![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3197853%2F74819e0ff09d05a84e22d13b58cc88c5%2F1d-cnn-shape.png?generation=1751381120080184&alt=media\")
I wanted to share a little diagram I made which I found useful when designing 1D-CNN architectures as sometimes input and output shapes can get a bit confusing.
\nIf we have:
\n(B, C_in, L_in) (batch size, input channels, input sequence length).(B, C_out, L_out) (batch size, output channels, output sequence length).(C_out, C_in, kernel_size)Disclaimer: the diagram numerical values are for illustration purposes only
\n
When determining a 1D convolution you are in charge of designing the C_out and kernel_size. This in turn will determine the L_out size by the formula:
\n$$
\nL_{out} = \\left\\lfloor \\frac{L_{in} + 2\\cdot \\text{pad} - \\text{dilation} \\cdot (\\text{kernel size} - 1) - 1}{\\text{stride}} + 1 \\right\\rfloor
\n$$
So, essentially, I try to think about the kernel shape as having C_out number of kernels of shape (C_in, kernel_size) where each channel in the kernel \"takes care\" of each channel in the input and I need C_out number of these kernels because I want the output to have C_out channels.
You can check tensor shapes with this simple code:
\ninput_tensor = torch.randn(, , )\nB, C_in, L = input_tensor.shape\n\nC_out = \nkernel_size = \n\nconv1d = nn.Conv1d(\n in_channels=C_in,\n out_channels=C_out,\n kernel_size=kernel_size,\n padding=,\n stride=,\n dilation=\n)\noutput_tensor = conv1d(input_tensor)\n\n()\n()\n()\nAlso, I am leaving this blog which has some great animations.
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3197853%2F283aac0afb7dd4067da393c16fe104b8%2F1d-cnn-animation-1.gif?generation=1751381663094261&alt=media\")
Happy Kaggling!
", "messages": [ { "id": "3238108", "postDate": "07/01/2025 14:57:46", "content": "I wanted to share a little diagram I made which I found useful when designing 1D-CNN architectures as sometimes input and output shapes can get a bit confusing.
\nIf we have:
\n(B, C_in, L_in) (batch size, input channels, input sequence length).(B, C_out, L_out) (batch size, output channels, output sequence length).(C_out, C_in, kernel_size)Disclaimer: the diagram numerical values are for illustration purposes only
\n
When determining a 1D convolution you are in charge of designing the C_out and kernel_size. This in turn will determine the L_out size by the formula:
\n$$
\nL_{out} = \\left\\lfloor \\frac{L_{in} + 2\\cdot \\text{pad} - \\text{dilation} \\cdot (\\text{kernel size} - 1) - 1}{\\text{stride}} + 1 \\right\\rfloor
\n$$
So, essentially, I try to think about the kernel shape as having C_out number of kernels of shape (C_in, kernel_size) where each channel in the kernel \"takes care\" of each channel in the input and I need C_out number of these kernels because I want the output to have C_out channels.
You can check tensor shapes with this simple code:
\ninput_tensor = torch.randn(, , )\nB, C_in, L = input_tensor.shape\n\nC_out = \nkernel_size = \n\nconv1d = nn.Conv1d(\n in_channels=C_in,\n out_channels=C_out,\n kernel_size=kernel_size,\n padding=,\n stride=,\n dilation=\n)\noutput_tensor = conv1d(input_tensor)\n\n()\n()\n()\nAlso, I am leaving this blog which has some great animations.
\n
Happy Kaggling!
