def test_lm_model_forward(self):
config = self._get_basic_config_and_input()
config["attn_layers"] = ["local", "lsh", "local", "lsh", "local", "lsh"]
config["num_buckets"] = [2, 4]
config["is_decoder"] = False
torch.manual_seed(0)
model = ReformerForMaskedLM(ReformerConfig(**config)).to(torch_device)
model.eval()
input_ids, attn_mask = self._get_input_ids_and_mask()
hidden_states = model(input_ids=input_ids, attention_mask=attn_mask)[0]
output_slice = hidden_states[1, -1, :5]
expected_output_slice = torch.tensor(
[0.0256, -0.0121, 0.0636, 0.0024, -0.0393], dtype=torch.float, device=torch_device,
)
self.assertTrue(torch.allclose(output_slice, expected_output_slice, atol=1e-3))
def masked_mlm():
from transformers import ReformerConfig, ReformerForMaskedLM
config = ReformerConfig.from_pretrained('google/reformer-enwik8')
config.is_decoder = False
model = ReformerForMaskedLM.from_pretrained('google/reformer-enwik8',
config=config)
sentence = sentence2 = "The quick brown fox jumps over the lazy dog."
input_ids, attention_masks = encode([sentence])
if True:
_input_ids, a = input_ids.clone(), attention_masks.clone()
for i in [19, 27, 37]:
a[0, i] = 0
sentence2 = sentence2[:i] + "%" + sentence2[i + 1:]
else:
_input_ids, a = input_ids, attention_masks
f = model.forward(input_ids=_input_ids,
position_ids=None,
attention_mask=a,
head_mask=None,
inputs_embeds=None,
num_hashes=None,
labels=_input_ids)
prediction = decode(torch.argmax(f.logits, 2))[0]
print(sentence2)
print(prediction)
def create_and_check_reformer_with_mlm(self, config, input_ids, input_mask, choice_labels):
config.is_decoder = False
model = ReformerForMaskedLM(config=config)
model.to(torch_device)
model.eval()
result = model(input_ids, attention_mask=input_mask, labels=input_ids)
self.parent.assertEqual(result.logits.shape, (self.batch_size, self.seq_length, self.vocab_size))
def create_and_check_reformer_model_with_lm_backward(self, config, input_ids, input_mask, choice_labels):
config.is_decoder = False
config.lsh_num_chunks_after = 1
model = ReformerForMaskedLM(config=config)
model.to(torch_device)
model.eval()
loss = model(input_ids, attention_mask=input_mask, labels=input_ids)[0]
loss.backward()
def create_and_check_reformer_no_chunking(self, config, input_ids, input_mask, choice_labels):
# force chunk length to be bigger than input_ids
config.lsh_attn_chunk_length = 2 * input_ids.shape[-1]
config.local_attn_chunk_length = 2 * input_ids.shape[-1]
config.lsh_num_chunks_after = 1
config.is_decoder = False
model = ReformerForMaskedLM(config=config)
model.to(torch_device)
model.eval()
output_logits = model(input_ids, attention_mask=input_mask)["logits"]
self.parent.assertTrue(output_logits.shape[1] == input_ids.shape[-1])
def create_and_check_reformer_with_mlm(self, config, input_ids, input_mask, choice_labels):
config.is_decoder = False
model = ReformerForMaskedLM(config=config)
model.to(torch_device)
model.eval()
loss, prediction_scores = model(input_ids, attention_mask=input_mask, labels=input_ids)
result = {
"loss": loss,
"prediction_scores": prediction_scores,
}
self.parent.assertListEqual(
list(result["prediction_scores"].size()), [self.batch_size, self.seq_length, self.vocab_size],
)
self.check_loss_output(result)
from transformers import ReformerConfig, ReformerForMaskedLM, ReformerTokenizer, LineByLineTextDataset,DataCollatorForLanguageModeling
from reformer_utils import encode, decode, CharTokenizer
import torch
from general_tools.utils import get_root
from pathlib import Path
ROOT = get_root("internn")
MSK = 44
config = ReformerConfig.from_pretrained('google/reformer-enwik8')
config.is_decoder = False
model = ReformerForMaskedLM.from_pretrained('google/reformer-enwik8', config=config)
sentence = "The quick brown fox jumps over the lazy dog."
input_ids, attention_masks = encode([sentence])
label_ids, _ = encode([sentence])
for idx in [10,21,26,32,35]:
input_ids[0,idx] = MSK
attention_masks[0,idx] = 0
f = model.forward(input_ids=input_ids,
position_ids=None,
attention_mask=attention_masks,
head_mask=None,
inputs_embeds=None,
num_hashes=None,
labels=label_ids)
loss = f.loss
prediction = decode(torch.argmax(f.logits, 2))[0]
print(prediction)
def create_and_check_reformer_feed_backward_chunking(self, config, input_ids, input_mask, choice_labels):
if not self.is_training:
return
# disable dropout
config.hidden_dropout_prob = 0
config.local_attention_probs_dropout_prob = 0
config.lsh_attention_probs_dropout_prob = 0
config.lsh_num_chunks_after = 1
config.is_decoder = False
torch.manual_seed(0)
model = ReformerForMaskedLM(config=config)
model.to(torch_device)
model.train()
model.zero_grad()
loss_no_chunk, output_no_chunk = model(input_ids, labels=input_ids, attention_mask=input_mask)[:2]
loss_no_chunk.backward()
grad_slice_word_no_chunk = model.reformer.embeddings.word_embeddings.weight.grad[0, :5]
grad_slice_position_factor_1_no_chunk = model.reformer.embeddings.position_embeddings.weights[0][1, 0, -5:]
grad_slice_position_factor_2_no_chunk = model.reformer.embeddings.position_embeddings.weights[1][0, 1, :5]
config.chunk_size_lm_head = 1
config.chunk_size_feed_forward = 1
torch.manual_seed(0)
model = ReformerForMaskedLM(config=config)
model.to(torch_device)
model.train()
model.zero_grad()
loss_chunk, output_chunk = model(input_ids, labels=input_ids, attention_mask=input_mask)[:2]
loss_chunk.backward()
grad_slice_word_chunk = model.reformer.embeddings.word_embeddings.weight.grad[0, :5]
grad_slice_position_factor_1_chunk = model.reformer.embeddings.position_embeddings.weights[0][1, 0, -5:]
grad_slice_position_factor_2_chunk = model.reformer.embeddings.position_embeddings.weights[1][0, 1, :5]
self.parent.assertTrue(torch.allclose(loss_chunk, loss_no_chunk, atol=1e-3))
self.parent.assertTrue(torch.allclose(grad_slice_word_no_chunk, grad_slice_word_chunk, atol=1e-3))
self.parent.assertTrue(
torch.allclose(grad_slice_position_factor_1_chunk, grad_slice_position_factor_1_no_chunk, atol=1e-3)
)
self.parent.assertTrue(
torch.allclose(grad_slice_position_factor_2_chunk, grad_slice_position_factor_2_no_chunk, atol=1e-3)
)
