def output_spec(self) -> Spec:
ret = {"tokens": lit_types.Tokens()}
ret["tokens_" + self.config.text_a_name] = lit_types.Tokens()
if self.config.text_b_name:
ret["tokens_" + self.config.text_b_name] = lit_types.Tokens()
if self.is_regression:
ret["score"] = lit_types.RegressionScore(
parent=self.config.label_name)
else:
ret["probas"] = lit_types.MulticlassPreds(
parent=self.config.label_name,
vocab=self.config.labels,
null_idx=self.config.null_label_idx)
ret["cls_emb"] = lit_types.Embeddings()
# Gradients, if requested.
if self.config.compute_grads:
ret["token_grad_" +
self.config.text_a_name] = lit_types.TokenGradients(
align="tokens_" + self.config.text_a_name)
if self.config.text_b_name:
ret["token_grad_" +
self.config.text_b_name] = lit_types.TokenGradients(
align="tokens_" + self.config.text_b_name)
# Attention heads, one field for each layer.
for i in range(self.model.config.num_hidden_layers):
ret[f"layer_{i}/attention"] = lit_types.AttentionHeads(
align=("tokens", "tokens"))
return ret
def output_spec(self):
spec = super().output_spec() # has 'output_text'
spec.update({
"input_tokens":
lit_types.Tokens(parent="input_text"),
"encoder_final_embedding":
lit_types.Embeddings(),
# If target text is given, the following will also be populated.
"target_tokens":
lit_types.Tokens(parent="target_text"),
"pred_tokens":
lit_types.TokenTopKPreds(align="target_tokens"),
})
if self.config.num_to_generate > 1:
spec["output_text"] = lit_types.GeneratedTextCandidates(
parent="target_text")
if self.config.output_attention:
# Add attention for each layer.
for i in range(self.num_layers):
spec[
f"encoder_layer_{i:d}_attention"] = lit_types.AttentionHeads(
align_in="input_tokens", align_out="input_tokens")
spec[
f"decoder_layer_{i:d}_attention"] = lit_types.AttentionHeads(
align_in="target_tokens", align_out="target_tokens")
return spec
def output_spec(self) -> Spec:
ret = {"tokens": lit_types.Tokens()}
ret["tokens_" + self.config.text_a_name] = lit_types.Tokens(
parent=self.config.text_a_name)
if self.config.text_b_name:
ret["tokens_" + self.config.text_b_name] = lit_types.Tokens(
parent=self.config.text_b_name)
if self.is_regression:
ret["score"] = lit_types.RegressionScore(parent=self.config.label_name)
else:
ret["probas"] = lit_types.MulticlassPreds(
parent=self.config.label_name,
vocab=self.config.labels,
null_idx=self.config.null_label_idx)
ret["cls_emb"] = lit_types.Embeddings()
# Average embeddings, one per layer including embeddings.
for i in range(1 + self.model.config.num_hidden_layers):
ret[f"layer_{i}/avg_emb"] = lit_types.Embeddings()
ret["cls_grad"] = lit_types.Gradients(
grad_for="cls_emb", grad_target_field_key="grad_class")
# The input_embs_ and grad_class fields are used for Integrated Gradients.
ret["input_embs_" + self.config.text_a_name] = lit_types.TokenEmbeddings(
align="tokens_" + self.config.text_a_name)
if self.config.text_b_name:
ret["input_embs_" + self.config.text_b_name] = lit_types.TokenEmbeddings(
align="tokens_" + self.config.text_b_name)
# Gradients, if requested.
if self.config.compute_grads:
ret["grad_class"] = lit_types.CategoryLabel(required=False,
vocab=self.config.labels)
ret["token_grad_" + self.config.text_a_name] = lit_types.TokenGradients(
align="tokens_" + self.config.text_a_name,
grad_for="input_embs_" + self.config.text_a_name,
grad_target_field_key="grad_class")
if self.config.text_b_name:
ret["token_grad_" + self.config.text_b_name] = lit_types.TokenGradients(
align="tokens_" + self.config.text_b_name,
grad_for="input_embs_" + self.config.text_b_name,
grad_target_field_key="grad_class")
# Attention heads, one field for each layer.
for i in range(self.model.config.num_hidden_layers):
ret[f"layer_{i+1}/attention"] = lit_types.AttentionHeads(
align_in="tokens", align_out="tokens")
return ret
def spec(self):
return {
"text":
lit_types.TextSegment(),
"tokens":
lit_types.Tokens(parent="text"),
"coref":