", "rawMarkdown": "I wanted to share a little diagram I made which I found useful when designing 1D-CNN architectures as sometimes input and output shapes can get a bit confusing.\n\nIf we have:\n- An input tensor of shape `(B, C_in, L_in)` (batch size, input channels, input sequence length).\n- The output will have shape `(B, C_out, L_out)` (batch size, output channels, output sequence length).\n- Then the 1D-CNN kernel will have weights of shape `(C_out, C_in, kernel_size)`\n\n*Disclaimer: the diagram numerical values are for illustration purposes only*\n\n\n\nWhen determining a 1D convolution you are in charge of designing the `C_out` and `kernel_size`. This in turn will determine the `L_out` size by the formula:\n$$\nL_{out} = \\left\\lfloor \\frac{L_{in} + 2\\cdot \\text{pad} - \\text{dilation} \\cdot (\\text{kernel size} - 1) - 1}{\\text{stride}} + 1 \\right\\rfloor\n$$\n\nSo, essentially, I try to think about the kernel shape as having `C_out` number of kernels of shape `(C_in, kernel_size)` where each channel in the kernel \"takes care\" of each channel in the input and I need `C_out` number of these kernels because I want the output to have `C_out` channels.\n\nYou can check tensor shapes with this simple code:\n```python\ninput_tensor = torch.randn(32, 8, 100)\nB, C_in, L = input_tensor.shape\n\nC_out = 12\nkernel_size = 3\n\nconv1d = nn.Conv1d(\n in_channels=C_in,\n out_channels=C_out,\n kernel_size=kernel_size,\n padding=4,\n stride=2,\n dilation=2\n)\noutput_tensor = conv1d(input_tensor)\n\nprint(f\"Input tensor shape: {input_tensor.shape}\")\nprint(f\"Output tensor shape: {output_tensor.shape}\")\nprint(f\"Kernel shape: {conv1d.weight.shape}\")\n```\n\nAlso, I am leaving [this blog](https://jinglescode.github.io/2020/11/01/how-convolutional-layers-work-deep-learning-neural-networks/) which has some great animations.\n\n\n\nHappy Kaggling!", "votes": null }, { "id": "3238144", "postDate": "07/01/2025 15:40:30", "content": "Let me add the link to my favorite explanation of convolutions: https://stackoverflow.com/a/44628011/5108365
", "rawMarkdown": "Let me add the link to my favorite explanation of convolutions: https://stackoverflow.com/a/44628011/5108365", "votes": null }, { "id": "3238208", "postDate": "07/01/2025 16:32:42", "content": "Thanks for sharing
", "rawMarkdown": "Thanks for sharing", "votes": null } ], "comments": [ { "id": 3238144, "author_name": "ogurtsov", "author_url": "", "post_date": "07/01/2025 15:40:30", "content": "Let me add the link to my favorite explanation of convolutions: https://stackoverflow.com/a/44628011/5108365
", "votes": null, "replies": [ { "id": 3238208, "author_name": "alejopaullier", "author_url": "", "post_date": "07/01/2025 16:32:42", "content": "Thanks for sharing
", "votes": null, "replies": [] } ] } ], "raw_markdown_by_id": { "3238108": "I wanted to share a little diagram I made which I found useful when designing 1D-CNN architectures as sometimes input and output shapes can get a bit confusing.\n\nIf we have:\n- An input tensor of shape `(B, C_in, L_in)` (batch size, input channels, input sequence length).\n- The output will have shape `(B, C_out, L_out)` (batch size, output channels, output sequence length).\n- Then the 1D-CNN kernel will have weights of shape `(C_out, C_in, kernel_size)`\n\n*Disclaimer: the diagram numerical values are for illustration purposes only*\n\n\n\nWhen determining a 1D convolution you are in charge of designing the `C_out` and `kernel_size`. This in turn will determine the `L_out` size by the formula:\n$$\nL_{out} = \\left\\lfloor \\frac{L_{in} + 2\\cdot \\text{pad} - \\text{dilation} \\cdot (\\text{kernel size} - 1) - 1}{\\text{stride}} + 1 \\right\\rfloor\n$$\n\nSo, essentially, I try to think about the kernel shape as having `C_out` number of kernels of shape `(C_in, kernel_size)` where each channel in the kernel \"takes care\" of each channel in the input and I need `C_out` number of these kernels because I want the output to have `C_out` channels.\n\nYou can check tensor shapes with this simple code:\n```python\ninput_tensor = torch.randn(32, 8, 100)\nB, C_in, L = input_tensor.shape\n\nC_out = 12\nkernel_size = 3\n\nconv1d = nn.Conv1d(\n in_channels=C_in,\n out_channels=C_out,\n kernel_size=kernel_size,\n padding=4,\n stride=2,\n dilation=2\n)\noutput_tensor = conv1d(input_tensor)\n\nprint(f\"Input tensor shape: {input_tensor.shape}\")\nprint(f\"Output tensor shape: {output_tensor.shape}\")\nprint(f\"Kernel shape: {conv1d.weight.shape}\")\n```\n\nAlso, I am leaving [this blog](https://jinglescode.github.io/2020/11/01/how-convolutional-layers-work-deep-learning-neural-networks/) which has some great animations.\n\n\n\nHappy Kaggling!", "3238144": "Let me add the link to my favorite explanation of convolutions: https://stackoverflow.com/a/44628011/5108365", "3238208": "Thanks for sharing" }, "source": "meta" }