---
tags:
- seismic
- first-break-picking
- segmentation
- pytorch
library_name: pytorch
---

# First-Break Picking HF Release

This directory is a compact release package for first-break-picking model
weights. It keeps only the best checkpoint and the exact training config for
each run.

## Contents

- 60 `best.pt` checkpoint files.
- 60 `config.yaml` files.
- No intermediate `epoch_*.pt` checkpoints.
- Current package size: about 4.1 GB.

Two experiment groups are included:

| Directory | Meaning | Runs |
| --- | --- | ---: |
| `first_break_picking/` | Models trained on all available first-break SEG-Y pairs together | 12 |
| `first_break_picking_single_dataset_grouped/by_dataset/` | Models trained separately on each individual SEG-Y pair | 48 |

## Directory Layout

```text
first_break_picking_hf_release/
  README.md
  first_break_picking/
    first_break_pick_<model>_geomseg_seed<seed>/
      config.yaml
      checkpoints/
        best.pt
  first_break_picking_single_dataset_grouped/
    by_dataset/
      <dataset>/
        <model>/
          seed<seed>/
            config.yaml
            checkpoints/
              best.pt
```

The multi-dataset runs use:

```text
models: unet, res_unet, atten_unet, dncnn_seg
seeds: 42, 43, 44
```

The single-dataset runs use:

```text
datasets: brunswick_valid, dongbei, halfmile_valid, lalor_valid
models: unet, res_unet, atten_unet, dncnn_seg
seeds: 42, 43, 44
```

## Task Definition

The benchmark treats first-break picking as binary step-mask segmentation.

- Input: single-channel SEG-Y amplitude patches.
- Label: binary step mask, with 0 before the first break and 1 from the
  first-break sample onward.
- Prediction: single-channel logits with the same spatial shape as the mask.
- Pick extraction: the first time index where `sigmoid(logit) >= 0.5`.

## Dataset Configuration

All configs point to:

```text
data.root: /home/dataset-local/dataset/first_break_picking/segy_with_masks
data.data_dir: data
data.label_dir: label
```

The available SEG-Y input files are:

```text
Brunswick_valid.sgy
Dongbei.segy
Halfmile_valid.sgy
Lalor_valid.sgy
```

For multi-dataset runs, `data.files: null`, so all SEG-Y pairs under `data/`
are used. For single-dataset runs, `data.files` contains exactly one input
SEG-Y filename.

Common data parameters:

| Parameter | Value |
| --- | --- |
| `label_threshold` | `0.5` |
| `prediction_threshold` | `0.5` |
| `validate_labels` | `true` |
| `label_check_traces` | `2048` |
| `max_patches_per_split` | `null` |
| `split.train` | `0.8` |
| `split.val` | `0.1` |
| `split.test` | `0.1` |
| `split.shuffle_ffids` | `true` |

Gather segmentation parameters:

| Parameter | Value |
| --- | --- |
| `gather_segment.enabled` | `true` |
| `gather_segment.line_id_header` | `INLINE_3D` |
| `gather_segment.infer_line_from_geometry` | `true` |
| `gather_segment.distance_floor` | `1000.0` |
| `gather_segment.median_multiplier` | `5.0` |

Patch and loader parameters:

| Parameter | Multi-dataset | Single-dataset |
| --- | ---: | ---: |
| `patch.trace` | `128` | `128` |
| `patch.time` | `512` | `512` |
| `patch.trace_stride` | `64` | `64` |
| `patch.time_stride` | `256` | `256` |
| `loader.batch_size` | `64` | `64` |
| `loader.num_workers` | `4` | `1` |
| `loader.pin_memory` | `true` | `true` |

## Preprocessing

| Parameter | Value |
| --- | --- |
| `normalize_mode` | `max_abs` |
| `normalize_scope` | `gather` |
| `clip_percentile` | `99.5` |
| `normalize_eps` | `1.0e-6` |

Only input amplitudes are normalized. Labels remain binary segmentation
targets; invalid or padded areas are ignored by the loss and metrics in the
training code.

