allenai · dirkgr · Feb 10, 2022 · Jan 31, 2022 · Feb 4, 2022 · Feb 4, 2022
diff --git a/CHANGELOG.md b/CHANGELOG.md
@@ -7,6 +7,10 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ## Unreleased
 
+### Fixed
+
+- Updated dependencies, especially around doc creation.
+
 ## [v2.9.0](https://github.com/allenai/allennlp/releases/tag/v2.9.0) - 2022-01-27
 
 ### Added

diff --git a/allennlp/commands/find_learning_rate.py b/allennlp/commands/find_learning_rate.py
@@ -281,7 +281,7 @@ def search_learning_rate(
         if linear_steps:
             current_lr = start_lr + (lr_update_factor * i)
         else:
-            current_lr = start_lr * (lr_update_factor ** i)
+            current_lr = start_lr * (lr_update_factor**i)
 
         for param_group in trainer.optimizer.param_groups:
             param_group["lr"] = current_lr

diff --git a/allennlp/fairness/bias_metrics.py b/allennlp/fairness/bias_metrics.py
@@ -240,7 +240,7 @@ def spearman_correlation(self, x: torch.Tensor, y: torch.Tensor):
 
         n = x.size(0)
         upper = 6 * torch.sum((x_rank - y_rank).pow(2))
-        down = n * (n ** 2 - 1.0)
+        down = n * (n**2 - 1.0)
         return 1.0 - (upper / down)
 
 

diff --git a/allennlp/modules/transformer/attention_module.py b/allennlp/modules/transformer/attention_module.py
@@ -134,16 +134,16 @@ def _normalize(self) -> None:
         self.query.weight.data.normal_(
             mean=0.0, std=(self.hidden_size * self.attention_head_size) ** -0.5
         )
-        self.key.weight.data.normal_(mean=0.0, std=self.hidden_size ** -0.5)
-        self.value.weight.data.normal_(mean=0.0, std=self.hidden_size ** -0.5)
+        self.key.weight.data.normal_(mean=0.0, std=self.hidden_size**-0.5)
+        self.value.weight.data.normal_(mean=0.0, std=self.hidden_size**-0.5)
 
         if hasattr(self, "output"):
             self.output.weight.data.normal_(
                 mean=0.0, std=(self.num_attention_heads * self.attention_head_size) ** -0.5
             )
 
         if hasattr(self, "relative_attention_bias"):
-            self.relative_attention_bias.weight.data.normal_(mean=0.0, std=self.hidden_size ** -0.5)
+            self.relative_attention_bias.weight.data.normal_(mean=0.0, std=self.hidden_size**-0.5)
 
     def _transpose_for_scores(self, x: torch.Tensor) -> torch.Tensor:
         new_x_shape = x.size()[:-1] + (

diff --git a/allennlp/modules/transformer/t5.py b/allennlp/modules/transformer/t5.py
@@ -63,9 +63,9 @@ class T5DenseReluDense(TransformerModule, FromParams):
     def __init__(self, hidden_size: int = 512, ff_size: int = 2048, dropout: float = 0.1):
         super().__init__()
         self.wi = nn.Linear(hidden_size, ff_size, bias=False)
-        self.wi.weight.data.normal_(mean=0.0, std=hidden_size ** -0.5)
+        self.wi.weight.data.normal_(mean=0.0, std=hidden_size**-0.5)
         self.wo = nn.Linear(ff_size, hidden_size, bias=False)
-        self.wo.weight.data.normal_(mean=0.0, std=ff_size ** -0.5)
+        self.wo.weight.data.normal_(mean=0.0, std=ff_size**-0.5)
         self.dropout = nn.Dropout(dropout)
 
     def forward(self, hidden_states) -> FloatT:
@@ -81,11 +81,11 @@ class T5DenseGatedGeluDense(TransformerModule, FromParams):
     def __init__(self, hidden_size: int = 512, ff_size: int = 2048, dropout: float = 0.1):
         super().__init__()
         self.wi_0 = nn.Linear(hidden_size, ff_size, bias=False)
-        self.wi_0.weight.data.normal_(mean=0.0, std=hidden_size ** -0.5)
+        self.wi_0.weight.data.normal_(mean=0.0, std=hidden_size**-0.5)
         self.wi_1 = nn.Linear(hidden_size, ff_size, bias=False)
-        self.wi_1.weight.data.normal_(mean=0.0, std=hidden_size ** -0.5)
+        self.wi_1.weight.data.normal_(mean=0.0, std=hidden_size**-0.5)
         self.wo = nn.Linear(ff_size, hidden_size, bias=False)
-        self.wo.weight.data.normal_(mean=0.0, std=ff_size ** -0.5)
+        self.wo.weight.data.normal_(mean=0.0, std=ff_size**-0.5)
         self.dropout = nn.Dropout(dropout)
         from allennlp.nn import Activation
 
