def play_video(): try: nonlocal frame_buffer, running, video_fps, is_webcam, num_frames, frames_displayed, vid_done video_frame_times = MovingAverage(100) frame_time_stabilizer = frame_time_target last_time = None stabilizer_step = 0.0005 progress_bar = ProgressBar(30, num_frames) while running: frame_time_start = time.time() if not frame_buffer.empty(): next_time = time.time() if last_time is not None: video_frame_times.add(next_time - last_time) video_fps = 1 / video_frame_times.get_avg() if out_path is None: cv2.imshow(path, frame_buffer.get()) else: out.write(frame_buffer.get()) frames_displayed += 1 last_time = next_time if out_path is not None: if video_frame_times.get_avg() == 0: fps = 0 else: fps = 1 / video_frame_times.get_avg() progress = frames_displayed / num_frames * 100 progress_bar.set_val(frames_displayed) print('\rProcessing Frames %s %6d / %6d (%5.2f%%) %5.2f fps ' % (repr(progress_bar), frames_displayed, num_frames, progress, fps), end='') # This is split because you don't want savevideo to require cv2 display functionality (see #197) if out_path is None and cv2.waitKey(1) == 27: # Press Escape to close running = False if not (frames_displayed < num_frames): running = False if not vid_done: buffer_size = frame_buffer.qsize() if buffer_size < args.video_multiframe: frame_time_stabilizer += stabilizer_step elif buffer_size > args.video_multiframe: frame_time_stabilizer -= stabilizer_step if frame_time_stabilizer < 0: frame_time_stabilizer = 0 new_target = frame_time_stabilizer if is_webcam else max(frame_time_stabilizer, frame_time_target) else: new_target = frame_time_target next_frame_target = max(2 * new_target - video_frame_times.get_avg(), 0) target_time = frame_time_start + next_frame_target - 0.001 # Let's just subtract a millisecond to be safe if out_path is None or args.emulate_playback: # This gives more accurate timing than if sleeping the whole amount at once while time.time() < target_time: time.sleep(0.001) else: # Let's not starve the main thread, now time.sleep(0.001) except: # See issue #197 for why this is necessary import traceback traceback.print_exc()
def evalvideo(net: Yolact, path: str): # If the path is a digit, parse it as a webcam index if path.isdigit(): vid = cv2.VideoCapture(int(path)) else: vid = cv2.VideoCapture(path) if not vid.isOpened(): print('Could not open video "%s"' % path) exit(-1) net = CustomDataParallel(net).cuda() transform = torch.nn.DataParallel(FastBaseTransform()).cuda() frame_times = MovingAverage() fps = 0 # The 0.8 is to account for the overhead of time.sleep frame_time_target = 0.8 / vid.get(cv2.CAP_PROP_FPS) def cleanup_and_exit(): print() pool.terminate() vid.release() cv2.destroyAllWindows() exit() def get_next_frame(vid): return [vid.read()[1] for _ in range(args.video_multiframe)] def transform_frame(frames): with torch.no_grad(): frames = [ torch.from_numpy(frame).cuda().float() for frame in frames ] return frames, transform(torch.stack(frames, 0)) def eval_network(inp): with torch.no_grad(): frames, imgs = inp return frames, net(imgs) def prep_frame(inp): with torch.no_grad(): frame, preds = inp return prep_display(preds, frame, None, None, undo_transform=False, class_color=True) extract_frame = lambda x, i: (x[0][i] if x[1][i] is None else x[0][i].to(x[ 1][i]['box'].device), [x[1][i]]) # Prime the network on the first frame because I do some thread unsafe things otherwise print('Initializing model... ', end='') eval_network(transform_frame(get_next_frame(vid))) print('Done.') # For each frame the sequence of functions it needs to go through to be processed (in reversed order) sequence = [prep_frame, eval_network, transform_frame] pool = ThreadPool(processes=len(sequence) + args.video_multiframe) active_frames = [] print() while vid.isOpened(): start_time = time.time() # Start loading the next frames from the disk next_frames = pool.apply_async(get_next_frame, args=(vid, )) # For each frame in our active processing queue, dispatch a job # for that frame using the current function in the sequence for frame in active_frames: frame['value'] = pool.apply_async(sequence[frame['idx']], args=(frame['value'], )) # For each frame whose job was the last in the sequence (i.e. for all final outputs) for frame in active_frames: if frame['idx'] == 0: # Wait here so that the frame has time to process and so that the video plays at the proper speed time.sleep(frame_time_target) cv2.imshow(path, frame['value'].get()) if cv2.waitKey(1) == 27: # Press Escape to close cleanup_and_exit() # Remove the finished frames from the processing queue active_frames = [x for x in active_frames if x['idx'] > 0] # Finish evaluating every frame in the processing queue and advanced their position in the sequence for frame in list(reversed(active_frames)): frame['value'] = frame['value'].get() frame['idx'] -= 1 if frame['idx'] == 0: # Split this up into individual threads for prep_frame since it doesn't support batch size active_frames += [{ 'value': extract_frame(frame['value'], i), 'idx': 0 } for i in range(1, args.video_multiframe)] frame['value'] = extract_frame(frame['value'], 0) # Finish loading in the next frames and add them to the processing queue active_frames.append({ 'value': next_frames.get(), 'idx': len(sequence) - 1 }) # Compute FPS frame_times.add(time.time() - start_time) fps = args.video_multiframe / frame_times.get_avg() print('\rAvg FPS: %.2f ' % fps, end='') cleanup_and_exit()
def evalvideo(net:Yolact, path:str, out_path:str=None): # If the path is a digit, parse it as a webcam index is_webcam = path.isdigit() # If the input image size is constant, this make things faster (hence why we can use it in a video setting). cudnn.benchmark = True if is_webcam: vid = cv2.VideoCapture(int(path)) else: vid = cv2.VideoCapture(path) if not vid.isOpened(): print('Could not open video "%s"' % path) exit(-1) target_fps = round(vid.get(cv2.CAP_PROP_FPS)) frame_width = round(vid.get(cv2.CAP_PROP_FRAME_WIDTH)) frame_height = round(vid.get(cv2.CAP_PROP_FRAME_HEIGHT)) if is_webcam: num_frames = float('inf') else: num_frames = round(vid.get(cv2.CAP_PROP_FRAME_COUNT)) net = CustomDataParallel(net).cuda() transform = torch.nn.DataParallel(FastBaseTransform()).cuda() frame_times = MovingAverage(100) fps = 0 frame_time_target = 1 / target_fps running = True fps_str = '' vid_done = False frames_displayed = 0 if out_path is not None: out = cv2.VideoWriter(out_path, cv2.VideoWriter_fourcc(*"mp4v"), target_fps, (frame_width, frame_height)) def cleanup_and_exit(): print() pool.terminate() vid.release() if out_path is not None: out.release() cv2.destroyAllWindows() exit() def get_next_frame(vid): frames = [] for idx in range(args.video_multiframe): frame = vid.read()[1] if frame is None: return frames frames.append(frame) return frames def transform_frame(frames): with torch.no_grad(): frames = [torch.from_numpy(frame).cuda().float() for frame in frames] return frames, transform(torch.stack(frames, 0)) def eval_network(inp): with torch.no_grad(): frames, imgs = inp num_extra = 0 while imgs.size(0) < args.video_multiframe: imgs = torch.cat([imgs, imgs[0].unsqueeze(0)], dim=0) num_extra += 1 out = net(imgs) if num_extra > 0: out = out[:-num_extra] return frames, out def prep_frame(inp, fps_str): with torch.no_grad(): frame, preds = inp return prep_display(preds, frame, None, None, undo_transform=False, class_color=True, fps_str=fps_str) frame_buffer = Queue() video_fps = 0 # All this timing code to make sure that def play_video(): try: nonlocal frame_buffer, running, video_fps, is_webcam, num_frames, frames_displayed, vid_done video_frame_times = MovingAverage(100) frame_time_stabilizer = frame_time_target last_time = None stabilizer_step = 0.0005 progress_bar = ProgressBar(30, num_frames) while running: frame_time_start = time.time() if not frame_buffer.empty(): next_time = time.time() if last_time is not None: video_frame_times.add(next_time - last_time) video_fps = 1 / video_frame_times.get_avg() if out_path is None: cv2.imshow(path, frame_buffer.get()) else: out.write(frame_buffer.get()) frames_displayed += 1 last_time = next_time if out_path is not None: if video_frame_times.get_avg() == 0: fps = 0 else: fps = 1 / video_frame_times.get_avg() progress = frames_displayed / num_frames * 100 progress_bar.set_val(frames_displayed) print('\rProcessing Frames %s %6d / %6d (%5.2f%%) %5.2f fps ' % (repr(progress_bar), frames_displayed, num_frames, progress, fps), end='') # This is split because you don't want savevideo to require cv2 display functionality (see #197) if out_path is None and cv2.waitKey(1) == 27: # Press Escape to close running = False if not (frames_displayed < num_frames): running = False if not vid_done: buffer_size = frame_buffer.qsize() if buffer_size < args.video_multiframe: frame_time_stabilizer += stabilizer_step elif buffer_size > args.video_multiframe: frame_time_stabilizer -= stabilizer_step if frame_time_stabilizer < 0: frame_time_stabilizer = 0 new_target = frame_time_stabilizer if is_webcam else max(frame_time_stabilizer, frame_time_target) else: new_target = frame_time_target next_frame_target = max(2 * new_target - video_frame_times.get_avg(), 0) target_time = frame_time_start + next_frame_target - 0.001 # Let's just subtract a millisecond to be safe if out_path is None or args.emulate_playback: # This gives more accurate timing than if sleeping the whole amount at once while time.time() < target_time: time.sleep(0.001) else: # Let's not starve the main thread, now time.sleep(0.001) except: # See issue #197 for why this is necessary import traceback traceback.print_exc() extract_frame = lambda x, i: (x[0][i] if x[1][i]['detection'] is None else x[0][i].to(x[1][i]['detection']['box'].device), [x[1][i]]) # Prime the network on the first frame because I do some thread unsafe things otherwise print('Initializing model... ', end='') first_batch = eval_network(transform_frame(get_next_frame(vid))) print('Done.') # For each frame the sequence of functions it needs to go through to be processed (in reversed order) sequence = [prep_frame, eval_network, transform_frame] pool = ThreadPool(processes=len(sequence) + args.video_multiframe + 2) pool.apply_async(play_video) active_frames = [{'value': extract_frame(first_batch, i), 'idx': 0} for i in range(len(first_batch[0]))] print() if out_path is None: print('Press Escape to close.') try: while vid.isOpened() and running: # Hard limit on frames in buffer so we don't run out of memory >.> while frame_buffer.qsize() > 100: time.sleep(0.001) start_time = time.time() # Start loading the next frames from the disk if not vid_done: next_frames = pool.apply_async(get_next_frame, args=(vid,)) else: next_frames = None if not (vid_done and len(active_frames) == 0): # For each frame in our active processing queue, dispatch a job # for that frame using the current function in the sequence for frame in active_frames: _args = [frame['value']] if frame['idx'] == 0: _args.append(fps_str) frame['value'] = pool.apply_async(sequence[frame['idx']], args=_args) # For each frame whose job was the last in the sequence (i.e. for all final outputs) for frame in active_frames: if frame['idx'] == 0: frame_buffer.put(frame['value'].get()) # Remove the finished frames from the processing queue active_frames = [x for x in active_frames if x['idx'] > 0] # Finish evaluating every frame in the processing queue and advanced their position in the sequence for frame in list(reversed(active_frames)): frame['value'] = frame['value'].get() frame['idx'] -= 1 if frame['idx'] == 0: # Split this up into individual threads for prep_frame since it doesn't support batch size active_frames += [{'value': extract_frame(frame['value'], i), 'idx': 0} for i in range(1, len(frame['value'][0]))] frame['value'] = extract_frame(frame['value'], 0) # Finish loading in the next frames and add them to the processing queue if next_frames is not None: frames = next_frames.get() if len(frames) == 0: vid_done = True else: active_frames.append({'value': frames, 'idx': len(sequence)-1}) # Compute FPS frame_times.add(time.time() - start_time) fps = args.video_multiframe / frame_times.get_avg() else: fps = 0 fps_str = 'Processing FPS: %.2f | Video Playback FPS: %.2f | Frames in Buffer: %d' % (fps, video_fps, frame_buffer.qsize()) if not args.display_fps: print('\r' + fps_str + ' ', end='') except KeyboardInterrupt: print('\nStopping...') cleanup_and_exit()
