Merge pull request #1077 from huggingface/pruning-save-and-load

Pruning changes so that deleted heads are kept on save/load

pytorch_transformers/modeling_bert.py

@@ -337,23 +337,30 @@ class BertAttention(nn.Module):
         super(BertAttention, self).__init__()
         self.self = BertSelfAttention(config)
         self.output = BertSelfOutput(config)
+        self.pruned_heads = set()
 
     def prune_heads(self, heads):
         if len(heads) == 0:
             return
         mask = torch.ones(self.self.num_attention_heads, self.self.attention_head_size)
+        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
         for head in heads:
+            # Compute how many pruned heads are before the head and move the index accordingly
+            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
             mask[head] = 0
         mask = mask.view(-1).contiguous().eq(1)
         index = torch.arange(len(mask))[mask].long()
+
         # Prune linear layers
         self.self.query = prune_linear_layer(self.self.query, index)
         self.self.key = prune_linear_layer(self.self.key, index)
         self.self.value = prune_linear_layer(self.self.value, index)
         self.output.dense = prune_linear_layer(self.output.dense, index, dim=1)
-        # Update hyper params
+
+        # Update hyper params and store pruned heads
         self.self.num_attention_heads = self.self.num_attention_heads - len(heads)
         self.self.all_head_size = self.self.attention_head_size * self.self.num_attention_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
 
     def forward(self, input_tensor, attention_mask, head_mask=None):
         self_outputs = self.self(input_tensor, attention_mask, head_mask)
@@ -531,12 +538,8 @@ class BertPreTrainedModel(PreTrainedModel):
     load_tf_weights = load_tf_weights_in_bert
     base_model_prefix = "bert"
 
-    def __init__(self, *inputs, **kwargs):
-        super(BertPreTrainedModel, self).__init__(*inputs, **kwargs)
-
-    def init_weights(self, module):
-        """ Initialize the weights.
-        """
+    def _init_weights(self, module):
+        """ Initialize the weights """
         if isinstance(module, (nn.Linear, nn.Embedding)):
             # Slightly different from the TF version which uses truncated_normal for initialization
             # cf https://github.com/pytorch/pytorch/pull/5617
@@ -649,7 +652,7 @@ class BertModel(BertPreTrainedModel):
         self.encoder = BertEncoder(config)
         self.pooler = BertPooler(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def _resize_token_embeddings(self, new_num_tokens):
         old_embeddings = self.embeddings.word_embeddings
@@ -758,7 +761,7 @@ class BertForPreTraining(BertPreTrainedModel):
         self.bert = BertModel(config)
         self.cls = BertPreTrainingHeads(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):
@@ -826,7 +829,7 @@ class BertForMaskedLM(BertPreTrainedModel):
         self.bert = BertModel(config)
         self.cls = BertOnlyMLMHead(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):
@@ -891,7 +894,7 @@ class BertForNextSentencePrediction(BertPreTrainedModel):
         self.bert = BertModel(config)
         self.cls = BertOnlyNSPHead(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, next_sentence_label=None,
                 position_ids=None, head_mask=None):
@@ -952,7 +955,7 @@ class BertForSequenceClassification(BertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, self.config.num_labels)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None,
                 position_ids=None, head_mask=None):
@@ -1056,7 +1059,7 @@ class BertForMultipleChoice(BertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, 1)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None,
                 position_ids=None, head_mask=None):
@@ -1124,7 +1127,7 @@ class BertForTokenClassification(BertPreTrainedModel):
         self.dropout = nn.Dropout(config.hidden_dropout_prob)
         self.classifier = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, labels=None,
                 position_ids=None, head_mask=None):
@@ -1198,7 +1201,7 @@ class BertForQuestionAnswering(BertPreTrainedModel):
         self.bert = BertModel(config)
         self.qa_outputs = nn.Linear(config.hidden_size, config.num_labels)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, start_positions=None,
                 end_positions=None, position_ids=None, head_mask=None):

pytorch_transformers/modeling_distilbert.py

@@ -174,12 +174,16 @@ class MultiHeadSelfAttention(nn.Module):
         self.v_lin = nn.Linear(in_features=config.dim, out_features=config.dim)
         self.out_lin = nn.Linear(in_features=config.dim, out_features=config.dim)
 
