Add xattn-perceiver-p05 (XAttn-Perceiver-p05)

xattn-perceiver-p05/README.md

@@ -0,0 +1,20 @@
+---
+license: apache-2.0
+tags:
+  - histopathology
+  - diffusion
+  - spatial-transcriptomics
+  - icml-2026-sd4h-workshop
+---
+
+# XAttn-Perceiver-p05
+
+EMA-only inference weights for the **XAttn-Perceiver-p05** row reported in the
+ICML 2026 SD4H workshop submission *Transcriptomics-Conditioned Virtual Tissue
+Synthesis via Diffusion Transformers*.
+
+- **Source checkpoint**: `step_2323000_ema.pt`
+- **Architecture**: see `training_config.yaml` in this folder.
+- **License**: Apache-2.0.
+
+See the umbrella repo README at `stmdit-anon/stmdit-checkpoints` for usage.

xattn-perceiver-p05/model.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85be3206cf0ce4625056a2f006d8c5767bbb941b8d70093516cd83849add75c8
+size 715792736

xattn-perceiver-p05/training_config.yaml

@@ -0,0 +1,119 @@
+# Training Configuration - PixArtGEXAttn-B with Perceiver Token Source
+# Dropout p=0.5 — equal weight across all 4 conditioning regimes.
+# bf16-mixed + torch.compile for faster training.
+
+output_dir: "/cluster/work/grlab/projects/projects2025-virtual-tissue-gen/scratch/10x_TuPro/PixCell-GE/training/pixart-ge-cf-B-xattn-perceiver-p05"
+device: "cuda"
+
+# ============================================================================
+# MODEL
+# ============================================================================
+
+model:
+  type: "pixart_ge_xattn"
+  variant: "B"  # 130M params: depth=12, hidden=768, heads=12
+  ge_encoder_type: "cancerfoundation"
+  ge_hidden_dim: 512
+  cf_model_dir: "/cluster/home/pvlachas/leomed-home/pretrained_model_weights/cancer-foundation"
+  cf_freeze_backbone: true
+  ge_token_source: "perceiver"
+  ge_num_tokens: 32
+  ge_perceiver_depth: 4
+  ge_xattn_fusion: "xattn_only"
+
+# ============================================================================
+# DATA
+# ============================================================================
+
+data:
+  features_dir: "/cluster/work/grlab/projects/projects2025-virtual-tissue-gen/scratch/10x_TuPro/feat-extraction/features_train"
+  load_gene_expression: true
+  load_gene_expression_binned: true
+  num_workers: 8
+  pin_memory: true
+  val_split: 0.1
+
+# ============================================================================
+# DIFFUSION
+# ============================================================================
+
+diffusion:
+  timesteps: 1000
+  beta_schedule: "linear"
+  image_size: 256
+  latent_size: 32
+
+# ============================================================================
+# TRAINING
+# ============================================================================
+
+training:
+  batch_size: 32
+  batch_size_val: 32
+  gradient_accumulation_steps: 4  # effective batch = 128
+  num_epochs: 1000
+  seed: 42
+  gradient_clip: 0.01
+  ema_rate: 0.9999
+
+  optimizer:
+    lr: 2e-5
+    weight_decay: 0.01
+    betas: [0.9, 0.999]
+
+  scheduler:
+    warmup_steps: 1000
+    min_lr_ratio: 0.1
+
+  classifier_free_guidance:
+    conditioning_schedule:
+      - mask: [uni, ge]   # full conditioning (UNI + GE active)
+        weight: 25
+      - mask: [ge]        # GE only (UNI dropped)
+        weight: 25
+      - mask: [uni]       # UNI only (GE dropped)
+        weight: 25
+      - mask: []          # unconditional (both dropped)
+        weight: 25
+
+  modality_monitor:
+    enabled: true
+    diagnostic_freq: 10
+    diagnostic_batch_size: 64
+
+  convergence:
+    monitor_timestep_range: [900, 1000]
+    patience: 50
+    min_epochs: 300
+
+# ============================================================================
+# DISTRIBUTED
+# ============================================================================
+
+distributed:
+  precision: "bf16-mixed"
+  compile_model: true
+
+# ============================================================================
+# CHECKPOINT
+# ============================================================================
+
+checkpoint:
+  save_every: 1000
+  resume: null
+
+# ============================================================================
+# LOGGING
+# ============================================================================
+
+logging:
+  log_every: 100
+  validate_every: 0
+  gpu_monitor: true
+  gpu_monitor_interval: 60.0
+  sample_every_epochs: 10
+  sample_every_steps: 0
+  num_samples: 16
+  sample_guidance_scale: 3.0
+  sample_num_steps: 20
+  sample_vae_path: "/cluster/home/pvlachas/leomed-home/pretrained_model_weights/stability-ai-stable-diffusion-3-5-large/models--stabilityai--stable-diffusion-3.5-large/snapshots/ceddf0a7fdf2064ea28e2213e3b84e4afa170a0f/vae"