Compute overall accuracy using number of correct predictions (#1341)

* Trying out accuracy fix

* Computing Accuracy using num_correct

* minor fixes

* Moving mutiplication by 100 to accuracy computation, training log fix

* code quality edits

Author: abhinavnmagic

src/sparseml/pytorch/torchvision/train.py

@@ -139,11 +139,17 @@ def train_one_epoch(
                 # Reset ema buffer to keep copying weights during warmup period
                 model_ema.n_averaged.fill_(0)
 
-        acc1, acc5 = utils.accuracy(output, target, topk=(1, 5))
+        acc1, num_correct_1, acc5, num_correct_5 = utils.accuracy(
+            output, target, topk=(1, 5)
+        )
         batch_size = image.shape[0]
         metric_logger.update(loss=loss.item(), lr=optimizer.param_groups[0]["lr"])
-        metric_logger.meters["acc1"].update(acc1.item(), n=batch_size)
-        metric_logger.meters["acc5"].update(acc5.item(), n=batch_size)
+        metric_logger.meters["acc1"].update(
+            acc1.item(), n=batch_size, total=num_correct_1
+        )
+        metric_logger.meters["acc5"].update(
+            acc5.item(), n=batch_size, total=num_correct_5
+        )
         metric_logger.meters["imgs_per_sec"].update(
             batch_size / (time.time() - start_time)
         )
@@ -181,13 +187,19 @@ def evaluate(
                 output = output[0]
             loss = criterion(output, target)
 
-            acc1, acc5 = utils.accuracy(output, target, topk=(1, 5))
+            acc1, num_correct_1, acc5, num_correct_5 = utils.accuracy(
+                output, target, topk=(1, 5)
+            )
             # FIXME need to take into account that the datasets
             # could have been padded in distributed setup
             batch_size = image.shape[0]
             metric_logger.update(loss=loss.item())
-            metric_logger.meters["acc1"].update(acc1.item(), n=batch_size)
-            metric_logger.meters["acc5"].update(acc5.item(), n=batch_size)
+            metric_logger.meters["acc1"].update(
+                acc1.item(), n=batch_size, total=num_correct_1
+            )
+            metric_logger.meters["acc5"].update(
+                acc5.item(), n=batch_size, total=num_correct_5
+            )
             num_processed_samples += batch_size
     # gather the stats from all processes
 

src/sparseml/pytorch/torchvision/utils.py

@@ -28,10 +28,13 @@ def __init__(self, window_size=20, fmt=None):
         self.count = 0
         self.fmt = fmt
 
-    def update(self, value, n=1):
+    def update(self, value, n=1, total=None):
         self.deque.append(value)
         self.count += n
-        self.total += value * n
+        if total is not None:
+            self.total += total
+        else:
+            self.total += value * n
 
     def synchronize_between_processes(self):
         """
@@ -210,6 +213,7 @@ def accuracy(output, target, topk=(1,)):
         for k in topk:
             correct_k = correct[:k].flatten().sum(dtype=torch.float32)
             res.append(correct_k * (100.0 / batch_size))
+            res.append(correct_k * 100.0)
         return res