feature/PrefGRPO-txt2img-cleaning #9
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
…g and reward computation - Allow batch processing with multiple prompts
…resentation output
…upport and advantage computation
…asses for GRPO alignment
…osses and calculate the mean at the end
…uler and EulerMaruyama methods for flow based models
…put_size; add PreComputedEmbedder class for handling precomputed embeddings; update SD3TextEmbedder to streamline initialization and embedding retrieval.
…mprove type casting for image and text feature encoding.
… timestep handling in Euler and EulerMaruyama samplers.
…ting methods - Fix Gaussian Diffusion in general
…asses to remove ModelInputGRPO and standardize on ModelInput, enhancing type consistency and clarity. - Refactor GRPOTrainer to take into account sampler use and grpo function merged into basic ones
…er abstract class
…n and improve parameter handling
…fusion classes to make data_shape optional, enhancing flexibility in model input handling.
…sk to embedding outputs + fix part of the code
…flect reverse flow matching process
LouisRouss
changed the title
…update context embedding handling
- add attnmask to precomputedembedder
…ling and attention mask support
…low, GaussianDiffusion, and DDT
…er context handling
LouisRouss
changed the title
- remove pre compute on dataset for context embedder
- Introduced RotaryPositionalEmbeddingNDim to support N-dimensional rotary embeddings. - Updated DiTAttention and MMDiTAttention classes to utilize the new rotary embedding structure. - Modified forward methods to accept precomputed cosine and sine values for rotary embeddings. - Enhanced PerceiverRotaryPositionalEmbedding to work with N-dimensional inputs. - Adjusted PerceiverAttention and PerceiverResampler to accommodate new rotary embedding logic. - Replaced RMSNorm with LayerNorm in DiTBlock and MMDiTBlock for consistency. - Updated MMDiT class to compute positional encodings for both text and image inputs. - Added PackedSwiGLU for improved MLP performance. - Fixed minor issues in the base trainer's model input handling.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Adds PrefGRPO training to Diffulab.
The implementation is done from scratch but I take inspiration from the official repository.
Tons of work left to do, draft PR just to open the subject.
This PR also: