def __init__(self, model="dimo", dataset_index=0):
# First, create the weighted sampler by analyzing the dataset and ascribing proper class weights
self.num_inner_nodes = 2
self.length_of_sequence = 16
pretrained_model_selection = args.pretrained_model
#pretrained_model_selection = "./results/viewpoint_models/vgg_viewpoint_experiments/v1/model_epoch_5.pth"
self.model = DimoAutoSequence(
pretrained_model=pretrained_model_selection,
num_inner_nodes=self.num_inner_nodes,
max_len=self.length_of_sequence)
self.model = nn.DataParallel(self.model, device_ids=[0, 1, 2,
3]).cuda()
self.model.load_state_dict(torch.load(args.saved_model))
mylogger.log("Playtype classifier ready")
print self.model
def __init__(self, model="dimo" ,dataset_index=0):
# First, create the weighted sampler by analyzing the dataset and ascribing proper class weights
self.num_inner_nodes = 2
self.length_of_sequence = 16
customset_preprocess = CustomMasterDatasetPlays(path = args.dataset_path,subset_type="training",dataset_index=dataset_index,categories=args.categories, retrieve_images=False)
self.processloader = torch.utils.data.DataLoader(dataset=customset_preprocess,batch_size=int(1),shuffle=False,num_workers=int(args.num_workers))
sample_plays = [] # when you start
for batch_idx, (imgs, play_type) in enumerate(self.processloader):
sample_plays.append(play_type.cpu().numpy()[0][0])
cat = 3
if args.categories == "unchunked":
cat = 13
if args.categories == "two":
cat = 2
self.cat = cat
class_presence = [0 for i in range(self.cat)]
for play in sample_plays:
class_presence[play] += 1
for i in range(cat):
class_presence[i] /= len(sample_plays)*1.0
class_weights = [0 for i in range(len(sample_plays))]
for i in range(len(sample_plays)):
class_weights[i] = 1.0/class_presence[sample_plays[i]]
m = sum(class_weights)
class_weights = [i*1.0/m for i in class_weights]
customset_train = CustomMasterDatasetPlays(path = args.dataset_path,subset_type="training",dataset_index=dataset_index,categories=args.categories,frame_select = self.length_of_sequence)
self.trainset = customset_train
mylogger.log("Loaded {} dataset with {} number of plays".format(customset_train.subset_type, customset_train.maxlength))
customset_excite = CustomMasterDatasetPlays(path = args.dataset_path,subset_type="exciting",dataset_index=dataset_index,categories=args.categories,frame_select = self.length_of_sequence)
self.exciteset = customset_excite
mylogger.log("Loaded {} dataset with {} number of plays".format("exciting", customset_train.maxlength))
customset_test = CustomMasterDatasetPlays(path = args.dataset_path,subset_type="testing",dataset_index=dataset_index,categories=args.categories,frame_select = self.length_of_sequence)
self.testset = customset_test
mylogger.log("Loaded {} dataset with {} number of plays".format(customset_test.subset_type, customset_test.maxlength))
# Finished with sampler weighting
sampler = torch.utils.data.sampler.WeightedRandomSampler(class_weights,len(self.processloader),replacement=True)
self.trainloader = torch.utils.data.DataLoader(dataset=customset_train,sampler=sampler,batch_size=int(args.batch_size),num_workers=int(args.num_workers))
self.train_acc_loader = torch.utils.data.DataLoader(dataset=customset_train,batch_size=int(args.batch_size),shuffle=False,num_workers=int(args.num_workers))
self.test_acc_loader = torch.utils.data.DataLoader(dataset=customset_test,batch_size=int(args.batch_size),shuffle=False,num_workers=int(args.num_workers))
self.exciteloader_train = torch.utils.data.DataLoader(dataset=customset_excite,batch_size=1,shuffle=False,num_workers=int(args.num_workers))
self.exciteloader = torch.utils.data.DataLoader(dataset=customset_test,batch_size=1,shuffle=True,num_workers=int(args.num_workers))
self.lr = 0.0001
mylogger.log("-dotted-line")
mylogger.log("Using worker count: {}".format(args.num_workers))
mylogger.log("Using epoch count: {}".format(args.epochs))
mylogger.log("Using dataset index: {}".format(args.dataset_index))
mylogger.log("Using model <{}>".format(args.pretrained_model))
mylogger.log("Using weight decay: {}".format(args.weight_decay))
if args.training:
mylogger.log("Training schedule type <training>")
mylogger.log("Using categories <{}>".format(args.categories))
mylogger.log("Using learning rate <{}>".format(self.lr))
pretrained_model_selection = args.pretrained_model
#pretrained_model_selection = "./results/viewpoint_models/vgg_viewpoint_experiments/v1/model_epoch_5.pth"
self.model = DimoAutoSequence(pretrained_model=pretrained_model_selection,num_inner_nodes=self.num_inner_nodes,max_len=self.length_of_sequence)
self.model = nn.DataParallel(self.model,device_ids=[0,1,2,3]).cuda()
if args.saved_model != None:
self.model.load_state_dict(torch.load(args.saved_model))
self.optimizer = optim.Adam(self.model.parameters(), weight_decay=float(args.weight_decay), lr= self.lr)
self.criterion = nn.NLLLoss().cuda()
mylogger.log("Playtype classifier ready")
print self.model
def __init__(self, model="dimo", dataset_index=0):