lit_types.EdgeLabels(align="tokens"),
# Metadata fields for filtering and analysis.
"occupation":
lit_types.CategoryLabel(),
"participant":
lit_types.CategoryLabel(),
"answer":
lit_types.CategoryLabel(vocab=ANSWER_VOCAB),
"someone":
lit_types.CategoryLabel(vocab=["True", "False"]),
"pronouns":
lit_types.CategoryLabel(vocab=list(PRONOUNS_BY_GENDER.values())),
"pronoun_type":
lit_types.CategoryLabel(vocab=["NOM", "POSS", "ACC"]),
"gender":
lit_types.CategoryLabel(vocab=[g.name for g in Gender]),
"pf_bls":
lit_types.Scalar(),
}
def test_find_spec_keys(self):
spec = {
"score": types.RegressionScore(),
"scalar_foo": types.Scalar(),
"text": types.TextSegment(),
"emb_0": types.Embeddings(),
"emb_1": types.Embeddings(),
"tokens": types.Tokens(),
"generated_text": types.GeneratedText(),
}
self.assertEqual(["score"], utils.find_spec_keys(spec,
types.RegressionScore))
self.assertEqual(["text", "tokens", "generated_text"],
utils.find_spec_keys(spec,
(types.TextSegment, types.Tokens)))
self.assertEqual(["emb_0", "emb_1"],
utils.find_spec_keys(spec, types.Embeddings))
self.assertEqual([], utils.find_spec_keys(spec, types.AttentionHeads))
# Check subclasses
self.assertEqual(
list(spec.keys()), utils.find_spec_keys(spec, types.LitType))
self.assertEqual(["text", "generated_text"],
utils.find_spec_keys(spec, types.TextSegment))
self.assertEqual(["score", "scalar_foo"],
utils.find_spec_keys(spec, types.Scalar))
def output_spec(self) -> lit_types.Spec:
return {
"tokens": lit_types.Tokens(),
"probas": lit_types.MulticlassPreds(parent="label",
vocab=self.LABELS),
"cls_emb": lit_types.Embeddings()
}
def __init__(self, model, tasks):
"""Initialize with Stanza model and a dictionary of tasks.
Args:
model: A Stanza model
tasks: A dictionary of tasks, grouped by task type.
Keys are the grouping, which should be one of:
('sequence', 'span', 'edge').
Values are a list of stanza task names as strings.
"""
self.model = model
# Store lists of task name strings by grouping
self.sequence_tasks = tasks["sequence"]
self.span_tasks = tasks["span"]
self.edge_tasks = tasks["edge"]
self._input_spec = {
"sentence": lit_types.TextSegment(),
}
self._output_spec = {
"tokens": lit_types.Tokens(),
}
# Output spec based on specified tasks
for task in self.sequence_tasks:
self._output_spec[task] = lit_types.SequenceTags(align="tokens")
for task in self.span_tasks:
self._output_spec[task] = lit_types.SpanLabels(align="tokens")
for task in self.edge_tasks:
self._output_spec[task] = lit_types.EdgeLabels(align="tokens")
def output_spec(self) -> lit_types.Spec:
return {
"tokens": lit_types.Tokens(),
"probas": lit_types.MulticlassPreds(parent="label", vocab=self._labels),
"cls_emb": lit_types.Embeddings(),
"token_grad_sentence": lit_types.TokenGradients(align="tokens")
}
def output_spec(self) -> lit_types.Spec:
return {
"tokens": lit_types.Tokens(),
"logits": lit_types.RegressionScore(),
"cls_emb": lit_types.Embeddings(),
"token_grad_sentence": lit_types.TokenGradients(align="tokens")
}
def output_spec(self):
return {
'top_layer_embs':
lit_types.TokenEmbeddings(),
'wpm_tokens':
lit_types.Tokens(),
'offsets':
lit_types.SubwordOffsets(align_in='tokens', align_out='wpm_tokens')
}
def output_spec(self):
