Image to LaTeX Converter: Project Report

The img2latex system uses the IM2LaTeX-100k dataset, which contains over 100,000 images of mathematical expressions paired with their corresponding LaTeX code. Our analysis of the dataset revealed:

Image Properties

• Width range: 128 - 800 pixels
• Height range: 32 - 800 pixels
• Aspect ratio range: 1.00 - 15.00
• File format: All images are RGB
• Pixel value range: 0.0 - 255.0 (uint8)
• Mean pixel value: 242.22
• Std dev of pixel values: 45.70

Preprocessing Strategy

Based on this analysis, our preprocessing pipeline includes:

• Scale-to-fit: Each image is scaled to fit a 64×512 canvas without cropping, then padded with white on the right and bottom
• Padding mask: Valid width and height are stored so the decoder cannot attend to empty canvas
• Channel Conversion:
  • CNN encoder: grayscale (1 channel)
  • ResNet encoder: RGB (3 channels)
• Cache: A one-time uint8 memmap of unique processed images avoids repeating decode and resize during training
• Normalization: Pixel values are scaled to [0, 1]

The tokenizer is fit on the train split only with min_freq=5, then sequences are truncated at 141 tokens. Batches are length-bucketed so padding waste stays low.

The img2latex system offers two encoder variants that share the same Transformer decoder:

1. CNN-Transformer Architecture

The default model (~4.45M parameters) consists of:

• Encoder: Three Conv-BN-ReLU-MaxPool blocks (filters 32, 64, 128) downsample the 64×512 image by 8×. A 1×1 convolution projects to d_model=256, then 2D sinusoidal positional encoding is added. The result is a spatial grid of 8×64 = 512 memory tokens plus a padding mask for empty canvas.
• Decoder: A 4-layer pre-norm Transformer decoder (8 heads, feedforward 1024, tied embeddings) with a boolean causal mask. Cross-attention reads the encoder grid, not a single collapsed vector.

2. ResNet-Transformer Architecture

The optional ResNet encoder replaces the CNN while keeping the same decoder:

• Encoder: ResNet-18 with the classification head and avg-pool removed (children()[:-2]). layer4 stride is set to 1 so the feature map stays a spatial grid (~/16) instead of collapsing to one vector.
• Decoder: The same Transformer decoder as the CNN model.

The previous CNN-LSTM / ResNet-LSTM stack collapsed each image to a single vector, so attention was a no-op. That, plus teacher-forced BLEU, is why the old report sat at 0.15 BLEU.

The training process implements several key strategies:

Optimization Setup

• Optimizer: AdamW, learning rate 3e-4, weight decay 1e-4
• Learning Rate Scheduling: Linear warmup for 5% of steps, then cosine decay
• Loss Function: Cross-entropy with label smoothing (0.1), pad ignored

Training Techniques

• Teacher Forcing: Used only for the training loss. Validation BLEU is computed from autoregressive generation on a fixed 256-sample subset
• Gradient Clipping: Norm-based clipping at 1.0
• Early Stopping: Stops after 8 epochs without a validation BLEU improvement
• Checkpointing: Best checkpoint is selected on validation BLEU, not validation loss

Hardware Acceleration

• Device Support: MPS on Apple Silicon (this run), CUDA, or CPU fallback
• Precision: fp32 on MPS (no GradScaler). Mixed precision is available on CUDA only
• DataLoader: num_workers=0 on macOS to avoid OpenMP / libomp collisions with PyTorch MPS

During inference, the model offers greedy decoding and batched beam search:

1. Greedy Search: Selects the most probable token at each step. Used for validation BLEU so scores stay comparable across epochs
2. Beam Search: Batched beam search with GNMT length penalty (beam size 5, length penalty 0.7). This is the default for the predictor CLI

Beam search is enabled in the predictor (it was previously short-circuited). Reported validation BLEU in this report uses greedy decode on 256 validation samples so the metric matches the checkpoint-selection criterion.

Experiments used the CNN-Transformer on Apple Silicon MPS. Checkpoints are selected on validation BLEU from autoregressive generation, not teacher-forced token accuracy. Key configuration:

Evaluation Metrics

Reported scores come from generated token sequences (greedy decode, 256 validation samples):

1. BLEU-4: Smoothed corpus BLEU (nltk method3)
2. Exact match: Generated tokens equal the reference after stripping START/END/PAD
3. Normalized edit distance: Mean Levenshtein / max(len); lower is better
4. Token accuracy: Teacher-forced diagnostic only; not used for checkpoint selection

Performance Results

Best checkpoint is epoch 14 of a 30-epoch schedule (img2latex_transformer_v1). BLEU is autoregressive. The legacy CNN-LSTM report of 0.1539 BLEU was teacher-forced and is not comparable.

Epoch	Val Loss	TF Acc	BLEU	Exact Match	Edit Dist
1	2.9760	0.4773	0.1165	0.0%	0.7295
5	1.5380	0.8461	0.6376	5.5%	0.2848
10	1.3741	0.8924	0.7508	15.2%	0.2093
14	1.2730	0.9152	0.7896	21.9%	0.1778
18	1.3295	0.9065	0.7880	21.1%	0.1770

Comparison to the Legacy LSTM

The previous CNN-LSTM / ResNet-LSTM numbers were teacher-forced and used encoders that collapsed the image to a single vector:

• CNN-Transformer (this run): generated BLEU 0.790, exact match 21.9%, edit distance 0.178 (epoch 14)
• Legacy CNN-LSTM (img2latex_v2): teacher-forced accuracy 62.56% and BLEU 0.1539; no checkpoint remained in the repo to re-score honestly
• Spatial features plus a Transformer decoder moved generated BLEU from ~0.15 to ~0.79 in 14 epochs on MPS

Error Analysis

Common error patterns included:

• Missing or incorrect brackets in nested expressions
• Confusion between similar-looking symbols (e.g., 1/l, 0/O)
• Incomplete transcription of complex subscripts and superscripts

Sample generations on a long nested integral show the remaining failure mode: the model recovers the overall structure (integrals, deltas, fractions) but still drops primes, confuses \mathrm{e} with \operatorname{e}, and loses subscripts inside nested fractions.