configs | incorporated in main code

mustafa1728 · Jun 1, 2021 · ab10546 · ab10546
1 parent 728c740
commit ab10546
Show file tree

Hide file tree

Showing 4 changed files with 237 additions and 184 deletions.
diff --git a/config.py b/config.py
@@ -3,7 +3,6 @@
 """
 
 import os
-from pickle import NONE
 
 from yacs.config import CfgNode as CN
 
@@ -13,53 +12,56 @@
 
 _C = CN()
 
-_C.TO_VALIDATE = False # choices = [True, False]
+_C.TO_VALIDATE = True # choices = [True, False]
+
 
 # -----------------------------------------------------------------------------
-# Dataset
+# Paths
 # -----------------------------------------------------------------------------
-
-# dataset paths
-_C.DATASET.PATH_DATA_ROOT = "data/" # directory where the feature pickles are stored. Depends on users
-_C.DATASET.PATH_LABELS_ROOT = "data/" # directory where the annotations are stored. Depends on users
-_C.DATASET.PATH_EXP_ROOT="model/action-model/" # directory where the checkpoints are to be stored. Depends on users
+_C.PATHS = CN()
+_C.PATHS.PATH_DATA_ROOT = "data/" # directory where the feature pickles are stored. Depends on users
+_C.PATHS.PATH_LABELS_ROOT = "annotations/" # directory where the annotations are stored. Depends on users
+_C.PATHS.PATH_EXP_ROOT="model/action-model/" # directory where the checkpoints are to be stored. Depends on users
 
 
-_C.DATASET.DATASET_SOURCE="source_train" # depends on users
-_C.DATASET.DATASET_TARGET="target_train" # depends on users
+_C.PATHS.DATASET_SOURCE="source_train" # depends on users
+_C.PATHS.DATASET_TARGET="target_train" # depends on users
 if _C.TO_VALIDATE:
-    _C.DATASET.VAL_DATASET_SOURCE="source_val" # depends on users
-    _C.DATASET.VAL_DATASET_TARGET="target_val" # depends on users
+    _C.PATHS.VAL_DATASET_SOURCE="source_val" # depends on users
+    _C.PATHS.VAL_DATASET_TARGET="target_val" # depends on users
 else:
-    _C.DATASET.VAL_DATASET_SOURCE= None 
-    _C.DATASET.VAL_DATASET_TARGET= None 
-_C.DATASET.NUM_SOURCE= 16115 # number of training data (source)
-_C.DATASET.NUM_TARGET= 26115 # number of training data (target)
+    _C.PATHS.VAL_DATASET_SOURCE= None 
+    _C.PATHS.VAL_DATASET_TARGET= None 
+_C.PATHS.NUM_SOURCE= 16115 # number of training data (source)
+_C.PATHS.NUM_TARGET= 26115 # number of training data (target)
 
-_C.DATASET.PATH_DATA_SOURCE=os.path.join(_C.DATASET.PATH_DATA_ROOT, _C.DATASET.DATASET_SOURCE)
-_C.DATASET.PATH_DATA_TARGET=os.path.join(_C.DATASET.PATH_DATA_ROOT, _C.DATASET.DATASET_TARGET)
+_C.PATHS.PATH_DATA_SOURCE=os.path.join(_C.PATHS.PATH_DATA_ROOT, _C.PATHS.DATASET_SOURCE)
+_C.PATHS.PATH_DATA_TARGET=os.path.join(_C.PATHS.PATH_DATA_ROOT, _C.PATHS.DATASET_TARGET)
 if _C.TO_VALIDATE:
-    _C.DATASET.PATH_VAL_DATA_SOURCE=os.path.join(_C.DATASET.PATH_DATA_ROOT, _C.DATASET.VAL_DATASET_SOURCE)
-    _C.DATASET.PATH_VAL_DATA_TARGET=os.path.join(_C.DATASET.PATH_DATA_ROOT, _C.DATASET.VAL_DATASET_TARGET)
+    _C.PATHS.PATH_VAL_DATA_SOURCE=os.path.join(_C.PATHS.PATH_DATA_ROOT, _C.PATHS.VAL_DATASET_SOURCE)
+    _C.PATHS.PATH_VAL_DATA_TARGET=os.path.join(_C.PATHS.PATH_DATA_ROOT, _C.PATHS.VAL_DATASET_TARGET)
 else:
-    _C.DATASET.PATH_VAL_DATA_SOURCE= None 
-    _C.DATASET.PATH_VAL_DATA_SOURCE= None 
+    _C.PATHS.PATH_VAL_DATA_SOURCE= None 
+    _C.PATHS.PATH_VAL_DATA_SOURCE= None 
 