+        self.pruned_heads = set()
+
     def prune_heads(self, heads):
         attention_head_size = self.dim // self.n_heads
         if len(heads) == 0:
             return
         mask = torch.ones(self.n_heads, attention_head_size)
+        heads = set(heads) - self.pruned_heads
         for head in heads:
+            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
             mask[head] = 0
         mask = mask.view(-1).contiguous().eq(1)
         index = torch.arange(len(mask))[mask].long()
@@ -191,6 +195,7 @@ class MultiHeadSelfAttention(nn.Module):
         # Update hyper params
         self.n_heads = self.n_heads - len(heads)
         self.dim = attention_head_size * self.n_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
 
     def forward(self, query, key, value, mask, head_mask = None):
         """
@@ -395,7 +400,7 @@ class DistilBertPreTrainedModel(PreTrainedModel):
     def __init__(self, *inputs, **kwargs):
         super(DistilBertPreTrainedModel, self).__init__(*inputs, **kwargs)
     
-    def init_weights(self, module):
+    def _init_weights(self, module):
         """ Initialize the weights.
         """
         if isinstance(module, nn.Embedding):
@@ -480,7 +485,7 @@ class DistilBertModel(DistilBertPreTrainedModel):
         self.embeddings = Embeddings(config)   # Embeddings
         self.transformer = Transformer(config) # Encoder
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def _resize_token_embeddings(self, new_num_tokens):
         old_embeddings = self.embeddings.word_embeddings
@@ -568,7 +573,7 @@ class DistilBertForMaskedLM(DistilBertPreTrainedModel):
         self.vocab_layer_norm = nn.LayerNorm(config.dim, eps=1e-12)
         self.vocab_projector = nn.Linear(config.dim, config.vocab_size)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
         self.mlm_loss_fct = nn.CrossEntropyLoss(ignore_index=-1)
@@ -642,7 +647,7 @@ class DistilBertForSequenceClassification(DistilBertPreTrainedModel):
         self.classifier = nn.Linear(config.dim, config.num_labels)
         self.dropout = nn.Dropout(config.seq_classif_dropout)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids,  attention_mask=None, labels=None, head_mask=None):
         distilbert_output = self.distilbert(input_ids=input_ids,
@@ -716,7 +721,7 @@ class DistilBertForQuestionAnswering(DistilBertPreTrainedModel):
         assert config.num_labels == 2
         self.dropout = nn.Dropout(config.qa_dropout)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         
     def forward(self, input_ids, attention_mask=None, start_positions=None, end_positions=None, head_mask=None):
         distilbert_output = self.distilbert(input_ids=input_ids,

pytorch_transformers/modeling_gpt2.py

@@ -233,22 +233,29 @@ class Attention(nn.Module):
         self.c_proj = Conv1D(n_state, nx)
         self.attn_dropout = nn.Dropout(config.attn_pdrop)
         self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.pruned_heads = set()
 
     def prune_heads(self, heads):
         if len(heads) == 0:
             return
         mask = torch.ones(self.n_head, self.split_size // self.n_head)
+        heads = set(heads) - self.pruned_heads  # Convert to set and emove already pruned heads
         for head in heads:
+            # Compute how many pruned heads are before the head and move the index accordingly
+            head = head - sum(1 if h < head else 0 for h in self.pruned_heads)
             mask[head] = 0
         mask = mask.view(-1).contiguous().eq(1)
         index = torch.arange(len(mask))[mask].long()
         index_attn = torch.cat([index, index + self.split_size, index + (2*self.split_size)])
+
         # Prune conv1d layers
         self.c_attn = prune_conv1d_layer(self.c_attn, index_attn, dim=1)
         self.c_proj = prune_conv1d_layer(self.c_proj, index, dim=0)
+
         # Update hyper params
         self.split_size = (self.split_size // self.n_head) * (self.n_head - len(heads))
         self.n_head = self.n_head - len(heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
 
     def _attn(self, q, k, v, head_mask=None):
         w = torch.matmul(q, k)
@@ -354,7 +361,7 @@ class GPT2PreTrainedModel(PreTrainedModel):
     def __init__(self, *inputs, **kwargs):
         super(GPT2PreTrainedModel, self).__init__(*inputs, **kwargs)
 
