim": 2,            #
+							"ultra_pe": False,       # additional position encoding. x -> (x, sin(x), cos(x))
+							"learn_pe": False,       # additional position encoding. x -> (trainable variable)
+							"unfold": False,         # unfold feat channel, make the feat dim be 3x3 times.
+							"local": False,          # enable local patch, 3x3 mapping near by the center point.
+							"no_aspp": True,         # final feat encoding add the aspp module.
+							"require_grad": True,	
+							"out_type": 'Normal',	 # out_type = 'Normal' / 'Conv' / 'Deconv'
+							"head_layers":[1024,512,256,256]}  # head layers is n+1, last layers is num_of_proposals  
+
+_C.MODEL.STRIDE = 8 					# the size of anchor map by image.shape/stride, ex: input 128x128, stride=8, anchor_map = 16x16
+_C.MODEL.ROW = 2 				# row, row number of anchor points
+_C.MODEL.LINE = 2 				# line, line number of anchor points
+_C.MODEL.FROZEN_WEIGHTS = None 	# frozen_weights, Path to the pretrained model. If set, only the mask head will be trained.
+
+# mixed to the unity loss kwargs
+_C.MODEL.POINT_LOSS_COEF = 0.0002 # point_loss_coef
+_C.MODEL.EOS_COEF = 0.5 # eos_coef, Relative classification weight of the no-object class
+
+_C.MODEL.LOSS = ['L2']
+_C.MODEL.WEIGHT_DICT = {'loss_ce': 1, 'loss_points': 0., 'loss_aux': 0.} 
+_C.MODEL.AUX_EN = False
+_C.MODEL.AUX_NUMBER = [1, 1] # the number of pos/neg anchors
+_C.MODEL.AUX_RANGE = [1, 4]  # the randomness range of auxiliary anchors
+_C.MODEL.AUX_kwargs = {'pos_coef': 1., 'neg_coef': 1., 'pos_loc': 0., 'neg_loc': 0.} 
+
+_C.RESUME = False # resume, resume from checkpoint
+_C.RESUME_PATH = '' # keep training weights.
+
+# -----------------------------------------------------------------------------
+# Dataset
+# -----------------------------------------------------------------------------
+_C.DATASETS = edict()
+_C.DATASETS.DATASET = 'SHHA' # dataset_file
+_C.DATASETS.DATA_ROOT = './dataset_path/' # data_root, path where the dataset is
+
+# -----------------------------------------------------------------------------
+# DATALOADER
+# -----------------------------------------------------------------------------
+_C.DATALOADER = edict()
+_C.DATALOADER.AUGUMENTATION = ['Normalize', 'Crop', 'Flip']
+_C.DATALOADER.CROP_SIZE = 128 		# radnom crip size for training
+_C.DATALOADER.CROP_NUMBER = 4 		# the number of training sample
+_C.DATALOADER.UPPER_BOUNDER = -1 	# the upper bounder of size
+_C.DATALOADER.NUM_WORKERS = 8 		# num_workers
+
+# ---------------------------------------------------------------------------- #
+# Solver
+# ---------------------------------------------------------------------------- #
+_C.SOLVER = edict()
+_C.SOLVER.BATCH_SIZE = 8 # batch_size\
+_C.SOLVER.START_EPOCH = 0 # start_epoch
+_C.SOLVER.EPOCHS = 3500  # epochs
+_C.SOLVER.LR = 1e-4    # lr
+_C.SOLVER.LR_BACKBONE = 1e-5   # lr_backbone
+_C.SOLVER.WEIGHT_DECAY = 1e-4  # weight_decay
+_C.SOLVER.LR_DROP = 3500 # lr_drop
+_C.SOLVER.CLIP_MAX_NORM = 0.1 # clip_max_norm, gradient clipping max norm
+
+_C.SOLVER.EVAL_FREQ = 5 # eval_freq, frequency of evaluation, default setting is evaluating in every 5 epoch
+_C.SOLVER.LOG_FREQ = 1 # log_freq, frequency of recording training.