def evalvideo(net: Yolact, path: str): # If the path is a digit, parse it as a webcam index is_webcam = path.isdigit() if is_webcam: vid = cv2.VideoCapture(int(path)) else: vid = cv2.VideoCapture(path) if not vid.isOpened(): print('Could not open video "%s"' % path) exit(-1) net = CustomDataParallel(net).cuda() transform = torch.nn.DataParallel(FastBaseTransform()).cuda() frame_times = MovingAverage(100) fps = 0 # The 0.8 is to account for the overhead of time.sleep frame_time_target = 1 / vid.get(cv2.CAP_PROP_FPS) running = True def cleanup_and_exit(): print() pool.terminate() vid.release() cv2.destroyAllWindows() exit() def get_next_frame(vid): return [vid.read()[1] for _ in range(args.video_multiframe)] def transform_frame(frames): with torch.no_grad(): frames = [ torch.from_numpy(frame).cuda().float() for frame in frames ] return frames, transform(torch.stack(frames, 0)) def eval_network(inp): with torch.no_grad(): frames, imgs = inp return frames, net(imgs) def prep_frame(inp): with torch.no_grad(): frame, preds = inp return prep_display(preds, frame, None, None, undo_transform=False, class_color=True) frame_buffer = Queue() video_fps = 0 # All this timing code to make sure that def play_video(): nonlocal frame_buffer, running, video_fps, is_webcam video_frame_times = MovingAverage(100) frame_time_stabilizer = frame_time_target last_time = None stabilizer_step = 0.0005 while running: frame_time_start = time.time() if not frame_buffer.empty(): next_time = time.time() if last_time is not None: video_frame_times.add(next_time - last_time) video_fps = 1 / video_frame_times.get_avg() cv2.imshow(path, frame_buffer.get()) last_time = next_time if cv2.waitKey(1) == 27: # Press Escape to close running = False buffer_size = frame_buffer.qsize() if buffer_size < args.video_multiframe: frame_time_stabilizer += stabilizer_step elif buffer_size > args.video_multiframe: frame_time_stabilizer -= stabilizer_step if frame_time_stabilizer < 0: frame_time_stabilizer = 0 new_target = frame_time_stabilizer if is_webcam else max( frame_time_stabilizer, frame_time_target) next_frame_target = max( 2 * new_target - video_frame_times.get_avg(), 0) target_time = frame_time_start + next_frame_target - 0.001 # Let's just subtract a millisecond to be safe # This gives more accurate timing than if sleeping the whole amount at once while time.time() < target_time: time.sleep(0.001) extract_frame = lambda x, i: (x[0][i] if x[1][i] is None else x[0][i].to(x[ 1][i]['box'].device), [x[1][i]]) # Prime the network on the first frame because I do some thread unsafe things otherwise print('Initializing model... ', end='') eval_network(transform_frame(get_next_frame(vid))) print('Done.') # For each frame the sequence of functions it needs to go through to be processed (in reversed order) sequence = [prep_frame, eval_network, transform_frame] pool = ThreadPool(processes=len(sequence) + args.video_multiframe + 2) pool.apply_async(play_video) active_frames = [] print() while vid.isOpened() and running: start_time = time.time() # Start loading the next frames from the disk next_frames = pool.apply_async(get_next_frame, args=(vid, )) # For each frame in our active processing queue, dispatch a job # for that frame using the current function in the sequence for frame in active_frames: frame['value'] = pool.apply_async(sequence[frame['idx']], args=(frame['value'], )) # For each frame whose job was the last in the sequence (i.e. for all final outputs) for frame in active_frames: if frame['idx'] == 0: frame_buffer.put(frame['value'].get()) # Remove the finished frames from the processing queue active_frames = [x for x in active_frames if x['idx'] > 0] # Finish evaluating every frame in the processing queue and advanced their position in the sequence for frame in list(reversed(active_frames)): frame['value'] = frame['value'].get() frame['idx'] -= 1 if frame['idx'] == 0: # Split this up into individual threads for prep_frame since it doesn't support batch size active_frames += [{ 'value': extract_frame(frame['value'], i), 'idx': 0 } for i in range(1, args.video_multiframe)] frame['value'] = extract_frame(frame['value'], 0) # Finish loading in the next frames and add them to the processing queue active_frames.append({ 'value': next_frames.get(), 'idx': len(sequence) - 1 }) # Compute FPS frame_times.add(time.time() - start_time) fps = args.video_multiframe / frame_times.get_avg() print( '\rProcessing FPS: %.2f | Video Playback FPS: %.2f | Frames in Buffer: %d ' % (fps, video_fps, frame_buffer.qsize()), end='') cleanup_and_exit()
# GPU net = net.cuda() torch.set_default_tensor_type('torch.cuda.FloatTensor') x = torch.zeros((1, 3, cfg.max_size, cfg.max_size)) y = net(x) for p in net.prediction_layers: print(p.last_conv_size) print() for k, a in y.items(): print(k + ': ', a.size(), torch.sum(a)) exit() net(x) # timer.disable('pass2') avg = MovingAverage() try: while True: timer.reset() with timer.env('everything else'): net(x) avg.add(timer.total_time()) print('\033[2J') # Moves console cursor to 0,0 timer.print_stats() print('Avg fps: %.2f\tAvg ms: %.2f ' % (1 / avg.get_avg(), avg.get_avg() * 1000)) except KeyboardInterrupt: pass
def evalvideo(net:Yolact, path:str): # If the path is a digit, parse it as a webcam index is_webcam = path.isdigit() if is_webcam: vid = cv2.VideoCapture(int(path)) else: vid = cv2.VideoCapture(path) if not vid.isOpened(): print('Could not open video "%s"' % path) exit(-1) vid.set(cv2.CAP_PROP_FRAME_HEIGHT, args.video_res[1]) vid.set(cv2.CAP_PROP_FRAME_WIDTH, args.video_res[0]) vid.set(cv2.CAP_PROP_FOURCC, cv2.VideoWriter_fourcc('H', '2', '6', '4')) W = vid.get(cv2.CAP_PROP_FRAME_WIDTH) H = vid.get(cv2.CAP_PROP_FRAME_HEIGHT) print('capture video size: w{} h{}'.format(W, H)) net = CustomDataParallel(net).cuda() transform = torch.nn.DataParallel(FastBaseTransform()).cuda() frame_times = MovingAverage(100) fps = 0 # The 0.8 is to account for the overhead of time.sleep frame_time_target = 1 / vid.get(cv2.CAP_PROP_FPS) running = True slide_num =0 bg_imgs = [] mask_alpha = args.mask_alpha bgvid = None if args.mask_bg_path is not None: if Path(args.mask_bg_path).suffix == '.pdf': pdfimages = pdf2image.convert_from_path(args.mask_bg_path) for pdfimage in pdfimages: cvimage = np.asarray(pdfimage) cvimage = cv2.cvtColor(cvimage, cv2.COLOR_RGB2BGR) bg_img = cv2.resize(cvimage, (int(W), int(H))) bg_img_gpu = torch.from_numpy( bg_img / 255.0 ).cuda().float() bg_imgs.append(bg_img_gpu) elif Path(args.mask_bg_path).suffix == '.mp4': bg_imgs = [] bgvid = cv2.VideoCapture(args.mask_bg_path) BG_W = bgvid.get(cv2.CAP_PROP_FRAME_WIDTH) BG_H = bgvid.get(cv2.CAP_PROP_FRAME_HEIGHT) print('background capture video size: w{} h{}'.format(BG_W, BG_H)) else: bg_img = cv2.resize(cv2.imread(args.mask_bg_path), (int(W), int(H))) bg_img_gpu = torch.from_numpy( bg_img / 255.0 ).cuda().float() bg_imgs.append(bg_img_gpu) def cleanup_and_exit(): print() pool.terminate() vid.release() cv2.destroyAllWindows() exit() def get_next_frame(vid): return [vid.read()[1] for _ in range(args.video_multiframe)] def rescale(frame): if args.rescale < 1.0: h, w = frame.shape[:2] new_img = np.zeros((h, w, 3), np.uint8) rescale_frame = cv2.resize(frame, dsize=(int(w * args.rescale), int(h * args.rescale))) new_img[h - int(h * args.rescale):h, 0:int(w * args.rescale)] = rescale_frame return new_img else: return frame def transform_frame(frames): with torch.no_grad(): frames = [ torch.from_numpy( rescale(frame) ).cuda().float() for frame in frames] return frames, transform(torch.stack(frames, 0)) def eval_network(inp): with torch.no_grad(): frames, imgs = inp return frames, net(imgs) def prep_frame(inp): nonlocal slide_num, bg_imgs, mask_alpha, bgvid with torch.no_grad(): frame, preds = inp return prep_display(preds, frame, None, None, undo_transform=False, class_color=True, mask_alpha=mask_alpha, bg_imgs=bg_imgs, slide_num=slide_num, bgvid=bgvid) frame_buffer = Queue() video_fps = 0 # All this timing code to make sure that def play_video(): nonlocal frame_buffer, running, video_fps, is_webcam, slide_num video_frame_times = MovingAverage(100) frame_time_stabilizer = frame_time_target last_time = None stabilizer_step = 0.0005 while running: frame_time_start = time.time() if not frame_buffer.empty(): next_time = time.time() if last_time is not None: video_frame_times.add(next_time - last_time) video_fps = 1 / video_frame_times.get_avg() cv2.imshow(path, frame_buffer.get()) last_time = next_time key_press = cv2.waitKey(1) & 0xff if key_press == 27: # Press Escape to close running = False elif key_press == ord('n'): if slide_num < len(bg_imgs) - 1: slide_num = slide_num + 1 elif key_press == ord('b'): if 0 < slide_num: slide_num = slide_num - 1 buffer_size = frame_buffer.qsize() if buffer_size < args.video_multiframe: frame_time_stabilizer += stabilizer_step elif buffer_size > args.video_multiframe: frame_time_stabilizer -= stabilizer_step if frame_time_stabilizer < 0: frame_time_stabilizer = 0 new_target = frame_time_stabilizer if is_webcam else max(frame_time_stabilizer, frame_time_target) next_frame_target = max(2 * new_target - video_frame_times.get_avg(), 0) target_time = frame_time_start + next_frame_target - 0.001 # Let's just subtract a millisecond to be safe # This gives more accurate timing than if sleeping the whole amount at once while time.time() < target_time: time.sleep(0.001) extract_frame = lambda x, i: (x[0][i] if x[1][i] is None else x[0][i].to(x[1][i]['box'].device), [x[1][i]]) # Prime the network on the first frame because I do some thread unsafe things otherwise print('Initializing model... ', end='') eval_network(transform_frame(get_next_frame(vid))) print('Done.') # For each frame the sequence of functions it needs to go through to be processed (in reversed order) sequence = [prep_frame, eval_network, transform_frame] pool = ThreadPool(processes=len(sequence) + args.video_multiframe + 2) pool.apply_async(play_video) active_frames = [] print() while vid.isOpened() and running: start_time = time.time() # Start loading the next frames from the disk next_frames = pool.apply_async(get_next_frame, args=(vid,)) # For each frame in our active processing queue, dispatch a job # for that frame using the current function in the sequence for frame in active_frames: frame['value'] = pool.apply_async(sequence[frame['idx']], args=(frame['value'],)) # For each frame whose job was the last in the sequence (i.e. for all final outputs) for frame in active_frames: if frame['idx'] == 0: frame_buffer.put(frame['value'].get()) # Remove the finished frames from the processing queue active_frames = [x for x in active_frames if x['idx'] > 0] # Finish evaluating every frame in the processing queue and advanced their position in the sequence for frame in list(reversed(active_frames)): frame['value'] = frame['value'].get() frame['idx'] -= 1 if frame['idx'] == 0: # Split this up into individual threads for prep_frame since it doesn't support batch size active_frames += [{'value': extract_frame(frame['value'], i), 'idx': 0} for i in range(1, args.video_multiframe)] frame['value'] = extract_frame(frame['value'], 0) # Finish loading in the next frames and add them to the processing queue active_frames.append({'value': next_frames.get(), 'idx': len(sequence)-1}) # Compute FPS frame_times.add(time.time() - start_time) fps = args.video_multiframe / frame_times.get_avg() print('\rProcessing FPS: %.2f | Video Playback FPS: %.2f | Frames in Buffer: %d ' % (fps, video_fps, frame_buffer.qsize()), end='') cleanup_and_exit()
