• 电脑设置、python 入门
  • NoMachine, ssh, python, conda, jupyter
  • 文件夹操作 pathlib, 图片操作 skimage
  • 数据增强 (data augmentation): imgaug
  • TensorBoard
• 机器学习简介
  • Linear Classifier
  • sklearn**.**model_selection**.**train_test_split()
  • Stochastic gradient descent
  • TensorFlow:
    • tensorflow_addons as tfa
    • tfa.image.rotate(), tf.image.random_flip_left_right()
    • **from** tensorflow.keras **import** Model, **from** tensorflow.keras.models **import** Sequential, **from** tensorflow.keras.layers **import** Input, Flatten, Dense, Activation, BatchNormalization, Conv2D, MaxPool2D, Softmax
    • tf**.**keras**.**losses**.**CategoricalCrossentropy(), tf**.**keras**.**optimizers**.**Adam(lr**=**1e-3, clipnorm**=**0.001)
    • linear_classifier **=**Model(...), linear_classifier.compile(), linear_classifier.fit(), linear_classifier.predict()
• 深度学习简介
  • Perceptron
  • Perceptron-based XOR gate
  • decision boundary of your model: np.meshgrid
• 图像恢复 (image restoration)
  • CARE network
  • Noise2Nosie, Noise2Void
• 图像翻译 (image translation)
  • micro-DL: a tool to generate and train U-net from config files.
• 图像语义分割 (image semantic segmentation): 比较详细，前两节有点水了。
  • **from** PIL **import** Image, **import** imageio, **from** torchvision **import** transforms
  • **from** torch.utils.data **import** Dataset, DataLoader, **import** torch.nn **as** nn, **from** torch.nn **import** functional **as** F, **from** torch.utils.tensorboard **import** SummaryWriter,
  • U-net on PyTorch
• 图像实例分割 (instance segmentation)
  1. Foreground segmentation:
     • Receptive Field of View: The term is borrowed from biology where it describes the “portion of sensory space that can elicit neuronal responses when stimulated” (wikipedia). Each output pixel can look at/depends on an input patch with that diameter centered at its position. Based on this patch, the network has to be able to make a decision about the prediction for the respective pixel.
     • Early Stopping to avoid overfitting: define an EarlyStopping class
  2. Instance Segmentation:
     • Ideas:
       • Three-class model (foreground, background, boundary),
       • Distance transform (label for each pixel is the distance to the closest boundary),
       • Edge affinity (consider not just the pixel but also its direct neighbors, predicts the probability that there is an edge, this is called affinity.) 听的时候懂了，回来看的时候没太看懂
       • Metric learning (learns to predict an embedding vector for each pixel.)
  3. Tile and Stitch：
     • 当需要处理的图片过大时，将图片切分成多个小图，分别预测之后拼接在一起。
     • 文中说图片尺寸不是 某个参数的整数倍的时候拼贴结果会不连续，但是代码注释中说等于这个整数倍的时候会不连续，晕。
     • https://arxiv.org/pdf/2101.05846.pdf
  4. 一个实例，epithelia cells
• 失败模式：极其之水，就是科普了一下训练参数错误的后果，以及一点对抗学习的内容
• 追踪：比较水，因为机器学习追踪的运算量极大，且主讲人感觉就是来做广告的，所以就直接用 CoLab 体验了一下就完事了。（就这还加州理工呢~）
• 知识提取：
  • 前面的基本上是从像素到像素的映射，这里的知识提取是从图片到标签的映射。
  • CycleGAN
  • Create a balanced Dataloader