+# ---------------------------------------------------------------------------- #
+# Matcher
+# ---------------------------------------------------------------------------- #
+_C.MATCHER = edict()
+_C.MATCHER.SET_COST_CLASS = 1. # set_cost_class, Class coefficient in the matching cost
+_C.MATCHER.SET_COST_POINT = 0.05 # set_cost_point, L1 point coefficient in the matching cost
+
+# -----------------------------------------------------------------------------
+# TEST
+# -----------------------------------------------------------------------------
+_C.TEST = edict()
+_C.TEST.THRESHOLD = 0.5
+
+################ modules ################
+def cfg_merge_a2b(a, b):
+    if type(a) is not edict and type(a) is not dict:
+        raise KeyError('a is not a edict.')
+
+    for k, v in a.items():
+        if k not in b:
+            raise KeyError('{} is not a valid config key'.format(k))
+
+        old_type = type(b[k]) # original type
+        if old_type is not type(v):
+            if isinstance(b[k], np.ndarray):
+                v = np.array(v, dtype=b[k].dtype)
+            elif not isinstance(v, dict) and not isinstance(b[k], type(None)):
+                raise ValueError(('Type mismatch ({} vs. {})'
+                                'for config key: {}').format(type(b[k]), type(v), k))
+        if type(v) is edict or type(v) is dict:
+            try:
+                cfg_merge_a2b(a[k], b[k])
+            except:
+                print(('Error under config key: {}').format(k))
+                raise
+        else:
+            if v == 'None':
+                b[k] = None
+            else:
+                b[k] = v
+
+    return b
+
+def cfg_from_file(filename):
+    import yaml
+    with open(filename, 'r') as f:
+        data = yaml.safe_load(f)
+    return data
+
+def cfg_from_list(args_opts):
+	from ast import literal_eval
+	print(args_opts)
+	if len(args_opts) == 0:
+		return None
+	assert len(args_opts)%2 == 0
+	for k, v in zip(args_opts[0::2], args_opts[1::2]):
+		key_list = k.split('.')
+		d = _C
+		for subkey in key_list[:-1]:
+			assert subkey in d
+			d = d[subkey]
+		subkey = key_list[-1]
+		assert subkey in d
+		try:
+			value = literal_eval(v)
+		except:
+			value = v
+		assert type(value) == type(d[subkey]), \
+            'type {} does not match original type {}'.format(type(value), type(d[subkey]))
+		d[subkey] = value
+	return d
+
+def merge_from_file(cfg, filename):
+    file_cfg = cfg_from_file(filename)
+    cfg = cfg_merge_a2b(file_cfg, cfg)
+    return cfg
+
+def merge_from_list(cfg, args):
+	args_cfg = cfg_from_list(args)
+	if args_cfg == None:
+		return cfg
+	else:
+		cfg = cfg_merge_a2b(args_cfg, cfg)
+		return cfg

apgcc/configs/SHHA_IFI.yml

@@ -0,0 +1,89 @@
+TAG: 'SHHA_AGPCC'
+SEED: 1229                                  # seed
+GPU_ID: 0                                   # gpu_id, the gpu used for training
+OUTPUT_DIR: './output_SHHA/temp/'           # output_dir, path where to save, empty for no saving
+VIS: False                                  # vis the predict sample
+
+# -----------------------------------------------------------------------------
+# MODEL
+# -----------------------------------------------------------------------------
+MODEL:
+  ENCODER: 'vgg16_bn'                       # backbone, select: ['vgg16', 'vgg16_bn', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152']
+  ENCODER_kwargs: {"last_pool": False}      # last layer downsample, False:feat4(H/16,W/16), True:feat4(H/32,W/32)                            
+  DECODER: 'IFI'                            # decoder, select: ['basic', 'IFI']
+  DECODER_kwargs: {"num_classes": 2,        # output num_classes, default:2 means confindence.
+                  "inner_planes": 64,       # default: basic: 256, IFI: 64
+                   "feat_layers":[3,4],     # control the number of decoder features. [1,2,3,4]
+                   "pos_dim": 32,           # if ultra_pe, it has to large than 2
+                   "ultra_pe": True,        # additional position encoding. x -> (x, sin(x), cos(x))
+                   "learn_pe": False,       # additional position encoding. x -> (trainable variable)
+                   "unfold": False,         # unfold feat channel, make the feat dim be 3x3 times.
+                   "local": True,           # enable local patch, 3x3 mapping near by the center point.
+                   "no_aspp": False,        # final feat encoding add the aspp module.