def train(): if not os.path.exists(args.save_folder): os.mkdir(args.save_folder) dataset = COCODetection(image_path=cfg.dataset.train_images, info_file=cfg.dataset.train_info, transform=SSDAugmentation(MEANS)) if args.validation_epoch > 0: setup_eval() val_dataset = COCODetection(image_path=cfg.dataset.valid_images, info_file=cfg.dataset.valid_info, transform=BaseTransform(MEANS)) # Parallel wraps the underlying module, but when saving and loading we don't want that yolact_net = Yolact() net = yolact_net net.train() # I don't use the timer during training (I use a different timing method). # Apparently there's a race condition with multiple GPUs. timer.disable_all() # Both of these can set args.resume to None, so do them before the check if args.resume == 'interrupt': args.resume = SavePath.get_interrupt(args.save_folder) elif args.resume == 'latest': args.resume = SavePath.get_latest(args.save_folder, cfg.name) if args.resume is not None: print('Resuming training, loading {}...'.format(args.resume)) yolact_net.load_weights(args.resume) if args.start_iter == -1: args.start_iter = SavePath.from_str(args.resume).iteration else: print('Initializing weights...') yolact_net.init_weights(backbone_path=args.save_folder + cfg.backbone.path) optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.decay) criterion = MultiBoxLoss(num_classes=cfg.num_classes, pos_threshold=cfg.positive_iou_threshold, neg_threshold=cfg.negative_iou_threshold, negpos_ratio=3) if args.cuda: cudnn.benchmark = True net = nn.DataParallel(net).cuda() criterion = nn.DataParallel(criterion).cuda() # net = net.cuda() # criterion = criterion.cuda() # criterion = criterion.cuda() # loss counters loc_loss = 0 conf_loss = 0 iteration = max(args.start_iter, 0) last_time = time.time() epoch_size = len(dataset) // args.batch_size num_epochs = math.ceil(cfg.max_iter / epoch_size) # Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index step_index = 0 data_loader = data.DataLoader(dataset, args.batch_size, num_workers=args.num_workers, shuffle=True, collate_fn=detection_collate, pin_memory=True) save_path = lambda epoch, iteration: SavePath(cfg.name, epoch, iteration).get_path(root=args.save_folder) time_avg = MovingAverage() global loss_types # Forms the print order loss_avgs = {k: MovingAverage(100) for k in loss_types} print('Begin training!') print() # try-except so you can use ctrl+c to save early and stop training try: for epoch in range(num_epochs): # Resume from start_iter if (epoch + 1) * epoch_size < iteration: continue for datum in data_loader: # Stop if we've reached an epoch if we're resuming from start_iter if iteration == (epoch + 1) * epoch_size: break # Stop at the configured number of iterations even if mid-epoch if iteration == cfg.max_iter: break # Change a config setting if we've reached the specified iteration changed = False for change in cfg.delayed_settings: if iteration >= change[0]: changed = True cfg.replace(change[1]) # Reset the loss averages because things might have changed for avg in loss_avgs: avg.reset() # If a config setting was changed, remove it from the list so we don't keep checking if changed: cfg.delayed_settings = [x for x in cfg.delayed_settings if x[0] > iteration] # Warm up by linearly interpolating the learning rate from some smaller value if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until: set_lr(optimizer, (args.lr - cfg.lr_warmup_init) * (iteration / cfg.lr_warmup_until) + cfg.lr_warmup_init) # Adjust the learning rate at the given iterations, but also if we resume from past that iteration while step_index < len(cfg.lr_steps) and iteration >= cfg.lr_steps[step_index]: step_index += 1 set_lr(optimizer, args.lr * (args.gamma ** step_index)) # Load training data # Note, for training on multiple gpus this will use the custom replicate and gather I wrote up there images, targets, masks, num_crowds = prepare_data(datum) # Forward Pass out = net(images) # Compute Loss optimizer.zero_grad() wrapper = ScatterWrapper(targets, masks, num_crowds) losses = criterion(out, wrapper, wrapper.make_mask()) losses = {k: v.mean() for k, v in losses.items()} # Mean here because Dataparallel # original loss = sum([losses[k] for k in losses]) # loss = sum([losses[k] for k in losses]) + losses['S'] * 10 # Huan # Backprop loss.backward() # Do this to free up vram even if loss is not finite if torch.isfinite(loss).item(): optimizer.step() # Add the loss to the moving average for bookkeeping for k in losses: loss_avgs[k].add(losses[k].item()) cur_time = time.time() elapsed = cur_time - last_time last_time = cur_time # Exclude graph setup from the timing information if iteration != args.start_iter: time_avg.add(elapsed) if iteration % 100 == 0: eta_str = \ str(datetime.timedelta(seconds=(cfg.max_iter - iteration) * time_avg.get_avg())).split('.')[0] total = sum([loss_avgs[k].get_avg() for k in losses]) loss_labels = sum([[k, loss_avgs[k].get_avg()] for k in loss_types if k in losses], []) print(('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) + ' T: %.3f || ETA: %s || timer: %.3f') % tuple([epoch, iteration] + loss_labels + [total, eta_str, elapsed]), flush=True) iteration += 1 if iteration % args.save_interval == 0 and iteration != args.start_iter: if args.keep_latest: latest = SavePath.get_latest(args.save_folder, cfg.name) print('Saving state, iter:', iteration) yolact_net.save_weights(save_path(epoch, iteration)) if args.keep_latest and latest is not None: if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval: print('Deleting old save...') os.remove(latest) # This is done per epoch if args.validation_epoch > 0: if epoch % args.validation_epoch == 0 and epoch > 0: compute_validation_map(yolact_net, val_dataset) except KeyboardInterrupt: print('Stopping early. Saving network...') # Delete previous copy of the interrupted network so we don't spam the weights folder SavePath.remove_interrupt(args.save_folder) yolact_net.save_weights(save_path(epoch, repr(iteration) + '_interrupt')) exit() yolact_net.save_weights(save_path(epoch, iteration))
if torch.isfinite(loss).item(): optimizer.step() for k in losses: loss_avgs[k].add(losses[k].item( )) # Add the loss to the moving average for bookkeeping grad_end = time.time() if not (i == 0 and epoch == start_epoch): iter_time = grad_end - temp batch_time.add(iter_time) temp = grad_end if iter % 10 == 0: cur_lr = optimizer.param_groups[0]['lr'] seconds = (cfg.max_iter - iter) * batch_time.get_avg() eta_str = str( datetime.timedelta(seconds=seconds)).split('.')[0] total = sum([loss_avgs[k].get_avg() for k in losses]) loss_labels = sum([[k, loss_avgs[k].get_avg()] for k in loss_types if k in losses], []) t_forward = forward_end - forward_start t_data = iter_time - (grad_end - forward_start) time_str = ' T: %.3f | lr: %.5f | t_data: %.3f | t_forward: %.3f | t_total: %.3f | ETA: %s' print( ('[%3d] %7d |' + (' %s: %.3f |' * len(losses)) + time_str) % tuple( [epoch, iter] + loss_labels + [total, cur_lr, t_data, t_forward, iter_time, eta_str]
if cuda: frame_origin = frame_origin.cuda() img_h, img_w = frame_origin.shape[0], frame_origin.shape[1] frame_trans = FastBaseTransform()(frame_origin.unsqueeze(0)) net_outs = net(frame_trans) nms_outs = NMS(net_outs, args.traditional_nms) results = after_nms(nms_outs, img_h, img_w, crop_masks=not args.no_crop, visual_thre=args.visual_thre) if cuda: torch.cuda.synchronize() temp = time_here time_here = time.time() if i > 0: frame_times.add(time_here - temp) fps = 1 / frame_times.get_avg() frame_numpy = draw_img(results, frame_origin, args, fps=fps) if args.real_time: cv2.imshow('Detection', frame_numpy) cv2.waitKey(1) else: video_writer.write(frame_numpy) progress = (i + 1) / num_frames * 100 progress_bar.set_val(i + 1) print(f'\rDetecting: {repr(progress_bar)} {i + 1} / {num_frames} ({progress:.2f}%) {fps:.2f} fps', end='') if not args.real_time: print(f'\n\nDone, saved in: results/videos/{name}')
def evalvideo(net:Yolact, path:str): # If the path is a digit, parse it as a webcam index is_webcam = path.isdigit() frame_count = 0 # first frame variables drawing = True start_pt = None end_pt = None coords_provided = False # if args.start_x != 0 and args.start_y !=0 and args.end_x != 0 and args.end_y != 0: # start_pt = [args.start_x, args.start_y] # end_pt = [args.end_x, args.end_y] # coords_provided = True full_frame = None if is_webcam: vid = cv2.VideoCapture(int(path)) else: vid = cv2.VideoCapture(path) if not vid.isOpened(): print('Could not open video "%s"' % path) exit(-1) net = CustomDataParallel(net).cuda() transform = torch.nn.DataParallel(FastBaseTransform()).cuda() frame_times = MovingAverage(100) print("frame_times:", frame_times) fps = 0 # The 0.8 is to account for the overhead of time.sleep frame_time_target = 1 / vid.get(cv2.CAP_PROP_FPS) print("vid.get(cv2.CAP_PROP_FPS):", vid.get(cv2.CAP_PROP_FPS)) print("frame_time_target:", frame_time_target) running = True current_pred = None pred = None def cleanup_and_exit(): print() pool.terminate() vid.release() cv2.destroyAllWindows() exit() def get_next_frame(vid, first_frame=False): if first_frame: nonlocal drawing, start_pt, end_pt, full_frame # load video here ret, frame = vid.read() # Adding Function Attached To Mouse Callback def draw(event,x,y,flags,params): nonlocal start_pt,end_pt,drawing # Left Mouse Button Down Pressed if(event==1): print("event 1 fired") if start_pt is None: start_pt = [x,y] else: end_pt = [x,y] cv2.destroyAllWindows() drawing = False print(x,y) # Making Window For The Image if not coords_provided: cv2.namedWindow("Window") # Adding Mouse CallBack Event cv2.setMouseCallback("Window",draw) cv2.imshow("Window",frame) cv2.waitKey(0) print(start_pt,end_pt) nonlocal frame_count frame_count += args.video_multiframe print('now processing', frame_count) frames = [] for _ in range(args.video_multiframe): ret, temp = vid.read() if ret == True: full_frame = temp.copy() frame = full_frame[start_pt[1]:end_pt[1], start_pt[0]:end_pt[0]].copy() frames.append(frame) return frames def transform_frame(frames): with torch.no_grad(): frames = [torch.from_numpy(frame).cuda().float() for frame in frames] return frames, transform(torch.stack(frames, 0)) def eval_network(inp): nonlocal current_pred nonlocal pred with torch.no_grad(): frames, imgs = inp if frame_count % 10 == 0 or frame_count == 1: pred = net(imgs) current_pred = make_copy(pred) return frames, current_pred def prep_frame(inp): with torch.no_grad(): frame, preds = inp processed = prep_display(preds, frame, None, None, undo_transform=False, class_color=True) full_frame[start_pt[1]:end_pt[1], start_pt[0]:end_pt[0]] = processed return full_frame frame_buffer = Queue() video_fps = 0 # All this timing code to make sure that def play_video(): nonlocal frame_buffer, running, video_fps, is_webcam video_frame_times = MovingAverage(100) print("video_frame_times:",video_frame_times) frame_time_stabilizer = frame_time_target