• Introduction
• Team Members (in alphabetical order)
• Achievements
• Final evaluation results
• ASR
  • Data Augmentation
  • Models
  • Training
  • Inference
• NLP
  • Data Transformation
  • Model
  • Training
    • Hyperparameters:
  • Inference
• VLM
  • Overview
  • Data Augmentation
    • YOLO
    • SigLIP
  • Synthetic Data
  • Models
    • Object Detectors
    • Upscalers
    • SigLIP
  • Training
    • YOLOv9e
    • YOLOv6-L6
    • SigLIP
  • Inference
    • YOLOv9e
    • Slicing Aided Hyper Inference (SAHI)
    • YOLOv6l6
    • Real-ESRGAN-x4v3
    • SigLIP
  • Hardware used
• Final words

TIL-AI 2024 comprised 3 tasks:

• Automatic speech recognition (ASR)
  Convert radio-distorted and noisy audio into text.
• Natural language processing (NLP)
  Convert text into structured data for controlling a turret.
• Vision-language modelling (VLM)
  Locate a flying object in an image based on a textual description of its appearance.

The three tasks were chained in the Finals to drive a DJI Robomaster's turret in a simulated environment.

• Billy Cao (L): NLP/VLM
• Ho Wing Yip: VLM
• Huang Qirui: ASR
• Marcus Wee: ASR
• Ooi Xuan Shan: VLM

• 3rd overall in Qualifiers
• 2nd in Semi-finals
• 7th in Finals

We hypothesise that our poor Finals performance was because we overfitted our VLM to the Qualifiers test set - i.e. we chose our checkpoints and made optimisations based almost solely on Qualifiers performance. It is likely that beyond some lower bound, increases in accuracy - even validation/test accuracy - become unrepresentative of model performance and robustness in the wild. (This was observed during our training process for the SOLIDER-REID modelin TIL 2023.)

We do not report speed score here. Due to our use of hardware-specific TensorRT optimisation, it is unrepresentative of the actual speed of our models in the Finals.

We tried denoising the given data, then training on the original noisy dataset combined with denoised data. However, the model ended up not performing very well, with an unexpected dip in accuracy. Hence, we scrapped the idea of training on denoised data.

For ASR data augmentation, we used the audiomentations library. We note that torch-audiomentations is an alternative that performs augmentations on GPU for a speed boost. Nonetheless, its support for some augmentations is lacking as of June 2024, so we did not use it.

Several combinations of augmentations were tried. We eventually settled on the following:

augment = Compose([
    HighShelfFilter(max_gain_db=6.0, p=0.3),
    LowShelfFilter(max_gain_db=6.0, p=0.3),

    # Limit rate to [0.9, 1.1] so audio remains recognizable
    TimeStretch(min_rate=0.9, max_rate=1.1, p=0.2),

    # Simulate radio transmission where a limited range of frequencies can be broadcast
    BandPassFilter(p=0.3),
])

"niner" is a special token that exists in this competition's data due to the convention of radio transmission. We tried adding it as an extra token to the tokenizer. There are marginal gains but it introduced issues for faster-whisper deployment so we did not pursue it.

Parakeet RNNT 0.6B gave a much worse leaderboard score despite a ~10x lower validation word error rate during training. Perhaps, Whisper has supreme robustness due to being trained on 680k hours of labelled data versus Parakeet's 64k hours.

More evaluation results can be found here.

Hyperparameters:

• Learning rate: 1e-5
• Warmup: 500 steps
• Epochs: 30
• adam_beta1: 0.9
• adam_beta2: 0.98
• warmup_ratio: 0.22
• weight_decay: 1e-4

Training code can be found in WhisperSmallAndMed.ipynb.

Best performing model weight and training log: https://huggingface.co/aliencaocao/whisper-medium.en-TIL24

To speed up inference using Whisper models, we used faster-whisper. It utilises ctranslate2, a fast inference engine for Transformer models, to increase inference speeds by more than 2x.

We also tried applying loudness normalisation to the audio clips before inference to increase accuracy score, and this was done using the pyloudnorm library. However, this loudness normalisation seemed to have no noticeable effect on our scores, but significantly slowed down model inference. This led us to conclude that our ASR models, namely Whisper Small and Medium, are already significantly robust to audio clips of varying loudness. This can also be seen in the finals where the raw whisper models were able to transcribe even the softest of clips.