-    def init_weights(self, module):
+    def _init_weights(self, module):
         """ Initialize the weights.
         """
         if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
@@ -453,7 +460,7 @@ class GPT2Model(GPT2PreTrainedModel):
         self.h = nn.ModuleList([Block(config.n_ctx, config, scale=True) for _ in range(config.n_layer)])
         self.ln_f = LayerNorm(config.n_embd, eps=config.layer_norm_epsilon)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.wte = self._get_resized_embeddings(self.wte, new_num_tokens)
@@ -584,7 +591,7 @@ class GPT2LMHeadModel(GPT2PreTrainedModel):
         self.transformer = GPT2Model(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):
@@ -708,7 +715,7 @@ class GPT2DoubleHeadsModel(GPT2PreTrainedModel):
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):

pytorch_transformers/modeling_openai.py

@@ -249,12 +249,15 @@ class Attention(nn.Module):
         self.c_proj = Conv1D(n_state, nx)
         self.attn_dropout = nn.Dropout(config.attn_pdrop)
         self.resid_dropout = nn.Dropout(config.resid_pdrop)
+        self.pruned_heads = set()
 
     def prune_heads(self, heads):
         if len(heads) == 0:
             return
         mask = torch.ones(self.n_head, self.split_size // self.n_head)
+        heads = set(heads) - self.pruned_heads
         for head in heads:
+            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
             mask[head] = 0
         mask = mask.view(-1).contiguous().eq(1)
         index = torch.arange(len(mask))[mask].long()
@@ -265,6 +268,7 @@ class Attention(nn.Module):
         # Update hyper params
         self.split_size = (self.split_size // self.n_head) * (self.n_head - len(heads))
         self.n_head = self.n_head - len(heads)
+        self.pruned_heads = self.pruned_heads.union(heads)
 
     def _attn(self, q, k, v, head_mask=None):
         w = torch.matmul(q, k)
@@ -363,10 +367,7 @@ class OpenAIGPTPreTrainedModel(PreTrainedModel):
     load_tf_weights = load_tf_weights_in_openai_gpt
     base_model_prefix = "transformer"
 
-    def __init__(self, *inputs, **kwargs):
-        super(OpenAIGPTPreTrainedModel, self).__init__(*inputs, **kwargs)
-
-    def init_weights(self, module):
+    def _init_weights(self, module):
         """ Initialize the weights.
         """
         if isinstance(module, (nn.Linear, nn.Embedding, Conv1D)):
@@ -456,7 +457,7 @@ class OpenAIGPTModel(OpenAIGPTPreTrainedModel):
         self.drop = nn.Dropout(config.embd_pdrop)
         self.h = nn.ModuleList([Block(config.n_ctx, config, scale=True) for _ in range(config.n_layer)])
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.tokens_embed = self._get_resized_embeddings(self.tokens_embed, new_num_tokens)
@@ -569,7 +570,7 @@ class OpenAIGPTLMHeadModel(OpenAIGPTPreTrainedModel):
         self.transformer = OpenAIGPTModel(config)
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):
@@ -676,7 +677,7 @@ class OpenAIGPTDoubleHeadsModel(OpenAIGPTPreTrainedModel):
         self.lm_head = nn.Linear(config.n_embd, config.vocab_size, bias=False)
         self.multiple_choice_head = SequenceSummary(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):

pytorch_transformers/modeling_roberta.py

@@ -168,7 +168,7 @@ class RobertaModel(BertModel):
         super(RobertaModel, self).__init__(config)
 
         self.embeddings = RobertaEmbeddings(config)
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, attention_mask=None, position_ids=None, head_mask=None):
         if input_ids[:, 0].sum().item() != 0:
@@ -220,7 +220,7 @@ class RobertaForMaskedLM(BertPreTrainedModel):
         self.roberta = RobertaModel(config)
         self.lm_head = RobertaLMHead(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):

pytorch_transformers/modeling_transfo_xl.py

@@ -853,9 +853,6 @@ class TransfoXLPreTrainedModel(PreTrainedModel):
     load_tf_weights = load_tf_weights_in_transfo_xl
     base_model_prefix = "transformer"
 
-    def __init__(self, *inputs, **kwargs):
-        super(TransfoXLPreTrainedModel, self).__init__(*inputs, **kwargs)
-
     def _init_weight(self, weight):
         if self.config.init == 'uniform':
             nn.init.uniform_(weight, -self.config.init_range, self.config.init_range)
@@ -865,7 +862,7 @@ class TransfoXLPreTrainedModel(PreTrainedModel):
     def _init_bias(self, bias):
         nn.init.constant_(bias, 0.0)
 
