Add autobatch feature for best batch-size estimation (#5092)

* Autobatch

* fix mem

* fix mem2

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Update train.py

* print result

* Cleanup print result

* swap fix in call

* to 64

* use total

* fix

* fix

* fix

* fix

* fix

* Update

* Update

* Update

* Update

* Update

* Update

* Update

* Cleanup printing

* Update final printout

* Update autobatch.py

* Update autobatch.py

* Update autobatch.py

utils/autobatch.py

@@ -0,0 +1,56 @@
+# YOLOv5 🚀 by Ultralytics, GPL-3.0 license
+"""
+Auto-batch utils
+"""
+
+from copy import deepcopy
+
+import numpy as np
+import torch
+from torch.cuda import amp
+
+from utils.general import colorstr
+from utils.torch_utils import profile
+
+
+def check_train_batch_size(model, imgsz=640):
+    # Check YOLOv5 training batch size
+    with amp.autocast():
+        return autobatch(deepcopy(model).train(), imgsz)  # compute optimal batch size
+
+
+def autobatch(model, imgsz=640, fraction=0.9, batch_size=16):
+    # Automatically estimate best batch size to use `fraction` of available CUDA memory
+    # Usage:
+    #     import torch
+    #     from utils.autobatch import autobatch
+    #     model = torch.hub.load('ultralytics/yolov5', 'yolov5s', autoshape=False)
+    #     print(autobatch(model))
+
+    prefix = colorstr('autobatch: ')
+    print(f'{prefix}Computing optimal batch size for --imgsz {imgsz}')
+    device = next(model.parameters()).device  # get model device
+    if device.type == 'cpu':
+        print(f'{prefix}CUDA not detected, using default CPU batch-size {batch_size}')
+        return batch_size
+
+    d = str(device).upper()  # 'CUDA:0'
+    t = torch.cuda.get_device_properties(device).total_memory / 1024 ** 3  # (GB)
+    r = torch.cuda.memory_reserved(device) / 1024 ** 3  # (GB)
+    a = torch.cuda.memory_allocated(device) / 1024 ** 3  # (GB)
+    f = t - (r + a)  # free inside reserved
+    print(f'{prefix}{d} {t:.3g}G total, {r:.3g}G reserved, {a:.3g}G allocated, {f:.3g}G free')
+
+    batch_sizes = [1, 2, 4, 8, 16]
+    try:
+        img = [torch.zeros(b, 3, imgsz, imgsz) for b in batch_sizes]
+        y = profile(img, model, n=3, device=device)
+    except Exception as e:
+        print(f'{prefix}{e}')
+
+    y = [x[2] for x in y if x]  # memory [2]
+    batch_sizes = batch_sizes[:len(y)]
+    p = np.polyfit(batch_sizes, y, deg=1)  # first degree polynomial fit
+    b = int((f * fraction - p[1]) / p[0])  # y intercept (optimal batch size)
+    print(f'{prefix}Using colorstr(batch-size {b}) for {d} {t * fraction:.3g}G/{t:.3g}G ({fraction * 100:.0f}%)')
+    return b

utils/torch_utils.py

@@ -126,7 +126,7 @@ def profile(input, ops, n=10, device=None):
                         _ = (sum([yi.sum() for yi in y]) if isinstance(y, list) else y).sum().backward()
                         t[2] = time_sync()
                     except Exception as e:  # no backward method
-                        print(e)
+                        # print(e)  # for debug
                         t[2] = float('nan')
                     tf += (t[1] - t[0]) * 1000 / n  # ms per op forward
                     tb += (t[2] - t[1]) * 1000 / n  # ms per op backward