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 spec(self):
return {
'span1_embs': lit_types.Embeddings(),
'span2_embs': lit_types.Embeddings(),
'label': lit_types.Scalar(),
'src_idx': lit_types.Scalar(),
'edge_idx': lit_types.Scalar(),
}
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 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) -> lit_types.Spec:
return {
"tokens": lit_types.Tokens(),
"probas": lit_types.MulticlassPreds(parent="label",
vocab=self.LABELS),
"cls_emb": lit_types.Embeddings()
}
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) -> 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 {
'probas':
lit_types.MulticlassPreds(
parent='label', vocab=['0', '1'], null_idx=0),
'cls_emb':
lit_types.Embeddings(),
'cls_grad':
lit_types.Gradients(grad_for='cls_emb', grad_target='grad_class'),
'grad_class':
lit_types.CategoryLabel()
}
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) -> 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 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 meta_spec(self) -> types.Spec:
return {
CLUSTER_ID_KEY: types.CategoryLabel(),
REPRESENTATION_KEY: types.Embeddings(),
}
def input_spec(self):
return {
'span1_embs': lit_types.Embeddings(),
'span2_embs': lit_types.Embeddings(),
'label': lit_types.Scalar(required=False), # in range [0,1]
}
def input_spec(self):
# 'x' denotes input features
return {"x": lit_types.Embeddings()}
def output_spec(self):
# 'z' denotes projected embeddings
return {"z": lit_types.Embeddings()}
def output_spec(self):
return {
"tokens": lit_types.Tokens(parent="text"),
"pred_tokens": lit_types.TokenTopKPreds(align="tokens"),
"cls_emb": lit_types.Embeddings(),
}
