[FAQ Bot] NEW: How many feature maps are produced by a PyTorch Conv2D layer, and how do

_questions/machine-learning-zoomcamp/module-8/034_ad7e934f1c_conv2d-output-size-pytorch.md

@@ -0,0 +1,44 @@
+---
+id: ad7e934f1c
+question: How many feature maps are produced by a PyTorch Conv2D layer, and how do
+  you compute the spatial output size?
+sort_order: 34
+---
+
+In PyTorch, a `Conv2d` layer outputs a 4D tensor of shape `(N, C_out, H_out, W_out)` where:
+
+- `N` is the batch size
+- `C_out` is the number of output channels (out_channels)
+- `H_out` and `W_out` are the spatial dimensions after the convolution
+
+The spatial size is computed (for dilation = 1) as:
+
+```
+H_out = floor((H_in - K_H + 2*P_H) / S_H) + 1
+W_out = floor((W_in - K_W + 2*P_W) / S_W) + 1
+```
+
+For the general case (including dilation `D`):
+
+```
+H_out = floor((H_in + 2*P_H - D*(K_H - 1) - 1) / S_H + 1)
+W_out = floor((W_in + 2*P_W - D*(K_W - 1) - 1) / S_W + 1)
+```
+
+If you have dilation = 1 (the common case), this reduces to the simplified formula above.
+
+Example for your parameters (assuming input has 3 channels and batch size 1):
+
+```
+import torch
+import torch.nn as nn
+
+conv = nn.Conv2d(in_channels=3, out_channels=16, kernel_size=(2,2), stride=(2,2), padding=(2,2))
+x = torch.randn(1, 3, 150, 150)
+y = conv(x)
+print(y.shape)  # torch.Size([1, 16, 77, 77])
+```
+
+Notes:
+- The number of feature maps equals `out_channels`.
+- If you only care about spatial size, apply the height/width formula above; you can also verify with a quick forward pass or by using a library like `torchsummary` to inspect the model.