-    def init_weights(self, m):
+    def _init_weights(self, m):
         """ Initialize the weights.
         """
         classname = m.__class__.__name__
@@ -1059,7 +1056,7 @@ class TransfoXLModel(TransfoXLPreTrainedModel):
             self.r_emb = nn.Parameter(torch.FloatTensor(
                     self.n_layer, self.max_klen, self.n_head, self.d_head))
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def _resize_token_embeddings(self, new_num_tokens):
         return self.word_emb
@@ -1306,7 +1303,7 @@ class TransfoXLLMHeadModel(TransfoXLPreTrainedModel):
         else:
             self.crit = ProjectedAdaptiveLogSoftmax(config.n_token, config.d_embed, config.d_model, 
                                                     config.cutoffs, div_val=config.div_val)
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):

pytorch_transformers/modeling_utils.py

@@ -104,6 +104,7 @@ class PretrainedConfig(object):
         self.output_attentions = kwargs.pop('output_attentions', False)
         self.output_hidden_states = kwargs.pop('output_hidden_states', False)
         self.torchscript = kwargs.pop('torchscript', False)
+        self.pruned_heads = kwargs.pop('pruned_heads', {})
 
     def save_pretrained(self, save_directory):
         """ Save a configuration object to the directory `save_directory`, so that it
@@ -200,6 +201,9 @@ class PretrainedConfig(object):
         # Load config
         config = cls.from_json_file(resolved_config_file)
 
+        if hasattr(config, 'pruned_heads'):
+            config.pruned_heads = dict((int(key), set(value)) for key, value in config.pruned_heads.items())
+
         # Update config with kwargs if needed
         to_remove = []
         for key, value in kwargs.items():
@@ -311,7 +315,7 @@ class PreTrainedModel(nn.Module):
         new_embeddings.to(old_embeddings.weight.device)
 
         # initialize all new embeddings (in particular added tokens)
-        self.init_weights(new_embeddings)
+        self._init_weights(new_embeddings)
 
         # Copy word embeddings from the previous weights
         num_tokens_to_copy = min(old_num_tokens, new_num_tokens)
@@ -355,14 +359,30 @@ class PreTrainedModel(nn.Module):
 
         return model_embeds
 
+    def init_weights(self):
+        """ Initialize and prunes weights if needed. """
+        # Initialize weights
+        self.apply(self._init_weights)
+
+        # Prune heads if needed
+        if self.config.pruned_heads:
+            self.prune_heads(self.config.pruned_heads)
+
     def prune_heads(self, heads_to_prune):
         """ Prunes heads of the base model.
 
             Arguments:
 
                 heads_to_prune: dict with keys being selected layer indices (`int`) and associated values being the list of heads to prune in said layer (list of `int`).
+                E.g. {1: [0, 2], 2: [2, 3]} will prune heads 0 and 2 on layer 1 and heads 2 and 3 on layer 2.
         """
         base_model = getattr(self, self.base_model_prefix, self)  # get the base model if needed
+
+        # save new sets of pruned heads as union of previously stored pruned heads and newly pruned heads
+        for layer, heads in heads_to_prune.items():
+            union_heads = set(self.config.pruned_heads.get(layer, [])) | set(heads)
+            self.config.pruned_heads[layer] = list(union_heads)  # Unfortunately we have to store it as list for JSON
+
         base_model._prune_heads(heads_to_prune)
 
     def save_pretrained(self, save_directory):

pytorch_transformers/modeling_xlm.py

@@ -271,13 +271,16 @@ class MultiHeadAttention(nn.Module):
         self.k_lin = nn.Linear(dim, dim)
         self.v_lin = nn.Linear(dim, dim)
         self.out_lin = nn.Linear(dim, dim)
+        self.pruned_heads = set()
 
     def prune_heads(self, heads):
         attention_head_size = self.dim // self.n_heads
         if len(heads) == 0:
             return
         mask = torch.ones(self.n_heads, attention_head_size)
+        heads = set(heads) - self.pruned_heads
         for head in heads:
+            head -= sum(1 if h < head else 0 for h in self.pruned_heads)
             mask[head] = 0
         mask = mask.view(-1).contiguous().eq(1)
         index = torch.arange(len(mask))[mask].long()
@@ -289,6 +292,7 @@ class MultiHeadAttention(nn.Module):
         # Update hyper params
         self.n_heads = self.n_heads - len(heads)
         self.dim = attention_head_size * self.n_heads
+        self.pruned_heads = self.pruned_heads.union(heads)
 
     def forward(self, input, mask, kv=None, cache=None, head_mask=None):
         """
@@ -383,7 +387,7 @@ class XLMPreTrainedModel(PreTrainedModel):
     def __init__(self, *inputs, **kwargs):
         super(XLMPreTrainedModel, self).__init__(*inputs, **kwargs)
 
