unorganised memo, notes, code, data, and writings of random topics
GPT-OSS is an open source, open weight model by OpenAI with Apache 2.0 license. It is an autoregressive mixture-of-experts (MoE) model that specialized in reasoning (chain of thought), tool use, and support structured output. Similar to Llama architecture, it uses pre-norm with RMS norm and each expert is a SwiGLU-based MLP.
I have a few Yongnuo RF603N II remote flash units. They are old but works. Their primary use is for remote flash over its hotshoe, but they also allows remote shutter release via the same 2.4GHz radio. It comes with a 10-pin Nikon cable that works for older high-end DSLRs (N1 connector, compatible with MC-30A or MC-36A shutter release). But Nikon Z5ii uses a headphone jack for shutter release.…
MTF stands for Modulation Transfer Function. It illustrates the optical quality of camera lens, a.k.a. “sharpness”. MTF chart is a curve of the transmittance of light (0 to 100%) against the distance from the centre of the image towards its edges (in millimetres).
This paper is a report on training the TinyLlama model, following the Llama 2 architecture but with 1.1B parameters. The objective is to experiment with the idea that a smaller model but trained with more data can match or overperform a larger model.
This is a paper to investigate the importance of the MLP sublayer in a decoder-only transformer model. In particular, the common design of a two-layer feed-forward network is challenged. Alternatives of 0 to 3 layers of feed-forward networks are compared for the standard error loss.
This is a paper of a new model for object detection, which the output is a bounding box and a class label. Compared to other object detection models, FCOS does not use predefined anchor boxes. This model is described as “anchor box free and proposal free”.
How to understand noise in digital camera photos? A common believe is that it is related to ISO. The higher the ISO, the stronger the noise. However, some also said ISO has nothing to do with noise. To be accurate, the only valid statement is that noise is a built-in property of the image sensor. It is complicated to tell how would that translate into the noise you see in the image.
This is a paper from Apple, targeted to create a backbone network that is fast enough to use on mobile devices. One characteristic of the network is that is decoupled the train-time and inference-time architecture. The trained model will be reparameterized for inference to make it more efficient.
The goal of the paper is to produce a promptable segmentation model. It is a model that supports flexible prompting and can output segmentation masks in real time. For any given segmentation prompt (e.g., a point in the image, and may be ambiguous), the model is expected to return a valid segmentation mask.
This is the paper that introduced the Vision Transformer (ViT), which proposed that transformers can be used for image classification replacing CNNs. Inspired by the success of transformer models in NLP, this paper explored the technique of using transformers to process 2D image data. The goal is to create a base model that can be adapted for various downstream tasks, just like what transformer…