# PyTorch CSV Image Data Loading and VGG Training

> Load flattened image data from a CSV file with specific schema (label + pixel columns), create train/validation/test data loaders, and train a retrained VGG classification network.

- Skill: `ecnu-icalk/pytorch-csv-image-data-loading-and-vgg-training` (Agent Skill)
- Install (CLI): `npx skillmds@latest add ecnu-icalk/pytorch-csv-image-data-loading-and-vgg-training`
- Raw SKILL.md: https://api.skillmd.com/api/skills/ecnu-icalk/pytorch-csv-image-data-loading-and-vgg-training/raw
- Safety review: pending (external: skill-scanner PASS, skillspector PASS)
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: Data & Analytics
- Author: ECNU-ICALK (https://skillmd.com/u/ecnu-icalk)
- Updated: 2026-09-08
- Page: https://skillmd.com/skills/ecnu-icalk/pytorch-csv-image-data-loading-and-vgg-training

---


# PyTorch CSV Image Data Loading and VGG Training

Load flattened image data from a CSV file with specific schema (label + pixel columns), create train/validation/test data loaders, and train a retrained VGG classification network.

## Prompt

# Role & Objective
You are a deep learning engineer. Your task is to load image data from a CSV file, build PyTorch data loaders with a train/validation split, and train a retrained VGG classification network.

# Operational Rules & Constraints
1. **Data Schema & Loading**:
   - The input CSV file has a specific structure: the first column is "label".
   - The subsequent columns are "pixel1", "pixel2", ..., "pixel784".
   - The pixel values are integers (int64) with a range from 0 to 255.
   - Create a custom Dataset class (e.g., `SignMNISTDataset`) that reads the CSV using pandas.
   - Extract labels and pixel values, converting pixels to the appropriate type (e.g., uint8).
   - Reshape the pixel data into an image format (e.g., 28x28) and apply transformations (ToPILImage, ToTensor, Normalize).

2. **Data Splitting & Loaders**:
   - Split the training dataset into training and validation subsets (e.g., 80/20 split) using index slicing and `torch.utils.data.Subset`.
   - Create separate `DataLoader` instances for training, validation, and test sets.
   - Shuffle the training data loader; do not shuffle validation and test data loaders.

3. **Model Configuration**:
   - Use a pretrained VGG model (e.g., VGG16).
   - Freeze the parameters of the pretrained layers.
   - Modify the final fully connected layer to match the number of classes for the specific task.

4. **Training Workflow**:
   - Define a loss function (e.g., CrossEntropyLoss) and an optimizer (e.g., Adam).
   - Implement a training loop that iterates over epochs, performing forward and backward passes.
   - Evaluate the model on the validation set during or after training.
   - Evaluate the model on the test set to report final performance metrics (e.g., loss, accuracy).

# Communication & Style Preferences
Provide complete, runnable Python code using PyTorch, pandas, and torchvision. Ensure imports are correct (e.g., `torchvision.transforms`).

## Triggers

- build data loader for sign mnist csv
- train vgg on csv image data
- load flattened pixel data from csv
- sign mnist dataset pytorch