bugfix for `plain_train_net.py` and lr scheduler step (#484)

README.md

@@ -24,7 +24,8 @@ Support many tasks beyond reid, such image retrieval and face recognition. See [
 - Can be used as a library to support [different projects](projects) on top of it. We'll open source more research projects in this way.
 - Remove [ignite](https://github.com/pytorch/ignite)(a high-level library) dependency and powered by [PyTorch](https://pytorch.org/).
 
-We write a [chinese blog](https://l1aoxingyu.github.io/blogpages/reid/2020/05/29/fastreid.html) about this toolbox.
+We write a [fastreid intro](https://l1aoxingyu.github.io/blogpages/reid/fastreid/2020/05/29/fastreid.html) 
+and [fastreid v1.0](https://l1aoxingyu.github.io/blogpages/reid/fastreid/2021/04/28/fastreid-v1.html) about this toolbox.
 
 ## Changelog
 

fastreid/engine/hooks.py

@@ -257,7 +257,7 @@ class LRScheduler(HookBase):
     def after_epoch(self):
         next_iter = self.trainer.iter + 1
         next_epoch = self.trainer.epoch + 1
-        if next_iter > self.trainer.warmup_iters and next_epoch >= self.trainer.delay_epochs:
+        if next_iter > self.trainer.warmup_iters and next_epoch > self.trainer.delay_epochs:
             self._scheduler["lr_sched"].step()
 
 

tools/plain_train_net.py

@@ -16,6 +16,7 @@ sys.path.append('.')
 
 from fastreid.config import get_cfg
 from fastreid.data import build_reid_test_loader, build_reid_train_loader
+from fastreid.evaluation.testing import flatten_results_dict
 from fastreid.engine import default_argument_parser, default_setup, launch
 from fastreid.modeling import build_model
 from fastreid.solver import build_lr_scheduler, build_optimizer
@@ -33,7 +34,7 @@ logger = logging.getLogger("fastreid")
 
 
 def get_evaluator(cfg, dataset_name, output_dir=None):
-    data_loader, num_query = build_reid_test_loader(cfg, dataset_name)
+    data_loader, num_query = build_reid_test_loader(cfg, dataset_name=dataset_name)
     return data_loader, ReidEvaluator(cfg, num_query, output_dir)
 
 
@@ -49,24 +50,28 @@ def do_test(cfg, model):
             )
             results[dataset_name] = {}
             continue
-        results_i = inference_on_dataset(model, data_loader, evaluator, flip_test=cfg.TEST.FLIP_ENABLED)
+        results_i = inference_on_dataset(model, data_loader, evaluator, flip_test=cfg.TEST.FLIP.ENABLED)
         results[dataset_name] = results_i
 
-    if comm.is_main_process():
-        assert isinstance(
-            results, dict
-        ), "Evaluator must return a dict on the main process. Got {} instead.".format(
-            results
-        )
-        print_csv_format(results)
+        if comm.is_main_process():
+            assert isinstance(
+                results, dict
+            ), "Evaluator must return a dict on the main process. Got {} instead.".format(
+                results
+            )
+            logger.info("Evaluation results for {} in csv format:".format(dataset_name))
+            results_i['dataset'] = dataset_name
+            print_csv_format(results_i)
 
-    if len(results) == 1: results = list(results.values())[0]
+    if len(results) == 1:
+        results = list(results.values())[0]
 
     return results
 
 
 def do_train(cfg, model, resume=False):
     data_loader = build_reid_train_loader(cfg)
+    data_loader_iter = iter(data_loader)
 
     model.train()
     optimizer = build_optimizer(cfg, model)
@@ -78,7 +83,7 @@ def do_train(cfg, model, resume=False):
         model,
         cfg.OUTPUT_DIR,
         save_to_disk=comm.is_main_process(),
-        optimizer=optimizer
+        optimizer=optimizer,
         **scheduler
     )
 
@@ -93,6 +98,10 @@ def do_train(cfg, model, resume=False):
     delay_epochs = cfg.SOLVER.DELAY_EPOCHS
 
     periodic_checkpointer = PeriodicCheckpointer(checkpointer, cfg.SOLVER.CHECKPOINT_PERIOD, max_epoch)
+    if len(cfg.DATASETS.TESTS) == 1:
+        metric_name = "metric"
+    else:
+        metric_name = cfg.DATASETS.TESTS[0] + "/metric"
 
     writers = (
         [
@@ -111,7 +120,8 @@ def do_train(cfg, model, resume=False):
     with EventStorage(start_iter) as storage:
         for epoch in range(start_epoch, max_epoch):
             storage.epoch = epoch
-            for data, _ in zip(data_loader, range(iters_per_epoch)):
+            for _ in range(iters_per_epoch):
+                data = next(data_loader_iter)
                 storage.iter = iteration
 
                 loss_dict = model(data)
@@ -128,9 +138,9 @@ def do_train(cfg, model, resume=False):
                 optimizer.step()
                 storage.put_scalar("lr", optimizer.param_groups[0]["lr"], smoothing_hint=False)
 
-                if iteration - start_iter > 5 and (
-                        (iteration + 1) % 200 == 0 or iteration == max_iter - 1
-                ):
+                if iteration - start_iter > 5 and \
+                        ((iteration + 1) % 200 == 0 or iteration == max_iter - 1) and \
+                        ((iteration + 1) % iters_per_epoch != 0):
                     for writer in writers:
                         writer.write()
 
@@ -143,18 +153,22 @@ def do_train(cfg, model, resume=False):
             for writer in writers:
                 writer.write()
 
-            if iteration > warmup_iters and (epoch + 1) >= delay_epochs:
+            if iteration > warmup_iters and (epoch + 1) > delay_epochs:
                 scheduler["lr_sched"].step()
 
             if (
                     cfg.TEST.EVAL_PERIOD > 0
                     and (epoch + 1) % cfg.TEST.EVAL_PERIOD == 0
-                    and epoch != max_iter - 1
+                    and iteration != max_iter - 1
             ):
-                do_test(cfg, model)
+                results = do_test(cfg, model)
                 # Compared to "train_net.py", the test results are not dumped to EventStorage
+            else:
+                results = {}
+            flatten_results = flatten_results_dict(results)
 
-            periodic_checkpointer.step(epoch)
+            metric_dict = dict(metric=flatten_results[metric_name] if metric_name in flatten_results else -1)
+            periodic_checkpointer.step(epoch, **metric_dict)
 
 
 def setup(args):
@@ -184,7 +198,9 @@ def main(args):
 
     distributed = comm.get_world_size() > 1
     if distributed:
-        model = DistributedDataParallel(model, delay_allreduce=True)
+        model = DistributedDataParallel(
+            model, device_ids=[comm.get_local_rank()], broadcast_buffers=False
+        )
 
     do_train(cfg, model, resume=args.resume)
     return do_test(cfg, model)