# First, create the weighted sampler by analyzing the dataset and ascribing proper class weights
print "Initializing playtype classifier"
final_length = 16
view = "both"
first_save = False
sampler = None
if first_save: # takes > 1 minute to even compute, compare with 5 seconds load time
customset_preprocess = MasterPlaysets(
path=args.dataset_path,
subset_type="training",
retrieve_images=False,
subsampling=None,
fill=None,
from_file="gt",
part="part",
view=view) # gt vs est, whole vs. part, 'both' vs '0' vs '1'
self.processloader = torch.utils.data.DataLoader(
dataset=customset_preprocess,
batch_size=int(1),
shuffle=False,
num_workers=int(args.num_workers))
class_presence = [0, 0, 0]
for batch_idx, (imgs, play_type) in enumerate(self.processloader):
class_presence[play_type.cpu().numpy()[0][0]] += 1
sample_weights = [
(1.0 / class_presence[play_type.cpu().numpy()[0][0]])
for (imgs, play_type) in self.processloader
]
sampler = torch.utils.data.sampler.WeightedRandomSampler(
sample_weights, len(self.processloader), replacement=True)
# mstd = MasterPlaysets(path = args.dataset_path,subset_type="training", retrieve_images=True, subsampling=None,fill=None, from_file="gt", part="part", view="both") # gt vs est, whole vs. part, 'both' vs '0' vs '1'
# self.processloader_weighted = torch.utils.data.DataLoader(sampler=sampler,dataset=mstd,batch_size=int(1),shuffle=False,num_workers=int(args.num_workers))
# customset_preprocess.compute_mean_std(self.processloader_weighted)
print "Loading datasets"
customset_train = MasterPlaysets(
path=args.dataset_path,
subset_type="training",
retrieve_images=True,
subsampling=None,
fill=None,
from_file="gt",
part="part",
view=view) # gt vs est, whole vs. part, 'both' vs '0' vs '1'
customset_test = MasterPlaysets(
path=args.dataset_path,
subset_type="testing",
retrieve_images=True,
subsampling=None,
fill=None,
from_file="gt",
part="part",
view=view) # gt vs est, whole vs. part, 'both' vs '0' vs '1'
customsets = [
customset_preprocess, customset_train, customset_test
]
pickle.dump(customsets, open("./junk/customset_v0.pkl", "wb"))
else:
customsets = pickle.load(open("./junk/customset_v0.pkl", "rb"))
customset_preprocess, customset_train, customset_test = customsets
self.processloader = torch.utils.data.DataLoader(
dataset=customset_preprocess,
batch_size=int(1),
shuffle=False,
num_workers=int(args.num_workers))
class_presence = [0, 0, 0]
for batch_idx, (imgs, play_type) in enumerate(self.processloader):
class_presence[play_type.cpu().numpy()[0][0]] += 1
sample_weights = [
(1.0 / class_presence[play_type.cpu().numpy()[0][0]])
for (imgs, play_type) in self.processloader
]
sampler = torch.utils.data.sampler.WeightedRandomSampler(
sample_weights, len(self.processloader), replacement=True)
customset_train.proper_length = final_length
customset_test.proper_length = final_length
self.trainloader = torch.utils.data.DataLoader(
dataset=customset_train,
batch_size=int(args.batch_size),
shuffle=False,
num_workers=int(args.num_workers))
self.train_acc_loader = torch.utils.data.DataLoader(
dataset=customset_train,
batch_size=int(args.batch_size),
shuffle=False,
num_workers=int(args.num_workers))
#self.trainloader = torch.utils.data.DataLoader(sampler=sampler,dataset=customset_train,batch_size=int(args.batch_size),num_workers=int(args.num_workers))
#self.train_acc_loader = torch.utils.data.DataLoader(sampler=sampler,dataset=customset_train,batch_size=int(args.batch_size),num_workers=int(args.num_workers))
#self.train_acc_loader = torch.utils.data.DataLoader(dataset=customset_train,batch_size=int(args.batch_size),shuffle=False,num_workers=int(args.num_workers))
self.test_acc_loader = torch.utils.data.DataLoader(
dataset=customset_test,
batch_size=int(args.batch_size),
shuffle=False,
num_workers=int(args.num_workers))
self.lr = 0.0001
mylogger.log("-dotted-line")
mylogger.log("Using worker count: {}".format(args.num_workers))
mylogger.log("Using epoch count: {}".format(args.epochs))
mylogger.log("Using model <{}>".format(args.pretrained_model))
mylogger.log("Using weight decay: {}".format(args.weight_decay))
if args.training:
mylogger.log("Training schedule type <training>")
mylogger.log("Using learning rate <{}>".format(self.lr))
pretrained_model_selection = args.pretrained_model
#pretrained_model_selection = "./results/viewpoint_models/vgg_viewpoint_experiments/v1/model_epoch_5.pth"
print "Building model"
self.num_inner_nodes = 2
self.length_of_sequence = final_length
self.model = DimoAutoSequence(
pretrained_model=pretrained_model_selection,
num_inner_nodes=self.num_inner_nodes,
max_len=self.length_of_sequence).cuda()
self.model = nn.DataParallel(self.model, device_ids=[0, 1, 2,
3]).cuda()
if args.saved_model != None:
self.model.load_state_dict(torch.load(args.saved_model))
self.optimizer = optim.Adam(self.model.parameters(),
weight_decay=float(args.weight_decay),
lr=self.lr)
self.criterion = nn.NLLLoss().cuda()
mylogger.log("Playtype classifier ready")
mylogger.log(self.model)