Any Copying from the work of another person is a violation of Carnegie Mellon University Policy on Academic Integrity.

To use

• Run get_data.sh to download both NIST36 and images used for text extraction
• Run fetch_flowers17.sh to download flowers 17 or fetch_flowers102.sh to download flowers 102
• run_q2.py validates the manual implementation of a fully-connected network
• run_q3.py trains the FC network from scratch, visualizes the weights learned, and computes the confusion matrix on test set
• run_q4.py extracts text from four images of handwritten characters by row, and classifies them with the FC network
• run_q5.py designs an autoencoder, trained with SGD with momentum, and reconstructs images in NIST36, which is then evaluated using Peak Signal-to-Noise Ratio
• run_q6_1.py trains and evaluates the following networks in PyTorch
  • A 2-layer FC on NIST36
  • A CNN (3 Conv layers followed by a FC) on NIST36
  • The same CNN on CIFAR-10
  • A CNN (4 Conv layers followed by a FC) on SUN
• run_q6_2.py fine tunes the last classifier layer of SqueezeNet on Flowers17 and compares its performance with a LeNet trained from scratch, in PyTorch

Results

• Confusion Matrix on NIST36 using the 2-layer FC Network

• Text Extraction By Row
  • Original -- Detected -- Extracted

• Autoencoder on NIST36

• Fine Tuning of SqueezeNet vs. Training LeNet from Scratch on Flowers 17