[SSD/PyT] Fixes

PyTorch/Detection/SSD/main.py

@@ -169,7 +169,7 @@ def train(train_loop_func, logger, args):
         loss_func.cuda()
 
     optimizer = torch.optim.SGD(tencent_trick(ssd300), lr=args.learning_rate,
-                                    momentum=args.momentum, weight_decay=args.weight_decay)
+                                momentum=args.momentum, weight_decay=args.weight_decay)
     scheduler = MultiStepLR(optimizer=optimizer, milestones=args.multistep, gamma=0.1)
     if args.amp:
         ssd300, optimizer = amp.initialize(ssd300, optimizer, opt_level='O2')
@@ -198,8 +198,8 @@ def train(train_loop_func, logger, args):
         acc = evaluate(ssd300, val_dataloader, cocoGt, encoder, inv_map, args)
         if args.local_rank == 0:
             print('Model precision {} mAP'.format(acc))
-
         return
+
     mean, std = generate_mean_std(args)
 
     for epoch in range(start_epoch, args.epochs):

PyTorch/Detection/SSD/src/logger.py

@@ -93,6 +93,7 @@ def log_summary(self):
         DLLogger.log((), self.summary)
 
     def update_iter(self, epoch, iteration, loss):
+        self.epoch = epoch
         self.train_iter = iteration
         self.train_loss_logger.update_iter(loss)
         if iteration % self.log_interval == 0:

PyTorch/Detection/SSD/src/train.py

@@ -20,7 +20,6 @@
 from apex import amp
 
 def train_loop(model, loss_func, epoch, optim, train_dataloader, val_dataloader, encoder, iteration, logger, args, mean, std):
-#     for nbatch, (img, _, img_size, bbox, label) in enumerate(train_dataloader):
     for nbatch, data in enumerate(train_dataloader):
         img = data[0][0][0]
         bbox = data[0][1][0]
@@ -82,8 +81,8 @@ def benchmark_train_loop(model, loss_func, epoch, optim, train_dataloader, val_d
     start_time = None
     # tensor for results
     result = torch.zeros((1,)).cuda()
-    for i, data in enumerate(loop(train_dataloader)):
-        if i >= args.benchmark_warmup:
+    for nbatch, data in enumerate(loop(train_dataloader)):
+        if nbatch >= args.benchmark_warmup:
             torch.cuda.synchronize()
             start_time = time.time()
 
@@ -109,6 +108,7 @@ def benchmark_train_loop(model, loss_func, epoch, optim, train_dataloader, val_d
             continue
         bbox, label = C.box_encoder(N, bbox, bbox_offsets, label, encoder.dboxes.cuda(), 0.5)
 
+        # output is ([N*8732, 4], [N*8732], need [N, 8732, 4], [N, 8732] respectively
         M = bbox.shape[0] // N
         bbox = bbox.view(N, M, 4)
         label = label.view(N, M)
@@ -141,13 +141,12 @@ def benchmark_train_loop(model, loss_func, epoch, optim, train_dataloader, val_d
         optim.step()
         optim.zero_grad()
 
-        if i >= args.benchmark_warmup + args.benchmark_iterations:
+        if nbatch >= args.benchmark_warmup + args.benchmark_iterations:
             break
 
-        if i >= args.benchmark_warmup:
+        if nbatch >= args.benchmark_warmup:
             torch.cuda.synchronize()
-            logger.update(args.batch_size, time.time() - start_time)
-
+            logger.update(args.batch_size*args.N_gpu, time.time() - start_time)
 
     result.data[0] = logger.print_result()
     if args.N_gpu > 1:
@@ -156,7 +155,6 @@ def benchmark_train_loop(model, loss_func, epoch, optim, train_dataloader, val_d
         print('Training performance = {} FPS'.format(float(result.data[0])))
 
 
-
 def loop(dataloader, reset=True):
     while True:
         for data in dataloader: