The reason your Keras model isn't throwing an error when different input sizes are passed to the `Dense` layer is due to how TensorFlow's `Dense` layer operates when connected to a 4D tensor (batch, height, width, channels).
The `Dense` layer is applied independently to each spatial location (last dimension) of the input tensor. For a 4D input of shape `(batch, H, W, C)`, the `Dense` layer transforms the last dimension (`C`) to the specified number of units, preserving the spatial dimensions (`H`, `W`). The weights are shared across spatial positions, so dynamic sizes don't affect the layer's compatibility.
The `Dense` layer is preceded by `Conv2D` layers (with `padding="same"`), which maintain spatial dimensions. The `Dense` layer simply acts as a pointwise (1x1) convolution. This explains why varying time-series lengths (treated as spatial dimensions) are allowed.
Why No Error? The `Dense` layer doesn’t require fixed `H` or `W` dimensions—it only cares about the last dimension (channels). The input/output shapes are:
And unlike traditional dense layers (which require fixed input sizes), your setup leverages the `Dense` layer like a `1x1 Conv2D`, making it compatible with dynamic shapes.
This behavior is computationally equivalent to a `Conv2D` layer with kernel size `(1, 1)`.
However the model is **fully convolutional** (including the `Dense` layer’s implicit behavior), so gradients propagate correctly regardless of input spatial dimensions.