-    def init_weights(self, module):
+    def _init_weights(self, module):
         """ Initialize the weights. """
         if isinstance(module, nn.Embedding):
             if self.config is not None and self.config.embed_init_std is not None:
@@ -559,7 +563,14 @@ class XLMModel(XLMPreTrainedModel):
             self.ffns.append(TransformerFFN(self.dim, self.hidden_dim, self.dim, config=config))
             self.layer_norm2.append(nn.LayerNorm(self.dim, eps=config.layer_norm_eps))
 
-        self.apply(self.init_weights)
+        if hasattr(config, "pruned_heads"):
+            pruned_heads = config.pruned_heads.copy().items()
+            config.pruned_heads = {}
+            for layer, heads in pruned_heads:
+                if self.attentions[int(layer)].n_heads == config.n_heads:
+                    self.prune_heads({int(layer): list(map(int, heads))})
+
+        self.init_weights()
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.embeddings = self._get_resized_embeddings(self.embeddings, new_num_tokens)
@@ -771,7 +782,7 @@ class XLMWithLMHeadModel(XLMPreTrainedModel):
         self.transformer = XLMModel(config)
         self.pred_layer = XLMPredLayer(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):
@@ -833,7 +844,7 @@ class XLMForSequenceClassification(XLMPreTrainedModel):
         self.transformer = XLMModel(config)
         self.sequence_summary = SequenceSummary(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, lengths=None, position_ids=None, langs=None, token_type_ids=None,
                 attention_mask=None, cache=None, labels=None, head_mask=None):
@@ -911,7 +922,7 @@ class XLMForQuestionAnswering(XLMPreTrainedModel):
         self.transformer = XLMModel(config)
         self.qa_outputs = SQuADHead(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, lengths=None, position_ids=None, langs=None, token_type_ids=None,
                 attention_mask=None, cache=None, start_positions=None, end_positions=None,

pytorch_transformers/modeling_xlnet.py

@@ -586,10 +586,7 @@ class XLNetPreTrainedModel(PreTrainedModel):
     load_tf_weights = load_tf_weights_in_xlnet
     base_model_prefix = "transformer"
 
-    def __init__(self, *inputs, **kwargs):
-        super(XLNetPreTrainedModel, self).__init__(*inputs, **kwargs)
-
-    def init_weights(self, module):
+    def _init_weights(self, module):
         """ Initialize the weights.
         """
         if isinstance(module, (nn.Linear, nn.Embedding)):
@@ -736,7 +733,7 @@ class XLNetModel(XLNetPreTrainedModel):
         self.layer = nn.ModuleList([XLNetLayer(config) for _ in range(config.n_layer)])
         self.dropout = nn.Dropout(config.dropout)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def _resize_token_embeddings(self, new_num_tokens):
         self.word_embedding = self._get_resized_embeddings(self.word_embedding, new_num_tokens)
@@ -1037,7 +1034,7 @@ class XLNetLMHeadModel(XLNetPreTrainedModel):
         self.transformer = XLNetModel(config)
         self.lm_loss = nn.Linear(config.d_model, config.n_token, bias=True)
 
-        self.apply(self.init_weights)
+        self.init_weights()
         self.tie_weights()
 
     def tie_weights(self):
@@ -1114,7 +1111,7 @@ class XLNetForSequenceClassification(XLNetPreTrainedModel):
         self.sequence_summary = SequenceSummary(config)
         self.logits_proj = nn.Linear(config.d_model, config.num_labels)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, input_mask=None, attention_mask=None,
                 mems=None, perm_mask=None, target_mapping=None,
@@ -1216,7 +1213,7 @@ class XLNetForQuestionAnswering(XLNetPreTrainedModel):
         self.end_logits = PoolerEndLogits(config)
         self.answer_class = PoolerAnswerClass(config)
 