+                   "require_grad": True,
+                   "out_type": 'Normal',    # out_type = 'Normal' / 'Conv' / 'Deconv'
+                   "head_layers":[1024,512,256,256]}  # head layers is n+1, last layers is num_of_proposals  
+
+  FROZEN_WEIGHTS: None # frozen_weights, Path to the pretrained model. If set, only the mask head will be trained.
+
+  STRIDE: 8 # the size of anchor map by image.shape/stride, ex: input 128x128, stride=8, anchor_map = 16x16
+  ROW: 2 # row, row number of anchor points
+  LINE: 2 # line, line number of anchor points
+  POINT_LOSS_COEF: 0.0002 # point_loss_coef
+  EOS_COEF: 0.5 # eos_coef, Relative classification weight of the no-object class
+
+  LOSS: ['L2']
+  WEIGHT_DICT: {'loss_ce': 1, 'loss_points': 0.0002, 'loss_aux': 0.2} 
+  AUX_EN: False      # auxiliary point is not used in the inference.
+  AUX_NUMBER: [2, 2] # the number of pos/neg anchors
+  AUX_RANGE: [2, 8]  # the RoI range of auxiliary anchors
+  AUX_kwargs: {'pos_coef': 1., 'neg_coef': 1., 'pos_loc': 0.0002, 'neg_loc': 0.0002} 
+
+RESUME: False # resume, resume from checkpoint
+RESUME_PATH: '' # keep training weights.
+
+# -----------------------------------------------------------------------------
+# Dataset
+# -----------------------------------------------------------------------------
+DATASETS:
+  DATASET: 'SHHA' # dataset_file
+  DATA_ROOT: '/mnt/191/c/CrowdCounting/Dataset/ShanghaiTech/PointData/part_A' # data_root, path where the dataset is
+
+# -----------------------------------------------------------------------------
+# DATALOADER
+# -----------------------------------------------------------------------------
+DATALOADER:
+  AUGUMENTATION: ['Normalize', 'Crop', 'Flip']
+  CROP_SIZE: 128      # radnom crip size for training
+  CROP_NUMBER: 4      # the number of training sample
+  UPPER_BOUNDER: -1   # the upper bounder of size
+  NUM_WORKERS: 0      # num_workers
+
+# ---------------------------------------------------------------------------- #
+# Solver
+# ---------------------------------------------------------------------------- #
+SOLVER:
+  BATCH_SIZE: 8 # batch_size
+  START_EPOCH: 0 # start_epoch
+  EPOCHS: 3500  # epochs
+  LR: 0.0001    # lr
+  LR_BACKBONE: 0.00001 # lr_backbone
+  WEIGHT_DECAY: 0.0001 # weight_decay
+  LR_DROP: 3500 # lr_drop
+  CLIP_MAX_NORM: 0.1 # clip_max_norm, gradient clipping max norm
+
+  EVAL_FREQ: 1 # eval_freq, frequency of evaluation, default setting is evaluating in every 5 epoch
+  LOG_FREQ: 1 # log_freq, frequency of recording training.
+
+# ---------------------------------------------------------------------------- #
+# Matcher
+# ---------------------------------------------------------------------------- #
+MATCHER:
+  SET_COST_CLASS: 1. # set_cost_class, Class coefficient in the matching cost
+  SET_COST_POINT: 0.05 # set_cost_point, L1 point coefficient in the matching cost
+
+# -----------------------------------------------------------------------------
+# TEST
+# -----------------------------------------------------------------------------
+TEST:
+  THRESHOLD: 0.5

apgcc/configs/SHHA_basic.yml

@@ -0,0 +1,83 @@
+TAG: 'SHHA_basic'
+SEED: 1229                                  # seed
+GPU_ID: 0                                   # gpu_id, the gpu used for training
+OUTPUT_DIR: './output_P2P/SHHA_P2P_mix_g2/' # output_dir, path where to save, empty for no saving
+VIS: False                                  # vis the predict sample
+
+# -----------------------------------------------------------------------------
+# MODEL
+# -----------------------------------------------------------------------------
+MODEL:
+  ENCODER: 'vgg16_bn'                       # backbone, select: ['vgg16', 'vgg16_bn', 'resnet18', 'resnet34', 'resnet50', 'resnet101', 'resnet152']
+  ENCODER_kwargs: {"last_pool": False}      # last layer downsample, False:feat4(H/16,W/16), True:feat4(H/32,W/32)
+  DECODER: 'basic'                          # decoder, select: ['basic', 'IFI']
+  DECODER_kwargs: {"num_classes": 2,        # output num_classes, default:2 means confindence.
