Skip to content

Supported Frameworks

epochix auto-detects the framework from your log output. The detection uses a "sniff" score — the parser with the highest confidence above the threshold wins. The universal fallback always fires if nothing else matches.

PyTorch Lightning

Auto-detection regex:

Epoch \d+/\d+:.*\|

Sample log line:

Epoch 5/50: 100%|=====>| 250/250 [00:12<00:00, loss=0.432, val_acc=0.867]

Detected metrics: Any key=value pair inside the progress bar line.

Keras / TensorFlow

Auto-detection: Epoch N/N header + progress bar [====>]

Sample:

Epoch 5/50
1563/1563 [==============================>] - 10s - loss: 0.423 - accuracy: 0.867 - val_loss: 0.401 - val_accuracy: 0.874

Detected metrics: All key: value pairs on the progress bar line.

HuggingFace Transformers

Auto-detection: JSON-like dict with 'loss' key.

Sample:

{'loss': 0.5123, 'learning_rate': 5e-05, 'epoch': 1.0}
{'eval_loss': 0.3421, 'eval_accuracy': 0.8765, 'epoch': 1.0}

Detected metrics: All numeric values in the dict.

Ultralytics YOLO

Auto-detection: Training row epoch/total GPU_mem box_loss cls_loss dfl_loss and validation row all N N precision recall mAP50 mAP50-95.

Sample:

      5/50     1.23G   0.456   0.234   0.123   128
                 all   5000   5000   0.712   0.654   0.678   0.432

Detected metrics: box_loss, cls_loss, dfl_loss, precision, recall, mAP50, mAP.

FastAI

Auto-detection: Tabular header train_loss valid_loss metric time.

Sample:

epoch  train_loss  valid_loss  accuracy   time
0      1.4321      1.2345      0.4567     00:12
1      0.9876      0.8765      0.5678     00:11

Accelerate

Auto-detection: JSON-like dict (same format as HuggingFace).

Universal fallback

Matches any log that contains at least one of:

  • key=value pairs
  • key: value pairs (lower confidence)
  • JSON fragments {...} containing numeric values

Confidence: 0.10 — it always fires as a last resort.

LLM fallback (opt-in)

For completely unrecognised formats, an optional LLM-powered extractor can be enabled:

# Ollama (local)
export EPOCHIX_LLM_URL=http://localhost:11434
epochix train.log

# OpenAI
export EPOCHIX_LLM_KEY=sk-...
epochix train.log

The LLM fallback batches 20 lines per call and asks the model to extract {key, value, epoch} tuples as JSON. It has a confidence of 0.55 — lower than all regex parsers.

Writing a custom parser

See Plugins.