-_C.DATASET.TRAIN_SOURCE_LIST=os.path.join(_C.DATASET.PATH_LABELS_ROOT, 'EPIC_100_uda_source_train.pkl') # '/domain_adaptation_source_train_pre-release_v3.pkl'
-_C.DATASET.TRAIN_TARGET_LIST=os.path.join(_C.DATASET.PATH_LABELS_ROOT, 'EPIC_100_uda_target_train_timestamps.pkl') # '/domain_adaptation_target_train_pre-release_v6.pkl'
+_C.PATHS.TRAIN_SOURCE_LIST=os.path.join(_C.PATHS.PATH_LABELS_ROOT, 'EPIC_100_uda_source_train.pkl') # '/domain_adaptation_source_train_pre-release_v3.pkl'
+_C.PATHS.TRAIN_TARGET_LIST=os.path.join(_C.PATHS.PATH_LABELS_ROOT, 'EPIC_100_uda_target_train_timestamps.pkl') # '/domain_adaptation_target_train_pre-release_v6.pkl'
 if _C.TO_VALIDATE:
-    _C.DATASET.VAL_SOURCE_LIST=os.path.join(_C.DATASET.PATH_LABELS_ROOT, "EPIC_100_uda_source_val.pkl")
-    _C.DATASET.VAL_TARGET_LIST=os.path.join(_C.DATASET.PATH_LABELS_ROOT, "EPIC_100_uda_target_val.pkl")
+    _C.PATHS.VAL_SOURCE_LIST=os.path.join(_C.PATHS.PATH_LABELS_ROOT, "EPIC_100_uda_source_val.pkl")
+    _C.PATHS.VAL_TARGET_LIST=os.path.join(_C.PATHS.PATH_LABELS_ROOT, "EPIC_100_uda_target_val.pkl")
 else:
-    _C.DATASET.VAL_SOURCE_LIST= None 
-    _C.DATASET.VAL_TARGET_LIST= None 
-_C.DATASET.VAL_LIST=os.path.join(_C.DATASET.PATH_LABELS_ROOT, "EPIC_100_uda_target_test_timestamps.pkl")
-_C.DATASET.PATH_EXP=os.path.join(_C.DATASET.PATH_EXP_ROOT, "Testexp")
+    _C.PATHS.VAL_SOURCE_LIST= None 
+    _C.PATHS.VAL_TARGET_LIST= None 
+_C.PATHS.VAL_LIST=os.path.join(_C.PATHS.PATH_LABELS_ROOT, "EPIC_100_uda_target_test_timestamps.pkl")
+_C.PATHS.PATH_EXP=os.path.join(_C.PATHS.PATH_EXP_ROOT, "Testexp")
 
-# dataset parameters
+
+# -----------------------------------------------------------------------------
+# Dataset
+# -----------------------------------------------------------------------------
 _C.DATASET = CN()
 _C.DATASET.DATASET = "epic" # dataset choices = [hmdb_ucf, hmdb_ucf_small, ucf_olympic]
-_C.DATASET.NUM_CLASSES = 97300
+_C.DATASET.NUM_CLASSES = "97,300"
 _C.DATASET.MODALITY = "ALL" # choices = [RGB ]
 _C.DATASET.FRAME_TYPE = "feature" # choices = [frame]
 _C.DATASET.NUM_SEGMENTS = 5 # sample frame # of each video for training
@@ -80,6 +82,10 @@
 _C.MODEL.WEIGHTED_CLASS_LOSS_DA = "N" # choices  = [Y, N]
 _C.MODEL.WEIGHTED_CLASS_LOSS = "N" # choices  = [Y, N]
 
+_C.MODEL.DROPOUT_I = 0.8
+_C.MODEL.DROPOUT_V = 0.8
+_C.MODEL.NO_PARTIALBN = True
+
 
 # DA configs
 if _C.MODEL.USE_TARGET == "none":
@@ -96,14 +102,17 @@
 _C.MODEL.USE_ATTN = "TransAttn" # choices  = [None, TransAttn, general]
 _C.MODEL.USE_ATTN_FRAME = None # choices  = [None, TransAttn, general]
 _C.MODEL.USE_BN = None # choices  = [None, AdaBN, AutoDIAL]
+_C.MODEL.N_ATTN = 1
+_C.MODEL.PLACE_DIS = ["Y", "Y", "N"]
+_C.MODEL.PLACE_ADV = ["Y", "Y", "Y"]
+
 
