Add VGG19 Architecture Support to PyDeepFlow #121
Description
Hello,
Following the successful implementation and integration of VGG16 into PyDeepFlow, I’d like to propose extending support by adding VGG19 as the next architecture. VGG19 builds directly on VGG16, introducing three additional convolutional layers for deeper feature extraction and slightly improved representational capacity.
Key differences and requirements for VGG19:
Same overall structure as VGG16 with an input size of 224×224 RGB.
16 convolutional layers grouped in five blocks:
Blocks 1 and 2 – 2 conv layers each
Blocks 3, 4, and 5 – 4 conv layers each (instead of 3 in VGG16)
Each conv layer uses 3×3 filters, stride 1, and padding 1.
MaxPooling2D is applied after each block (2×2 window, stride 2).
The fully connected section remains identical:
Two dense layers of 4096 neurons each, followed by a softmax output layer.
Reuse existing components: ConvLayer, ReLU, MaxPooling2D, Flatten, Dense layers, and ModelCheckpoint.
Since the VGG16 infrastructure is already in place, adapting it for VGG19 should be straightforward . Implementing VGG19 would extend another standard architecture option for transfer learning and benchmarking.
Please confirm if I can proceed with this addition.
Thanks!