We also tried denoising the data prior to inference as another attempt to raise our accuracy from the whisper model. However, we noticed early on that denoising the audio clips on inference was doing a disservice to the whisper model. Denoising on inference caused our accuracy to take a hit, while the performance tanked due to the time required to process every input. This is also likely due to whisper being trained on 680k audio clips and being robust to noise. It is also possible that denoising the clips, introduced audio artifacting in some clips and causing inferences to fail.

Inference code can be found in ASRManager.py.

The task is to convert instructions on operating a turret into structured data with the following fields: heading (a 3-digit number), tool (the weapon to use), target (description of the target's color and type).

We approached the task as a function calling task common in LLM agents. We are aware of much smaller scale/faster pipelines available but decided to go the most time saving way of "whacking" a large model to solve everything so we can focus on the hard VLM task.

System prompt: Convert the given instruction to turret into a function call to the control_turret function. Strictly ONLY one call is needed. If there are multiple instructions, use the FIRST one only.

Function calling task require a function definition to be provided to the LLM, in JSON string and appended to the system prompt. After tuning by evaluating zero-shot on training data, ours is:

{"name": "control_turret",
"description": "Control the turret by giving it heading, tool to use and target description",
"parameters": {"type": "object", "properties": {"heading":
{"type": "string", "description": "Heading of target in three arabic numbers and multiples of five (005 to 360). Give None if not specified."},
"tool": {"type": "string", "description": "Tool to use or deploy. Give None if not specified."},
"target": {"type": "string", "description": "Description of the target or enemy, exclude any quantifiers like 'the' or 'a'. It is a phrase that describe the appearance of the target like its color and type. Include ONLY the appearance and NOTHING else like its heading. Give None if not specified."}},
"required": ["heading", "tool", "target"]}}

As sometimes the command can contain repetitions and model can be confused, we did the following:

1. Check for “repeat” in text and attempt to extract information on first half, retry on full text if failed to extract all 3 fields. When retrying, omit the Give None if not specified. part of the prompt.
2. Any matched tool will be included in prompt so model do not choose it again as target (e.g. missiles)

To further reduce inference time, simple regex-based rules were used:

1. Regex to match 3-word sequence of heading e.g. "One One Zero"
2. Regex to match 7 known weapons in training data: {'anti-air artillery', 'drone catcher', 'electromagnetic pulse', 'emp', 'interceptor jets', 'machine gun', 'surface-to-air missiles'}

Gorilla-OpenFunctions-v2 by UC Berkely, SOTA open-source model on Berkeley Function-Calling Leaderboard, fine-tuned from LLaMA2-7B.

Using the above data transformations, the pretrained model scored 0.9978597402597402 on leaderboard.

More evaluation results can be found here.

We then fine-tuned the model using Unsloth. Training took 34 minutes on Colab's free T4 GPU. We used Rank-Stabilized LoRA (Low-Rank Adaptation of Large Language Models).

• LoRA rank: 16
• LoRA alpha: 16
• optimizer: AdamW 8bit
• initial LR: 2e-4
• warmup: 10 steps
• LR schedule: cosine annealing decay
• effective batchsize: 8
• epoch: 1
• weight decay: 0.01

Training code can be found in train.ipynb.

Model weight: https://huggingface.co/aliencaocao/gorilla-openfunctions-v2-TIL24-r16-a16-ctx768-v2

LLMs may seem extreme for this task for its sheer size and inference cost, but it CAN be deployed efficiently with inference costs matchihng smaller models like BERT.

We used ExLlamaV2 quantization and inference engine. It was chosen because it does all matrix ops in FP16 which Tesla T4 GPU in submission runtime excels at. It is also specifically optimized for consumer Nvidia GPU which is used in the Finals (RTX 4070Ti Super).

Our model was quantized to average of 5.0 bit with 6-bit head precision, using default calibration data. Anything below 5.0 bit severely reduces model performance even with custom calibration data.

We made tweaks to further reduce VRAM I/O bottlenecks to speed up inference:

• Discard logits in VRAM instead of copying back to CPU (2x+ speedup), controlled by config.max_output_len = 1
• restricting KV cache pool to short context length (512) and batch size (4 in qualifiers, 1 in finals), controlled by config.max_seq_len = 512 and config.max_batch_size = 4

Our trained model has 0.9993333333333333 leaderboard score without any tricks / regex mentioned above, just to test fine-tuning effectiveness. In finals, we still employ those tricks.

Model peak VRAM usage: 5.7GB

Time taken to process 1 sample at bs=1: 0.36s. With bs=4, it is about 2 times faster per sample.

Inference code can be found in NLPManager.py.