last_time = None stabilizer_step = 0.0005 while running: frame_time_start = time.time() if not frame_buffer.empty(): next_time = time.time() if last_time is not None: video_frame_times.add(next_time - last_time) video_fps = 1 / video_frame_times.get_avg() cv2.imshow(path, frame_buffer.get()) last_time = next_time if cv2.waitKey(1) == 27: # Press Escape to close running = False buffer_size = frame_buffer.qsize() if buffer_size < args.video_multiframe: frame_time_stabilizer += stabilizer_step elif buffer_size > args.video_multiframe: frame_time_stabilizer -= stabilizer_step if frame_time_stabilizer < 0: frame_time_stabilizer = 0 new_target = frame_time_stabilizer if is_webcam else max(frame_time_stabilizer, frame_time_target) next_frame_target = max(2 * new_target - video_frame_times.get_avg(), 0) target_time = frame_time_start + next_frame_target - 0.001 # Let's just subtract a millisecond to be safe # This gives more accurate timing than if sleeping the whole amount at once while time.time() < target_time: time.sleep(0.001) extract_frame = lambda x, i: (x[0][i] if x[1][i] is None else x[0][i].to(x[1][i]['box'].device), [x[1][i]]) # Prime the network on the first frame because I do some thread unsafe things otherwise print('Initializing model... ', end='') eval_network(transform_frame(get_next_frame(vid, True))) print('Done.') # For each frame the sequence of functions it needs to go through to be processed (in reversed order) sequence = [prep_frame, eval_network, transform_frame] pool = ThreadPool(processes=len(sequence) + args.video_multiframe + 2) pool.apply_async(play_video) active_frames = [] print() while vid.isOpened() and running: start_time = time.time() # Start loading the next frames from the disk next_frames = pool.apply_async(get_next_frame, args=(vid,)) # print(next_frames.get()) # For each frame in our active processing queue, dispatch a job # for that frame using the current function in the sequence for frame in active_frames: frame['value'] = pool.apply_async(sequence[frame['idx']], args=(frame['value'],)) # For each frame whose job was the last in the sequence (i.e. for all final outputs) for frame in active_frames: if frame['idx'] == 0: frame_buffer.put(frame['value'].get()) # Remove the finished frames from the processing queue active_frames = [x for x in active_frames if x['idx'] > 0] # Finish evaluating every frame in the processing queue and advanced their position in the sequence for frame in list(reversed(active_frames)): frame['value'] = frame['value'].get() frame['idx'] -= 1 if frame['idx'] == 0: # Split this up into individual threads for prep_frame since it doesn't support batch size active_frames += [{'value': extract_frame(frame['value'], i), 'idx': 0} for i in range(1, args.video_multiframe)] frame['value'] = extract_frame(frame['value'], 0) # Finish loading in the next frames and add them to the processing queue #if len(next_frames.get())>0: active_frames.append({'value': next_frames.get(), 'idx': len(sequence)-1}) # Compute FPS frame_times.add(time.time() - start_time) fps = args.video_multiframe / frame_times.get_avg() print('\rProcessing FPS: %.2f | Video Playback FPS: %.2f | Frames in Buffer: %d ' % (fps, video_fps, frame_buffer.qsize()), end='') cleanup_and_exit()
def savevideo(net:Yolact, in_path:str, out_path:str): vid = cv2.VideoCapture(in_path) target_fps = round(vid.get(cv2.CAP_PROP_FPS)) frame_width = round(vid.get(cv2.CAP_PROP_FRAME_WIDTH)) frame_height = round(vid.get(cv2.CAP_PROP_FRAME_HEIGHT)) num_frames = round(vid.get(cv2.CAP_PROP_FRAME_COUNT)) print("target_fps:{} frame_width:{} frame_height:{} num_frames:{}".format(target_fps, frame_width, frame_height, num_frames)) out = cv2.VideoWriter(out_path, cv2.VideoWriter_fourcc(*"mp4v"), target_fps, (frame_width, frame_height)) frame_freq = 10 transform = FastBaseTransform() frame_times = MovingAverage() progress_bar = ProgressBar(30, num_frames) preds = None drawing = True start_pt = None end_pt = None # load video here ret, frame = vid.read() # Adding Function Attached To Mouse Callback def draw(event,x,y,flags,params): nonlocal start_pt,end_pt,drawing # Left Mouse Button Down Pressed if(event==1): print("event 1 fired") if start_pt is None: start_pt = [x,y] else: end_pt = [x,y] cv2.destroyAllWindows() drawing = False print(x,y) # Making Window For The Image cv2.namedWindow("Window") # Adding Mouse CallBack Event cv2.setMouseCallback("Window",draw) cv2.imshow("Window",frame) cv2.waitKey(0) print(start_pt,end_pt) try: for i in range(num_frames): timer.reset() with timer.env('Video'): # process only 10th frame # care to be taken that the first frame is read always ret, full_frame = vid.read() if ret == True: frame = full_frame[start_pt[1]:end_pt[1], start_pt[0]:end_pt[0]].copy() # print("Full_frame Shape:", full_frame.shape, "frame shape:", frame.shape) frame = torch.from_numpy(frame).cuda().float() # need to adjust for multi frame if i%frame_freq == 0: batch = transform(frame.unsqueeze(0)) preds = net(batch) current_preds = make_copy(preds) processed= prep_display(current_preds, frame, None, None, undo_transform=False, class_color=True) full_frame[start_pt[1]:end_pt[1], start_pt[0]:end_pt[0]] = processed # print("Full_frame Shape:", full_frame.shape, "frame shape:", processed.shape) out.write(full_frame) if i > 1: frame_times.add(timer.total_time()) fps = 1 / frame_times.get_avg() progress = (i+1) / num_frames * 100 progress_bar.set_val(i+1) print('\rProcessing Frames %s %6d / %6d (%5.2f%%) %5.2f fps ' % (repr(progress_bar), i+1, num_frames, progress, fps), end='') except KeyboardInterrupt: print('Stopping early.') vid.release() out.release() print()
def train(args, cfg, option, DataSet): if args.exp_name is not None: args.save_folder = os.path.join(args.save_folder, args.exp_name) args.log_folder = os.path.join(args.log_folder, args.exp_name) if not os.path.exists(args.save_folder): os.makedirs(args.save_folder, exist_ok=True) if not os.path.exists(args.log_folder): os.makedirs(args.log_folder, exist_ok=True) if True: dataset = DataSet(image_path=cfg.dataset.train_images, mask_out_ch=cfg.gt_inst_ch, info_file=cfg.dataset.train_info, option=cfg.dataset, transform=SSDAugmentation(cfg, MEANS), running_mode='train') else: dataset = DataSet(image_path=cfg.dataset.valid_images, mask_out_ch=cfg.gt_inst_ch, info_file=cfg.dataset.valid_info, option=cfg.dataset, transform=SSDAugmentation(cfg, MEANS), running_mode='train') # Parallel wraps the underlying module, but when saving and loading we don't want that dvis_net = DVIS(cfg) net = dvis_net net.train() if args.log: log = Log(cfg.name, args.log_folder, dict(args._get_kwargs()), overwrite=(args.resume is None), log_gpu_stats=args.log_gpu) # I don't use the timer during training (I use a different timing method). # Apparently there's a race condition with multiple GPUs, so disable it just to be safe. timer.disable_all() # Both of these can set args.resume to None, so do them before the check if args.resume == 'interrupt': args.resume = SavePath.get_interrupt(args.save_folder) elif args.resume == 'latest': args.resume = SavePath.get_latest(args.save_folder, cfg.name) if args.resume is not None: print('Resuming training, loading {}...'.format(args.resume)) dvis_net.load_weights(args.resume, load_firstLayer=option['model_1stLayer_en'], load_lastLayer=option['model_lastLayer_en']) if args.start_iter == -1: args.start_iter = SavePath.from_str(args.resume).iteration else: print('Initializing weights...') dvis_net.init_weights(backbone_path=args.save_folder + cfg.backbone.path) #optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum, # weight_decay=args.decay) optimizer = optim.SGD([{ 'params': net.backbone.parameters(), 'lr': args.lr * option['bkb_lr_alpha'] }, { 'params': net.fpn.parameters(), 'lr': args.lr * option['fpn_lr_alpha'] }, { 'params': net.proto_net.parameters(), 'lr': args.lr * option['proto_net_lr_alpha'] }], lr=args.lr, momentum=args.momentum, weight_decay=args.decay) criterion = LossEvaluate(option, class_weights=cfg.dataset.sem_weights) if args.batch_alloc is not None: args.batch_alloc = [int(x) for x in args.batch_alloc.split(',')] if sum(args.batch_alloc) != args.batch_size: print( 'Error: Batch allocation (%s) does not sum to batch size (%s).' % (args.batch_alloc, args.batch_size)) exit(-1) net = NetLoss(net, criterion) net = CustomDataParallel(net) if args.cuda: net = net.cuda() # Initialize everything if not cfg.freeze_bn: dvis_net.freeze_bn() # Freeze bn so we don't kill our means # loss counters loc_loss = 0 conf_loss = 0 iteration = max(args.start_iter, 0) last_time = time.time() epoch_size = len(dataset) // args.batch_size num_epochs = math.ceil(cfg.max_iter / epoch_size) # Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index step_index = 0 data_loader = data.DataLoader(dataset, args.batch_size, num_workers=args.num_workers, shuffle=False, collate_fn=detection_collate, pin_memory=True) writer = SummaryWriter(log_dir=args.log_folder) save_path = lambda epoch, iteration: SavePath( cfg.name, epoch, iteration).get_path(root=args.save_folder) time_avg = MovingAverage() loss_keys = [ 'binary', 'pi', 'l1', 'regul', 'iou', 'classify', 'eval_prec', 'eval_rec', 'eval_acc' ] vis_keys = ['preds', 'gts', 'rgb', 'wghts', 'grad'] loss_avgs = {k: MovingAverage(100) for k in loss_keys} print('Begin training!') # try-except so you can use ctrl+c to save early and stop training try: log_loss = dict() for epoch in range(num_epochs): # Resume from start_iter if (epoch + 1) * epoch_size < iteration: continue for datum in data_loader: # Stop if we've reached an epoch if we're resuming from start_iter if iteration == (epoch + 1) * epoch_size: break # Stop at the configured number of iterations even if mid-epoch if iteration == cfg.max_iter: break if iteration < 99: iteration += 1 continue # Change a config setting if we've reached the specified iteration changed = False for change in cfg.delayed_settings: if iteration >= change[0]: changed = True cfg.replace(change[1]) # Reset the loss averages because things might have changed for avg in loss_avgs: avg.reset() # If a config setting was changed, remove it from the list so we don't keep checking if changed: cfg.delayed_settings = [ x for x in cfg.delayed_settings if x[0] > iteration ] # Warm up by linearly interpolating the learning rate from some smaller value if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until: set_lr(optimizer, (args.lr - cfg.lr_warmup_init) * (iteration / cfg.lr_warmup_until) + cfg.lr_warmup_init) # Adjust the learning rate at the given iterations, but also if we resume from past that iteration while step_index < len( cfg.lr_steps ) and iteration >= cfg.lr_steps[step_index]: step_index += 1 set_lr(optimizer, args.lr * (args.gamma**step_index)) # Zero the grad to get ready to compute gradients optimizer.zero_grad() # Forward Pass + Compute loss at the same time (see CustomDataParallel and NetLoss0) ret = net(datum) # Mean here because Dataparallel and do Backprop losses = {k: ret[k].mean() for k in loss_keys if k in ret} det_loss_keys = [k for k in loss_keys if k in losses] all_loss = sum([losses[k] for k in det_loss_keys]) for k in det_loss_keys: loss_avgs[k].add(losses[k].item()) # backward and optimize if args.show_gradients == True: ret['preds_0'].retain_grad() all_loss.backward(retain_graph=True) ret['grad'] = ret['preds_0'].grad[:, 0, :, :] else: all_loss.backward( ) # Do this to free up vram even if loss is not finite if torch.isfinite(all_loss).item(): optimizer.step() ret['preds'] = torch.nn.ReLU()(ret['preds']) vis_imgs = {k: ret[k] for k in vis_keys if k in ret} cur_time = time.time() elapsed = cur_time - last_time last_time = cur_time # Exclude graph setup from the timing information if iteration != args.start_iter: time_avg.add(elapsed) if iteration % 10 == 0: eta_str = str( datetime.timedelta(seconds=(cfg.max_iter - iteration) * time_avg.get_avg())).split('.')