Made requested changes to descriptions.

notebooks/02_quant_activation_overview.ipynb

@@ -711,9 +711,9 @@
    "cell_type": "markdown",
    "metadata": {},
    "source": [
-    "We can see that the per-tensor scale parameter has calibrated itself to provide a full quantization range of 3, matching that of the most extreme channel. \n",
+    "We can see that the per-tensor scale parameter has calibrated itself to provide a full quantization range of 3, matching that of the channel with the largest dynamic range. \n",
     "\n",
-    "We can take a look at the `QuantReLU` object, and in particular look at what the ` scaling_impl` object is composed of. It is responsible for gathering statistics for determining the quantization parameters, and we can see that its ` stats_input_view_shape_impl` attribute is set to be an instance of `OverTensorView`. This is defined [here](https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/core/function_wrapper/shape.py#L78), and serves to flatten out the observed tensor into a 1D tensor and, in this case, use the `AbsPercentile` observer to calculate the quantization parameters during the gathering statistics stage of QAT."
+    "We can take a look at the `QuantReLU` object, and in particular look at what the `scaling_impl` object is composed of. It is responsible for gathering statistics for determining the quantization parameters, and we can see that its `stats_input_view_shape_impl` attribute is set to be an instance of `OverTensorView`. This is defined [here](https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/core/function_wrapper/shape.py#L78), and serves to flatten out the observed tensor into a 1D tensor and, in this case, use the `AbsPercentile` observer to calculate the quantization parameters during the gathering statistics stage of QAT."
    ]
   },
   {
@@ -782,7 +782,8 @@
    "source": [
     "Next, we initialise a new `QuantRelU` instance, but this time we specify that we desire per-channel quantization i.e. `scaling_per_output_channel=True`. This will implictly call `scaling_stats_input_view_shape_impl`, defined [here](https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/quant/solver/common.py#L184), and will change the `QuantReLU` from using a per-tensor view when gathering stats to a per output channel view ([`OverOutputChannelView`](https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/core/function_wrapper/shape.py#L52)). This simply permutes the tensor into a 2D tensor, with dim 0 equal to the number of output channels.\n",
     "\n",
-    "To accomplish this, we also need to give it some extra information: `scaling_stats_permute_dims` and `per_channel_broadcastable_shape`. `scaling_stats_permute_dims` is responsible for defining how we do the permutation. `per_channel_broadcastable_shape` simply represents what the dimensions of the quantization parameters will be, i.e. there should be one parameter per output channel.\n",
+    "To accomplish this, we also need to give it some extra information: `scaling_stats_permute_dims` and `per_channel_broadcastable_shape`. `scaling_stats_permute_dims` is responsible for defining how we do the permutation. `per_channel_broadcastable_shape` is necessary to understand along which dimensions the scale factor has to be broadcasted, so that the scale factor values are applied along the channel dimensions of the input.\n",
+    "By default, PyTorch will broadcast along the first rightmost dimension for which the shapes of the two tensors match. To make sure that we apply the scale factor in our desired output channel dimension, we need to tell PyTorch how to correctly broadcast the scale factors. Therefore the scale factor will have as many dimensions as the input tensors, with all the shapes equal to 1 apart from the channel dimension.\n",
     "\n",
     "Below, we can see that in the per-channel ` QuantReLU` instance, the `stats_input_view_shape_impl` is now ` OverOutputChannelView`, and uses a `PermuteDims` [instance](https://github.com/Xilinx/brevitas/blob/200456825f3b4b8db414f2b25b64311f82d3991a/src/brevitas/core/function_wrapper/shape.py#L21) to do the permutation of the tensor to, in this case, a 2D tensor. "
    ]