spec = {
"input_tokens": lit_types.Tokens(parent="input_text"),
"generation": lit_types.GeneratedText(parent="target_text"),
"encoder_final_embedding": lit_types.Embeddings(),
# If target text is given, the following will also be populated.
"target_tokens": lit_types.Tokens(parent="target_text"),
"pred_tokens": lit_types.TokenTopKPreds(align="target_tokens"),
"rougeL": lit_types.Scalar(),
}
if self.config.output_attention:
# Add attention for each layer.
for i in range(self.num_layers):
spec[f"encoder_layer_{i:d}_attention"] = lit_types.AttentionHeads(
align=("input_tokens", "input_tokens"))
spec[f"decoder_layer_{i:d}_attention"] = lit_types.AttentionHeads(
align=("target_tokens", "target_tokens"))
return spec
def output_spec(self) -> lit_types.Spec:
return {
"src_tokens":
lit_types.Tokens(parent="src_text"),
"trg_text":
lit_types.GeneratedText(parent="ref_text"),
"trg_tokens":
lit_types.Tokens(parent="trg_text"),
"attention":
lit_types.AttentionHeads(align_in="src_tokens",
align_out="trg_tokens"),
"pred_tokens":
lit_types.TokenTopKPreds(align="trg_tokens", parent="trg_text"),
"encoder_final_embedding":
lit_types.Embeddings(),
"ter":
lit_types.Scalar(),
"chrf3":
lit_types.Scalar(),
}
def spec(self) -> lit_types.Spec:
"""Should match MLM's input_spec()."""
return {
'input_text':
lit_types.TextSegment(),
'target_text':
lit_types.TextSegment(),
'input_tokens':
lit_types.Tokens(required=False),
'gece_tags':
lit_types.SequenceTags(align='input_tokens', required=False)
}
def output_spec(self):
spec = {
# the "parent" keyword tells LIT which field in the input spec we should
# compare this to when computing metrics.
"pred_tokens": lit_types.TokenTopKPreds(align="tokens"),
"tokens": lit_types.Tokens(parent="text"), # all tokens
}
# Add attention and embeddings from each layer.
for i in range(self.num_layers):
spec[f"layer_{i:d}_attention"] = lit_types.AttentionHeads(
align_in="tokens", align_out="tokens")
spec[f"layer_{i:d}_avg_embedding"] = lit_types.Embeddings()
return spec
def output_spec(self):
return {'probas': lit_types.MulticlassPreds(
parent='label',
vocab=['0', '1'],
null_idx=0),
'input_embs': lit_types.TokenEmbeddings(align='tokens'),
'input_embs_grad': lit_types.TokenGradients(align='tokens',
grad_for='input_embs',
grad_target='grad_class'
),
'tokens': lit_types.Tokens(),
'grad_class': lit_types.CategoryLabel(vocab=['0', '1'])
}
def output_spec(self):
# TODO(lit-dev): also return the embeddings for each span on datasets
# with a fixed number of targets; for Winogender this would be
# {occupation, other participant, pronoun}
return {
'tokens':
lit_types.Tokens(parent='text'),
'coref':
lit_types.EdgeLabels(align='tokens'),
'pred_answer':
lit_types.MulticlassPreds(vocab=winogender.ANSWER_VOCAB,
parent='answer'),
}
def input_spec(self) -> Spec:
ret = {}
ret[self.config.text_a_name] = lit_types.TextSegment()
ret["tokens_" + self.config.text_a_name] = lit_types.Tokens(
parent=self.config.text_a_name, required=False)
if self.config.text_b_name:
ret[self.config.text_b_name] = lit_types.TextSegment()
ret["tokens_" + self.config.text_b_name] = lit_types.Tokens(
parent=self.config.text_b_name, required=False)
if self.is_regression:
ret[self.config.label_name] = lit_types.RegressionScore(required=False)
else:
ret[self.config.label_name] = lit_types.CategoryLabel(
required=False, vocab=self.config.labels)
# The input_embs_ and grad_class fields are used for Integrated Gradients.
ret["input_embs_" + self.config.text_a_name] = lit_types.TokenEmbeddings(
align="tokens", required=False)
if self.config.text_b_name:
ret["input_embs_" + self.config.text_b_name] = lit_types.TokenEmbeddings(
align="tokens", required=False)
ret["grad_class"] = lit_types.CategoryLabel(required=False,
vocab=self.config.labels)
return ret
def output_spec(self) -> lit_types.Spec:
output = {
"input_tokens":
lit_types.Tokens(parent="input_text"),
"predicted":
lit_types.GeneratedText(parent='target_text'),
'layer_average':
lit_types.AttentionHeads(align=('input_tokens', 'input_tokens'))
}
for layer in range(self.ATTENTION_LAYERS):
output['layer{}'.format(layer)] = lit_types.AttentionHeads(
align=('input_tokens', 'input_tokens'))
return output
def test_compatibility_optionals(self):
"""Test with optionals in the model spec."""
mspec = model.ModelSpec(input={
"text":
types.TextSegment(),
"tokens":
types.Tokens(parent="text", required=False),
"label":
types.CategoryLabel(vocab=["0", "1"], required=False),
},
output={})
dspec = {
"text": types.TextSegment(),
"label": types.CategoryLabel(vocab=["0", "1"]),
}
self.assertTrue(mspec.is_compatible_with_dataset(dspec))
def output_spec(self) -> lit_types.Spec:
"""Give the output specifications."""
