Chapter 5: Preprocessing for Pretrained Models

“You can’t transfer knowledge if the inputs don’t speak the same language. Preprocessing isn’t optional—it’s the protocol.”

---

Why This Chapter Matters

Pretrained CNNs—like those from torchvision.models or keras.applications—have been trained on massive datasets like ImageNet. But they were trained with specific input formats in mind:

• A fixed input shape (often 224×224)

• Normalized using specific mean and std

• RGB ordering (not grayscale or BGR)

• Scaled values in a specific range (e.g., [0, 1], [-1, 1])

If you deviate from this, even slightly, your fine-tuned model might:

• Output garbage predictions

• Fail to converge during training

• Seem to overfit instantly

This chapter teaches you how to match preprocessing exactly to each pretrained model so you can extract their full power—safely.

---

Conceptual Breakdown

🔹 Pretrained Model Expectations

Model	Expected Input Shape	Pixel Range	Normalization
ResNet, VGG	224×224×3	[0.0 – 1.0]	Mean: [0.485, 0.456, 0.406] Std: [0.229, 0.224, 0.225]
MobileNetV2	224×224×3	[-1.0 – 1.0]	preprocess_input() scales it
EfficientNet	224×224×3 or 240×240	[0.0 – 255.0]	preprocess_input() handles it

Rule of thumb: Always check the docs or source code of the model you’re using.

---

🔹 Why You Can’t Just Use .ToTensor() or /255.0

Because pretrained models were trained on data that was already:

• Normalized using dataset-wide statistics

• Possibly scaled to [-1, 1] or whitened

• Fed in a specific channel order

If you skip or mismatch the normalization, you're effectively corrupting the input distribution—and the model’s learned filters won’t match.

---

🔹 Matching PyTorch Pretrained Models

Most PyTorch models in torchvision.models use:

• [0, 1] range from ToTensor()

• Mean-std normalization with ImageNet stats

  transforms.Normalize(mean=[0.485, 0.456, 0.406],
                     std=[0.229, 0.224, 0.225])

📌 Always use this transform after ToTensor().

---

🔹 Matching TensorFlow/Keras Pretrained Models

Each model has a helper function:

Model Family	Preprocessing Function
ResNet, VGG	keras.applications.resnet50.preprocess_input()
MobileNetV2	keras.applications.mobilenet_v2.preprocess_input()
EfficientNetB0	keras.applications.efficientnet.preprocess_input()

These handle:

• Mean/std normalization

• Scaling to [-1, 1] if needed

• RGB channel order

📌 These functions expect raw pixel values (0–255 float), not normalized.

---

PyTorch Implementation

from torchvision import models, transforms
from PIL import Image

# Load pretrained ResNet
model = models.resnet50(pretrained=True)
model.eval()

# Preprocessing pipeline
preprocess = transforms.Compose([
    transforms.Resize((224, 224)),
    transforms.ToTensor(),  # scales to [0,1]
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

# Load image and preprocess
img = Image.open("dog.jpg").convert("RGB")
tensor = preprocess(img).unsqueeze(0)  # shape: [1, 3, 224, 224]

---

TensorFlow Implementation

from tensorflow.keras.applications import MobileNetV2
from tensorflow.keras.applications.mobilenet_v2 import preprocess_input
from tensorflow.keras.preprocessing import image
import numpy as np

# Load pretrained MobileNetV2
model = MobileNetV2(weights="imagenet")

# Load and preprocess
img = image.load_img("dog.jpg", target_size=(224, 224))
img_array = image.img_to_array(img)  # shape: [224, 224, 3]
img_array = np.expand_dims(img_array, axis=0)  # [1, 224, 224, 3]
img_array = preprocess_input(img_array)  # scales to [-1, 1]

---

Framework Comparison Table

Framework	Step	PyTorch	TensorFlow
Preprocessing	Built-in model normalization	transforms.Normalize()	keras.applications.*.preprocess_input()
Channel Order	Input Format	[C, H, W]	[H, W, C]
Expected Range	Pixel Values	[0, 1] + mean/std	[0, 255] → auto-scaled internally
Input Shape	Default for pretrained	[1, 3, 224, 224]	[1, 224, 224, 3]

---

Mini-Exercise

Try loading the same image in both PyTorch and TensorFlow. Use:

• resnet50 in PyTorch

• ResNet50 in Keras

• Apply the correct preprocessing for each framework.

• Feed the tensor into the model and extract the top-1 class prediction.

• Confirm both models give similar results.

Bonus: Print the difference between preprocessed tensors in PyTorch vs TensorFlow (after matching shape order).

---