 # ---------------------------------------------------------------------------- #
 # Hyperparameters
 # ---------------------------------------------------------------------------- #
 _C.HYPERPARAMETERS = CN()
 _C.HYPERPARAMETERS.ALPHA = 0
 _C.HYPERPARAMETERS.BETA = [0.75, 0.75, 0.5]
-_C.HYPERPARAMETERS.N_ATTN = 1
 _C.HYPERPARAMETERS.GAMMA = 0.003 # U->H: 0.003 | H->U: 0.3
 _C.HYPERPARAMETERS.MU = 0
 
@@ -117,21 +126,51 @@
 _C.TRAINER.ARCH = "TBN" # choices  = [resnet50]
 _C.TRAINER.USE_TARGET = "uSv" # choices  = [uSv, Sv, none]
 _C.TRAINER.SHARE_PARAMS = "Y" # choices  = [Y, N]
+_C.TRAINER.PRETRAIN_SOURCE = False
+_C.TRAINER.VERBOSE = True
 
 # Learning configs
+_C.TRAINER.LOSS_TYPE = 'nll'
 _C.TRAINER.LR = 0.003
 _C.TRAINER.LR_DECAY = 10
 _C.TRAINER.LR_ADAPTIVE = None # choices = [None, loss, dann]
 _C.TRAINER.LR_STEPS = [10, 20]
+_C.TRAINER.MOMENTUM = 0.9
+_C.TRAINER.WEIGHT_DECAY = 0.0001
 _C.TRAINER.BATCH_SIZE = [128, 128*(_C.DATASET.NUM_TARGET/_C.DATASET.NUM_SOURCE), 128]
 _C.TRAINER.OPTIMIZER_NAME = "SGD" # choices = [SGD, Adam]
-_C.TRAINER.GD = 20
+_C.TRAINER.CLIP_GRADIENT = 20
 
 _C.TRAINER.PRETRAINED = None
+_C.TRAINER.RESUME = ""
+_C.TRAINER.RESUME_HP = ""
 
 _C.TRAINER.MIN_EPOCHS = 25
 _C.TRAINER.MAX_EPOCHS = 30
 
+_C.TRAINER.ACCELERATOR = "ddp"
+
+
+
+_C.PATHS.EXP_PATH = os.path.join(_C.DATASET.PATH_EXP + '_' + _C.TRAINER.OPTIMIZER_NAME + '-share_params_' + _C.MODEL.SHARE_PARAMS + '-lr_' + str(_C.TRAINER.LR) + '-bS_' + str(_C.TRAINER.BATCH_SIZE[0]), _C.DATASET.DATASET + '-'+ str(_C.DATASET.NUM_SEGMENTS) + '-seg-disDA_' + _C.MODEL.DIS_DA + '-alpha_' + str(_C.HYPERPARAMETERS.ALPHA) + '-advDA_' + _C.MODEL.ADV_DA + '-beta_' + str(_C.HYPERPARAMETERS.BETA[0])+ '_'+ str(_C.HYPERPARAMETERS.BETA[1])+'_'+ str(_C.HYPERPARAMETERS.BETA[2])+"_gamma_" + str(_C.HYPERPARAMETERS.GAMMA) + "_mu_" + str(_C.HYPERPARAMETERS.MU))
+
+
+# ---------------------------------------------------------------------------- #
+# Tester
+# ---------------------------------------------------------------------------- #
+_C.TESTER = CN()
+
+_C.TESTER.TEST_TARGET_DATA = os.path.join(_C.PATHS.PATH_DATA_ROOT, "target_test")
+
+_C.TESTER.WEIGHTS = os.path.join(_C.EXP_PATH , "checkpoint.pth.tar")
+_C.TESTER.NOUN_WEIGHTS = None
+_C.TESTER.RESULT_JSON = "test.json"
+_C.TESTER.TEST_SEGMENTS = 5 # sample frame # of each video for testing
+_C.TESTER.SAVE_SCORES = os.path.join(_C.EXP_PATH , "scores")
+_C.TESTER.SAVE_CONFUSION = os.path.join(_C.EXP_PATH , "confusion_matrix")
+
+_C.TESTER.VERBOSE = True
+
 # ---------------------------------------------------------------------------- #
 # Miscellaneous configs
 # ---------------------------------------------------------------------------- #
@@ -146,10 +185,10 @@
 _C.TRAINER.PF = 50
 _C.TRAINER.SF = 50
 _C.TRAINER.COPY_LIST = ["N", "N"]
+_C.TRAINER.SAVE_MODEL = True
 
