update to transformers v4.2.1

jiant/proj/main/modeling/heads.py

@@ -108,8 +108,10 @@ class BertMLMHead(BaseMLMHead):
     def __init__(self, hidden_size, vocab_size, layer_norm_eps=1e-12, hidden_act="gelu"):
         super().__init__()
         self.dense = nn.Linear(hidden_size, hidden_size)
-        self.transform_act_fn = transformers.modeling_bert.ACT2FN[hidden_act]
-        self.LayerNorm = transformers.modeling_bert.BertLayerNorm(hidden_size, eps=layer_norm_eps)
+        self.transform_act_fn = transformers.models.bert.modeling_bert.ACT2FN[hidden_act]
+        self.LayerNorm = transformers.models.bert.modeling_bert.BertLayerNorm(
+            hidden_size, eps=layer_norm_eps
+        )
 
         self.decoder = nn.Linear(hidden_size, vocab_size, bias=False)
         self.bias = nn.Parameter(torch.zeros(vocab_size), requires_grad=True)
@@ -132,7 +134,9 @@ class RobertaMLMHead(BaseMLMHead):
     def __init__(self, hidden_size, vocab_size, layer_norm_eps=1e-12):
         super().__init__()
         self.dense = nn.Linear(hidden_size, hidden_size)
-        self.layer_norm = transformers.modeling_bert.BertLayerNorm(hidden_size, eps=layer_norm_eps)
+        self.layer_norm = transformers.models.bert.modeling_bert.BertLayerNorm(
+            hidden_size, eps=layer_norm_eps
+        )
 
         self.decoder = nn.Linear(hidden_size, vocab_size, bias=False)
         self.bias = nn.Parameter(torch.zeros(vocab_size), requires_grad=True)
@@ -143,7 +147,7 @@ def __init__(self, hidden_size, vocab_size, layer_norm_eps=1e-12):
 
     def forward(self, unpooled):
         x = self.dense(unpooled)
-        x = transformers.modeling_bert.gelu(x)
+        x = transformers.models.bert.modeling_bert.gelu(x)
         x = self.layer_norm(x)
 
         # project back to size of vocabulary with bias
@@ -161,7 +165,7 @@ def __init__(self, hidden_size, embedding_size, vocab_size, hidden_act="gelu"):
         self.bias = nn.Parameter(torch.zeros(vocab_size), requires_grad=True)
         self.dense = nn.Linear(hidden_size, embedding_size)
         self.decoder = nn.Linear(embedding_size, vocab_size)
-        self.activation = transformers.modeling_bert.ACT2FN[hidden_act]
+        self.activation = transformers.models.bert.modeling_bert.ACT2FN[hidden_act]
 
         # Need a link between the two variables so that the bias is correctly resized with
         # `resize_token_embeddings`

jiant/proj/main/modeling/taskmodels.py

@@ -334,7 +334,12 @@ def get_output_from_encoder(encoder, input_ids, segment_ids, input_mask) -> Enco
 
 
 def get_output_from_standard_transformer_models(encoder, input_ids, segment_ids, input_mask):
-    output = encoder(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask)
+    output = encoder(
+        input_ids=input_ids,
+        token_type_ids=segment_ids,
+        attention_mask=input_mask,
+        return_dict=False,
+    )
     pooled, unpooled, other = output[1], output[0], output[2:]
     return pooled, unpooled, other
 
@@ -347,7 +352,7 @@ def get_output_from_bart_models(encoder, input_ids, input_mask):
     # sentence representation is the final decoder state.
     # That's what we use for `unpooled` here.
     dec_last, dec_all, enc_last, enc_all = encoder(
-        input_ids=input_ids, attention_mask=input_mask, output_hidden_states=True,
+        input_ids=input_ids, attention_mask=input_mask, output_hidden_states=True, return_dict=False
     )
     unpooled = dec_last
 
@@ -361,7 +366,12 @@ def get_output_from_bart_models(encoder, input_ids, input_mask):
 
 
 def get_output_from_electra(encoder, input_ids, segment_ids, input_mask):
-    output = encoder(input_ids=input_ids, token_type_ids=segment_ids, attention_mask=input_mask)
+    output = encoder(
+        input_ids=input_ids,
+        token_type_ids=segment_ids,
+        attention_mask=input_mask,
+        return_dict=False,
+    )
     unpooled = output[0]
     pooled = unpooled[:, 0, :]
     return pooled, unpooled, output

jiant/tasks/lib/templates/squad_style/core.py

@@ -5,7 +5,7 @@
 from dataclasses import dataclass
 from typing import Union, List, Dict, Optional
 
-from transformers.tokenization_bert import whitespace_tokenize
+from transformers.models.bert.tokenization_bert import whitespace_tokenize
 
 from jiant.tasks.lib.templates.squad_style import utils as squad_utils
 from jiant.shared.constants import PHASE

jiant/tasks/lib/templates/squad_style/utils.py

@@ -6,7 +6,7 @@
 from dataclasses import dataclass
 from typing import List, Dict
 
-from transformers.tokenization_bert import BasicTokenizer
+from transformers.models.bert.tokenization_bert import BasicTokenizer
 from jiant.utils.display import maybe_tqdm
 
 

requirements-no-torch.txt

@@ -13,6 +13,6 @@ seqeval==0.0.12
 scikit-learn==0.22.2.post1
 scipy==1.4.1
 sentencepiece==0.1.86
-tokenizers==0.8.1.rc2
+tokenizers==0.9.4
 tqdm==4.46.0
-transformers==3.1.0
+transformers==4.2.1

setup.py

@@ -72,10 +72,10 @@
         "scikit-learn == 0.22.2.post1",
         "scipy == 1.4.1",
         "sentencepiece == 0.1.86",
-        "tokenizers == 0.8.1.rc2",
+        "tokenizers == 0.9.4",
         "torch >= 1.5.0",
         "tqdm == 4.46.0",
-        "transformers == 3.1.0",
+        "transformers == 4.2.1",
         "torchvision == 0.6.0",
     ],
     extras_require=extras,