This task is on using description of a flying object to locate it in an image. We tried multiple approaches:

• MM-Grounding-DINO: trained model giving 0.509 map (we got scammed)
• Co-Det: 0.61 map zero-shot but too slow to train
• OWLv2-large-ensemble: 0.629 map zero-shot BUT training on both own-implemented PyTorch code and Google’s Official JAX code lead to loss explosion
• A few that inference could not run / already very bad on training data
  • APE - works on author's HF space, but somehow triggered a severe memory leak on GCP that filled all of CPU RAM
  • UNINEXT
  • ONE-PEACE
  • OFA
  • GLEE-Pro
  • PolyFormer
  • HIPIE
  • LLaVA 1.6 (LLaMA-3 8B) w/ prompting
  • Phi3-vision w/ prompting

Details can be found here

We only document in detail the multi-stage approach that we used eventually. The full evaluation results and comments are here. The writeup below is a summary.

Our pipeline:

1. YOLOv6-L6 trained on single-class detection of targets in general, using synthetic data too
2. Extract the bboxes as detected by YOLO, optionally using SAHI (Slicing Aided Hyper Inference). SAHI helped for v9e but not v6l6.
3. Run each extracted bbox through Real-ESRGAN x4v3 model to upscale 4x
4. Feed each bbox into a SigLIP and get similarity score VS caption (1/image)
5. Choose the box with the highest similarity score for each caption

We used the following augmentations for YOLOv9e and YOLOv6-L6 training:

T = [
    A.GaussNoise(var_limit=2500, p=0.5),
    A.ISONoise(p=0.5),
    A.Flip(p=0.5),
    A.Blur(p=0.1),
    A.MedianBlur(p=0.1),
    A.ToGray(p=0.1),
    A.CLAHE(p=0.1),
    A.RandomBrightnessContrast(p=0.5),
    A.RandomGamma(p=0.2),
    A.ImageCompression(quality_lower=75, p=0.5),
]

This is V3 of the augmentations. Evaluation of previous 2 versions can be found here.

Note for YOLOv9e, we had to modify the source of Ultralytics directly to modify its augmentations.

Augmentations for YOLOv6 are defined here.

We used the following augmentations for SigLIP training:

T = [
    A.Resize(image_size, image_size, interpolation=cv2.INTER_LANCZOS4),
    A.GaussNoise(var_limit=(500, 5000), p=1.0, per_channel=True),
    A.ISONoise(p=1.0, color_shift=(0.02, 0.07)),
    A.MultiplicativeNoise(p=1.0),
    A.AdvancedBlur(blur_limit=(3, 11), p=0.3),
    A.Flip(p=0.5),
    A.RandomRotate90(p=0.5),
    A.CLAHE(p=0.2),
    A.RandomBrightnessContrast(brightness_limit=0.2, contrast_limit=0.5, p=0.5),
    A.RandomGamma(p=0.2),
    A.Perspective(p=0.5),
    A.ImageCompression(quality_range=(25, 75), p=0.8),
    A.Normalize(mean=mean, std=std),
    ToTensorV2()
]

This is V2 of the augmentations. Evaluation of V1 can be found here.

We also train SigLIP on upscaled data as it is shown to have great accuracy benefits, but reduced robustness to false positive bboxes. Over all, it is still by far a worthy trade-off.

We generated synthetic data for VLM task by cropping and pasting flying objects from online onto random images, and resizing them and add augmentations.

Our synthetic data consists of 2338 images, 9214 bboxes, 69 unique objects.

Code for generating the dataset can be found here

Sample:

Full dataset is open sourced: https://huggingface.co/datasets/aliencaocao/described-flying-objects

We initially used YOLOv9e for qualifiers, then switched to YOLOv6l6 as we find it perform better both on our synthetic data and leaderboard, while being 3x faster to train and 2x faster to infer. It also converges with 2x less epoch (~30 VS ~70 on V9e).

Best leaderboard score for YOLOv9e is 0.905 with SAHI and TTA (slow). Best leaderboard score for YOLOv6l6 is 0.911 for single-model, 0.913 for 2x YOLOv6l6 WBF, but both without SAHI/TTA. In finals, we used single model in favour of speed.

We used Real-ESRGAN x4v3 for upscaling. It is a very lightweight model at 6.2MB. While larger models like RealESRNet_x4plus are better, they are too slow to run inference, since the model cannot be batched without padding (which severely hurts its quality). In local evaluations, the lightweight version performed just as well.

Effect of upscaling:

Interestingly, upscaling has always been worse on local clean validations set, likely due to the image not being noisy and the upscaling artifacts are not learned by the model trained on non-upscaled data.

