A tensor’s shape gives its size along each axis. For example, a tensor holding three points with two coordinates each has shape (3, 2): three rows, two values per row. In PyTorch, inspect this with `points.shape`. section 3.2.3
For images, axis order matters as much as axis size. PyTorch image tensors commonly use C × H × W: channels, height, width. An RGB image that is 32 pixels high and 32 pixels wide therefore has shape (3, 32, 32). Adding a batch puts the number of images first, giving N × C × H × W. For example, `torch.zeros(len(png_files), 3, 256, 256, dtype=torch.uint8)` makes room for a batch of 256 × 256 RGB images. section 4.1.3 section 7.1.3
Put channels first
An image array may arrive as H × W × C. Convert it with `torch.from_numpy(img_arr)`, then use `permute(2, 0, 1)` to place channels first. A resulting shape of `torch.Size([3, 720, 1280])` describes three channels, 720 rows, and 1280 columns. Alternatively, `torchvision.transforms.ToTensor()` converts a PIL image or NumPy array to a C × H × W tensor; a 32 × 32 RGB image becomes `(3, 32, 32)`. To display that tensor with Matplotlib, permute its axes back to H × W × C. section 4.1.3 section 7.1.3
Shape also helps organize non-image data. Hourly records can be grouped into days with `view(-1, 24, bikes.shape[1])`, producing (730, 24, 17). Transposing axes 1 and 2 produces (730, 17, 24): days × variables × hours. The choice keeps each variable’s daily readings together. Shape is not the whole storage story: a transpose can change strides without reallocating the underlying data. Shape also determines whether tensors can participate in broadcasting. section 4.4.2 section 3.9.1 section 3.4
# In[3]: img = torch.from_numpy(img_arr) out = img.permute(2, 0, 1) out.shape # Out[3]: torch.Size([3, 720, 1280])
The listing shows the axis permutation and its resulting C × H × W shape. section 4.1.3
# In[5]: daily_bikes = daily_bikes.transpose(1, 2) daily_bikes.shape, daily_bikes.stride() # Out[5]: (torch.Size([730, 17, 24]), (408, 1, 17))
The listing shows how transposing a daily time series changes both its shape and strides. section 4.4.2