[0] total = sum([ loss_avgs[k].get_avg() for k in det_loss_keys if 'eval' not in k ]) loss_labels = sum( [[k, loss_avgs[k].get_avg()] for k in loss_keys if k in det_loss_keys], []) print(('[%3d] %7d ||' + (' %s: %.3f |' * len(det_loss_keys)) + ' T: %.3f || ETA: %s || timer: %.3f') % tuple([epoch, iteration] + loss_labels + [total, eta_str, elapsed]), flush=True) if args.log: log_step = 50 // args.batch_size for k in det_loss_keys: if k not in log_loss: log_loss[k] = loss_avgs[k].get_avg() else: log_loss[k] += loss_avgs[k].get_avg() if iteration % log_step == log_step - 1: for k in det_loss_keys: writer.add_scalar(k + '_loss', log_loss[k] / float(log_step), iteration / log_step) log_loss[k] = 0 log_fig_step = 100 if iteration % log_fig_step == log_fig_step - 1: if 'davis' in args.dataset: vis_imgs['rgb'] = vis_imgs['rgb'][:, :3, :, :] fig = plot_tfboard_figure( cfg, vis_imgs, show_grad=args.show_gradients) writer.add_figure('prediction _ grad', fig, global_step=iteration / log_fig_step) iteration += 1 if iteration % args.save_interval == 0 and iteration != args.start_iter: if args.keep_latest: latest = SavePath.get_latest(args.save_folder, cfg.name) print('Saving state, iter:', iteration) dvis_net.save_weights(save_path(epoch, iteration)) if args.keep_latest and latest is not None: if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval: print('Deleting old save...') os.remove(latest) del ret, vis_imgs, losses # end of batch run # end of epoch except KeyboardInterrupt: if args.interrupt: print('Stopping early. Saving network...') # Delete previous copy of the interrupted network so we don't spam the weights folder SavePath.remove_interrupt(args.save_folder) writer.close() dvis_net.save_weights( save_path(epoch, repr(iteration) + '_interrupt')) exit() writer.close() dvis_net.save_weights(save_path(epoch, iteration))
def train(rank, args): if args.num_gpus > 1: multi_gpu_rescale(args) if rank == 0: if not os.path.exists(args.save_folder): os.mkdir(args.save_folder) # fix the seed for reproducibility seed = args.random_seed + rank torch.manual_seed(seed) np.random.seed(seed) random.seed(seed) # set up logger setup_logger(output=os.path.join(args.log_folder, cfg.name), distributed_rank=rank) logger = logging.getLogger("yolact.train") w = SummaryHelper(distributed_rank=rank, log_dir=os.path.join(args.log_folder, cfg.name)) w.add_text("argv", " ".join(sys.argv)) logger.info("Args: {}".format(" ".join(sys.argv))) import git with git.Repo(search_parent_directories=True) as repo: w.add_text("git_hash", repo.head.object.hexsha) logger.info("git hash: {}".format(repo.head.object.hexsha)) if args.num_gpus > 1: try: logger.info("Initializing torch.distributed backend...") dist.init_process_group(backend='nccl', init_method=args.dist_url, world_size=args.num_gpus, rank=rank) except Exception as e: logger.error("Process group URL: {}".format(args.dist_url)) raise e misc.barrier() if torch.cuda.device_count() > 1: logger.info('Multiple GPUs detected! Turning off JIT.') collate_fn = detection_collate if cfg.dataset.name == 'YouTube VIS': dataset = YoutubeVIS(image_path=cfg.dataset.train_images, info_file=cfg.dataset.train_info, configs=cfg.dataset, transform=SSDAugmentationVideo(MEANS)) if cfg.dataset.joint == 'coco': joint_dataset = COCODetection( image_path=cfg.joint_dataset.train_images, info_file=cfg.joint_dataset.train_info, transform=SSDAugmentation(MEANS)) joint_collate_fn = detection_collate if args.validation_epoch > 0: setup_eval() val_dataset = YoutubeVIS(image_path=cfg.dataset.valid_images, info_file=cfg.dataset.valid_info, configs=cfg.dataset, transform=BaseTransformVideo(MEANS)) collate_fn = collate_fn_youtube_vis elif cfg.dataset.name == 'FlyingChairs': dataset = FlyingChairs(image_path=cfg.dataset.trainval_images, info_file=cfg.dataset.trainval_info) collate_fn = collate_fn_flying_chairs else: dataset = COCODetection(image_path=cfg.dataset.train_images, info_file=cfg.dataset.train_info, transform=SSDAugmentation(MEANS)) if args.validation_epoch > 0: setup_eval() val_dataset = COCODetection(image_path=cfg.dataset.valid_images, info_file=cfg.dataset.valid_info, transform=BaseTransform(MEANS)) # Set cuda device early to avoid duplicate model in master GPU if args.cuda: torch.cuda.set_device(rank) # Parallel wraps the underlying module, but when saving and loading we don't want that yolact_net = Yolact() net = yolact_net net.train() # I don't use the timer during training (I use a different timing method). # Apparently there's a race condition with multiple GPUs. # use timer for experiments timer.disable_all() # Both of these can set args.resume to None, so do them before the check if args.resume == 'interrupt': args.resume = SavePath.get_interrupt(args.save_folder) elif args.resume == 'latest': args.resume = SavePath.get_latest(args.save_folder, cfg.name) if args.resume is not None: logger.info('Resuming training, loading {}...'.format(args.resume)) yolact_net.load_weights(args.resume, args=args) if args.start_iter == -1: args.start_iter = SavePath.from_str(args.resume).iteration else: logger.info('Initializing weights...') yolact_net.init_weights(backbone_path=args.save_folder + cfg.backbone.path) if cfg.flow.train_flow: criterion = OpticalFlowLoss() else: criterion = MultiBoxLoss(num_classes=cfg.num_classes, pos_threshold=cfg.positive_iou_threshold, neg_threshold=cfg.negative_iou_threshold, negpos_ratio=3) if args.cuda: net.cuda(rank) if misc.is_distributed_initialized(): net = nn.parallel.DistributedDataParallel( net, device_ids=[rank], output_device=rank, broadcast_buffers=False, find_unused_parameters=True) optimizer = optim.SGD(filter(lambda x: x.requires_grad, net.parameters()), lr=args.lr, momentum=args.momentum, weight_decay=args.decay) # loss counters iteration = max(args.start_iter, 0) w.set_step(iteration) last_time = time.time() epoch_size = len(dataset) // args.batch_size // args.num_gpus num_epochs = math.ceil(cfg.max_iter / epoch_size) # Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index step_index = 0 from data.sampler_utils import InfiniteSampler, build_batch_data_sampler infinite_sampler = InfiniteSampler(dataset, seed=args.random_seed, num_replicas=args.num_gpus, rank=rank, shuffle=True) train_sampler = build_batch_data_sampler(infinite_sampler, images_per_batch=args.batch_size) data_loader = data.DataLoader( dataset, num_workers=args.num_workers, collate_fn=collate_fn, multiprocessing_context="fork" if args.num_workers > 1 else None, batch_sampler=train_sampler) data_loader_iter = iter(data_loader) if cfg.dataset.joint: joint_infinite_sampler = InfiniteSampler(joint_dataset, seed=args.random_seed, num_replicas=args.num_gpus, rank=rank, shuffle=True) joint_train_sampler = build_batch_data_sampler( joint_infinite_sampler, images_per_batch=args.batch_size) joint_data_loader = data.DataLoader( joint_dataset, num_workers=args.num_workers, collate_fn=joint_collate_fn, multiprocessing_context="fork" if args.num_workers > 1 else None, batch_sampler=joint_train_sampler) joint_data_loader_iter = iter(joint_data_loader) save_path = lambda epoch, iteration: SavePath( cfg.name, epoch, iteration).get_path(root=args.save_folder) time_avg = MovingAverage() data_time_avg = MovingAverage(10) global loss_types # Forms the print order loss_avgs = {k: MovingAverage(100) for k in loss_types} def backward_and_log(prefix, net_outs, targets, masks, num_crowds, extra_loss=None): optimizer.zero_grad() out = net_outs["pred_outs"] losses = criterion(out, targets, masks, num_crowds) losses = {k: v.mean() for k, v in losses.items()} # Mean here because Dataparallel if extra_loss is not None: assert type(extra_loss) == dict losses.update(extra_loss) loss = sum([losses[k] for k in losses]) # Backprop loss.backward() # Do this to free up vram even if loss is not finite if torch.isfinite(loss).item(): optimizer.step() # Add the loss to the moving average for bookkeeping for k in losses: loss_avgs[k].add(losses[k].item()) w.add_scalar('{prefix}/{key}'.format(prefix=prefix, key=k), losses[k].item()) return losses logger.info('Begin training!') # try-except so you can use ctrl+c to save early and stop training try: for epoch in range(num_epochs): # Resume from start_iter if (epoch + 1) * epoch_size < iteration: continue while True: data_start_time = time.perf_counter() datum = next(data_loader_iter) data_end_time = time.perf_counter() data_time = data_end_time - data_start_time if iteration != args.start_iter: data_time_avg.add(data_time) # Stop if we've reached an epoch if we're resuming from start_iter if iteration == (epoch + 1) * epoch_size: break # Stop at the configured number of iterations even if mid-epoch if iteration == cfg.max_iter: break # Change a config setting if we've reached the specified iteration changed = False for change in cfg.delayed_settings: if iteration >= change[0]: changed = True cfg.replace(change[1]) # Reset the loss averages because things might have changed for avg in loss_avgs: avg.reset() # If a config setting was changed, remove it from the list so we don't keep checking if changed: cfg.delayed_settings = [ x for x in cfg.delayed_settings if x[0] > iteration ] # Warm up by linearly interpolating the learning rate from some smaller value if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until and cfg.lr_warmup_init < args.lr: set_lr(optimizer, (args.lr - cfg.lr_warmup_init) * (iteration / cfg.lr_warmup_until) + cfg.lr_warmup_init) elif cfg.lr_schedule == 'cosine': set_lr( optimizer, args.lr * ((math.cos(math.pi * iteration / cfg.max_iter) + 1.) * .5)) # Adjust the learning rate at the given iterations, but also if we resume from past that iteration while cfg.lr_schedule == 'step' and step_index < len( cfg.lr_steps ) and iteration >= cfg.lr_steps[step_index]: step_index += 1 set_lr(optimizer, args.lr * (args.gamma**step_index)) global lr w.add_scalar('meta/lr', lr) if cfg.dataset.name == "FlyingChairs": imgs_1, imgs_2, flows = prepare_flow_data(datum) net_outs = net(None, extras=(imgs_1, imgs_2)) # Compute Loss optimizer.zero_grad() losses = criterion(net_outs, flows) losses = {k: v.mean() for k, v in losses.items() } # Mean here because Dataparallel loss = sum([losses[k] for k in losses]) # Backprop loss.backward( ) # Do this to free up vram even if loss is not finite if torch.isfinite(loss).item(): optimizer.step() # Add the loss to the moving average for bookkeeping for k in losses: loss_avgs[k].add(losses[k].item()) w.add_scalar('loss/%s' % k, losses[k].item()) elif cfg.dataset.joint or not cfg.dataset.is_video: if cfg.dataset.joint: joint_datum = next(joint_data_loader_iter) # Load training data # Note, for training on multiple gpus this will use the custom replicate and gather I wrote up there images, targets, masks, num_crowds = prepare_data( joint_datum) else: images, targets, masks, num_crowds = prepare_data( datum) extras = { "backbone": "full", "interrupt": False, "moving_statistics": { "aligned_feats": [] } } net_outs = net(images, extras=extras) run_name = "joint" if cfg.dataset.joint else "compute" losses = backward_and_log(run_name, net_outs, targets, masks, num_crowds) # Forward Pass if cfg.dataset.is_video: # reference frames references = [] moving_statistics = {"aligned_feats": [], "conf_hist": []} for idx, frame in enumerate(datum[:0:-1]): images, annots = frame extras = { "backbone": "full", "interrupt": True, "keep_statistics": True, "moving_statistics": moving_statistics } with torch.no_grad(): net_outs = net(images, extras=extras) moving_statistics["feats"] = net_outs["feats"] moving_statistics["lateral"] = net_outs["lateral"] keys_to_save = ("outs_phase_1", "outs_phase_2") for key in set(net_outs.keys()) - set(keys_to_save): del net_outs[key] references.append(net_outs) # key frame with annotation, but not compute full backbone frame = datum[0] images, annots = frame frame = ( images, annots, ) images, targets, masks, num_crowds = prepare_data(frame) extras = { "backbone": "full", "interrupt": not cfg.flow.base_backward, "moving_statistics": moving_statistics } gt_net_outs = net(images, extras=extras) if cfg.flow.base_backward: losses = backward_and_log("compute", gt_net_outs, targets, masks, num_crowds) keys_to_save = ("outs_phase_1", "outs_phase_2") for key in set(gt_net_outs.keys()) - set(keys_to_save): del gt_net_outs[key] # now do the warp if len(references) > 0: reference_frame = references[0] extras = { "backbone": "partial", "moving_statistics": moving_statistics } net_outs = net(images, extras=extras) extra_loss = yolact_net.extra_loss( net_outs, gt_net_outs) losses = backward_and_log("warp", net_outs, targets, masks, num_crowds, extra_loss=extra_loss) cur_time = time.time() elapsed = cur_time - last_time last_time = cur_time w.add_scalar('meta/data_time', data_time) w.add_scalar('meta/iter_time', elapsed) # Exclude graph setup from the timing information if iteration != args.start_iter: time_avg.add(elapsed) if iteration % 10 == 0: eta_str = str( datetime.timedelta(seconds=(cfg.max_iter - iteration) * time_avg.get_avg())).split('.')