[FIXED] Pytorch: RuntimeError: The size of tensor a (2) must match the size of tensor b (4) at non-singleton dimension 1

Issue

The pytorch program needs to return the result of the "Rock, Paper, Scissors" game. Input is given as one-hot tensors: [ [1,0,0], [0,1,0] ] ([1,0,0]-rock, [0,1,0] - scissors) output must be: [1, 0] (first player win, second player lose). What's wrong with this code?

import torch
from torch import nn
import torch.utils.data as data
torch.manual_seed(42)

input = [[1, 0, 0], [0, 1, 0]], [[1, 0, 0], [0, 0, 1]], [[0, 1, 0], [0, 1, 0]], [[0, 0, 1], [1, 0, 0]]
input = torch.tensor(input, dtype=torch.float32)
result = [[1, 0], [0, 1], [1, 1], [1, 0]]
result = torch.tensor(result, dtype=torch.float32)

class LinearRegression(nn.Module):
    def __init__(self, num_inputs, num_outputs):
        super().__init__()
        self.linear = nn.Linear(num_inputs, num_outputs)
        self.act_fn = nn.Sigmoid()
    def forward(self, x):
        x = self.linear(x)
        x = self.act_fn(x)
        return x


model = LinearRegression(num_inputs=3, num_outputs=2)
print(model)
# Training loop
for name, param in model.named_parameters():
    print(f"Parameter {name}, shape {param.shape}")

for x in input:
  print(model(x))

print(model(input))
print('####################')
model.train()
optimizer = torch.optim.SGD(model.parameters(), lr=0.3)
lossfunc = nn.MSELoss()
# Training loop
for _ in range(1000):
    res = model(input)
    res = res.squeeze(dim=1)

    loss = lossfunc(res, result)
    # print(loss)

    ## Backpropagation
   
    optimizer.zero_grad()
    loss.backward()

    optimizer.step()

print(model(input))

Solution

you have the following problem: You want your model to get as input the rock/paper/scissors of player1 and the rock/paper/scissors of player2. Since you do one-hot encoding you want to input 6 values into your neural network. 3 for player1 and 3 for player2. But you are trying to

Hey, a dense/feed-forward neural networl can only take ONE vector as input. But you are trying to pass two vectors of size three [[1, 0, 0], [0, 1, 0]]. What you can/should do instead is to concat both vectors. That means the first 3 values is the one-hot encoded result of player1 and the last 3 values are the one-hot encoded result of player2. Like this: [1, 0, 0, 0, 1, 0].

Therefore simply change the input data to this:

input = [[1, 0, 0, 0, 1, 0]], [[1, 0, 0, 0, 0, 1]], [[0, 1, 0, 0, 1, 0]], [[0, 0, 1, 1, 0, 0]]

and the model to this:

model = LinearRegression(num_inputs=6, num_outputs=2)

And it works!

Answered By - Theodor Peifer