HongChuang/ultralytics-ascend
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions 3

New train profile argument for loggers (#2862)

Browse source 
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
This commit is contained in:
Glenn Jocher 2023-05-28 03:51:49 +02:00 • committed by GitHub
parent 0bdd4ad379
commit 6391c60089
No known key found for this signature in database
GPG key ID: 4AEE18F83AFDEB23
17 changed files with 76 additions and 47 deletions
Download patch file Download diff file Expand all files Collapse all files

35
ultralytics/yolo/utils/benchmarks.py
View file

@ -4,7 +4,7 @@ Benchmark a YOLO model formats for speed and accuracy
Usage:
from ultralytics.yolo.utils.benchmarks import ProfileModels, benchmark
ProfileModels(['yolov8n.yaml', 'yolov8s.yaml'])
ProfileModels(['yolov8n.yaml', 'yolov8s.yaml']).profile()
run_benchmarks(model='yolov8n.pt', imgsz=160)
Format | `format=argument` | Model
@ -163,13 +163,13 @@ class ProfileModels:
profile(): Profiles the models and prints the result.
"""
def __init__(self, paths: list, num_timed_runs=100, num_warmup_runs=10, imgsz=640, trt=True):
def __init__(self, paths: list, num_timed_runs=100, num_warmup_runs=10, imgsz=640, trt=True, device=None):
self.paths = paths
self.num_timed_runs = num_timed_runs
self.num_warmup_runs = num_warmup_runs
self.imgsz = imgsz
self.trt = trt # run TensorRT profiling
self.profile() # run profiling
self.device = device or torch.device(0 if torch.cuda.is_available() else 'cpu')
def profile(self):
files = self.get_files()
@ -179,15 +179,16 @@ class ProfileModels:
return
table_rows = []
device = 0 if torch.cuda.is_available() else 'cpu'
output = []
for file in files:
engine_file = file.with_suffix('.engine')
if file.suffix in ('.pt', '.yaml'):
model = YOLO(str(file))
model.fuse() # to report correct params and GFLOPs in model.info()
model_info = model.info()
if self.trt and device == 0 and not engine_file.is_file():
engine_file = model.export(format='engine', half=True, imgsz=self.imgsz, device=device)
onnx_file = model.export(format='onnx', half=True, imgsz=self.imgsz, simplify=True, device=device)
if self.trt and self.device.type != 'cpu' and not engine_file.is_file():
engine_file = model.export(format='engine', half=True, imgsz=self.imgsz, device=self.device)
onnx_file = model.export(format='onnx', half=True, imgsz=self.imgsz, simplify=True, device=self.device)
elif file.suffix == '.onnx':
model_info = self.get_onnx_model_info(file)
onnx_file = file
@ -197,8 +198,10 @@ class ProfileModels:
t_engine = self.profile_tensorrt_model(str(engine_file))
t_onnx = self.profile_onnx_model(str(onnx_file))
table_rows.append(self.generate_table_row(file.stem, t_onnx, t_engine, model_info))
output.append(self.generate_results_dict(file.stem, t_onnx, t_engine, model_info))
self.print_table(table_rows)
return output
def get_files(self):
files = []
@ -219,7 +222,7 @@ class ProfileModels:
# return (num_layers, num_params, num_gradients, num_flops)
return 0.0, 0.0, 0.0, 0.0
def iterative_sigma_clipping(self, data, sigma=2, max_iters=5):
def iterative_sigma_clipping(self, data, sigma=2, max_iters=3):
data = np.array(data)
for _ in range(max_iters):
mean, std = np.mean(data), np.std(data)
@ -235,13 +238,13 @@ class ProfileModels:
# Warmup runs
model = YOLO(engine_file)
input_data = np.random.rand(self.imgsz, self.imgsz, 3).astype(np.float32)
input_data = np.random.rand(self.imgsz, self.imgsz, 3).astype(np.float32) # must be FP32
for _ in range(self.num_warmup_runs):
model(input_data, verbose=False)
# Timed runs
run_times = []
for _ in tqdm(range(self.num_timed_runs * 30), desc=engine_file):
for _ in tqdm(range(self.num_timed_runs * 50), desc=engine_file):
results = model(input_data, verbose=False)
run_times.append(results[0].speed['inference']) # Convert to milliseconds
@ -255,6 +258,7 @@ class ProfileModels:
# Session with either 'TensorrtExecutionProvider', 'CUDAExecutionProvider', 'CPUExecutionProvider'
sess_options = ort.SessionOptions()
sess_options.graph_optimization_level = ort.GraphOptimizationLevel.ORT_ENABLE_ALL
sess_options.intra_op_num_threads = 8 # Limit the number of threads
sess = ort.InferenceSession(onnx_file, sess_options, providers=['CPUExecutionProvider'])
input_tensor = sess.get_inputs()[0]
@ -289,13 +293,22 @@ class ProfileModels:
sess.run([output_name], {input_name: input_data})
run_times.append((time.time() - start_time) * 1000) # Convert to milliseconds
run_times = self.iterative_sigma_clipping(np.array(run_times), sigma=2, max_iters=3) # sigma clipping
run_times = self.iterative_sigma_clipping(np.array(run_times), sigma=2, max_iters=5) # sigma clipping
return np.mean(run_times), np.std(run_times)
def generate_table_row(self, model_name, t_onnx, t_engine, model_info):
layers, params, gradients, flops = model_info
return f'| {model_name:18s} | {self.imgsz} | - | {t_onnx[0]:.2f} ± {t_onnx[1]:.2f} ms | {t_engine[0]:.2f} ± {t_engine[1]:.2f} ms | {params / 1e6:.1f} | {flops:.1f} |'
def generate_results_dict(self, model_name, t_onnx, t_engine, model_info):
layers, params, gradients, flops = model_info
return {
'model/name': model_name,
'model/parameters': params,
'model/GFLOPs': round(flops, 3),
'model/speed_ONNX(ms)': round(t_onnx[0], 3),
'model/speed_TensorRT(ms)': round(t_engine[0], 3)}
def print_table(self, table_rows):
gpu = torch.cuda.get_device_name(0) if torch.cuda.is_available() else 'GPU'
header = f'| Model | size<br><sup>(pixels) | mAP<sup>val<br>50-95 | Speed<br><sup>CPU ONNX<br>(ms) | Speed<br><sup>{gpu} TensorRT<br>(ms) | params<br><sup>(M) | FLOPs<br><sup>(B) |'