CNN Input Output Shape

Refer to this link.

Input Shape

You need to ensure that all images are the same dimensions. Usually the images are squares with dimensions by powers of 2.

• 4D Array of Image

  • (batch_size, height, width, depth)

Displaying Input Image

from PIL import Image
import matplotlib.pyplot as plt

### Displaying image from direct path ###
image = Image.open(img_path).convert("RGB")
plt.imshow(image) 

### Displaying image from http ###
import requests 
response = requests.get(http_address)
image = Image.open(BytesIO(response.content)).convert("RGB")
plt.imshow(image) 

### Displaying multiple images in one row ###
f, axarr = plt.subplots(1,2) # plt.subplots(# of rows, # of columns)
axarr[0].imshow(Image.open(STYLE_IMG).convert("RGB")) 
axarr[1].imshow(Image.open(CONTENT_IMG).convert("RGB"))

## Could also do the above it like this ##
fig, (ax1, ax2) = plt.subplots(1, 2)
ax1.imshow(im_convert(content))
ax2.imshow(im_convert(style))

### Displaying tensor as an image ###
image = tensor.to("cpu").clone().detach()
image = image.numpy().squeeze()
image = image.transpose(1,2,0)

# These normalization values must match your respective model
image = image * np.array((0.229, 0.224, 0.225)) + np.array((0.485, 0.456, 0.406))
image = image.clip(0, 1) # now the image can be treated like an Image.open()

Batch Size

Specifying batch size in CNN

model.add(Conv2D(64, kernel_size=1, batch_input_shape=(16, 10, 10, 3)))

Specifying batch size when fitting (default batch size will be None)

model.add(Conv2D(64, kernel_size=1, input_shape=(10, 10, 3))) # This is not 3D Array! Batch_size is just set to None as default
...
model.fit(X, y, epochs=10, batch_size=16)

Depth of an image

• Number of colour channel

  • RGB: $Depth = 3$

  • Greyscale: $Depth = 1$

Output Shape

• 4D Array

  • batch_size always stays the same as input (even when Flattening)

  • The rest change based on filter, kernal size, padding, etc.

Changing dimensions of Output Shape

• Usually, Dense layers are added on top of Conv layers for classification, but Dense layers take in a 2D array

  • So you must Flatten before adding a Dense layer to change the dimensions to 2D array.