ammesatyajit

class Encoder(nn.Module):
    def __init__(self, 
                 input_dim, 
                 hid_dim, 
                 n_layers, 
                 n_heads, 
                 pf_dim,
                 dropout, 
                 device,
                 max_length = 100):

ammesatyajit

class MultiHeadAttentionLayer(nn.Module):
    def __init__(self, hid_dim, n_heads, dropout, device):
        super().__init__()
        
        assert hid_dim % n_heads == 0
        
        self.hid_dim = hid_dim
        self.n_heads = n_heads
        self.head_dim = hid_dim // n_heads
        

ammesatyajit

class PositionwiseFeedforwardLayer(nn.Module):
    def __init__(self, hid_dim, pf_dim, dropout):
        super().__init__()
        
        self.fc_1 = nn.Linear(hid_dim, pf_dim)
        self.fc_2 = nn.Linear(pf_dim, hid_dim)
        
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, x):

ammesatyajit

class EncoderLayer(nn.Module):
    def __init__(self, hid_dim, n_heads, pf_dim,  dropout, device):
        super().__init__()
        
        self.self_attn_layer_norm = nn.LayerNorm(hid_dim)
        self.ff_layer_norm = nn.LayerNorm(hid_dim)
        self.self_attention = MultiHeadAttentionLayer(hid_dim, n_heads, dropout, device)
        self.positionwise_feedforward = PositionwiseFeedforwardLayer(hid_dim, pf_dim, dropout)
        self.dropout = nn.Dropout(dropout)
        

ammesatyajit

device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')

BATCH_SIZE = 128

train_iterator, valid_iterator, test_iterator = BucketIterator.splits(
    (train_data, valid_data, test_data), 
     batch_size = BATCH_SIZE,
     device = device)

ammesatyajit

train_data, valid_data, test_data = IWSLT.splits(exts = ('.en', '.de'), fields = (SRC, TRG))

SRC.build_vocab(train_data, min_freq = 2)
TRG.build_vocab(train_data, min_freq = 2)

ammesatyajit

SRC = Field(tokenize = tokenize_en, 
            init_token = '<sos>', 
            eos_token = '<eos>', 
            lower = True, 
            batch_first = True)

TRG = Field(tokenize = tokenize_de, 
            init_token = '<sos>', 
            eos_token = '<eos>', 
            lower = True, 

ammesatyajit

def tokenize_de(text):
    """
    Tokenizes German text from a string into a list of strings
    """
    return [tok.text for tok in spacy_de.tokenizer(text)]

def tokenize_en(text):
    """
    Tokenizes English text from a string into a list of strings
    """

ammesatyajit

!python -m spacy download en
!python -m spacy download de

spacy_de = spacy.load('de')
spacy_en = spacy.load('en')

ammesatyajit

SEED = 1234

random.seed(SEED)
np.random.seed(SEED)
torch.manual_seed(SEED)
torch.cuda.manual_seed(SEED)
torch.backends.cudnn.deterministic = True