Spaces:
Sleeping
Sleeping
File size: 4,201 Bytes
f672f40 f9d964d f672f40 f9d964d f672f40 f9d964d f672f40 f9d964d f672f40 f9d964d f672f40 f9d964d f672f40 f9d964d f672f40 f9d964d |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 |
import torch
from src.models.decoder import TransformerDecoder
from src.models.encoder import TransformerEncoder
from src.models.heads import ClassificationHead, LMHead, TokenClassificationHead
from src.models.multitask import MultiTaskModel
def test_multitask_encoder_classification_forward_and_loss():
torch.manual_seed(0)
vocab_size = 30
d_model = 32
num_layers = 2
num_heads = 4
d_ff = 64
batch_size = 3
seq_len = 8
num_labels = 5
enc = TransformerEncoder(
vocab_size=vocab_size,
d_model=d_model,
num_layers=num_layers,
num_heads=num_heads,
d_ff=d_ff,
dropout=0.0,
max_len=seq_len,
pad_token_id=0,
)
mt = MultiTaskModel(encoder=enc)
head = ClassificationHead(d_model=d_model, num_labels=num_labels, pooler="mean", dropout=0.0)
mt.add_head("sentiment", head)
input_ids = torch.randint(1, vocab_size, (batch_size, seq_len), dtype=torch.long)
labels = torch.randint(0, num_labels, (batch_size,), dtype=torch.long)
logits = mt.forward("sentiment", {"input_ids": input_ids})
assert logits.shape == (batch_size, num_labels)
loss, logits2 = mt.forward(
"sentiment", {"input_ids": input_ids, "labels": labels}, return_loss=True
)
assert loss.item() >= 0
# grads
loss.backward()
grads = [p.grad for p in mt.parameters() if p.requires_grad]
assert any(g is not None for g in grads)
def test_multitask_seq2seq_lm_forward_and_loss():
torch.manual_seed(1)
vocab_size = 40
d_model = 32
num_layers = 2
num_heads = 4
d_ff = 64
batch_size = 2
src_len = 7
tgt_len = 6
enc = TransformerEncoder(
vocab_size=vocab_size,
d_model=d_model,
num_layers=num_layers,
num_heads=num_heads,
d_ff=d_ff,
dropout=0.0,
max_len=src_len,
pad_token_id=0,
)
dec = TransformerDecoder(
vocab_size=vocab_size,
d_model=d_model,
num_layers=num_layers,
num_heads=num_heads,
d_ff=d_ff,
dropout=0.0,
max_len=tgt_len,
pad_token_id=0,
)
mt = MultiTaskModel(encoder=enc, decoder=dec)
lm_head = LMHead(d_model=d_model, vocab_size=vocab_size, tie_embedding=None)
mt.add_head("summarize", lm_head)
src_ids = torch.randint(1, vocab_size, (batch_size, src_len), dtype=torch.long)
# for training: provide decoder inputs (typically shifted right) and labels
tgt_ids = torch.randint(1, vocab_size, (batch_size, tgt_len), dtype=torch.long)
labels = tgt_ids.clone()
logits = mt.forward("summarize", {"src_ids": src_ids, "tgt_ids": tgt_ids})
assert logits.shape == (batch_size, tgt_len, vocab_size)
loss, logits2 = mt.forward(
"summarize", {"src_ids": src_ids, "tgt_ids": tgt_ids, "labels": labels}, return_loss=True
)
assert loss.item() >= 0
loss.backward()
grads = [p.grad for p in mt.parameters() if p.requires_grad]
assert any(g is not None for g in grads)
def test_token_classification_forward_and_loss():
torch.manual_seed(2)
vocab_size = 20
d_model = 24
num_layers = 2
num_heads = 4
d_ff = 64
batch_size = 2
seq_len = 5
num_labels = 7
enc = TransformerEncoder(
vocab_size=vocab_size,
d_model=d_model,
num_layers=num_layers,
num_heads=num_heads,
d_ff=d_ff,
dropout=0.0,
max_len=seq_len,
pad_token_id=0,
)
mt = MultiTaskModel(encoder=enc)
head = TokenClassificationHead(d_model=d_model, num_labels=num_labels, dropout=0.0)
mt.add_head("ner", head)
input_ids = torch.randint(1, vocab_size, (batch_size, seq_len), dtype=torch.long)
labels = torch.randint(0, num_labels, (batch_size, seq_len), dtype=torch.long)
logits = mt.forward("ner", {"input_ids": input_ids})
assert logits.shape == (batch_size, seq_len, num_labels)
loss, logits2 = mt.forward("ner", {"input_ids": input_ids, "labels": labels}, return_loss=True)
assert loss.item() >= 0
loss.backward()
grads = [p.grad for p in mt.parameters() if p.requires_grad]
assert any(g is not None for g in grads)
|