DIM=256,
classes=opt.n_classes)
# HCN pre-trained model
if opt.temporal_length == 25:
fc7_dim = 2048
elif opt.temporal_length == 60:
fc7_dim = 4096
if opt.num_per == 1:
hcn_model = HCN4(in_channel=3,
num_person=1,
num_class=49,
fc7_dim=fc7_dim).cuda()
elif opt.num_per == 2:
hcn_model = HCN(in_channel=3,
num_person=2,
num_class=49,
fc7_dim=fc7_dim).cuda()
if opt.act_pretrain == 1:
#hcn_model_path = '/fast-stripe/workspaces/deval/synthetic-data/hcn/CCC_models/clall_len60_per1/model_state.pth'
if opt.real_per == 100:
hcn_model_path = '/fast-stripe/workspaces/deval/synthetic-data/hcn/int_act_val_split_len60_per1/model_state.pth'
if opt.real_per == 25:
hcn_model_path = '/fast-stripe/workspaces/deval/synthetic-data/hcn/clall_len60_per1_realper25/model_state.pth'
hcn_model.load_state_dict(
torch.load(hcn_model_path)['best_test_model'], strict=False)
print("Loaded action model", hcn_model_path, "| Best test per",
torch.load(hcn_model_path)['best_test_perf'],
"| Best test epoch",
torch.load(hcn_model_path)['best_test_epoch'])
else:
images_dir = os.path.join('/fast-stripe/workspaces/deval/synthetic-data/wgan_gp_action_class',str(class_num),opt.expt,'images')
videos_dir = os.path.join('/fast-stripe/workspaces/deval/synthetic-data/wgan_gp_action_class',str(class_num),opt.expt,'videos')
models_dir = os.path.join('/fast-stripe/workspaces/deval/synthetic-data/wgan_gp_action_class',str(class_num),opt.expt,'models')
os.makedirs(images_dir, exist_ok=True)
os.makedirs(ckpt_dir, exist_ok=True)
os.makedirs(videos_dir,exist_ok=True)
#HCN Pre-trained model
if opt.temporal_length == 25:
fc7_dim = 2048
elif opt.temporal_length == 60:
fc7_dim = 4096
if opt.num_per == 1:
hcn_model = HCN1per_feat_1(in_channel=3,num_person=1,num_class=49,fc7_dim=fc7_dim).cuda()
elif opt.num_per == 2:
hcn_model = HCN(in_channel=3,num_person=2,num_class=49,fc7_dim=fc7_dim).cuda()
hcn_model_path = '/fast-stripe/workspaces/deval/synthetic-data/hcn/CCC_models/clall_len60_per1/model_state.pth'
hcn_model.load_state_dict(torch.load(hcn_model_path)['best_test_model'], strict=False)
print ("Loaded model",hcn_model_path,"| Best test per",torch.load(hcn_model_path)['best_test_perf'],"| Best test epoch",torch.load(hcn_model_path)['best_test_epoch'])
# Loss weight for gradient penalty
lambda_gp = opt.lambda_gp
# Initialize generator and discriminator
if opt.arch == 'mlp':
if opt.feat_concat == 0:
generator = Generator(latent_dim=opt.latent_dim,img_shape=img_shape)
elif opt.feat_concat == 512:
generator = Generator_concat512(latent_dim=opt.latent_dim,img_shape=img_shape)
discriminator = Discriminator(img_shape=img_shape)
elif opt.arch == 'cnn':
# Initialize generator and discriminator
if opt.arch == 'mlp':
generator = Generator(latent_dim = opt.latent_dim, img_shape=img_shape, n_classes=opt.n_classes)
discriminator = Discriminator(img_shape = img_shape, n_classes=opt.n_classes)
elif opt.arch == 'cnn':
generator = Generator_cnn(latent_dim=opt.latent_dim,DIM=256,classes=opt.n_classes)
# HCN pre-trained model
if opt.temporal_length == 25:
fc7_dim = 2048
elif opt.temporal_length == 60:
