MNIST Trainer bug fix

[longforu]

Potential bug fix for MNIST trainer.

When cloning the repo and running it on the latest numpy, this error occurs:

Traceback (most recent call last):
  File "/Users/longtran/Programs/teenygrad/mnist.py", line 90, in <module>
    train(model, X_train, Y_train, optimizer, steps=100)
    ~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/longtran/Programs/teenygrad/mnist.py", line 22, in train
    loss = lossfn(out, y)
  File "/Users/longtran/Programs/teenygrad/mnist.py", line 10, in <lambda>
    def train(model, X_train, Y_train, optim, steps, BS=128, lossfn=lambda out,y: out.sparse_categorical_crossentropy(y),
                                                                                  ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~^^^
  File "/Users/longtran/Programs/teenygrad/teenygrad/tensor.py", line 794, in sparse_categorical_crossentropy
    loss_mask = Y != ignore_index
                ^^^^^^^^^^^^^^^^^
  File "/Users/longtran/Programs/teenygrad/teenygrad/tensor.py", line 753, in __ne__
    def __ne__(self, x) -> Tensor: return (self<x) + (self>x)   # type: ignore
                                           ^^^^^^
  File "/Users/longtran/Programs/teenygrad/teenygrad/tensor.py", line 749, in __lt__
    def __lt__(self, x) -> Tensor: return mlops.Less.apply(*self._broadcasted(x, False))
                                                            ~~~~~~~~~~~~~~~~~^^^^^^^^^^
  File "/Users/longtran/Programs/teenygrad/teenygrad/tensor.py", line 661, in _broadcasted
    y = Tensor(y, device=self.device, requires_grad=False, dtype=self.dtype if self.dtype != dtypes.bool and self.dtype.__class__ is not ImageDType else dtypes.float32)
  File "/Users/longtran/Programs/teenygrad/teenygrad/tensor.py", line 62, in __init__
    data = LazyBuffer.loadop(LoadOps.CONST, tuple(), dtype or Tensor.default_type, device, data)
  File "/Users/longtran/Programs/teenygrad/teenygrad/lazy.py", line 35, in loadop
    elif op == LoadOps.CONST: return LazyBuffer(np.full(shape, arg, dtype=dtype.np))
                                                ~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/Users/longtran/Programs/teenygrad/env/lib/python3.13/site-packages/numpy/_core/numeric.py", line 387, in full
    multiarray.copyto(a, fill_value, casting='unsafe')
    ~~~~~~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
OverflowError: Python integer -1 out of bounds for uint8

This is because there is a comparison between an unsigned int type tensor with -1 which cannot be broadcasted into a signed int. This can be fixed with a cast.

The other change is to the loss function. Cross entropy loss should be - logprob. Without this change the trainer did not converge for me.

[syzygy137]

I got this error too

[0danylo]

This fixed it for me as well. Could help for people looking at this repo.