Update documentation

iamgroot42 · Feb 7, 2024 · ae8841c · ae8841c
1 parent 9be6591
commit ae8841c
Showing 1 changed file with 36 additions and 18 deletions.
diff --git a/docs/models.html b/docs/models.html
@@ -124,13 +124,16 @@ <h1 class="title">Module <code>mimir.models</code></h1>
             target_ids = input_ids.clone()
             target_ids[:, :-trg_len] = -100
 
-            outputs = self.model(input_ids, labels=target_ids)
-            logits = outputs.logits
+            logits = self.model(input_ids, labels=target_ids).logits.cpu()
             shift_logits = logits[..., :-1, :].contiguous()
             probabilities = torch.nn.functional.log_softmax(shift_logits, dim=-1)
-            shift_labels = target_ids[..., 1:].contiguous()
+            shift_labels = target_ids[..., 1:].cpu().contiguous()
             labels_processed = shift_labels[0]
 
+            del input_ids
+            del target_ids
+
+
             for i, token_id in enumerate(labels_processed):
                 if token_id != -100:
                     probability = probabilities[0, i, token_id].item()
@@ -365,7 +368,7 @@ <h1 class="title">Module <code>mimir.models</code></h1>
             attention_mask = tokenized.attention_mask
             assert attention_mask.size() == label_batch.size()
 
-            needs_sliding = label_batch.size(1) &gt; self.max_length
+            needs_sliding = label_batch.size(1) &gt; self.max_length // 2
             if not needs_sliding:
                 label_batch = label_batch.to(self.device)
                 attention_mask = attention_mask.to(self.device)
@@ -387,8 +390,8 @@ <h1 class="title">Module <code>mimir.models</code></h1>
                 target_ids[:, :-trg_len] = -100
                 # target_ids[attention_mask == 0] = -100
 
-                outputs = self.model(input_ids, labels=target_ids, attention_mask=mask)
-                logits = outputs.logits
+                logits = self.model(input_ids, labels=target_ids, attention_mask=mask).logits.cpu()
+                target_ids = target_ids.cpu()
                 shift_logits = logits[..., :-1, :].contiguous()
                 probabilities = torch.nn.functional.log_softmax(shift_logits, dim=-1)
                 shift_labels = target_ids[..., 1:].contiguous()
@@ -398,6 +401,9 @@ <h1 class="title">Module <code>mimir.models</code></h1>
                         if token_id != -100 and token_id != self.tokenizer.pad_token_id:
                             probability = probabilities[i, j, token_id].item()
                             all_prob[i].append(probability)
+
+                del input_ids
+                del mask
 
             # average over each sample to get losses
             batch_losses = [-np.mean(all_prob[idx]) for idx in range(label_batch.size(0))]
@@ -818,7 +824,7 @@ <h3>Inherited members</h3>
             attention_mask = tokenized.attention_mask
             assert attention_mask.size() == label_batch.size()
 
-            needs_sliding = label_batch.size(1) &gt; self.max_length
+            needs_sliding = label_batch.size(1) &gt; self.max_length // 2
             if not needs_sliding:
                 label_batch = label_batch.to(self.device)
                 attention_mask = attention_mask.to(self.device)
@@ -840,8 +846,8 @@ <h3>Inherited members</h3>
                 target_ids[:, :-trg_len] = -100
                 # target_ids[attention_mask == 0] = -100
 
-                outputs = self.model(input_ids, labels=target_ids, attention_mask=mask)
-                logits = outputs.logits
+                logits = self.model(input_ids, labels=target_ids, attention_mask=mask).logits.cpu()
+                target_ids = target_ids.cpu()
                 shift_logits = logits[..., :-1, :].contiguous()
                 probabilities = torch.nn.functional.log_softmax(shift_logits, dim=-1)
                 shift_labels = target_ids[..., 1:].contiguous()
@@ -851,6 +857,9 @@ <h3>Inherited members</h3>
                         if token_id != -100 and token_id != self.tokenizer.pad_token_id:
                             probability = probabilities[i, j, token_id].item()
                             all_prob[i].append(probability)
+
+                del input_ids
+                del mask
 
             # average over each sample to get losses
             batch_losses = [-np.mean(all_prob[idx]) for idx in range(label_batch.size(0))]
@@ -1047,7 +1056,7 @@ <h3>Methods</h3>
         attention_mask = tokenized.attention_mask
         assert attention_mask.size() == label_batch.size()
 
-        needs_sliding = label_batch.size(1) &gt; self.max_length
+        needs_sliding = label_batch.size(1) &gt; self.max_length // 2
         if not needs_sliding:
             label_batch = label_batch.to(self.device)
             attention_mask = attention_mask.to(self.device)
@@ -1069,8 +1078,8 @@ <h3>Methods</h3>
             target_ids[:, :-trg_len] = -100
             # target_ids[attention_mask == 0] = -100
 
-            outputs = self.model(input_ids, labels=target_ids, attention_mask=mask)
-            logits = outputs.logits
+            logits = self.model(input_ids, labels=target_ids, attention_mask=mask).logits.cpu()
+            target_ids = target_ids.cpu()
             shift_logits = logits[..., :-1, :].contiguous()
             probabilities = torch.nn.functional.log_softmax(shift_logits, dim=-1)
             shift_labels = target_ids[..., 1:].contiguous()
@@ -1080,6 +1089,9 @@ <h3>Methods</h3>
                     if token_id != -100 and token_id != self.tokenizer.pad_token_id:
                         probability = probabilities[i, j, token_id].item()
                         all_prob[i].append(probability)
+
+            del input_ids
+            del mask
 
         # average over each sample to get losses
         batch_losses = [-np.mean(all_prob[idx]) for idx in range(label_batch.size(0))]
@@ -1391,13 +1403,16 @@ <h3>Inherited members</h3>
             target_ids = input_ids.clone()
             target_ids[:, :-trg_len] = -100
 
-            outputs = self.model(input_ids, labels=target_ids)
-            logits = outputs.logits
+            logits = self.model(input_ids, labels=target_ids).logits.cpu()
             shift_logits = logits[..., :-1, :].contiguous()
             probabilities = torch.nn.functional.log_softmax(shift_logits, dim=-1)
-            shift_labels = target_ids[..., 1:].contiguous()
+            shift_labels = target_ids[..., 1:].cpu().contiguous()
             labels_processed = shift_labels[0]
 
+            del input_ids
+            del target_ids
+
+
             for i, token_id in enumerate(labels_processed):
                 if token_id != -100:
                     probability = probabilities[0, i, token_id].item()
@@ -1601,13 +1616,16 @@ <h3>Methods</h3>
         target_ids = input_ids.clone()
         target_ids[:, :-trg_len] = -100
 
-        outputs = self.model(input_ids, labels=target_ids)
-        logits = outputs.logits
+        logits = self.model(input_ids, labels=target_ids).logits.cpu()
         shift_logits = logits[..., :-1, :].contiguous()
         probabilities = torch.nn.functional.log_softmax(shift_logits, dim=-1)
-        shift_labels = target_ids[..., 1:].contiguous()
+        shift_labels = target_ids[..., 1:].cpu().contiguous()
         labels_processed = shift_labels[0]
 
+        del input_ids
+        del target_ids
+
+
         for i, token_id in enumerate(labels_processed):
             if token_id != -100:
                 probability = probabilities[0, i, token_id].item()