 
 
-_C.DATASET.EXP_PATH = os.path.join(_C.DATASET.PATH_EXP + '_' + _C.TRAINER.OPTIMIZER_NAME + '-share_params_' + _C.MODEL.SHARE_PARAMS + '-lr_' + str(_C.TRAINER.LR) + '-bS_' + str(_C.TRAINER.BATCH_SIZE[0]), _C.DATASET.DATASET + '-'+ str(_C.DATASET.NUM_SEGMENTS) + '-seg-disDA_' + _C.MODEL.DIS_DA + '-alpha_' + str(_C.HYPERPARAMETERS.ALPHA) + '-advDA_' + _C.MODEL.ADV_DA + '-beta_' + str(_C.HYPERPARAMETERS.BETA[0])+ '_'+ str(_C.HYPERPARAMETERS.BETA[1])+'_'+ str(_C.HYPERPARAMETERS.BETA[2])+"_gamma_" + str(_C.HYPERPARAMETERS.GAMMA) + "_mu_" + str(_C.HYPERPARAMETERS.MU))
 
 
 

diff --git a/main_lightning.py b/main_lightning.py
@@ -1,21 +1,19 @@
+import os
 import numpy as np
 import time
+import argparse
 
 import torch
 import torch.nn.parallel
 import torch.optim
 
-from colorama import init
-from colorama import Fore, Back, Style
-
-
 from pytorch_lightning import Trainer
 from pytorch_lightning.callbacks import ModelCheckpoint
 from tensorboardX import SummaryWriter
 
 from utils.loss import *
-from utils.opts import parser
 from utils.model_init import initialise_trainer
+from config import get_cfg_defaults
 from utils.data_loaders import get_train_data_loaders, get_val_data_loaders
 from utils.logging import *
 
@@ -25,47 +23,53 @@
 torch.manual_seed(1)
 torch.cuda.manual_seed_all(1)
 
-init(autoreset=True)
-
-best_prec1 = 0
 gpu_count = torch.cuda.device_count()
 log_info( "Number of GPUS available: " + str(gpu_count))
 
+def arg_parse():
+    """Parsing arguments"""
+    parser = argparse.ArgumentParser(description="TA3N Domain Adaptation")
+    parser.add_argument("--cfg", required=True, help="path to config file", type=str)
+    parser.add_argument("--gpus", default="0", help="gpu id(s) to use", type=str)
+    parser.add_argument("--resume", default="", type=str)
+    args = parser.parse_args()
+    return args
+
 def main():
-	args = parser.parse_args()
+	args = arg_parse()
+	cfg = get_cfg_defaults()
+	cfg.merge_from_file(args.cfg)
+	cfg.freeze()
 
-	path_exp = args.exp_path + args.modality + '/'
+	# log_info(str(cfg))
+
+	path_exp = os.path.join(cfg.PATHS.EXP_PATH, cfg.DATASET.MODALITY)
 
 	#========== model init ========#
 
 	log_info('Initialising model......')
-	model = initialise_trainer(args)
+	model = initialise_trainer(cfg)
 
 	#========== log files init ========#
 
-	open_log_files(args)
+	open_log_files(cfg)
 
 	#========== Data loading ========#
 
 	log_info('Loading data......')
-
-	if args.use_opencv:
-		log_debug("use opencv functions")
-
-	source_loader, target_loader = get_train_data_loaders(args)
+	source_loader, target_loader = get_train_data_loaders(cfg)
 
-	to_validate  = args.val_source_data != "none" and args.val_target_data != "none" 
-	if(to_validate):
+	if(cfg.TO_VALIDATE):
 		log_info('Loading validation data......')
-		source_loader_val, target_loader_val = get_val_data_loaders(args)
+		source_loader_val, target_loader_val = get_val_data_loaders(cfg)
 
 	#========== Callbacks and checkpoints ========#
 
-	if args.train_metric == "all":
+	if cfg.TRAINER.TRAIN_METRIC == "all":
 		monitor = "Prec@1 Action"
-	elif args.train_metric == "noun":
+	elif cfg.TRAINER.TRAIN_METRIC == "noun":
 		monitor = "Prec@1 Noun"
-	elif args.train_metric == "verb":
+	elif cfg.TRAINER.TRAIN_METRIC == "verb":
 		monitor = "Prec@1 Verb" 
 	else:
 		log_error("invalid metric to train")
@@ -81,12 +85,18 @@ def main():
 