## Models

| Model | Parameters |
| --- | --- |
| `unet` | `in_channels=1`, `out_channels=1`, `base_channels=32`, `depth=4` |
| `res_unet` | `in_channels=1`, `out_channels=1`, `base_channels=32`, `depth=4` |
| `atten_unet` | `in_channels=1`, `out_channels=1`, `base_channels=32`, `depth=4` |
| `dncnn_seg` | `in_channels=1`, `out_channels=1`, `depth=17`, `base_channels=64`, `kernel_size=3` |

## Training Parameters

| Parameter | Value |
| --- | --- |
| `train.epochs` | `20` |
| `train.grad_clip` | `1.0` |
| `train.log_interval` | `20` |
| `train.eval_interval` | `1` |
| `train.ckpt_interval` | `1` |
| `train.vis_interval` | `1` |
| `train.resume` | `null` |
| `optimizer` | `adamw` |
| `optimizer.lr` | `1.0e-4` |
| `optimizer.weight_decay` | `1.0e-5` |
| `scheduler` | `cosine` |
| `scheduler.min_lr` | `1.0e-6` |

Loss:

```text
type: bce_dice
bce_weight: 0.5
dice_weight: 0.5
smooth: 1.0
pos_weight: null
```

Metrics:

```text
dice
iou
f1
HitRate1px
HitRate3px
HitRate5px
HitRate7px
HitRate9px
MeanAbsoluteError
RootMeanSquaredError
MeanBiasError
GatherCoverage
```

All metric thresholds are `0.5`.

## Loading A Checkpoint

Use the matching `config.yaml` next to each `best.pt` to reconstruct the model
and preprocessing settings.

```python
import torch

checkpoint_path = "first_break_picking/first_break_pick_unet_geomseg_seed42/checkpoints/best.pt"
checkpoint = torch.load(checkpoint_path, map_location="cpu")
print(checkpoint.keys())
```

The `experiment.output_dir` values inside `config.yaml` are the original
training output paths. They are preserved for reproducibility and do not need
to match this release directory.

## Notes For Hugging Face Upload

This folder is intended to be uploaded as a model-weight release package. The
raw SEG-Y dataset should be uploaded separately because it is much larger and
has a different structure:

```text
segy_with_masks/
  data/
  label/
```

For this release package, upload the entire `first_break_picking_hf_release/`
directory so each `best.pt` remains next to its exact `config.yaml`.

---

# 初至拾取 Hugging Face 发布包

该目录是初至拾取模型权重的精简发布包。每一次实验只保留最优模型
checkpoint 和对应的完整训练配置。

## 内容概览

- 60 个 `best.pt` checkpoint 文件。
- 60 个 `config.yaml` 配置文件。
- 不包含中间训练轮次的 `epoch_*.pt` checkpoint。
- 当前发布包大小约 4.1 GB。

包含两类实验：

| 目录 | 含义 | 运行数量 |
| --- | --- | ---: |
| `first_break_picking/` | 使用全部初至拾取 SEG-Y 数据联合训练的模型 | 12 |
| `first_break_picking_single_dataset_grouped/by_dataset/` | 在单个 SEG-Y 数据集上分别训练的模型 | 48 |

## 目录结构

```text
first_break_picking_hf_release/
  README.md
  first_break_picking/
    first_break_pick_<model>_geomseg_seed<seed>/
      config.yaml
      checkpoints/
        best.pt
  first_break_picking_single_dataset_grouped/
    by_dataset/
      <dataset>/
        <model>/
          seed<seed>/
            config.yaml
            checkpoints/
              best.pt
```

联合训练实验包含：

```text
模型: unet, res_unet, atten_unet, dncnn_seg
随机种子: 42, 43, 44
```

单数据集训练实验包含：

```text
数据集: brunswick_valid, dongbei, halfmile_valid, lalor_valid
模型: unet, res_unet, atten_unet, dncnn_seg
随机种子: 42, 43, 44
```

## 任务定义

该 benchmark 将初至拾取建模为二值 step-mask 分割任务。

- 输入：单通道 SEG-Y 振幅 patch。
- 标签：二值 step mask，初至之前为 0，从初至采样点开始为 1。
- 预测：与标签空间尺寸相同的单通道 logits。
- 拾取点提取：取 `sigmoid(logit) >= 0.5` 的第一个时间采样点作为初至位置。

## 数据配置

所有配置文件都指向：

```text
data.root: /home/dataset-local/dataset/first_break_picking/segy_with_masks
data.data_dir: data
data.label_dir: label
```

可用的 SEG-Y 输入文件为：

```text
Brunswick_valid.sgy
Dongbei.segy
Halfmile_valid.sgy
Lalor_valid.sgy
```

