def __init__(self,
vocab: List[str],
hidden_features: int,
enc_layers=1,
dec_layers=1,
nhead=1,
dropout=0.1):
super().__init__()
self.letter_to_token, _ = tokenize_vocab(vocab)
self.pos_encoder = PositionalEncoding(hidden_features, dropout)
self.decoder = nn.Embedding(len(vocab), hidden_features)
self.pos_decoder = PositionalEncoding(hidden_features, dropout)
self.transformer = nn.Transformer(d_model=hidden_features,
nhead=nhead,
num_encoder_layers=enc_layers,
num_decoder_layers=dec_layers,
dim_feedforward=hidden_features * 4,
dropout=dropout,
activation='relu')
self.src_mask = None
self.trg_mask = None
self.memory_mask = None
self.fc = nn.Linear(hidden_features, len(vocab), bias=False)
self._initialize_weights()
def __init__(self, hidden_features, vocab, dropout_p=0.1):
super(BahdanauAttnDecoderRNN, self).__init__()
self.hidden_size = hidden_features
self.output_size = len(vocab)
self.letter_to_token, _ = tokenize_vocab(vocab)
self.dropout_p = dropout_p
self.embedding = nn.Linear(self.output_size,
self.hidden_size,
bias=False)
self.attn_combine = nn.Linear(self.hidden_size * 2, self.hidden_size)
self.dropout = nn.Dropout(self.dropout_p)
self.gru = nn.GRU(self.hidden_size, self.hidden_size)
self.out = nn.Linear(self.hidden_size, self.output_size)
self.vat = nn.Linear(hidden_features, 1)
def __init__(
self,
vocab: List[str],
sequence_size: int,
backbone_cfg: CfgT = dict(),
encoder_cfg: Optional[CfgT] = None,
decoder_cfg: Optional[CfgT] = None,
head_cfg: Optional[CfgT] = None,
loss_cfgs: Optional[CfgT] = None,
metric_cfgs: List[CfgT] = list(),
train_transforms_cfg: CfgT = dict(),
val_transforms_cfg: CfgT = dict(),
train_dataset_cfg: CfgT = dict(),
val_dataset_cfg: CfgT = dict(),
train_dataloader_cfg: CfgT = dict(),
val_dataloader_cfg: CfgT = dict(),
optimizer_cfg: CfgT = dict(),
scheduler_cfg: Optional[CfgT] = None,
scheduler_update_params: Optional[CfgT] = None,
):
super(BaseOCRPLModule, self).__init__()
self.backbone_cfg = backbone_cfg
self.head_cfg = head_cfg
self.encoder_cfg = encoder_cfg
self.decoder_cfg = decoder_cfg
self.loss_cfgs = loss_cfgs
self.metric_cfgs = metric_cfgs
self.train_transforms_cfg = train_transforms_cfg
self.val_transforms_cfg = val_transforms_cfg
self.train_dataset_cfg = train_dataset_cfg
self.val_dataset_cfg = val_dataset_cfg
self.train_dataloader_cfg = train_dataloader_cfg
self.val_dataloader_cfg = val_dataloader_cfg
self.optimizer_cfg = optimizer_cfg
self.scheduler_cfg = scheduler_cfg
self.scheduler_update_params = scheduler_update_params
self.save_hyperparameters()
self.scheduler = None # Can be useful for LR logging in the progress bar
self.sequence_size = sequence_size
self.letter_to_token, self.token_to_letter = tokenize_vocab(vocab)
self._build_models()
def __init__(self, vocab, seq_size):
super().__init__(always_apply=True, p=1)
self.letter_to_token, _ = tokenize_vocab(vocab)
self.seq_size = seq_size
def __init__(self, letters):
super().__init__(compute_on_step=True)
_, self.token_to_letter = tokenize_vocab(letters)
self.add_state('dists', default=torch.tensor(0.), dist_reduce_fx="sum")
self.add_state('total_seq', default=torch.tensor(0), dist_reduce_fx="sum")