 	#========== Actual Training ========#
 
-	trainer = Trainer(min_epochs=20, max_epochs=30, callbacks=[checkpoint_callback], gpus = gpu_count, accelerator='ddp')
+	trainer = Trainer(
+		min_epochs=cfg.TRAINER.MIN_EPOCHS, 
+		max_epochs=cfg.TRAINER.MAX_EPOCHS, 
+		callbacks=[checkpoint_callback], 
+		gpus = args.gpus, 
+		accelerator=cfg.TRAINER.ACCELERATOR
+	)
 
 	log_info('Starting training......')	
 	start_train = time.time()
 
-	if(to_validate):
+	if(cfg.TO_VALIDATE):
 		trainer.fit(model, (source_loader, target_loader), (source_loader_val, target_loader_val))
 	else:
 		trainer.fit(model, (source_loader, target_loader))
@@ -95,14 +105,14 @@ def main():
 
 	#========== Logging ========#
 
-	write_log_files('total time: {:.3f} '.format(end_train - start_train), best_prec1)
+	write_log_files('total time: {:.3f} '.format(end_train - start_train), model.best_prec1)
 	model.writer_train.close()
 	model.writer_val.close()
 
 	log_info('Training complete')
 	log_info('Total training time:' + str(end_train - start_train))
 
-	if(to_validate):
+	if(cfg.TO_VALIDATE):
 		log_info('Validation scores:\n |  Prec@1 Verb: ' + str(model.prec1_verb_val) + "\n |  Prec@1 Noun: " + str(model.prec1_noun_val)+ "\n |  Prec@1 Action: " + str(model.prec1_val) + "\n |  Prec@5 Verb: " + str(model.prec5_verb_val) + "\n |  Prec@5 Noun: " + str(model.prec5_noun_val) + "\n |  Prec@5 Action: " + str(model.prec5_val) + "\n |  Loss total: " + str(model.losses_val))
 
 

diff --git a/test_models_lightning.py b/test_models_lightning.py
@@ -1,4 +1,4 @@
-import numpy as np
+import argparse
 import json
 from json import encoder
 
@@ -9,37 +9,50 @@
 
 from pytorch_lightning import Trainer
 
-from colorama import init
-from colorama import Fore, Back, Style
-
-from utils.opts_test import parser
 from utils.model_init import initialise_tester
 from utils.data_loaders import get_test_data_loaders
-
+from config import get_cfg_defaults
 from utils.logging import *
 
 encoder.FLOAT_REPR = lambda o: format(o, '.3f')
-init(autoreset=True)
 
-def main():
+def arg_parse():
+	"""Parsing arguments"""
+	parser = argparse.ArgumentParser(description="TA3N Domain Adaptation Testing")
+	parser.add_argument("--cfg", required=True, help="path to config file", type=str)
+	parser.add_argument("--gpus", default="0", help="gpu id(s) to use", type=str)
+	parser.add_argument("--ckpt", default=None, help="pre-trained parameters for the model (ckpt files)", type=str)
 	args = parser.parse_args()
+	return args
+
+def main():
+	args = arg_parse()
+	cfg = get_cfg_defaults()
+	cfg.merge_from_file(args.cfg)
+	cfg.freeze()
 
 	#========== model init ========#
 
 	log_info('Preparing the model......')
-	verb_net, noun_net = initialise_tester(args)
+	verb_net, noun_net = initialise_tester(cfg)
 
 	#========== Data loading ========#
 
 	log_info('Loading data......')
-	data_loader = get_test_data_loaders(args)
-	log_info('Data loaded from: ' + args.test_target_data+".pkl")
+	data_loader = get_test_data_loaders(cfg)
+	log_info('Data loaded from: ' + cfg.TESTER.TEST_TARGET_DATA+".pkl")
 
 	#========== Actual Testing ========#
 	log_info('starting validation......')
-	trainer = Trainer(gpus = torch.cuda.device_count())
-
-	trainer.test(model = verb_net, test_dataloaders=data_loader, ckpt_path=args.weights, verbose = True)
+	trainer = Trainer(gpus = args.gpus)
+
+	if args.ckpt is None:
+		ckpt_path = cfg.TESTER.WEIGHTS
+	else:
+		ckpt_path = args.ckpt
+	trainer.test(model = verb_net, test_dataloaders=data_loader, ckpt_path=ckpt_path, verbose = cfg.TESTER.VERBOSE)
+	if noun_net is not None:
+		trainer.test(model = noun_net, test_dataloaders=data_loader, ckpt_path=ckpt_path, verbose = cfg.TESTER.VERBOSE)
 
 	log_info('validation complete')