Martin Görner
Martin Görner
Here is an excerpt from a model where LoRA was enabled on query and value layers: ``` decoder_block_0/pre_attention_norm/scale (2048,) PartitionSpec() decoder_block_0/attention/query/kernel (8, 2048, 256) PartitionSpec(None, 'model') query/lora_kernel_a (8, 2048, 4)...
Repro Colab: https://colab.research.google.com/drive/1TzvwkSY_EteBQuhNjki2i3Kv37EPfY-Y?usp=sharing The use case for using ExportArchive with a JAX model is when using jax2tf manually. Manual use of jax2tf is an important CUJ for two reasons: -...
Repro Colab: https://colab.research.google.com/drive/1ACGyVwTT-IeaeKBRIphVFxB3NPZHrdP5?usp=sharing When using keras.export.ExportArchive: For most models, using ExportArchive.track(model) is no longer necessary as tf.function auto-track their weights. Note: reloading with keras.layers.TFSMLayer() also works fine. However in that...
Something similar to [this](https://ai.google.dev/tutorials/python_quickstart#chat_conversations) with a ChatSession that helps with the context. This feature was originally requested by Paul Mooney (Kaggle)
Keras.io example: https://keras.io/examples/nlp/data_parallel_training_with_keras_nlp/ Merged PR: https://github.com/keras-team/keras-io/pull/1395 This example is good on the whole but it would be much better with proper batch size and learning rate scaling. Without this, using...
Page where the problem occurs: https://keras.io/api/keras_cv/models/ screenshot of problem: 
Add instance segmentation to the existing YOLOv8 object detection model. This can be a Keras.io example first and get finalized as a KerasCV contribution in a second step. Numerical validation...
Looking through the friction log: https://docs.google.com/document/d/1xoq2axs1QHWvRjKRQP-L8HSlputy3N3R9Q5Glp-lN4g/edit#bookmark=id.262un4qwr0if The most important tasks for now are: - name change -> RandomZoomAndCrop - crop box params change: remove target_size, split crop size into crrop_width,...
Source file https://github.com/keras-team/keras-cv/blob/master/keras_cv/training/contrastive/contrastive_trainer.py A "linear probe" is a linear classifier trained on top of a frozen (typically contrastively pretrained) backbone. It is usually trained after the contrastive pretraining, to evaluate...