[0] if torch.cuda.is_available(): max_mem_mb = torch.cuda.max_memory_allocated( ) / 1024.0 / 1024.0 # torch.cuda.reset_max_memory_allocated() else: max_mem_mb = None logger.info("""\ eta: {eta} epoch: {epoch} iter: {iter} \ {losses} {loss_total} \ time: {time} data_time: {data_time} lr: {lr} {memory}\ """.format(eta=eta_str, epoch=epoch, iter=iteration, losses=" ".join([ "{}: {:.3f}".format(k, loss_avgs[k].get_avg()) for k in losses ]), loss_total="T: {:.3f}".format( sum([loss_avgs[k].get_avg() for k in losses])), data_time="{:.3f}".format(data_time_avg.get_avg()), time="{:.3f}".format(elapsed), lr="{:.6f}".format(lr), memory="max_mem: {:.0f}M".format(max_mem_mb))) if rank == 0 and iteration % 100 == 0: if cfg.flow.train_flow: import flowiz as fz from layers.warp_utils import deform_op tgt_size = (64, 64) flow_size = flows.size()[2:] vis_data = [] for pred_flow in net_outs: vis_data.append(pred_flow) deform_gt = deform_op(imgs_2, flows) flows_pred = [ F.interpolate(x, size=flow_size, mode='bilinear', align_corners=False) for x in net_outs ] deform_preds = [ deform_op(imgs_2, x) for x in flows_pred ] vis_data.append( F.interpolate(flows, size=tgt_size, mode='area')) vis_data = [ F.interpolate(flow[:1], size=tgt_size) for flow in vis_data ] vis_data = [ fz.convert_from_flow( flow[0].data.cpu().numpy().transpose( 1, 2, 0)).transpose( 2, 0, 1).astype('float32') / 255 for flow in vis_data ] def convert_image(image): image = F.interpolate(image, size=tgt_size, mode='area') image = image[0] image = image.data.cpu().numpy() image = image[::-1] image = image.transpose(1, 2, 0) image = image * np.array(STD) + np.array(MEANS) image = image.transpose(2, 0, 1) image = image / 255 image = np.clip(image, -1, 1) image = image[::-1] return image vis_data.append(convert_image(imgs_1)) vis_data.append(convert_image(imgs_2)) vis_data.append(convert_image(deform_gt)) vis_data.extend( [convert_image(x) for x in deform_preds]) vis_data_stack = np.stack(vis_data, axis=0) w.add_images("preds_flow", vis_data_stack) elif cfg.flow.warp_mode == "flow": import flowiz as fz tgt_size = (64, 64) vis_data = [] for pred_flow, _, _ in net_outs["preds_flow"]: vis_data.append(pred_flow) vis_data = [ F.interpolate(flow[:1], size=tgt_size) for flow in vis_data ] vis_data = [ fz.convert_from_flow( flow[0].data.cpu().numpy().transpose( 1, 2, 0)).transpose( 2, 0, 1).astype('float32') / 255 for flow in vis_data ] input_image = F.interpolate(images, size=tgt_size, mode='area') input_image = input_image[0] input_image = input_image.data.cpu().numpy() input_image = input_image.transpose(1, 2, 0) input_image = input_image * np.array( STD[::-1]) + np.array(MEANS[::-1]) input_image = input_image.transpose(2, 0, 1) input_image = input_image / 255 input_image = np.clip(input_image, -1, 1) vis_data.append(input_image) vis_data_stack = np.stack(vis_data, axis=0) w.add_images("preds_flow", vis_data_stack) iteration += 1 w.set_step(iteration) if rank == 0 and iteration % args.save_interval == 0 and iteration != args.start_iter: if args.keep_latest: latest = SavePath.get_latest(args.save_folder, cfg.name) logger.info('Saving state, iter: {}'.format(iteration)) yolact_net.save_weights(save_path(epoch, iteration)) if args.keep_latest and latest is not None: if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval: logger.info('Deleting old save...') os.remove(latest) misc.barrier() # This is done per epoch if args.validation_epoch > 0: if epoch % args.validation_epoch == 0 and epoch > 0: if rank == 0: compute_validation_map(yolact_net, val_dataset) misc.barrier() except KeyboardInterrupt: misc.barrier() if args.interrupt_no_save: logger.info('No save on interrupt, just exiting...') elif rank == 0: print('Stopping early. Saving network...') # Delete previous copy of the interrupted network so we don't spam the weights folder SavePath.remove_interrupt(args.save_folder) yolact_net.save_weights( save_path(epoch, repr(iteration) + '_interrupt')) return if rank == 0: yolact_net.save_weights(save_path(epoch, iteration))
def evaluate(net, dataset, max_num=-1, during_training=False, benchmark=False, cocoapi=False, traditional_nms=False): frame_times = MovingAverage() dataset_size = len(dataset) if max_num < 0 else min(max_num, len(dataset)) dataset_indices = list(range(len(dataset))) dataset_indices = dataset_indices[:dataset_size] progress_bar = ProgressBar(40, dataset_size) if benchmark: timer.disable('Data loading') else: # For each class and iou, stores tuples (score, isPositive) # Index ap_data[type][iouIdx][classIdx] ap_data = { 'box': [[APDataObject() for _ in cfg.dataset.class_names] for _ in iou_thresholds], 'mask': [[APDataObject() for _ in cfg.dataset.class_names] for _ in iou_thresholds] } make_json = Make_json() for i, image_idx in enumerate(dataset_indices): timer.reset() with timer.env('Data loading'): img, gt, gt_masks, h, w, num_crowd = dataset.pull_item(image_idx) batch = Variable(img.unsqueeze(0)) if cuda: batch = batch.cuda() with timer.env('Network forward'): net_outs = net(batch) nms_outs = NMS(net_outs, traditional_nms) if benchmark: prep_benchmark(nms_outs, h, w) else: prep_metrics(ap_data, nms_outs, gt, gt_masks, h, w, num_crowd, dataset.ids[image_idx], make_json, cocoapi) # First couple of images take longer because we're constructing the graph. # Since that's technically initialization, don't include those in the FPS calculations. fps = 0 if i > 1 and not during_training: frame_times.add(timer.total_time()) fps = 1 / frame_times.get_avg() progress = (i + 1) / dataset_size * 100 progress_bar.set_val(i + 1) print('\rProcessing: %s %d / %d (%.2f%%) %.2f fps ' % (repr(progress_bar), i + 1, dataset_size, progress, fps), end='') if benchmark: print('\n\nStats for the last frame:') timer.print_stats() avg_seconds = frame_times.get_avg() print('Average: %5.2f fps, %5.2f ms' % (1 / frame_times.get_avg(), 1000 * avg_seconds)) else: if cocoapi: make_json.dump() print(f'\nJson files dumped, saved in: {json_path}.') return table = calc_map(ap_data) print(table) return table
def evaluate(net: Yolact, dataset, train_mode=False): net.detect.use_fast_nms = args.fast_nms cfg.mask_proto_debug = args.mask_proto_debug if args.image is not None: if ':' in args.image: inp, out = args.image.split(':') evalimage(net, inp, out) else: evalimage(net, args.image) return elif args.images is not None: inp, out = args.images.split(':') evalimages(net, inp, out) return elif args.video is not None: if ':' in args.video: inp, out = args.video.split(':') savevideo(net, inp, out) else: evalvideo(net, args.video) return frame_times = MovingAverage() dataset_size = len(dataset) if args.max_images < 0 else min( args.max_images, len(dataset)) progress_bar = ProgressBar(30, dataset_size) print() if not args.display and not args.benchmark: # For each class and iou, stores tuples (score, isPositive) # Index ap_data[type][iouIdx][classIdx] ap_data = { 'box': [[APDataObject() for _ in cfg.dataset.class_names] for _ in iou_thresholds], 'mask': [[APDataObject() for _ in cfg.dataset.class_names] for _ in iou_thresholds] } detections = Detections() else: timer.disable('Load Data') dataset_indices = list(range(len(dataset))) if args.shuffle: random.shuffle(dataset_indices) elif not args.no_sort: # Do a deterministic shuffle based on the image ids # # I do this because on python 3.5 dictionary key order is *random*, while in 3.6 it's # the order of insertion. That means on python 3.6, the images come in the order they are in # in the annotations file. For some reason, the first images in the annotations file are # the hardest. To combat this, I use a hard-coded hash function based on the image ids # to shuffle the indices we use. That way, no matter what python version or how pycocotools # handles the data, we get the same result every time. hashed = [badhash(x) for x in dataset.ids] dataset_indices.sort(key=lambda x: hashed[x]) dataset_indices = dataset_indices[:dataset_size] try: # Main eval loop for it, image_idx in enumerate(dataset_indices): timer.reset() with timer.env('Load Data'): img, gt, gt_masks, h, w, num_crowd = dataset.pull_item( image_idx) # Test flag, do not upvote if cfg.mask_proto_debug: with open('scripts/info.txt', 'w') as f: f.write(str(dataset.ids[image_idx])) np.save('scripts/gt.npy', gt_masks) batch = Variable(img.unsqueeze(0)) if args.cuda: batch = batch with timer.env('Network Extra'): preds = net(batch) # Perform the meat of the operation here depending on our mode. if args.display: img_numpy = prep_display(preds, img, h, w) elif args.benchmark: prep_benchmark(preds, h, w) else: prep_metrics(ap_data, preds, img, gt, gt_masks, h, w, num_crowd, dataset.ids[image_idx], detections) # First couple of images take longer because we're constructing the graph. # Since that's technically initialization, don't include those in the FPS calculations. if it > 1: frame_times.add(timer.total_time()) if args.display: if it > 1: print('Avg FPS: %.4f' % (1 / frame_times.get_avg())) plt.imshow(img_numpy) plt.title(str(dataset.ids[image_idx])) plt.show() elif not args.no_bar: if it > 1: fps = 1 / frame_times.get_avg() else: fps = 0 progress = (it + 1) / dataset_size * 100 progress_bar.set_val(it + 1) print( '\rProcessing Images %s %6d / %6d (%5.2f%%) %5.2f fps ' % (repr(progress_bar), it + 1, dataset_size, progress, fps), end='') if not args.display and not args.benchmark: print() if args.output_coco_json: print('Dumping detections...') if args.output_web_json: detections.dump_web() else: detections.dump() else: if not train_mode: print('Saving data...') with open(args.ap_data_file, 'wb') as f: pickle.dump(ap_data, f) return calc_map(ap_data) elif args.benchmark: print() print() print('Stats for the last frame:') timer.print_stats() avg_seconds = frame_times.get_avg() print('Average: %5.2f fps, %5.2f ms' % (1 / frame_times.get_avg(), 1000 * avg_seconds)) except KeyboardInterrupt: print('Stopping...')
