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)
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))