訓練 LSTM 的問題

[i923002]

在訓練 LSTM 時，python main.py --validate -c applications/FootballAction/train_proposal/configs/lstm_football.yaml -o output_dir=applications/FootballAction/checkpoints/LSTM 請問各位大神下列要如何解決？搞了好久了都沒結果

W0322 08:23:56.238795 4752 gpu_resources.cc:61] Please NOTE: device: 0, GPU Compute Capability: 8.9, Driver API Version: 12.0, Runtime API Version: 11.7 W0322 08:23:56.240795 4752 gpu_resources.cc:91] device: 0, cuDNN Version: 8.4. [03/22 08:23:56] Training in fp32 mode. Traceback (most recent call last): File "main.py", line 142, in main() File "main.py", line 138, in main profiler_options=args.profiler_options) File "D:\Paddle\PaddleVideo\paddlevideo\tasks\train.py", line 237, in train_model outputs = model(data, mode='train') File "C:\Users\Lee\anaconda3\envs\paddlevideo\lib\site-packages\paddle\fluid\dygraph\layers.py", line 1012, in call return self.forward(*inputs, kwargs) File "D:\Paddle\PaddleVideo\paddlevideo\modeling\framework\recognizers\base.py", line 49, in forward return self.train_step(data_batch) File "D:\Paddle\PaddleVideo\paddlevideo\modeling\framework\recognizers\recognizer1d.py", line 83, in train_step output_logit, output_iou = self.forward_net(imgs) File "D:\Paddle\PaddleVideo\paddlevideo\modeling\framework\recognizers\recognizer1d.py", line 72, in forward_net lstm_logit, lstm_output = self.head(imgs) File "C:\Users\Lee\anaconda3\envs\paddlevideo\lib\site-packages\paddle\fluid\dygraph\layers.py", line 1012, in call return self.forward(*inputs, *kwargs) File "D:\Paddle\PaddleVideo\paddlevideo\modeling\heads\attention_lstm_head.py", line 232, in forward lstmout, = m(inputs=inputs[i][0], sequence_length=inputs[i][1]) File "C:\Users\Lee\anaconda3\envs\paddlevideo\lib\site-packages\paddle\fluid\dygraph\layers.py", line 1012, in call return self.forward(inputs, kwargs) File "C:\Users\Lee\anaconda3\envs\paddlevideo\lib\site-packages\paddle\nn\layer\rnn.py", line 1073, in forward for _ in range(self.statecomponents) File "C:\Users\Lee\anaconda3\envs\paddlevideo\lib\site-packages\paddle\nn\layer\rnn.py", line 1073, in for in range(self.state_components) File "C:\Users\Lee\anaconda3\envs\paddlevideo\lib\site-packages\paddle\fluid\layers\tensor.py", line 909, in fill_constant_batch_size_like input_dim_idx, output_dim_idx, place) NotImplementedError: (Unimplemented) Unsupported backend Undefined when casting it to paddle place type. (at ..\paddle\phi\core\compat\convert_utils.cc:103)

[yangming915]

也是碰到这个问题，我一行行代码查下去，最后输入tensor.py中的fill_constant_batch_size_like(input, shape, dtype, value, input_dim_idx=0, output_dim_idx=0, force_cpu=False):函数的几个变量的类型和shape如下： input：<class 'paddle.Tensor'> [64, 200, 2048] dtype：f32 shape：(2, -1, 2048) value:<class 'int'> 0 place:<class 'paddle.fluid.libpaddle.CUDAPlace'>，实在看不出哪里不对，看起来都没错。若是有疑问的，这个函数的input参数为什么使用了关键字input，就是和input一样的。

[Danvae]

也是碰到这个问题，我一行行代码查下去，最后输入tensor.py中的fill_constant_batch_size_like(input, shape, dtype, value, input_dim_idx=0, output_dim_idx=0, force_cpu=False):函数的几个变量的类型和shape如下： input：<class 'paddle.Tensor'> [64, 200, 2048] dtype：f32 shape：(2, -1, 2048) value:<class 'int'> 0 place:<class 'paddle.fluid.libpaddle.CUDAPlace'>，实在看不出哪里不对，看起来都没错。若是有疑问的，这个函数的input参数为什么使用了关键字input，就是和input一样的。

同样遇到这个问题，请问你解决了吗？