+                   "inner_planes": 256,     # basic: 256, IFI: 64
+                   "feat_layers":[3, 4]}    # control the number of decoder features. [1,2,3,4] # notice!! change Line, Row
+
+  FROZEN_WEIGHTS: None # frozen_weights, Path to the pretrained model. If set, only the mask head will be trained.
+
+  STRIDE: 8 # the size of anchor map by image.shape/stride, ex: input 128x128, stride=8, anchor_map = 16x16
+  ROW: 2 # row, row number of anchor points
+  LINE: 2 # line, line number of anchor points
+  POINT_LOSS_COEF: 0.0002 # point_loss_coef
+  EOS_COEF: 0.5 # eos_coef, Relative classification weight of the no-object class
+
+  LOSS: ['L2']
+  WEIGHT_DICT: {'loss_ce': 1, 'loss_points': 0.0002, 'loss_aux': 0.} 
+  AUX_EN: False      # auxiliary point is not used in the inference.
+  AUX_NUMBER: [1, 1] # the number of pos/neg anchors
+  AUX_RANGE: [1, 4]  # the RoI range of auxiliary anchors
+  AUX_kwargs: {'pos_coef': 1., 'neg_coef': 1., 'pos_loc': 0.0002, 'neg_loc': 0.0002} 
+
+RESUME: False # resume, resume from checkpoint
+RESUME_PATH: '' # keep training weights.
+
+# -----------------------------------------------------------------------------
+# Dataset
+# -----------------------------------------------------------------------------
+DATASETS:
+  DATASET: 'SHHA' # dataset_file
+  DATA_ROOT: '/work/sylab607/CIH/CrowdCounting/Dataset/ShanghaiTech/TorchData/shanghaitech_part_A' # data_root, path where the dataset is
+  # /work/sylab607/CIH/CrowdCounting/Dataset/ShanghaiTech/PointData/part_A/
+  # /work/sylab607/CIH/CrowdCounting/Dataset/ShanghaiTech/TorchData/shanghaitech_part_A
+
+# -----------------------------------------------------------------------------
+# DATALOADER
+# -----------------------------------------------------------------------------
+DATALOADER:
+  AUGUMENTATION: ['Normalize', 'Crop', 'Flip']
+  CROP_SIZE: 128      # radnom crip size for training
+  CROP_NUMBER: 4      # the number of training sample
+  UPPER_BOUNDER: -1   # the upper bounder of size
+  NUM_WORKERS: 0      # num_workers
+
+# ---------------------------------------------------------------------------- #
+# Solver
+# ---------------------------------------------------------------------------- #
+SOLVER:
+  BATCH_SIZE: 8 # batch_size\
+  START_EPOCH: 0 # start_epoch
+  EPOCHS: 3500  # epochs
+  LR_DROP: 3500 # lr_drop
+  LR: 0.0001    # lr
+  LR_BACKBONE: 0.00001 # lr_backbone
+  WEIGHT_DECAY: 0.0001 # weight_decay
+  LR_DROP: 3500 # lr_drop
+  CLIP_MAX_NORM: 0.1 # clip_max_norm, gradient clipping max norm
+
+  EVAL_FREQ: 1 # eval_freq, frequency of evaluation, default setting is evaluating in every 5 epoch
+  LOG_FREQ: 1 # log_freq, frequency of recording training.
+
+# ---------------------------------------------------------------------------- #
+# Matcher
+# ---------------------------------------------------------------------------- #
+MATCHER:
+  SET_COST_CLASS: 1. # set_cost_class, Class coefficient in the matching cost
+  SET_COST_POINT: 0.05 # set_cost_point, L1 point coefficient in the matching cost
+
+# -----------------------------------------------------------------------------
+# TEST
+# -----------------------------------------------------------------------------
+TEST:
+  THRESHOLD: 0.5

apgcc/datasets/__init__.py

@@ -0,0 +1,4 @@
+from .build import loading_data
+
+def build_dataset(cfg):
+    return loading_data(cfg)