We extensively compared OpenAI CLIPs, LAION CLIPs, EVA-CLIPs, Metaclips and Google SigLIPs, SigLIP clearly stands out as the winner as:

• SigLIP is trained on sigmoid loss which operates on only 1 pair of image-text labels, compared to normal CLIPs trained on softmax which learns 1-to-many best
• This ensures its learnt representations are best suited for 1-to-1 similarity scoring
• Normal CLIPs fails at 1-to-1 as softmax will normalize score to 1.0. This heavily affects its representation quality for this task (after removal of softmax function to get raw logits)
• Zero shot ImageNet performance is worth considering but not strictly correlated to this task e.g. CLIP-H-LAION2B underperform SigLIP zero-shot

We tested both google/siglip-large-patch16-384 and google/siglip-so400m-patch14-384. Before training on upscaled data, so400m performed better likely due to its higher parameter count and thus better generalization. However, after training on upscaled data, so400m tends to overfit early and consistently underperformed large. Training on upscaled data likely made the task a lot easier and large is less prone to overfitting thus ended up performing better.

• YOLOv9e trained on 640px input size with augmentations
• Inferencing at 1600px improves score despite model trained at 640px
• Continued training for ~10 epoches on 1600px improves further
• A clear characteristic of small object detection tasks where input resolution matters greatly

YOLOv9e training code can be found in train.ipynb

To launch YOLOv6-L6 training (on 5 gpus), run the following:

cd vlm/multistage/YOLOv6
CUDA_VISIBLE_DEVICES=0,1,2,3,4 python -m torch.distributed.launch \
    --nproc_per_node 5 \
    --master_port 27322 \
    tools/train.py \
    --batch-size 52 \
    --bs_per_gpu 16 \
    --conf-file configs/yolov6l6_finetune.py \
    --data-path data/dataset.yaml \
    --specific-shape \
    --height 870 \
    --width 1520 \
    --epochs 200 \
    --eval-final-only \
    --save_ckpt_on_last_n_epoch 200 \
    --output-dir runs/train \
    --name bs52_effbs64_dsta+xs_expanded \
    --device 0,1,2,3,4 \
    --workers 32 \
    --check-images \
    --check-labels

To improve compute efficiency, we used rectangular training by setting input to 1520x870, same as DSTA training data and test data. Note that it will be upscaled to 1536x896 due to it must be divisible by strides=32.

HF transformers does not support training SigLIP yet but it is simply due to missing loss function. We adapted it from JAX implementation to transformers:

eye = torch.eye(logits_per_text.size(0), device=logits_per_text.device)
m1_diag1 = -torch.ones_like(logits_per_text) + 2 * eye
loglik = torch.nn.functional.logsigmoid(m1_diag1 * logits_per_text)
nll = -torch.sum(loglik, dim=-1)
loss = nll.mean()

For ease of use, our modified modeling_siglip.py can be found here as our training script imports it locally. A PR has been made to transformers too.

Here is a summarized abaltion study for SigLIP:

• Qualifier end: 0.864 (5 epoch on original data)
• Continued for 5 more epoch: 0.874
• Inference on upscaled boxes: 0.881
• Train & inference on upscaled boxes: 0.899
• Trained on DSTA+synthetic: 0.91 → 0.913

Hyperparameters:

• Learning rate: 2e-5 (tried higher but it is much worse)
• Warmup: 10%
• Weight decay: 1e-4 (tried 0 following SigLIP paper's finetune recommendations but performs worse)
• Effective batch size: 960 (tried smaller but it is much worse, confirming various contrastive learning paper's observation that generally cranking up batch size is a free gain. SigLIP paper used sigmoid to use 1M BS!)
• AdamW beta1: 0.9
• AdamW beta2: 0.99 (reduced from default 0.999 to increase stability at higher BS as recommended by SigLIP paper)
• Epochs: 5 or 10 (see below)
• LR schedule: cosine annealing decay
• Optimizer: AdamW Fused

Training was done in full BF16 with torch.compile which speeds up training by 30% at cost of a 4 minute compile time.

SigLIP training code can be found in HF_train.py.

