Test problem

[thiago092]

First congratulations for the great work. As shown in the picture above I'm having problems with the shape of the network.

RuntimeError: shape '[-1, 128, 1024]' is invalid for input of size 65536

Would you know to help me what is happening, since I have not changed any parameters of the network.

[HReynaud]

Hello, could you share your main.py file so I can reproduce your error ?

[thiago092]

Hi, I haven't made any changes to yours.

 Network.train(  dataset_path=dataset_path,  # path to the dataset folder containing the "Videos" foldes and "FileList.csv" file
                    num_epochs=5,               # number of epoch to train
                    device=[0,1,2,3],                 # "cpu" or gpu ids, ex [0] or [0,1] or [2] etc
                    batch_size=2,               # batch size
                    seed=0,                     # random seed for reproducibility
                    run_test=False,             # run test loop after each epoch
                    lr = 1e-5,                  # learning rate
                    modelname="UVT_repeat_reg",         # name of the folder where weight files will be stored
                    latent_dim=1024,            # embedding dimension
                    lr_step_period=3,           # number of epoch before dividing the learning rate by 10
                    ds_max_length = 128,        # maximum number of frame during training
                    ds_min_spacing = 10,        # minimum number of frame during training
                    DTmode = 'repeat',          # data preprocessing method: 'repeat' (mirroring) / 'full' (entire video) / 'sample' (single heartbeat with random amounf of additional frames)
                    SDmode = 'reg',             # SD branch network type: reg (regression) or cla (classification)
                    num_hidden_layers = 16,     # Number of Transformers
                    intermediate_size = 8192,   # size of the main MLP inside of the Transformers
                    rm_branch = None,           # select branch to not train: None, 'SD', 'EF'
                    use_conv = False,           # use convolutions instead of MLP for the regressors - worse results
                    attention_heads = 16        # number of attention heads in each Transformer
                    )
    
    # Parameters must match train-time parameters, or the weight files wont load
    Network.test(   dataset_path=dataset_path,  # Path to the dataset folder containing the "Videos" foldes and "FileList.csv" file
                    SDmode='reg',               # SD branch network type: reg (regression) or cla (classification)
                    use_full_videos=True,       # Use full video (no preprocessing other than intensity scaling)
                    latent_dim=1024,            # embedding dimension
                    num_hidden_layers=16,       # Number of Transformers
                    intermediate_size=8192,     # Size of the main MLP inside of the Transformers
                    model_path="./output/UVT_repeat_reg",# path of trained weight
                    device=[0,1,2,3]
                    )

[HReynaud]

The problem comes from the test function not supporting multiple GPUs.
In train.py, I have put the call to the test function outside the run_test condition, which leads to this error.
Please put the call to function test on line 160 in train.py, under if run_test:.
This should solve your issue while letting you use multiple gpus for training.

[thiago092]

I managed to run only with 1 gpu and batch 1, I could not run with several gpu with the modifications.