ResNet-50 Layer Shapes — Complete Shape Trace
ResNet-50: Input 224 → Conv7×7/s2 → 112 → MaxPool/s2 → 56 → Layer1 → 56 → Layer2 → 28 → Layer3 → 14 → Layer4 → 7 → AvgPool → 1 → FC → 1000
Complete Shape Trace
Input: (batch, 3, 224, 224)
Conv1: 7x7, s2 (batch, 64, 112, 112) # floor((224-7+2*3)/2)+1 = 112
BN + ReLU (batch, 64, 112, 112)
MaxPool: 3x3, s2 (batch, 64, 56, 56) # floor((112-3+2*1)/2)+1 = 56
Layer1 (3 blocks): (batch, 256, 56, 56) # no spatial change
Layer2 (4 blocks): (batch, 512, 28, 28) # first block has stride=2
Layer3 (6 blocks): (batch, 1024, 14, 14) # first block has stride=2
Layer4 (3 blocks): (batch, 2048, 7, 7) # first block has stride=2
AdaptiveAvgPool2d: (batch, 2048, 1, 1) # global average pooling
Flatten: (batch, 2048)
FC: (batch, 1000) # classification headBottleneck Block (Used in ResNet-50)
Each "block" in Layer1-4 is a bottleneck with three convolutions:
# Example: one block in Layer3
# Input: (batch, 512, 28, 28)
Conv1x1: reduce channels (batch, 256, 28, 28) # 1x1 conv
Conv3x3: spatial processing (batch, 256, 28, 28) # 3x3 conv
Conv1x1: expand channels (batch, 1024, 28, 28) # 1x1 conv
+ skip connection (with 1x1 projection if needed)PyTorch Verification
import torchvision.models as models
import torch
model = models.resnet50()
x = torch.randn(1, 3, 224, 224)
# Trace through each stage
x = model.conv1(x); print(f"conv1: {x.shape}") # [1, 64, 112, 112]
x = model.bn1(x); x = model.relu(x)
x = model.maxpool(x); print(f"pool: {x.shape}") # [1, 64, 56, 56]
x = model.layer1(x); print(f"layer1: {x.shape}") # [1, 256, 56, 56]
x = model.layer2(x); print(f"layer2: {x.shape}") # [1, 512, 28, 28]
x = model.layer3(x); print(f"layer3: {x.shape}") # [1, 1024, 14, 14]
x = model.layer4(x); print(f"layer4: {x.shape}") # [1, 2048, 7, 7]
x = model.avgpool(x); print(f"avgpool:{x.shape}") # [1, 2048, 1, 1]
x = torch.flatten(x, 1)
x = model.fc(x); print(f"fc: {x.shape}") # [1, 1000]