# GPU net = net.cuda() torch.set_default_tensor_type('torch.cuda.FloatTensor') x = torch.zeros((1, 3, cfg.max_size, cfg.max_size)) y = net(x) for p in net.prediction_layers: print(p.last_conv_size) print() for k, a in y.items(): print(k + ': ', a.size(), torch.sum(a)) exit() net(x) # timer.disable('pass2') avg = MovingAverage() try: while True: timer.reset() with timer.env('everything else'): net(x) avg.add(timer.total_time()) print('\033[2J') # Moves console cursor to 0,0 timer.print_stats() print('Avg fps: %.2f\tAvg ms: %.2f ' % (1/avg.get_avg(), avg.get_avg()*1000)) except KeyboardInterrupt: pass
def train(): if not os.path.exists(args.save_folder): os.mkdir(args.save_folder) dataset = COCODetection(image_path=cfg.dataset.train_images, info_file=cfg.dataset.train_info, transform=SSDAugmentation(MEANS)) if args.validation_epoch > 0: setup_eval() val_dataset = COCODetection(image_path=cfg.dataset.valid_images, info_file=cfg.dataset.valid_info, transform=BaseTransform(MEANS)) # Parallel wraps the underlying module, but when saving and loading we don't want that yolact_net = Yolact() net = yolact_net net.train() if args.log: log = Log(cfg.name, args.log_folder, dict(args._get_kwargs()), overwrite=(args.resume is None), log_gpu_stats=args.log_gpu) # I don't use the timer during training (I use a different timing method). # Apparently there's a race condition with multiple GPUs, so disable it just to be safe. timer.disable_all() # Both of these can set args.resume to None, so do them before the check if args.resume == 'interrupt': args.resume = SavePath.get_interrupt(args.save_folder) elif args.resume == 'latest': args.resume = SavePath.get_latest(args.save_folder, cfg.name) if args.resume is not None: print('Resuming training, loading {}...'.format(args.resume)) yolact_net.load_weights(args.resume) if args.start_iter == -1: args.start_iter = SavePath.from_str(args.resume).iteration else: print('Initializing weights...') yolact_net.init_weights(backbone_path=args.save_folder + cfg.backbone.path) optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum, weight_decay=args.decay) criterion = MultiBoxLoss(num_classes=cfg.num_classes, pos_threshold=cfg.positive_iou_threshold, neg_threshold=cfg.negative_iou_threshold, negpos_ratio=cfg.ohem_negpos_ratio) if args.batch_alloc is not None: args.batch_alloc = [int(x) for x in args.batch_alloc.split(',')] if sum(args.batch_alloc) != args.batch_size: print( 'Error: Batch allocation (%s) does not sum to batch size (%s).' % (args.batch_alloc, args.batch_size)) exit(-1) net = CustomDataParallel(NetLoss(net, criterion)) if args.cuda: net = net.cuda() # Initialize everything if not cfg.freeze_bn: yolact_net.freeze_bn() # Freeze bn so we don't kill our means yolact_net(torch.zeros(1, 3, cfg.max_size, cfg.max_size).cuda()) if not cfg.freeze_bn: yolact_net.freeze_bn(True) # loss counters loc_loss = 0 conf_loss = 0 iteration = max(args.start_iter, 0) last_time = time.time() epoch_size = len(dataset) // args.batch_size num_epochs = math.ceil(cfg.max_iter / epoch_size) # Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index step_index = 0 data_loader = data.DataLoader(dataset, args.batch_size, num_workers=args.num_workers, shuffle=True, collate_fn=detection_collate, pin_memory=True) save_path = lambda epoch, iteration: SavePath( cfg.name, epoch, iteration).get_path(root=args.save_folder) time_avg = MovingAverage() global loss_types # Forms the print order loss_avgs = {k: MovingAverage(100) for k in loss_types} print('Begin training!') print() # try-except so you can use ctrl+c to save early and stop training try: for epoch in range(num_epochs): # Resume from start_iter if (epoch + 1) * epoch_size < iteration: continue for datum in data_loader: # Stop if we've reached an epoch if we're resuming from start_iter if iteration == (epoch + 1) * epoch_size: break # Stop at the configured number of iterations even if mid-epoch if iteration == cfg.max_iter: break # Change a config setting if we've reached the specified iteration changed = False for change in cfg.delayed_settings: if iteration >= change[0]: changed = True cfg.replace(change[1]) # Reset the loss averages because things might have changed for avg in loss_avgs: avg.reset() # If a config setting was changed, remove it from the list so we don't keep checking if changed: cfg.delayed_settings = [ x for x in cfg.delayed_settings if x[0] > iteration ] # Warm up by linearly interpolating the learning rate from some smaller value if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until: set_lr(optimizer, (args.lr - cfg.lr_warmup_init) * (iteration / cfg.lr_warmup_until) + cfg.lr_warmup_init) # Adjust the learning rate at the given iterations, but also if we resume from past that iteration while step_index < len( cfg.lr_steps ) and iteration >= cfg.lr_steps[step_index]: step_index += 1 set_lr(optimizer, args.lr * (args.gamma**step_index)) # Zero the grad to get ready to compute gradients optimizer.zero_grad() # Forward Pass + Compute loss at the same time (see CustomDataParallel and NetLoss) losses = net(datum) losses = {k: (v).mean() for k, v in losses.items() } # Mean here because Dataparallel loss = sum([losses[k] for k in losses]) # no_inf_mean removes some components from the loss, so make sure to backward through all of it # all_loss = sum([v.mean() for v in losses.values()]) # Backprop loss.backward( ) # Do this to free up vram even if loss is not finite if torch.isfinite(loss).item(): optimizer.step() # Add the loss to the moving average for bookkeeping for k in losses: loss_avgs[k].add(losses[k].item()) cur_time = time.time() elapsed = cur_time - last_time last_time = cur_time # Exclude graph setup from the timing information if iteration != args.start_iter: time_avg.add(elapsed) if iteration % 10 == 0: eta_str = str( datetime.timedelta(seconds=(cfg.max_iter - iteration) * time_avg.get_avg())).split('.')[0] total = sum([loss_avgs[k].get_avg() for k in losses]) loss_labels = sum([[k, loss_avgs[k].get_avg()] for k in loss_types if k in losses], []) print(('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) + ' T: %.3f || ETA: %s || timer: %.3f') % tuple([epoch, iteration] + loss_labels + [total, eta_str, elapsed]), flush=True) if args.log: precision = 5 loss_info = { k: round(losses[k].item(), precision) for k in losses } loss_info['T'] = round(loss.item(), precision) if args.log_gpu: log.log_gpu_stats = (iteration % 10 == 0 ) # nvidia-smi is sloooow log.log('train', loss=loss_info, epoch=epoch, iter=iteration, lr=round(cur_lr, 10), elapsed=elapsed) log.log_gpu_stats = args.log_gpu iteration += 1 if iteration % args.save_interval == 0 and iteration != args.start_iter: if args.keep_latest: latest = SavePath.get_latest(args.save_folder, cfg.name) print('Saving state, iter:', iteration) yolact_net.save_weights(save_path(epoch, iteration)) if args.keep_latest and latest is not None: if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval: print('Deleting old save...') os.remove(latest) # This is done per epoch if args.validation_epoch > 0: if epoch % args.validation_epoch == 0 and epoch > 0: compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None) # Compute validation mAP after training is finished compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None) except KeyboardInterrupt: if args.interrupt: print('Stopping early. Saving network...') # Delete previous copy of the interrupted network so we don't spam the weights folder SavePath.remove_interrupt(args.save_folder) yolact_net.save_weights( save_path(epoch, repr(iteration) + '_interrupt')) exit() yolact_net.save_weights(save_path(epoch, iteration))
def train(): if not os.path.exists(args.save_folder): os.mkdir(args.save_folder) dataset = COCODetection(image_path=cfg.dataset.train_images, info_file=cfg.dataset.train_info, transform=SSDAugmentation(MEANS)) if args.validation_epoch > 0: setup_eval() val_dataset = COCODetection(image_path=cfg.dataset.valid_images, info_file=cfg.dataset.valid_info, transform=BaseTransform(MEANS)) # Parallel wraps the underlying module, but when saving and loading we don't want that yolact_net = Yolact() net = yolact_net net.train() print('\n--- Generator created! ---') # NOTE # I maunally set the original image size and seg size as 138 # might change in the future, for example 550 if cfg.pred_seg: dis_size = 138 dis_net = Discriminator_Dcgan(i_size=dis_size, s_size=dis_size) # set the dis net's initial parameter values dis_net.apply(gan_init) dis_net.train() print('--- Discriminator created! ---\n') if args.log: log = Log(cfg.name, args.log_folder, dict(args._get_kwargs()), overwrite=(args.resume is None), log_gpu_stats=args.log_gpu) # I don't use the timer during training (I use a different timing method). # Apparently there's a race condition with multiple GPUs, so disable it just to be safe. timer.disable_all() # Both of these can set args.resume to None, so do them before the check if args.resume == 'interrupt': args.resume = SavePath.get_interrupt(args.save_folder) elif args.resume == 'latest': args.resume = SavePath.get_latest(args.save_folder, cfg.name) if args.resume is not None: print('Resuming training, loading {}...'.format(args.resume)) yolact_net.load_weights(args.resume) if args.start_iter == -1: args.start_iter = SavePath.from_str(args.resume).iteration else: print('Initializing weights...') yolact_net.init_weights(backbone_path=args.save_folder + cfg.backbone.path) # optimizer_gen = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum, # weight_decay=args.decay) # if cfg.pred_seg: # optimizer_dis = optim.SGD(dis_net.parameters(), lr=cfg.dis_lr, momentum=args.momentum, # weight_decay=args.decay) # schedule_dis = ReduceLROnPlateau(optimizer_dis, mode = 'min', patience=6, min_lr=1E-6) # NOTE: Using the Ranger Optimizer for the generator optimizer_gen = Ranger(net.parameters(), lr=args.lr, weight_decay=args.decay) if cfg.pred_seg: optimizer_dis = optim.SGD(dis_net.parameters(), lr=cfg.dis_lr) schedule_dis = ReduceLROnPlateau(optimizer_dis, mode='min', patience=6, min_lr=1E-6) criterion = MultiBoxLoss(num_classes=cfg.num_classes, pos_threshold=cfg.positive_iou_threshold, neg_threshold=cfg.negative_iou_threshold, negpos_ratio=cfg.ohem_negpos_ratio, pred_seg=cfg.pred_seg) criterion_dis = nn.BCELoss() if args.batch_alloc is not None: args.batch_alloc = [int(x) for x in args.batch_alloc.split(',')] if sum(args.batch_alloc) != args.batch_size: print( 'Error: Batch allocation (%s) does not sum to batch size (%s).' % (args.batch_alloc, args.batch_size)) exit(-1) net = CustomDataParallel(NetLoss(net, criterion, pred_seg=cfg.pred_seg)) if args.cuda: net = net.cuda() if cfg.pred_seg: dis_net = dis_net.cuda() # Initialize everything if not cfg.freeze_bn: yolact_net.freeze_bn() # Freeze bn so we don't kill our means yolact_net(torch.zeros(1, 3, cfg.max_size, cfg.max_size).cuda()) if not cfg.freeze_bn: yolact_net.freeze_bn(True) # loss