Fix mlp mnist explicit gradients (#385)

* style(mlp_mnist): remove trailing spaces

* doc(mlp_mnist): add further user output about progress

* feat(mlp_mnist): use explicit gradients

* feat(mlp_mnist): use same loss function during train process and reporting

* build(mlp_mnist): compat Flux = 0.13.9 and MLDatasets = 0.7; update Manifest

* doc(README): reflect update of mlp_mnist to use explicit gradients

* chore(mlp_mnist): ADAM -> Adam

Co-authored-by: Michael Abbott <[email protected]>

* chore(mlp_mnist): define and use const loss; aggregate loss within loss_and_accuracy by sum (instead of mean) to restore original behaviour

Co-authored-by: Michael Abbott <[email protected]>

Author: noxthot

README.md

@@ -65,7 +65,7 @@ Flux v0.13 is the latest right now, marked with ☀️; models upgraded to use
 
 **Vision**
 * MNIST
-  * [Simple multi-layer perceptron](vision/mlp_mnist) ☀️ v0.13
+  * [Simple multi-layer perceptron](vision/mlp_mnist) ☀️ v0.13 +
   * [Simple ConvNet (LeNet)](vision/conv_mnist) ☀️ v0.13
   * [Variational Auto-Encoder](vision/vae_mnist) ☀️ v0.13 +
   * [Deep Convolutional Generative Adversarial Networks](vision/dcgan_mnist) ☀️ v0.13 +