联合训练实验中，`data.files: null`，表示使用 `data/` 下所有 SEG-Y
数据对。单数据集实验中，`data.files` 只包含一个输入 SEG-Y 文件名。

通用数据参数：

| 参数 | 取值 |
| --- | --- |
| `label_threshold` | `0.5` |
| `prediction_threshold` | `0.5` |
| `validate_labels` | `true` |
| `label_check_traces` | `2048` |
| `max_patches_per_split` | `null` |
| `split.train` | `0.8` |
| `split.val` | `0.1` |
| `split.test` | `0.1` |
| `split.shuffle_ffids` | `true` |

炮集与接收线切分参数：

| 参数 | 取值 |
| --- | --- |
| `gather_segment.enabled` | `true` |
| `gather_segment.line_id_header` | `INLINE_3D` |
| `gather_segment.infer_line_from_geometry` | `true` |
| `gather_segment.distance_floor` | `1000.0` |
| `gather_segment.median_multiplier` | `5.0` |

patch 与 DataLoader 参数：

| 参数 | 联合训练 | 单数据集训练 |
| --- | ---: | ---: |
| `patch.trace` | `128` | `128` |
| `patch.time` | `512` | `512` |
| `patch.trace_stride` | `64` | `64` |
| `patch.time_stride` | `256` | `256` |
| `loader.batch_size` | `64` | `64` |
| `loader.num_workers` | `4` | `1` |
| `loader.pin_memory` | `true` | `true` |

## 预处理

| 参数 | 取值 |
| --- | --- |
| `normalize_mode` | `max_abs` |
| `normalize_scope` | `gather` |
| `clip_percentile` | `99.5` |
| `normalize_eps` | `1.0e-6` |

只对输入振幅做归一化。标签保持二值分割目标；无效区域或 padding 区域在训练代码中会被 loss 和 metrics 忽略。

## 模型

| 模型 | 参数 |
| --- | --- |
| `unet` | `in_channels=1`, `out_channels=1`, `base_channels=32`, `depth=4` |
| `res_unet` | `in_channels=1`, `out_channels=1`, `base_channels=32`, `depth=4` |
| `atten_unet` | `in_channels=1`, `out_channels=1`, `base_channels=32`, `depth=4` |
| `dncnn_seg` | `in_channels=1`, `out_channels=1`, `depth=17`, `base_channels=64`, `kernel_size=3` |

## 训练参数

| 参数 | 取值 |
| --- | --- |
| `train.epochs` | `20` |
| `train.grad_clip` | `1.0` |
| `train.log_interval` | `20` |
| `train.eval_interval` | `1` |
| `train.ckpt_interval` | `1` |
| `train.vis_interval` | `1` |
| `train.resume` | `null` |
| `optimizer` | `adamw` |
| `optimizer.lr` | `1.0e-4` |
| `optimizer.weight_decay` | `1.0e-5` |
| `scheduler` | `cosine` |
| `scheduler.min_lr` | `1.0e-6` |

损失函数：

```text
type: bce_dice
bce_weight: 0.5
dice_weight: 0.5
smooth: 1.0
pos_weight: null
```

评价指标：

```text
dice
iou
f1
HitRate1px
HitRate3px
HitRate5px
HitRate7px
HitRate9px
MeanAbsoluteError
RootMeanSquaredError
MeanBiasError
GatherCoverage
```

所有指标阈值均为 `0.5`。

## 加载 Checkpoint

使用每个 `best.pt` 旁边对应的 `config.yaml` 来重建模型和预处理设置。

```python
import torch

checkpoint_path = "first_break_picking/first_break_pick_unet_geomseg_seed42/checkpoints/best.pt"
checkpoint = torch.load(checkpoint_path, map_location="cpu")
print(checkpoint.keys())
```

`config.yaml` 中的 `experiment.output_dir` 是原始训练时的输出路径。这里保留这些路径是为了可复现性，不要求它们与当前发布目录一致。

## Hugging Face 上传说明

该目录适合作为模型权重发布包上传。原始 SEG-Y 数据集体积更大，结构也不同，建议单独上传：

```text
segy_with_masks/
  data/
  label/
```

上传当前权重发布包时，应上传整个 `first_break_picking_hf_release/`
目录，以保证每个 `best.pt` 都和对应的 `config.yaml` 保持在一起。