Exemplo n.º 1
0
 def __init__(self, model=None, dataset=None):
     self.model = model
     self.data_loader = data.WavLoader(dataset)
     self.device = None
     self.torch_rng_state = t.get_rng_state()
     if t.cuda.is_available():
         self.torch_cuda_rng_states = t.cuda.get_rng_state_all()
     else:
         self.torch_cuda_rng_states = None
Exemplo n.º 2
0
 def __init__(self, step=0, model=None, dataset=None, optim=None):
     self.model = model
     self.data_loader = data.WavLoader(dataset)
     self.optim = optim
     self.step = step
     self.device = None
     self.torch_rng_state = torch.get_rng_state()
     if torch.cuda.is_available():
         self.torch_cuda_rng_states = torch.cuda.get_rng_state_all()
     else:
         self.torch_cuda_rng_states = None
Exemplo n.º 3
0
    def load(self, ckpt_file, dat_file):
        sinfo = torch.load(ckpt_file)

        # This is the required order for model and data init
        self.model = pickle.loads(sinfo['model'])
        # ignore the pickled dataset characteristics
        dataset = data.MfccInference(pickle.loads(sinfo['dataset']))
        dataset.load_data(dat_file)
        self.model.post_init(dataset)
        self.model.load_state_dict(sinfo['model_state_dict'])
        dataset.post_init(self.model)
        self.data_loader = data.WavLoader(dataset)
Exemplo n.º 4
0
    def load(self, ckpt_file, dat_file):
        sinfo = torch.load(ckpt_file)

        # This is the required order for model and data init
        self.model = pickle.loads(sinfo['model'])
        dataset = pickle.loads(sinfo['dataset'])
        dataset.load_data(dat_file)
        self.model.post_init(dataset)
        self.model.load_state_dict(sinfo['model_state_dict'])
        dataset.post_init(self.model)

        self.data_loader = data.WavLoader(dataset)
        self.optim = torch.optim.Adam(self.model.parameters())
        self.optim.load_state_dict(sinfo['optim'])
        self.step = sinfo['step']
        self.torch_rng_state = sinfo['rand_state']
        self.torch_cuda_rng_states = sinfo['cuda_rand_states']
Exemplo n.º 5
0
    def load(self, ckpt_file, dat_file):
        ckpt = t.load(ckpt_file)

        # This is the required order for model and data init
        self.model = pickle.loads(ckpt['model'])

        # win batch of 1 is inference mode
        self.model.override(n_win_batch=1)

        # ignore the pickled dataset characteristics
        dataset = data.MfccInference(pickle.loads(ckpt['dataset']), dat_file)

        # dataset.load_data(dat_file)
        self.model.post_init(dataset)
        sub_state = {
            k: v
            for k, v in ckpt['model_state_dict'].items()
            if '_lead' not in k and 'left_wing_size' not in k
        }
        self.model.load_state_dict(sub_state, strict=False)
        dataset.post_init(self.model)
        self.data_loader = data.WavLoader(dataset)
Exemplo n.º 6
0
 def __init__(self, model=None, dataset=None):
     self.model = model
     self.data_loader = data.WavLoader(dataset)
     self.device = None