File size: 3,787 Bytes
497c818 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 | # Minimal Code Package For My PixelDiT Three-Control Network
This folder contains extracted useful code from the current project. It is not just a prose document. It is a small Python package that implements the core innovations:
- independent `depth / seg / edge` control branches
- strict single-condition hard selection
- multi-condition layer-wise gated fusion
- DDP-safe mode sampling
- inactive branch gradient masking
- single-control and three-control dataset loading
- multi-condition cycle loss dispatch
- SoftCanny image-cycle edge consistency
## Files
```text
minimal_my_network/
__init__.py
independent_gated_control.py
datasets.py
losses.py
README.md
```
## Core Model Code
Use:
```python
from minimal_my_network import IndependentBranchGatedFusion
```
Create fusion module:
```python
fusion = IndependentBranchGatedFusion(
hidden_size=1536,
num_layers=14,
init_gate_logits=(0.5, 0.0, -0.5),
control_structure_inject=(True, True, False),
alpha_inject=2.0,
)
```
Inside a PixelDiT block loop, after you compute branch tokens:
```python
x = fusion(
hidden=x,
layer_idx=inject_idx,
branch_tokens=[depth_tokens, seg_tokens, edge_tokens],
keep_mask=control_keep, # [B, 3]
branch_structure_maps=[depth_struct, seg_struct, edge_struct],
)
```
Behavior is exactly:
```text
depth-only: uses depth branch only, gate ignored
seg-only: uses seg branch only, gate ignored
edge-only: uses edge branch only, gate ignored
multi-control: masked softmax gate over active branches only
```
## Training Utilities
```python
from minimal_my_network import (
apply_multi_control_mode,
sample_control_mode_ddp,
mask_inactive_control_grads,
)
```
Sample one mode per step:
```python
mode = sample_control_mode_ddp(
modes=("depth", "seg", "edge", "depth_seg", "depth_edge", "seg_edge", "depth_seg_edge"),
probs=(0.15, 0.15, 0.15, 0.12, 0.12, 0.12, 0.19),
enable_dropout=True,
device=device,
)
```
Apply sampled mode:
```python
control, control_keep = apply_multi_control_mode(control, mode, num_controls=3)
```
After backward:
```python
mask_inactive_control_grads(model, mode)
```
## Dataset Code
Three-control dataset:
```python
from minimal_my_network.datasets import PixelThreeControlDataset, subdir_range
ds = PixelThreeControlDataset(
image_root="data/blip/extracted",
depth_root="data/blip_depth_da3_nested_giant_large_1_1",
seg_root="data/blip_sam2_large_extracted",
edge_root="data/blip_edge",
subdirs=subdir_range(0, 199),
)
```
Single-control dataset:
```python
from minimal_my_network.datasets import PixelSingleControlDataset, subdir_range
seg_ds = PixelSingleControlDataset(
image_root="data/blip/extracted",
control_root="data/blip_sam2_large_extracted",
control_type="seg",
subdirs=subdir_range(0, 199),
)
```
## Loss Code
```python
from minimal_my_network import MultiConditionCycleLoss, SoftCannyImagePyramidCycleLoss
edge_loss = SoftCannyImagePyramidCycleLoss(
gaussian_kernel=11,
threshold_min=0.2745,
threshold_max=0.5882,
temperature=0.03,
)
cycle = MultiConditionCycleLoss(
depth_cycle_loss=depth_loss,
seg_cycle_loss=seg_loss,
edge_cycle_loss=edge_loss,
depth_weight=1.0,
seg_weight=1.0,
edge_weight=1.0,
)
```
Call:
```python
loss = cycle(
gen_image_m11,
depth_01=depth,
seg_01=seg,
gt_image_m11=gt_image_m11,
control_mode=mode,
)
```
## What Is Not Included
This folder intentionally does not copy the full PixelDiT backbone. You should keep using the original backbone from:
```text
pixdit_core/pixeldit.py
pixdit_core/pixeldit_t2i_control.py
```
This minimal package contains the transferable innovation code that Codex can reuse in another implementation.
|