def _convert_transformer_inputs(model, wps: WordpieceBatch, is_train):
# Adapter for the PyTorchWrapper. See https://thinc.ai/docs/usage-frameworks
kwargs = {
"input_ids": xp2torch(wps.input_ids),
"attention_mask": xp2torch(wps.attention_mask),
}
if wps.token_type_ids is not None:
kwargs["token_type_ids"] = xp2torch(wps.token_type_ids)
return ArgsKwargs(args=(), kwargs=kwargs), lambda dX: []
def to_hf_dict(self) -> Dict:
"""Return a dict similar to the format produced by the Huggingface
tokenizer, converting arrays to pytorch tensors as well.
"""
output = {
"input_ids": xp2torch(self.input_ids),
"attention_mask": xp2torch(self.attention_mask),
"input_texts": self.strings,
}
if self.token_type_ids is not None:
output["token_type_ids"] = xp2torch(self.token_type_ids)
return output
def _convert_transformer_inputs(model, wps: WordpieceBatch, is_train):
# Adapter for the HFWrapper. See https://thinc.ai/docs/usage-frameworks
hf_device = model.shims[0]._hfmodel.transformer.device
kwargs = {
# Note: remove conversion to long when PyTorch >= 1.8.0.
"input_ids": xp2torch(wps.input_ids).long().to(device=hf_device),
"attention_mask": xp2torch(wps.attention_mask).to(device=hf_device),
}
if wps.token_type_ids is not None:
kwargs["token_type_ids"] = (xp2torch(
wps.token_type_ids).long().to(device=hf_device))
return ArgsKwargs(args=(), kwargs=kwargs), lambda dX: []
def backprop(d_tokvecs: List[Floats2d]) -> ArgsKwargs:
# Restore entries for bos and eos markers.
row = model.ops.alloc2f(1, d_tokvecs[0].shape[1])
d_tokvecs = [model.ops.xp.vstack((row, arr, row)) for arr in d_tokvecs]
return ArgsKwargs(
args=(torch_tokvecs,),
kwargs={"grad_tensors": xp2torch(model.ops.pad(d_tokvecs))},
)
def test_pytorch_roundtrip_conversion():
import torch
xp_tensor = numpy.zeros((2, 3), dtype="f")
torch_tensor = xp2torch(xp_tensor)
assert isinstance(torch_tensor, torch.Tensor)
new_xp_tensor = torch2xp(torch_tensor)
assert numpy.array_equal(xp_tensor, new_xp_tensor)
def unsplit_by_doc(self,
arrays: List[List[Floats3d]]) -> "FullTransformerBatch":
xp = get_array_module(arrays[0][0])
return FullTransformerBatch(
spans=self.spans,
tokens=self.tokens,
tensors=[xp2torch(xp.vstack(x)) for x in transpose_list(arrays)],
align=self.align,
)
def unsplit_by_doc(self,
arrays: List[List[Floats3d]]) -> "FullTransformerBatch":
"""Return a new FullTransformerBatch from a split batch of activations,
using the current object's spans, wordpieces and alignment.
This is used during the backward pass, in order to construct the gradients
to pass back into the transformer model.
"""
xp = get_array_module(arrays[0][0])
return FullTransformerBatch(
spans=self.spans,
wordpieces=self.wordpieces,
tensors=[xp2torch(xp.vstack(x)) for x in transpose_list(arrays)],
align=self.align,
)
def unsplit_by_doc(self,
arrays: List[List[Floats3d]]) -> "FullTransformerBatch":
"""Return a new FullTransformerBatch from a split batch of activations,
using the current object's spans, wordpieces and alignment.
This is used during the backward pass, in order to construct the gradients
to pass back into the transformer model.
"""
xp = get_array_module(arrays[0][0])
# construct a dummy ModelOutput with the tensor values
model_output = ModelOutput()
for i, x in enumerate(transpose_list(arrays)):
model_output[f"output_{i}"] = xp2torch(xp.vstack(x))
return FullTransformerBatch(
spans=self.spans,
wordpieces=self.wordpieces,
model_output=model_output,
align=self.align,
)