def run_live(args):
print(args)
args['text_language'] = args['language']
args.pop('language', None)
# load args
with open(os.path.join(MM_CODES_PATH, 'config.json'), 'r') as fp:
config = json.load(fp)
model_args = config["MODELS"][args['modelName']]['args'][
args['datasetName']]
dataset_args = config["DATASETS"][args['datasetName']]
if "need_data_aligned" in model_args and model_args['need_data_aligned']:
dataset_args = dataset_args['aligned']
else:
dataset_args = dataset_args['unaligned']
# load args
args = Storage(dict(**model_args, **dataset_args, **args))
args.device = 'cpu'
print(args)
# load model
setup_seed(args.seed)
model = AMIO(args)
pretrained_model_path = os.path.join(MODEL_TMP_SAVE,
args.pre_trained_model)
# if os.path.exists(pretrained_model_path):
model.load_state_dict(torch.load(pretrained_model_path))
# data pre
dp = MLive(args.live_working_dir, args.transcript, args.text_language)
dp.dataPre()
text, audio, video = dp.getEmbeddings(args.seq_lens, args.feature_dims)
text = torch.Tensor(text).unsqueeze(0)
audio = torch.Tensor(audio).unsqueeze(0)
video = torch.Tensor(video).unsqueeze(0)
print(text.size(), audio.size(), video.size())
if args.need_normalized:
audio = torch.mean(audio, dim=1, keepdims=True)
video = torch.mean(video, dim=1, keepdims=True)
# predict
model.eval()
with torch.no_grad():
outputs = model(text, audio, video)
if 'tasks' not in args:
args.tasks = 'M'
annotation_dict = {v: k for k, v in args.annotations.items()}
ret = {}
for m in args.tasks:
cur_output = outputs[m].detach().squeeze().numpy()
cur_output = np.clip(cur_output, 1e-8, 10)
cur_output = softmax(cur_output)
sentiment = np.argmax(cur_output)
# json cannot serialize float32
probs = {
annotation_dict[i]: str(round(v, 4))
for i, v in enumerate(cur_output)
}
ret[m] = {'model': args.modelName, "probs": probs}
print(ret)
return ret
"""
测试MULT中使用的ATTENTION机制的特点，是否是强变强、其它变弱？
"""
import torch
from config.config_run import Config
from data.load_data import MMDataLoader
from models.AMIO import AMIO
from run import parse_args
from trains.ATIO import ATIO
model_path = "/home/zhuchuanbo/paper_code/results/model_saves/mult-sims-M.pth"
# 进行参数配置
configs = Config(parse_args()).get_config()
device = torch.device('cuda:%d' % configs.gpu_ids[0])
configs.device = device
# 定义并且加载模型
dataloader = MMDataLoader(configs)
model = AMIO(configs).to(device)
model.load_state_dict(torch.load(model_path))
model.eval()
atio = ATIO().get_train(configs)
results = atio.do_test(model, dataloader['test'], mode="TEST")
import torchvision.models as models