fc7_dim = 4096
if opt.num_per == 1:
hcn_model = HCN4(in_channel=3,num_person=1,num_class=49,fc7_dim=fc7_dim).cuda()
elif opt.num_per == 2:
hcn_model = HCN(in_channel=3,num_person=2,num_class=49,fc7_dim=fc7_dim).cuda()
if opt.act_pretrain == 1:
#hcn_model_path = '/fast-stripe/workspaces/deval/synthetic-data/hcn/CCC_models/clall_len60_per1/model_state.pth'
hcn_model_path = '/fast-stripe/workspaces/deval/synthetic-data/hcn/int_act_val_split_len60_per1/model_state.pth'
hcn_model.load_state_dict(torch.load(hcn_model_path)['best_test_model'], strict=False)
print ("Loaded action model",hcn_model_path,"| Best test per",torch.load(hcn_model_path)['best_test_perf'],"| Best test epoch",torch.load(hcn_model_path)['best_test_epoch'])
else:
print ("Not using any pretrained action classifer, training from scratch")
if opt.gen_pretrain == 1:
gen_model_path = '/fast-stripe/workspaces/deval/synthetic-data/cgan/img25_changelabels_per1_ep5k/models/3000/model_state_100.pth'
generator.load_state_dict(torch.load(gen_model_path)['last_state_dict_G'], strict=True)
print ("Loaded generator model",gen_model_path)
else:
print ("Not using any pretrained generator, training from scratch")
os.makedirs(videos_dir, exist_ok=True)
os.makedirs(models_dir, exist_ok=True)
#HCN Pre-trained model
if opt.temporal_length == 25:
fc7_dim = 2048
elif opt.temporal_length == 60:
fc7_dim = 4096
if opt.num_per == 1:
hcn_model = HCN1per_feat_1(in_channel=3,
num_person=1,
num_class=49,
fc7_dim=fc7_dim).cuda()
elif opt.num_per == 2:
hcn_model = HCN(in_channel=3,
num_person=2,
num_class=49,
fc7_dim=fc7_dim).cuda()
hcn_model_path = '/fast-stripe/workspaces/deval/synthetic-data/hcn/CCC_models/clall_len60_per1/model_state.pth'
hcn_model.load_state_dict(torch.load(hcn_model_path)['best_test_model'],
strict=False)
print("Loaded model", hcn_model_path, "| Best test per",
torch.load(hcn_model_path)['best_test_perf'], "| Best test epoch",
torch.load(hcn_model_path)['best_test_epoch'])
# Loss functions
adversarial_loss = torch.nn.MSELoss()
triplet_loss_torch = torch.nn.TripletMarginLoss(
margin=opt.margin) #2.0 default
cosine_embedding_loss = torch.nn.CosineEmbeddingLoss(
margin=opt.margin) #0.5 default
#adversarial_loss = torch.nn.BCEWithLogitsLoss()
if args.temporal_length == 25:
fc7_dim = 2048
elif args.temporal_length == 60:
fc7_dim = 4096
if args.feat_map == 0:
if args.arch == 'hcn':
if args.num_person == 1:
net = HCN4(
in_channel=3, num_person=1, num_class=49, fc7_dim=fc7_dim
).to(
device
) # HCN4 - FC8 HCN with 512 as fc; HCN1per_256_fc9 - FC9 with 256&512 as FCs
elif args.num_person == 2:
net = HCN(in_channel=3,
num_person=2,
num_class=49,
fc7_dim=fc7_dim).to(device)
elif args.arch == 'hcn_ensemble':
net = ensemble_model(arch=args.arch,
num_person=args.num_person,
num_classes=49,
fc7_dim=fc7_dim,
device=device)
elif args.feat_map == 1 or args.feat_map == 2:
if args.arch == 'hcn':
if args.num_person == 1:
net = HCN1per_feat256_fc9(
in_channel=3, num_person=1, num_class=49, fc7_dim=fc7_dim
).to(
device
) # HCN1per_feat_1 - FC8 with 512 feat maps; HCN1per_feat256_fc9 - 256 feature maps