ret = {
"tokens": lit_types.Tokens(),
"probas": lit_types.MulticlassPreds(parent="label", vocab=self.LABELS),
"cls_emb": lit_types.Embeddings()
}
# Gradients, if requested.
if self.compute_grads:
ret["token_grad_sentence"] = lit_types.TokenGradients(align="tokens")
# Attention heads, one field for each layer.
for i in range(self.model.config.num_hidden_layers):
ret[f"layer_{i}/attention"] = lit_types.AttentionHeads(align=("tokens", "tokens"))
return ret
def test_compatibility_optionals_mismatch(self):
"""Test with optionals that don't match metadata."""
mspec = model.ModelSpec(input={
"text":
types.TextSegment(),
"tokens":
types.Tokens(parent="text", required=False),
"label":
types.CategoryLabel(vocab=["0", "1"], required=False),
},
output={})
dspec = {
"text": types.TextSegment(),
# This label field doesn't match the one the model expects.
"label": types.CategoryLabel(vocab=["foo", "bar"]),
}
self.assertFalse(mspec.is_compatible_with_dataset(dspec))
def output_spec(self):
# TODO(lit-dev): also return the embeddings for each span on datasets
# with a fixed number of targets; for Winogender this would be
# {occupation, other participant, pronoun}
return {
'tokens':
lit_types.Tokens(parent='text'),
'coref':
lit_types.EdgeLabels(align='tokens'),
'pred_answer':
lit_types.MulticlassPreds(vocab=winogender.ANSWER_VOCAB,
parent='answer'),
# TODO(b/172975096): allow plotting of scalars from input data,
# so we don't need to add this to the predictions.
'pf_bls':
lit_types.Scalar(),
}
def input_spec(self):
return {
'text':
lit_types.TextSegment(),
'tokens':
lit_types.Tokens(parent='text'),
'coref':
lit_types.EdgeLabels(align='tokens'),
# Index of predicted (single) edge for Winogender
'answer':
lit_types.CategoryLabel(vocab=winogender.ANSWER_VOCAB,
required=False),
# TODO(b/172975096): allow plotting of scalars from input data,
# so we don't need to add this to the predictions.
'pf_bls':
lit_types.Scalar(required=False),
}
def output_spec(self) -> lit_types.Spec:
spec = {
"tokens":
lit_types.Tokens(),
"bio_tags":
lit_types.SequenceTags(align="tokens"),
"token_ids":
lit_types.SequenceTags(align="tokens"),
"grads":
lit_types.TokenGradients(align="tokens"),
"probas":
lit_types.MulticlassPreds(parent="bio_tags", vocab=self.LABELS)
}
for i in range(self.model.config.num_hidden_layers):
spec[f'layer_{i}/attention'] = lit_types.AttentionHeads(
align=("tokens", "tokens"))
return spec
def config_spec(self) -> types.Spec:
return {
NUM_EXAMPLES_KEY:
types.TextSegment(default=str(NUM_EXAMPLES_DEFAULT)),
MAX_FLIPS_KEY:
types.TextSegment(default=str(MAX_FLIPS_DEFAULT)),
TOKENS_TO_IGNORE_KEY:
types.Tokens(default=TOKENS_TO_IGNORE_DEFAULT),
PREDICTION_KEY:
types.FieldMatcher(spec="output",
types=["MulticlassPreds", "RegressionScore"]),
REGRESSION_THRESH_KEY:
types.TextSegment(default=str(REGRESSION_THRESH_DEFAULT)),
FIELDS_TO_HOTFLIP_KEY:
types.MultiFieldMatcher(spec="input",
types=["Tokens"],
select_all=True),
}
def input_spec(self):
return {
"text": lit_types.TextSegment(),
"tokens": lit_types.Tokens(required=False),
}
def input_spec(self):
return {'tokens': lit_types.Tokens()}
def input_spec(self):
return {
"text": lit_types.TextSegment(),
"tokens": lit_types.Tokens(mask_token="[MASK]", required=False),
}
def output_spec(self):
return {
"tokens": lit_types.Tokens(parent="text"),
"pred_tokens": lit_types.TokenTopKPreds(align="tokens"),
"cls_emb": lit_types.Embeddings(),
}
def spec(self):
return {
'text': lit_types.TextSegment(),
'tokens': lit_types.Tokens(parent='text'),
'coref': lit_types.EdgeLabels(align='tokens'),
}
