args_dict = {
'ENV_ID': 'Acrobot-v1',
'SEED': 1,
'NB_FRAMES': 10000,
'BATCH_SIZE': 32,
'DISCOUNT': 0.99,
'TARGET_UPDATE_STEPS': 100,
'LEARNING_RATE': 1e-3,
'REPLAY_BUFFER_SIZE': 1000,
'MIN_REPLAY_BUFFER_SIZE': 100,
'EPSILON_START': 1,
'EPSILON_END': 0.1,
'EPSILON_DECAY_DURATION': 5000,
}
# Allow changing hyperparameters from command-line arguments
args = get_args(default_args=args_dict)
# Create wrapped environment
env = make_env(args.ENV_ID)
# Set Seed
set_seed(env, args.SEED)
# GPU or CPU
device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu')
# Create agent
agent = Agent(env, device, args)
# Train agent for args.NB_FRAMES
agent.train()
tdc.eval()
cmc.eval()
plot_tsne(tsne_loaders, tdc, cmc, device)
tdc.train()
cmc.train()
tsne_end_t = time.time()
print(
"[pretrain] Updated t-SNE figure (Duration: {}s)".format(
tsne_end_t - tsne_start_t
)
)
if __name__ == "__main__":
# Parse arguments
args = get_args()
# Setup wandb
wandb.init(project="youtube")
wandb.config.update(args)
print("[pretrain] Wandb setup complete.")
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
dataset = TDCCMCDataset(
filenames=["./data/6zXXZvVvTFs", "./data/2AYaxTiWKoY", "./data/sYbBgkP9aMo"],
trims=[(960, 9960), (550, 9550), (1, 9901)],
crops=[(35, 50, 445, 300), (0, 13, 640, 335), (5, 22, 475, 341)],
)
loader = DataLoader(
dataset, batch_size=args.BATCH_SIZE, num_workers=0, pin_memory=True