Table 1 Detailed architecture of the proposed student model.
From: Knowledge distillation-based lightweight MobileNet model for diabetic retinopathy classification
Layer Type | Output Shape | Parameters |
|---|---|---|
Input Layer | (512, 512, 3) | 0 |
Conv2D (3\(\times\)3, 32 filters, stride 2) | (256, 256, 32) | 896 |
Batch Normalization | (256, 256, 32) | 128 |
ReLU Activation | (256, 256, 32) | 0 |
Block 1 (64 filters, stride 1) | ||
|---|---|---|
Depthwise Conv2D (3\(\times\)3) | (256, 256, 32) | 320 |
Batch Normalization | (256, 256, 32) | 128 |
ReLU Activation | (256, 256, 32) | 0 |
Pointwise Conv2D (1\(\times\)1, 64 filters) | (256, 256, 64) | 2,112 |
Batch Normalization | (256, 256, 64) | 256 |
ReLU Activation | (256, 256, 64) | 0 |
Block 2 (64 filters, stride 2) | ||
|---|---|---|
Depthwise Conv2D (3\(\times\)3) | (128, 128, 64) | 640 |
Batch Normalization | (128, 128, 64) | 256 |
ReLU Activation | (128, 128, 64) | 0 |
Pointwise Conv2D (1\(\times\)1, 64 filters) | (128, 128, 64) | 4,160 |
Batch Normalization | (128, 128, 64) | 256 |
ReLU Activation | (128, 128, 64) | 0 |
Block 3 (128 filters, stride 1) | ||
|---|---|---|
Depthwise Conv2D (3\(\times\)3) | (128, 128, 64) | 640 |
Batch Normalization | (128, 128, 64) | 256 |
ReLU Activation | (128, 128, 64) | 0 |
Pointwise Conv2D (1\(\times\)1, 128 filters) | (128, 128, 128) | 8,320 |
Batch Normalization | (128, 128, 128) | 512 |
ReLU Activation | (128, 128, 128) | 0 |
Block 4 (128 filters, stride 2) | ||
|---|---|---|
Depthwise Conv2D (3\(\times\)3) | (64, 64, 128) | 1,280 |
Batch Normalization | (64, 64, 128) | 512 |
ReLU Activation | (64, 64, 128) | 0 |
Pointwise Conv2D (1\(\times\)1, 128 filters) | (64, 64, 128) | 16,512 |
Batch Normalization | (64, 64, 128) | 512 |
ReLU Activation | (64, 64, 128) | 0 |
Block 5 (128 filters, stride 1) | ||
|---|---|---|
Depthwise Conv2D (3\(\times\)3) | (64, 64, 128) | 1,280 |
Batch Normalization | (64, 64, 128) | 512 |
ReLU Activation | (64, 64, 128) | 0 |
Pointwise Conv2D (1\(\times\)1, 128 filters) | (64, 64, 128) | 16,512 |
Batch Normalization | (64, 64, 128) | 512 |
ReLU Activation | (64, 64, 128) | 0 |
Fully Connected Layers | ||
|---|---|---|
Global Average Pooling 2D | (128,) | 0 |
Dense (128 units, ReLU) | (128,) | 16,512 |
Dense (2 units) | (2,) | 258 |
Dense (3 units) | (3,) | 387 |
Parameter Summary | ||
|---|---|---|
Dense (2 units) | Trainable | 71,362 |
Non-trainable | 1,920 | |
Total | 73,282 | |
Dense (3 units) | Trainable | 71,491 |
Non-trainable | 1,920 | |
Total | 73,411 | |
