RuntimeError: size mismatch, m1: [5 x 10], m2: [5 x 32] at /pytorch/aten/src/TH/generic/THTensorMath.cpp

1

I need your help. Running the code below throws:

RuntimeError: size mismatch, m1: [5 x 10], m2: [5 x 32] at /pytorch/aten/src/TH/generic/THTensorMath.cpp

I looked at similar questions but they are image related and suggest flattening the input, I tried them with no luck.

I’m using Python 3.6.8 and torch 1.1.0

**code sample:

state = [[307, 1588204935.0, 1.0869, 1.08708, 1.08659, 1.08662, 1.08708, 1.08724, 1.08674, 1.08677],
          [370, 1588204920.0, 1.08668, 1.08709, 1.08661, 1.08693, 1.08682, 1.08724, 1.08677, 1.08708],
          [243, 1588204905.0, 1.08637, 1.08671, 1.08624, 1.08669, 1.08651, 1.08686, 1.08639, 1.08683],
          [232, 1588204890.0, 1.08614, 1.08656, 1.08609, 1.08636, 1.08628, 1.0867, 1.08626, 1.0865],
          [349, 1588204875.0, 1.086, 1.08623, 1.08585, 1.08614, 1.08614, 1.08638, 1.08597, 1.0863]]

def predict(state, state_dim=5, action_dim=3, hidden_dim=32, lr=0.05):
    """ Compute Q values for all actions using the DQL. """
    criterion = torch.nn.MSELoss()
    model = torch.nn.Sequential(torch.nn.Linear(state_dim, hidden_dim),
                                     torch.nn.LeakyReLU(),
                                     torch.nn.Linear(hidden_dim, hidden_dim*2),
                                     torch.nn.LeakyReLU(),
                                     torch.nn.Linear(hidden_dim*2, action_dim))

    optimizer = torch.optim.Adam(model.parameters(), lr)

    with torch.no_grad():
        return model(torch.Tensor(state)) # Error throw here

predict(state).tolist()
2

Hi Masilive!

The short answer is that you have the dimensions of your input
swapped. You would want something like, for example, “m1: [10 x 5]”.

Pytorch will interpret state as a batch of 5 samples, each of which
is a vector of 10 values.

But the first layer of your Sequential is a Linear that is expecting an
input vector that consists of only 5 values.

Remember, pytorch models always work on batches of samples (even
if the batch size is 1). So the first dimension of state is, in fact, your
batch size.

How to fix this depends on what you want your state tensor to mean.
Is it supposed to be a batch of 5 length-10 vectors (as it is being treated
now), or is it supposed to be a batch of 10 length-5 vectors? Or
something else?

Good luck.

K. Frank

3

Hello Frank, thanks for your response. I want it to a batch of 5 length with 10 vectors.

Warm regards,

Masilive

4

Thanks Frank.
I changed my state_dim form 5 to 10

def predict(state_, state_dim=10, action_dim=3, hidden_dim=32, lr=0.05):