@@ -964,7 +964,7 @@ def _get_lm_logits(self, decoder_last_hidden_state: FloatT) -> FloatT:
         # TODO: HF only does this when does this when embeddings are tied.
         # Currently tied embeddings is the only option we have, but if make
         # that configurable then we should put this in an 'if' block.
-        sequence_output = sequence_output * (self.model_dim ** -0.5)
+        sequence_output = sequence_output * (self.model_dim**-0.5)
         # Shape: (batch_size, target_length, vocab_size)
         logits = self.lm_head(sequence_output)
         return logits

diff --git a/allennlp/nn/beam_search.py b/allennlp/nn/beam_search.py
@@ -515,7 +515,7 @@ def score(
         lengths += is_end_token.long()
 
         # shape: (batch_size, beam_size)
-        average_log_probs = log_probabilities / (lengths ** self.length_penalty)
+        average_log_probs = log_probabilities / (lengths**self.length_penalty)
         return average_log_probs
 
 

diff --git a/allennlp/training/learning_rate_schedulers/cosine.py b/allennlp/training/learning_rate_schedulers/cosine.py
@@ -109,8 +109,8 @@ def get_values(self):
             self._cycle_counter = 0
             self._last_restart = step
 
-        base_lrs = [lr * self.eta_mul ** self._n_restarts for lr in self.base_values]
-        self._cycle_len = int(self.t_initial * self.t_mul ** self._n_restarts)
+        base_lrs = [lr * self.eta_mul**self._n_restarts for lr in self.base_values]
+        self._cycle_len = int(self.t_initial * self.t_mul**self._n_restarts)
 
         lrs = [
             self.eta_min

diff --git a/allennlp/training/learning_rate_schedulers/slanted_triangular.py b/allennlp/training/learning_rate_schedulers/slanted_triangular.py
@@ -86,7 +86,7 @@ def __init__(
             for i in range(len(self.base_values) - 1, -1, -1):
                 param_group = optimizer.param_groups[i]
                 if param_group["params"]:
-                    param_group["lr"] = self.base_values[i] * decay_factor ** exponent
+                    param_group["lr"] = self.base_values[i] * decay_factor**exponent
                     self.base_values[i] = param_group["lr"]
                     exponent += 1
         # set up for the first batch

diff --git a/allennlp/training/metrics/fbeta_measure.py b/allennlp/training/metrics/fbeta_measure.py
@@ -196,7 +196,7 @@ def get_metric(self, reset: bool = False):
             pred_sum = pred_sum.sum()  # type: ignore
             true_sum = true_sum.sum()  # type: ignore
 
-        beta2 = self._beta ** 2
+        beta2 = self._beta**2
         # Finally, we have all our sufficient statistics.
         precision = nan_safe_tensor_divide(tp_sum, pred_sum)
         recall = nan_safe_tensor_divide(tp_sum, true_sum)

diff --git a/dev-requirements.txt b/dev-requirements.txt
@@ -7,7 +7,7 @@ flake8
 mypy==0.931
 
 # Automatic code formatting
-black==21.12b0
+black==22.1.0
 
 # Allows generation of coverage reports with pytest.
 pytest-cov
@@ -34,11 +34,10 @@ pytest-benchmark
 ruamel.yaml
 
 # Generating markdown files from Python modules.
-git+https://github.com/NiklasRosenstein/pydoc-markdown.git@f0bf8af1db4f11581c19d206d4ed1ab34b4854c1
-nr.databind.core<0.0.17
-nr.interface<0.0.6
+pydoc-markdown>=4.0.0,<5.0.0
 databind.core
 
-mkdocs==1.1.2
+mkdocs==1.2.3
 mkdocs-material>=5.5.0,<8.2.0
 markdown-include==0.6.0
 

diff --git a/scripts/ai2_internal/run_with_beaker.py b/scripts/ai2_internal/run_with_beaker.py
@@ -12,7 +12,7 @@
 
 # This has to happen before we import spacy (even indirectly), because for some crazy reason spacy
 # thought it was a good idea to set the random seed on import...
-random_int = random.randint(0, 2 ** 32)
+random_int = random.randint(0, 2**32)
 
 sys.path.insert(
     0, os.path.dirname(os.path.abspath(os.path.join(os.path.join(__file__, os.pardir), os.pardir)))