counters loc_loss = 0 conf_loss = 0 iteration = max(args.start_iter, 0) last_time = time.time() epoch_size = len(dataset) // args.batch_size num_epochs = math.ceil(cfg.max_iter / epoch_size) # Which learning rate adjustment step are we on? lr' = lr * gamma ^ step_index step_index = 0 data_loader = data.DataLoader(dataset, args.batch_size, num_workers=args.num_workers, shuffle=True, collate_fn=detection_collate, pin_memory=True) # NOTE val_loader = data.DataLoader(val_dataset, args.batch_size, num_workers=args.num_workers * 2, shuffle=True, collate_fn=detection_collate, pin_memory=True) save_path = lambda epoch, iteration: SavePath( cfg.name, epoch, iteration).get_path(root=args.save_folder) time_avg = MovingAverage() global loss_types # Forms the print order # TODO: global command can modify global variable inside of the function. loss_avgs = {k: MovingAverage(100) for k in loss_types} print('Begin training!') print() # try-except so you can use ctrl+c to save early and stop training try: for epoch in range(num_epochs): # Resume from start_iter if (epoch + 1) * epoch_size < iteration: continue for datum in data_loader: # Stop if we've reached an epoch if we're resuming from start_iter if iteration == (epoch + 1) * epoch_size: break # Stop at the configured number of iterations even if mid-epoch if iteration == cfg.max_iter: break # Change a config setting if we've reached the specified iteration changed = False for change in cfg.delayed_settings: if iteration >= change[0]: changed = True cfg.replace(change[1]) # Reset the loss averages because things might have changed for avg in loss_avgs: avg.reset() # If a config setting was changed, remove it from the list so we don't keep checking if changed: cfg.delayed_settings = [ x for x in cfg.delayed_settings if x[0] > iteration ] # Warm up by linearly interpolating the learning rate from some smaller value if cfg.lr_warmup_until > 0 and iteration <= cfg.lr_warmup_until: set_lr(optimizer_gen, (args.lr - cfg.lr_warmup_init) * (iteration / cfg.lr_warmup_until) + cfg.lr_warmup_init) # Adjust the learning rate at the given iterations, but also if we resume from past that iteration while step_index < len( cfg.lr_steps ) and iteration >= cfg.lr_steps[step_index]: step_index += 1 set_lr(optimizer_gen, args.lr * (args.gamma**step_index)) # Zero the grad to get ready to compute gradients optimizer_gen.zero_grad() # NOTE if cfg.pred_seg: # ====== GAN Train ====== optimizer_dis.zero_grad() dis_net.zero_grad() # train the gen and dis in different iteration count it_alter_period = iteration % (cfg.gen_iter + cfg.dis_iter) if it_alter_period >= cfg.gen_iter: freeze_pretrain(yolact_net, freeze=True) freeze_pretrain(net, freeze=True) freeze_pretrain(dis_net, freeze=False) if it_alter_period == (cfg.gen_iter + 1): print('--- Generator freeze ---') print('--- Discriminator training ---') # ----- Discriminator part ----- # seg_list is the prediction mask # can be regarded as generated images from YOLACT # pred_list is the prediction label # seg_list dim: list of (138,138,instances) # pred_list dim: list of (instances) losses, seg_list, pred_list = net(datum) seg_clas, mask_clas, b, seg_size = seg_mask_clas( seg_list, pred_list, datum) # input image size is [b, 3, 550, 550] # downsample to [b, 3, seg_h, seg_w] image = interpolate(torch.stack(datum[0]), size=seg_size, mode='bilinear', align_corners=False) # Because in the discriminator training, we do not # want the gradient flow back to the generator part # we detach seg_clas (mask_clas come the data, does not have grad) seg_input = seg_clas.clone().detach() output_pred = dis_net(img=image, seg=seg_input) output_grou = dis_net(img=image, seg=mask_clas) if iteration % (cfg.gen_iter + cfg.dis_iter) == 0: print( f'Probability of fake is fake: {output_pred.mean().item():.2f}' ) print( f'Probability of real is real: {output_grou.mean().item():.2f}' ) # 0 for Fake/Generated # 1 for True/Ground Truth fake_label = torch.zeros(b) real_label = torch.ones(b) # Advice of practical implementation # from https://arxiv.org/abs/1611.08408 # loss_pred = -criterion_dis(output_pred,target=real_label) loss_pred = criterion_dis(output_pred, target=fake_label) loss_grou = criterion_dis(output_grou, target=real_label) loss_dis = loss_pred + loss_grou # TODO: Grid Search this one lambda_dis = cfg.lambda_dis loss_dis = lambda_dis * loss_dis # Backprop of the discriminator loss_dis.backward() optimizer_dis.step() else: freeze_pretrain(yolact_net, freeze=False) freeze_pretrain(net, freeze=False) freeze_pretrain(dis_net, freeze=True) if it_alter_period == 0: print('--- Generator training ---') print('--- Discriminator freeze ---') # ----- Generator part ----- net.zero_grad() losses, seg_list, pred_list = net(datum) seg_clas, mask_clas, b, seg_size = seg_mask_clas( seg_list, pred_list, datum) image = interpolate(torch.stack(datum[0]), size=seg_size, mode='bilinear', align_corners=False) # Perform forward pass of all-fake batch through D # NOTE that seg_clas CANNOT detach, in order to flow the # gradient back to the generator output = dis_net(img=image, seg=seg_clas) # Since the log(1-D(G(x))) not provide sufficient gradients # We want log(D(G(x)) instead, this can be achieve by # use the real_label as target. # This step is crucial for the information of discriminator # to go into the generator. # Calculate G's loss based on this output real_label = torch.ones(b) loss_gen = criterion_dis(output, target=real_label) if not it_alter_period >= cfg.gen_iter: # since the dis is already freeze, the gradients will only # record the YOLACT loss_gen.backward() # Do this to free up vram even if loss is not finite losses = { k: (v).mean() for k, v in losses.items() } # Mean here because Dataparallel loss = sum([losses[k] for k in losses]) if torch.isfinite(loss).item(): # since the optimizer_gen is for YOLACT only # only the gen will be updated optimizer_gen.step() else: # ====== Normal YOLACT Train ====== # Forward Pass + Compute loss at the same time (see CustomDataParallel and NetLoss) losses = net(datum) losses = {k: (v).mean() for k, v in losses.items() } # Mean here because Dataparallel loss = sum([losses[k] for k in losses]) # no_inf_mean removes some components from the loss, so make sure to backward through all of it # all_loss = sum([v.mean() for v in losses.values()]) # Backprop loss.backward( ) # Do this to free up vram even if loss is not finite if torch.isfinite(loss).item(): optimizer_gen.step() # Add the loss to the moving average for bookkeeping for k in losses: loss_avgs[k].add(losses[k].item()) cur_time = time.time() elapsed = cur_time - last_time last_time = cur_time # Exclude graph setup from the timing information if iteration != args.start_iter: time_avg.add(elapsed) if iteration % 10 == 0: eta_str = str( datetime.timedelta(seconds=(cfg.max_iter - iteration) * time_avg.get_avg())).split('.')[0] total = sum([loss_avgs[k].get_avg() for k in losses]) loss_labels = sum([[k, loss_avgs[k].get_avg()] for k in loss_types if k in losses], []) if cfg.pred_seg and (it_alter_period >= cfg.gen_iter): print( ('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) + ' T: %.3f || Dis: %.2f || ETA: %s || timer: %.3f') % tuple([epoch, iteration] + loss_labels + [total, loss_dis, eta_str, elapsed]), flush=True) else: print(('[%3d] %7d ||' + (' %s: %.3f |' * len(losses)) + ' T: %.3f || ETA: %s || timer: %.3f') % tuple([epoch, iteration] + loss_labels + [total, eta_str, elapsed]), flush=True) # Loss Key: # - B: Box Localization Loss # - C: Class Confidence Loss # - M: Mask Loss # - P: Prototype Loss # - D: Coefficient Diversity Loss # - E: Class Existence Loss # - S: Semantic Segmentation Loss # - T: Total loss # -Dis:Discriminator Loss if args.log: precision = 5 loss_info = { k: round(losses[k].item(), precision) for k in losses } loss_info['T'] = round(loss.item(), precision) if args.log_gpu: log.log_gpu_stats = (iteration % 10 == 0 ) # nvidia-smi is sloooow log.log('train', loss=loss_info, epoch=epoch, iter=iteration, lr=round(cur_lr, 10), elapsed=elapsed) log.log_gpu_stats = args.log_gpu iteration += 1 if iteration % args.save_interval == 0 and iteration != args.start_iter: if args.keep_latest: latest = SavePath.get_latest(args.save_folder, cfg.name) print('Saving state, iter:', iteration) yolact_net.save_weights(save_path(epoch, iteration)) if args.keep_latest and latest is not None: if args.keep_latest_interval <= 0 or iteration % args.keep_latest_interval != args.save_interval: print('Deleting old save...') os.remove(latest) # This is done per epoch if args.validation_epoch > 0: # NOTE: Validation loss # if cfg.pred_seg: # net.eval() # dis_net.eval() # cfg.gan_eval = True # with torch.no_grad(): # for datum in tqdm(val_loader, desc='GAN Validation'): # losses, seg_list, pred_list = net(datum) # losses, seg_list, pred_list = net(datum) # # TODO: warp below as a function # seg_list = [v.permute(2,1,0).contiguous() for v in seg_list] # b = len(seg_list) # batch size # _, seg_h, seg_w = seg_list[0].size() # seg_clas = torch.zeros(b, cfg.num_classes-1, seg_h, seg_w) # mask_clas = torch.zeros(b, cfg.num_classes-1, seg_h, seg_w) # target_list = [target for target in datum[1][0]] # mask_list = [interpolate(mask.unsqueeze(0), size = (seg_h,seg_w),mode='bilinear', \ # align_corners=False).squeeze() for mask in datum[1][1]] # for idx in range(b): # for i, (pred, i_target) in enumerate(zip(pred_list[idx], target_list[idx])): # seg_clas[idx, pred, ...] += seg_list[idx][i,...] # mask_clas[idx, i_target[-1].long(), ...] += mask_list[idx][i,...] # seg_clas = torch.clamp(seg_clas, 0, 1) # image = interpolate(torch.stack(datum[0]), size = (seg_h,seg_w), # mode='bilinear',align_corners=False) # real_label = torch.ones(b) # output_pred = dis_net(img = image, seg = seg_clas) # output_grou = dis_net(img = image, seg = mask_clas) # loss_pred = -criterion_dis(output_pred,target=real_label) # loss_grou = criterion_dis(output_grou,target=real_label) # loss_dis = loss_pred + loss_grou # losses = { k: (v).mean() for k,v in losses.items() } # loss = sum([losses[k] for k in losses]) # val_loss = loss - cfg.lambda_dis*loss_dis # schedule_dis.step(loss_dis) # lr = [group['lr'] for group in optimizer_dis.param_groups] # print(f'Discriminator lr: {lr[0]}') # net.train() if epoch % args.validation_epoch == 0 and epoch > 0: cfg.gan_eval = False dis_net.eval() compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None) # Compute validation mAP after training is finished compute_validation_map(epoch, iteration, yolact_net, val_dataset, log if args.log else None) except KeyboardInterrupt: if args.interrupt: print('Stopping early. Saving network...') # Delete previous copy of the interrupted network so we don't spam the weights folder SavePath.remove_interrupt(args.save_folder) yolact_net.save_weights( save_path(epoch, repr(iteration) + '_interrupt')) exit() yolact_net.save_weights(save_path(epoch, iteration))