fix SqueezeBertForMaskedLM (huggingface#8479)

model_cards/squeezebert/squeezebert-uncased/README.md

@@ -22,7 +22,7 @@ The authors found that SqueezeBERT is 4.3x faster than `bert-base-uncased` on a
 The model is pretrained using the Masked Language Model (MLM) and Sentence Order Prediction (SOP) tasks.
 (Author's note: If you decide to pretrain your own model, and you prefer to train with MLM only, that should work too.)
 
-The SqueezeBERT paper presents 2 approaches to finetuning the model:
+From the SqueezeBERT paper:
 > We pretrain SqueezeBERT from scratch (without distillation) using the [LAMB](https://arxiv.org/abs/1904.00962) optimizer, and we employ the hyperparameters recommended by the LAMB authors: a global batch size of 8192, a learning rate of 2.5e-3, and a warmup proportion of 0.28. Following the LAMB paper's recommendations, we pretrain for 56k steps with a maximum sequence length of 128 and then for 6k steps with a maximum sequence length of 512.
 
 ## Finetuning

src/transformers/modeling_squeezebert.py

@@ -373,6 +373,53 @@ def forward(self, hidden_states):
         return pooled_output
 
 
+class SqueezeBertPredictionHeadTransform(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.dense = nn.Linear(config.hidden_size, config.hidden_size)
+        if isinstance(config.hidden_act, str):
+            self.transform_act_fn = ACT2FN[config.hidden_act]
+        else:
+            self.transform_act_fn = config.hidden_act
+        self.LayerNorm = nn.LayerNorm(config.hidden_size, eps=config.layer_norm_eps)
+
+    def forward(self, hidden_states):
+        hidden_states = self.dense(hidden_states)
+        hidden_states = self.transform_act_fn(hidden_states)
+        hidden_states = self.LayerNorm(hidden_states)
+        return hidden_states
+
+
+class SqueezeBertLMPredictionHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.transform = SqueezeBertPredictionHeadTransform(config)
+
+        # The output weights are the same as the input embeddings, but there is
+        # an output-only bias for each token.
+        self.decoder = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        self.bias = nn.Parameter(torch.zeros(config.vocab_size))
+
+        # Need a link between the two variables so that the bias is correctly resized with `resize_token_embeddings`
+        self.decoder.bias = self.bias
+
+    def forward(self, hidden_states):
+        hidden_states = self.transform(hidden_states)
+        hidden_states = self.decoder(hidden_states)
+        return hidden_states
+
+
+class SqueezeBertOnlyMLMHead(nn.Module):
+    def __init__(self, config):
+        super().__init__()
+        self.predictions = SqueezeBertLMPredictionHead(config)
+
+    def forward(self, sequence_output):
+        prediction_scores = self.predictions(sequence_output)
+        return prediction_scores
+
+
 class SqueezeBertPreTrainedModel(PreTrainedModel):
     """
     An abstract class to handle weights initialization and a simple interface for downloading and loading pretrained
@@ -594,16 +641,19 @@ def forward(
 
 @add_start_docstrings("""SqueezeBERT Model with a `language modeling` head on top. """, SQUEEZEBERT_START_DOCSTRING)
 class SqueezeBertForMaskedLM(SqueezeBertPreTrainedModel):
+
+    authorized_missing_keys = [r"predictions.decoder.bias"]
+
     def __init__(self, config):
         super().__init__(config)
 
         self.transformer = SqueezeBertModel(config)
-        self.lm_head = nn.Linear(config.embedding_size, config.vocab_size)
+        self.cls = SqueezeBertOnlyMLMHead(config)
 
         self.init_weights()
 
     def get_output_embeddings(self):
-        return self.lm_head
+        return self.cls.predictions.decoder
 
     @add_start_docstrings_to_model_forward(SQUEEZEBERT_INPUTS_DOCSTRING.format("(batch_size, sequence_length)"))
     @add_code_sample_docstrings(
@@ -646,7 +696,7 @@ def forward(
         )
 
         sequence_output = outputs[0]
-        prediction_scores = self.lm_head(sequence_output)
+        prediction_scores = self.cls(sequence_output)
 
         masked_lm_loss = None
         if labels is not None: