Merge branch 'main' of https://github.com/PaddleCV-SIG/iann

README.md

@@ -29,11 +29,7 @@ python iann/__main__.py
 - [x] 解决图片放大后缩小，顶部定格显示不能居中的问题
 - [x] 解决放大图像后平移速度变快
 - [x] 解决图像切换后窗口缩放没有清除（顺便缩放到显示框最佳）
-
-## 新提交
-
-1. 模型未加载时直接打开文件夹会报错，发现在打开文件夹直接加载模型前就访问了controller的属性，而模型加载时才创建的controller；这里改了一下代码的顺序。
-2. 直接加载模型会报错，发现在加载模型时默认打开的图像，当未选择时图像为None；这里加了一个判断，如果直接加载模型，当图像为None就不加载。
-3. 每次加载图像时根据图像大小重新设定scene的区域，这样缩小图像和切换不会造成不能居中的问题。
-4. 添加最佳缩放，每次切换图片后，自适应缩放到最佳显示。
-5. update_image时判断是否应该缩放重置，避免在点击点时也进行了图像缩放重置。
+- [ ] 训练代码整理和训练配置的抽出（初步实现）
+- [ ] 重新审查修改按键图标，确保图标完整清晰
+- [ ] 界面配色以paddle蓝色为主，同时做一个ico窗口图标
+- [ ] 标签可修改？

iann/data/mdiy.py

@@ -0,0 +1,44 @@
+from pathlib import Path
+
+import cv2
+import numpy as np
+
+from .base import ISDataset
+
+
+class MyDataset(ISDataset):
+    def __init__(self, dataset_path, folder_name,
+                 images_dir_name, masks_dir_name,
+                 **kwargs):
+        super(MyDataset, self).__init__(**kwargs)
+
+        self.dataset_path = Path(dataset_path) / folder_name
+        self._images_path = self.dataset_path / images_dir_name
+        self._insts_path = self.dataset_path / masks_dir_name
+
+        self.dataset_samples = [x.name for x in sorted(self._images_path.glob('*.*'))]
+        self._masks_paths = {x.stem: x for x in self._insts_path.glob('*.*')}
+
+    def get_sample(self, index):
+        image_name = self.dataset_samples[index]
+        image_path = str(self._images_path / image_name)
+        mask_path = str(self._masks_paths[image_name.split('.')[0]])
+
+        image = cv2.imread(image_path)
+        image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
+        instances_mask = np.max(cv2.imread(mask_path).astype(np.int32), axis=2)
+        instances_mask[instances_mask > 0] = 1
+
+        instances_ids = [1]
+
+        instances_info = {
+            x: {'ignore': False}
+            for x in instances_ids
+        }
+
+        return {
+            'image': image,
+            'instances_mask': instances_mask,
+            'instances_info': instances_info,
+            'image_id': index
+        }

iann/model/loss.py

@@ -0,0 +1,133 @@
+import numpy as np
+import paddle
+import paddle.nn as nn
+import paddle.nn.functional as F
+import util.util as U
+
+
+class NormalizedFocalLossSigmoid(nn.Layer):
+    def __init__(self, axis=-1, alpha=0.25, gamma=2,
+                 from_logits=False, batch_axis=0,
+                 weight=None, size_average=True, detach_delimeter=True,
+                 eps=1e-12, scale=1.0,
+                 ignore_label=-1):
+        super(NormalizedFocalLossSigmoid, self).__init__()
+        self._axis = axis
+        self._alpha = alpha
+        self._gamma = gamma
+        self._ignore_label = ignore_label
+        self._weight = weight if weight is not None else 1.0
+        self._batch_axis = batch_axis
+
+        self._scale = scale
+        self._from_logits = from_logits
+        self._eps = eps
+        self._size_average = size_average
+        self._detach_delimeter = detach_delimeter
+        self._k_sum = 0
+
+    def forward(self, pred, label, sample_weight=None):
+        one_hot = label > 0
+        sample_weight = label != self._ignore_label
+        sample_weight = sample_weight.astype('float32')
+
+        if not self._from_logits:
+            pred = F.sigmoid(pred)
+
+        x = sample_weight * 0.5
+        y = (1 - self._alpha) * sample_weight
+        alpha = paddle.where(one_hot, x, y)
+        pt = paddle.where(one_hot, pred, 1 - pred)
+        sample_weight = sample_weight.astype('bool')
+        pt = paddle.where(sample_weight, pt, paddle.ones_like(pt))
+        beta = (1 - pt) ** self._gamma
+        sample_weight = sample_weight.astype('float32')
+        sw_sum = paddle.sum(sample_weight, axis=(-2, -1), keepdim=True)
+        beta_sum = paddle.sum(beta, axis=(-2, -1), keepdim=True)
+        mult = sw_sum / (beta_sum + self._eps)
+
+        if self._detach_delimeter:
+            mult = mult.detach()
+
+        beta = beta * mult
+        ignore_area = paddle.sum((label == self._ignore_label).astype('float32'), axis=tuple(range(1, len(label.shape)))).numpy()
+        sample_mult = paddle.mean(mult, axis=tuple(range(1, len(mult.shape)))).numpy()
+        if np.any(ignore_area == 0):
+            self._k_sum = 0.9 * self._k_sum + 0.1 * sample_mult[ignore_area == 0].mean()
+        loss = -alpha * beta * paddle.log(paddle.mean(pt + self._eps))
+        loss = self._weight * (loss * sample_weight)
+
+        if self._size_average:
+            bsum = paddle.sum(sample_weight, axis=U.get_dims_with_exclusion(len(sample_weight.shape), self._batch_axis))
+            loss = paddle.sum(loss, axis=U.get_dims_with_exclusion(len(loss.shape), self._batch_axis)) / (
+                    bsum + self._eps)
+        else:
+            loss = paddle.sum(loss, axis=U.get_dims_with_exclusion(len(loss.shape), self._batch_axis))
+
+        return self._scale * loss
+
+
+class FocalLoss(nn.Layer):
+    def __init__(self, axis=-1, alpha=0.25, gamma=2,
+                 from_logits=False, batch_axis=0,
+                 weight=None, num_class=None,
+                 eps=1e-9, size_average=True, scale=1.0):
+        super(FocalLoss, self).__init__()
+        self._axis = axis
+        self._alpha = alpha
+        self._gamma = gamma
+        self._weight = weight if weight is not None else 1.0
+        self._batch_axis = batch_axis
+
+        self._scale = scale
+        self._num_class = num_class
+        self._from_logits = from_logits
+        self._eps = eps
+        self._size_average = size_average
+
+    def forward(self, pred, label, sample_weight=None):
+        if not self._from_logits:
+            pred = F.sigmoid(pred)
+
+        one_hot = label > 0
+        pt = paddle.where(one_hot, pred, 1 - pred)
+        t = label != -1
+        alpha = paddle.where(one_hot, self._alpha * t, (1 - self._alpha) * t)
+        beta = (1 - pt) ** self._gamma
+
+        loss = -alpha * beta * paddle.log(paddle.min(pt + self._eps, paddle.ones(1, dtype='float32')))
+        sample_weight = label != -1
+
+        loss = self._weight * (loss * sample_weight)
+
+        if self._size_average:
+            tsum = paddle.sum(label == 1, axis=U.get_dims_with_exclusion(len(label.shape), self._batch_axis))
+            loss = paddle.sum(loss, axis=U.get_dims_with_exclusion(len(loss.shape), self._batch_axis)) / (
+                    tsum + self._eps)
+        else:
+            loss = paddle.sum(loss, axis=U.get_dims_with_exclusion(len(loss.shape), self._batch_axis))
+
+        return self._scale * loss
+
+
+class SigmoidBinaryCrossEntropyLoss(nn.Layer):
+    def __init__(self, from_sigmoid=False, weight=None, batch_axis=0, ignore_label=-1):
+        super(SigmoidBinaryCrossEntropyLoss, self).__init__()
+        self._from_sigmoid = from_sigmoid
+        self._ignore_label = ignore_label
+        self._weight = weight if weight is not None else 1.0
+        self._batch_axis = batch_axis
+
+    def forward(self, pred, label):
+        label = label.reshape(pred.shape)
+        sample_weight = label != self._ignore_label
+        label = paddle.where(sample_weight, label, paddle.zeros_like(label))
+
+        if not self._from_sigmoid:
+            loss = F.relu(pred) - pred * label + F.softplus(-paddle.abs(pred))
+        else:
+            eps = 1e-12
+            loss = -(paddle.log(pred + eps) * label + paddle.log(1. - pred + eps) * (1. - label))
+        loss = self._weight * (loss * sample_weight)
+        return paddle.mean(loss, axis=U.get_dims_with_exclusion(len(loss.shape), self._batch_axis))
+

iann/train/README.md

@@ -0,0 +1,81 @@
+# 训练iann可用的自定义模型
+
+目前已经可以通过简单的配置完成模型训练了，但其中有些设置还不能通过配置文件进行修改。
+
+## 一、数据组织
+
+在需要训练自己的数据集时，目前需要将数据集构造为如下格式，直接放在datasets文件夹中。文件名可以根据要求来进行设置，只需要在配置文件中设定好即可，图像和标签与平时使用的分割图像的用法相同。
+
+```
+datasets
+    |
+    ├── train_data
+    |       ├── img
+    |       |    └── filename_1.jpg
+    |       └── gt
+    |            └── filename_1.png
+    |
+    └── eval_data
+            ├── img
+            |    └── filename_1.jpg
+            └── gt
+                 └── filename_1.png
+```
+
+## 二、训练
+
+直接运行ritm_train.py即可开始训练。
+
+```python
+%cd train
+! python ritm_train.py --config train_config.yaml
+```
+
+目前一些简单的参数已经可以在yaml配置文件中进行自定义设置，不过现阶段仍然不够灵活，可能出现各种问题。
+
+```
+iters: 100000  # 训练轮数
+batch_size: 16  # bs大小
+save_interval: 1000  # 保存间隔
+log_iters: 10  # 打印log的间隔
+worker: 4  # 子进程数
+save_dir: model_output  # 保存路径
+use_vdl: False  # 是否使用vdl
+
+dataset:
+  dataset_path: iann/train/datasets  # 数据集所在路径
+  image_name: img  # 图像文件夹的名称
+  label_name: gt  # 标签文件夹的名称
+
+train_dataset:  # 训练数据
+  crop_size: [320, 480]  # 裁剪大小
+  folder_name: train_data  # 训练数据文件夹的名称
+
+val_dataset:  # 验证数据
+  folder_name: val_data  # 验证数据文件夹的名称
+
+optimizer:
+  type: adam  # 优化器，目前仅可以选择‘adam’和‘sgd’
+
+learning_rate:
+  value_1: 5e-5  # 需要设置两个学习率
+  value_2: 5e-6
+  decay:
+    type: poly  # 学习率衰减，目前仅支持‘poly’，可以修改下面的参数
+    steps: 1000
+    power: 0.9
+    end_lr: 0.0
+
+model:
+  type: deeplab  # 模型名称，目前支持‘hrnet’、‘deeplab’以及‘shufflenet’
+  backbone: resnet18  # 下面的参数是模型对应的参数，可在源码中查看
+  is_ritm: True
+  weights: None  # 加载权重的路径
+```
+
+
+
+### * 说明
+
+1. 这里有个坑，数据不能有没有标签的纯背景，这样找不到正样点训练就会卡住，并且还不报错。
+