We used Weights & Biases for logging: https://wandb.ai/aliencaocao/TIL2024-SigLIP

Qualifiers and semi-finals submission model is https://wandb.ai/aliencaocao/TIL2024-SigLIP/runs/3c00nigo which is resumed from https://wandb.ai/aliencaocao/TIL2024-SigLIP/runs/os657bxe. This is trained for 5epochs then continued for another 5 epochs (effectively 2 cycles of cosine annealing schedule) on full DSTA data. It scored 0.905 with YOLOv9e w/ SAHI and 0.91 with 2xYOLOv6l6 WBF. Weights: https://huggingface.co/aliencaocao/siglip-large-epoch5-augv2-upscale_0.892_cont_5ep_0.905-TIL24

Finals submission/best score model is https://wandb.ai/aliencaocao/TIL2024-SigLIP/runs/ffkw9nka. This is trained for 5 epochs only on DSTA + synthetic data. This was used for finals. It scored our best score 0.913 with YOLOv6l6, compared to 0.91 on the other model. Weights: https://huggingface.co/aliencaocao/siglip-large-epoch5v3_merged-TIL24

Due to a bug with HuggingFace Accelerate/FDSP, we had to save the checkpoint only once and thus set save_steps to 0.999.

Due to a mysterious issue, we constantly CUDA OOMed when running validation, possibility due to a VRAM leak. Thus we had to disable validation and rely soley on leaderboard score for iterating.

• Test-time Augmentation improves score at cost of inference speed.
• A lower confidence of 0.1 was used to prevent false negatives. This gave the best result on leaderboard.
• SAHI (Slicing Aided Hyper Inference) improves significantly at large cost of speed: (0.896 → 0.905), however must be paired with higher confidence score (0.1 vs 0.5), else, it is a lot worse than without SAHI, likely due to way too many FP within each slice.
• Weighted Box Fusion (WBF) did not improve score.

We modified the original SAHI to:

1. Add support for batched inference on slices to greatly speedup inference
2. Add support for TensorRT for Ultralytics-based/so-called YOLOv8 models (exported using Ultralytics)
3. Add support for YOLOv6 (requires yolov6 repo on sys.path)

Our modified fork is here.

• Did not use SAHI as it degraded score marginally, likely due to us reaching the limit of yolov6l6.

  • We initially hypothesised that YOLOv6, having been fine-tuned on 1520x870 images, was unable to generalise to the 760x870 slices output by SAHI. As such, we fine-tuned a new YOLOv6 on randomly selected 760x870 patches of the input images, and tried using this new model for sliced inference. However, this did not improve performance on the qualifiers test data.
  • YOLOv6 sliced training may be enabled by setting self.sahi_random_crop = True here.

• Weighted Box Fusion (WBF) between 2 YOLOv6l6 models improved score from 0.911 to 0.913 but did not get used in finals (in hindsight, should very much have!)

• Converted model to use Nvidia TensorRT in FP16 on a RTX 4070Ti Super (same as finals) with fixed batch size (1) and input shape (1536x896) for maximum speed. Conversion script with profiling code is here. It uses torch2trt. Using fixed input shape allows TRT to apply the most aggressive optimizations.

Real-ESRGAN models does a pre-pad to images before upscaling to avoid border artifacts. The default value is 10 for the x4v3 model but it cannot be used on images with less than 10px on either side. 10px is better than 1px pre padding (0.892 vs 0.887).

Originally we ignored these smaller images, but evaluations showed that having a conditional padding where pre-pad=1px for these images is still beneficial (0.892 vs 0.894). We generated the training data for SigLIP using the same way.

We converted model to use Nvidia TensorRT in FP16 on a RTX 4070Ti Super (same as finals). Conversion script is here, also using torch2trt.

We split the model into text and vision encoders and converted them separately for maximum performance.

Image encoder:

• dynamic batch size (min1, optimal 10, max 20)
• fixed input shape (384x384)

Text encoder:

• fixed batch size (1)
• fixed input length (64) as all SigLIP models are trained with input texts padded to 64 tokens.

Then we calculate the image-text similarity using dot product on normalized embeddings since it is cheaper but functionally equal to cosine-similarity. We also omit the sigmoid activation as it is not needed for image reranking task like this.

For our entire pipeline, the end2end time taken to process 1 sample at bs=1 is 0.15s. Without TensorRT acceleration, it takes 1.8s.

We thank our generous senior as most of our models are trained on a 5xRTX 3090 node owned by him.

We also made use of Kaggle's free 2xT4 and TPUv4-8 pod for training.

We like to thank Ryan for his hardwork SOLOing the tech side of the competition and support during the competition, as well as DSTA for organizing the largest yearly ML competition in Singapore. Although we kind of trolled ourselves in finals, we still believe in the work and effort we have done, and that our learnings and findings are valuable, and attribute our failure to just luck (coping).

As four-year TIL participants, we are glad to see it gains so much traction in 2024 and our previous works (TIL-2023, TIL-2022) has helped and inspired not only ML beginners but also the organizer himself (Ryan).

For any questions please open a Discussion.

Please email [email protected] for unreleased weights.