-        self.apply(self.init_weights)
+        self.init_weights()
 
     def forward(self, input_ids, token_type_ids=None, input_mask=None, attention_mask=None,
                 mems=None, perm_mask=None, target_mapping=None,

pytorch_transformers/tests/modeling_common_test.py

@@ -213,12 +213,12 @@ class CommonTestCases:
             if not self.test_pruning:
                 return
 
+            for model_class in self.all_model_classes:
                 config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
 
                 if "head_mask" in inputs_dict:
                     del inputs_dict["head_mask"]
 
-            for model_class in self.all_model_classes:
                 config.output_attentions = True
                 config.output_hidden_states = False
                 model = model_class(config=config)
@@ -237,6 +237,120 @@ class CommonTestCases:
                 self.assertEqual(
                     attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
 
+        def test_head_pruning_save_load_from_pretrained(self):
+            if not self.test_pruning:
+                return
+
+            for model_class in self.all_model_classes:
+                config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+                if "head_mask" in inputs_dict:
+                    del inputs_dict["head_mask"]
+
+                config.output_attentions = True
+                config.output_hidden_states = False
+                model = model_class(config=config)
+                model.eval()
+                heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
+                                -1: [0]}
+                model.prune_heads(heads_to_prune)
+                directory = "pruned_model"
+                if not os.path.exists(directory):
+                    os.makedirs(directory)
+                model.save_pretrained(directory)
+                model = model_class.from_pretrained(directory)
+
+                outputs = model(**inputs_dict)
+                attentions = outputs[-1]
+                self.assertEqual(attentions[0].shape[-3], 1)
+                self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
+                self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
+
+                shutil.rmtree(directory)
+
+        def test_head_pruning_save_load_from_config_init(self):
+            if not self.test_pruning:
+                return
+
+            for model_class in self.all_model_classes:
+                config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+                if "head_mask" in inputs_dict:
+                    del inputs_dict["head_mask"]
+
+                config.output_attentions = True
+                config.output_hidden_states = False
+
+                heads_to_prune = {0: list(range(1, self.model_tester.num_attention_heads)),
+                                 -1: [0]}
+                config.pruned_heads = heads_to_prune
+
+                model = model_class(config=config)
+                model.eval()
+
+                outputs = model(**inputs_dict)
+                attentions = outputs[-1]
+
+                self.assertEqual(attentions[0].shape[-3], 1)
+                self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads)
+                self.assertEqual(attentions[-1].shape[-3], self.model_tester.num_attention_heads - 1)
+
+        def test_head_pruning_integration(self):
+            if not self.test_pruning:
+                return
+
+            for model_class in self.all_model_classes:
+                config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()
+
+                if "head_mask" in inputs_dict:
+                    del inputs_dict["head_mask"]
+
+                config.output_attentions = True
+                config.output_hidden_states = False
+
+                heads_to_prune = {0: [0], 1: [1, 2]}
+                config.pruned_heads = heads_to_prune
+
+                model = model_class(config=config)
+                model.eval()
+
+                outputs = model(**inputs_dict)
+                attentions = outputs[-1]
+
+                self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
+                self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
+                self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
+                self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
+
+                directory = "pruned_model"
+
+                if not os.path.exists(directory):
+                    os.makedirs(directory)
+                model.save_pretrained(directory)
+                model = model_class.from_pretrained(directory)
+                shutil.rmtree(directory)
+
+                outputs = model(**inputs_dict)
+                attentions = outputs[-1]
+
+                self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads - 1)
+                self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
+                self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads)
+                self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
+
+                heads_to_prune = {0: [0], 2: [1, 2]}
+                model.prune_heads(heads_to_prune)
+
+                outputs = model(**inputs_dict)
+                attentions = outputs[-1]
+
+                self.assertEqual(attentions[0].shape[-3], self.model_tester.num_attention_heads -1)
+                self.assertEqual(attentions[1].shape[-3], self.model_tester.num_attention_heads - 2)
+                self.assertEqual(attentions[2].shape[-3], self.model_tester.num_attention_heads - 2)
+                self.assertEqual(attentions[3].shape[-3], self.model_tester.num_attention_heads)
+
+                self.assertDictEqual(model.config.pruned_heads, {0: [0], 1: [1, 2], 2: [1, 2]})
+
 
         def test_hidden_states_output(self):
             config, inputs_dict = self.model_tester.prepare_config_and_inputs_for_common()