init work on training pipeline

Author: dmitryduev

scripts/train.py

@@ -0,0 +1,255 @@
+#!/usr/bin/env python
+import fire
+import matplotlib.pyplot as plt
+import numpy as np
+
+# import pandas as pd
+# import pathlib
+# from sklearn.model_selection import train_test_split
+# import sys
+import tensorflow as tf
+
+# from tqdm.auto import tqdm
+
+import tails.models
+from tails.utils import log
+
+
+def fnr_vs_fpr(predictions, ground_truth):
+    rbbins = np.arange(-0.0001, 1.0001, 0.0001)
+    h_b, e_b = np.histogram(predictions[ground_truth == 0], bins=rbbins, density=True)
+    h_b_c = np.cumsum(h_b)
+    h_r, e_r = np.histogram(predictions[ground_truth == 1], bins=rbbins, density=True)
+    h_r_c = np.cumsum(h_r)
+
+    # h_b, e_b
+    print(sum(ground_truth == 0), sum(ground_truth == 1))
+
+    fig = plt.figure(figsize=(9, 4), dpi=200)
+    ax = fig.add_subplot(111)
+
+    rb_thres = np.array(list(range(len(h_b)))) / len(h_b)
+
+    ax.plot(
+        rb_thres,
+        h_r_c / np.max(h_r_c),
+        label="False Negative Rate (FNR)",
+        linewidth=1.5,
+    )
+    ax.plot(
+        rb_thres,
+        1 - h_b_c / np.max(h_b_c),
+        label="False Positive Rate (FPR)",
+        linewidth=1.5,
+    )
+
+    mmce = (h_r_c / np.max(h_r_c) + 1 - h_b_c / np.max(h_b_c)) / 2
+    ax.plot(
+        rb_thres,
+        mmce,
+        "--",
+        label="Mean misclassification error",
+        color="gray",
+        linewidth=1.5,
+    )
+
+    ax.set_xlim([-0.05, 1.05])
+
+    ax.set_xticks(np.arange(0, 1.1, 0.1))
+    ax.set_yticks(np.arange(0, 1.1, 0.1))
+
+    # vals = ax.get_yticks()
+    # ax.set_yticklabels(['{:,.0%}'.format(x) for x in vals])
+
+    ax.set_yscale("log")
+    ax.set_ylim([5e-4, 1])
+    vals = ax.get_yticks()
+    ax.set_yticklabels(
+        ["{:,.1%}".format(x) if x < 0.01 else "{:,.0%}".format(x) for x in vals]
+    )
+
+    # thresholds:
+    # thrs = [0.5, ]
+    thrs = [0.5, 0.7]
+    for t in thrs:
+        m_t = rb_thres < t
+        fnr = np.array(h_r_c / np.max(h_r_c))[m_t][-1]
+        fpr = np.array(1 - h_b_c / np.max(h_b_c))[m_t][-1]
+        print(t, fnr * 100, fpr * 100)
+        # ax.vlines(t_1, 0, 1.1)
+        ax.vlines(t, 0, max(fnr, fpr))
+        ax.text(
+            t - 0.05,
+            max(fnr, fpr) + 0.01,
+            f" {fnr*100:.1f}% FNR\n {fpr*100:.1f}% FPR",
+            fontsize=10,
+        )
+
+    ax.set_xlabel("$p_c$ score threshold")
+    ax.set_ylabel("Cumulative percentage")
+    ax.legend(loc="upper center")
+    ax.grid(True, which="major", linewidth=0.5)
+    ax.grid(True, which="minor", linewidth=0.3)
+    plt.tight_layout()
+    plt.show()
+
+
+class TailsLoss(tf.keras.losses.BinaryCrossentropy):
+    def __init__(self, w_1: float = 1, w_2: float = 1, **kwargs):
+        super(TailsLoss, self).__init__(**kwargs)
+        self.w_1 = w_1
+        self.w_2 = w_2
+
+    def call(self, y_true, y_pred):
+        output = tf.convert_to_tensor(y_pred[..., 0])
+        target = tf.cast(y_true[..., 0], output.dtype)
+
+        # l_1: binary crossentropy for the label
+        l_1 = super(TailsLoss, self).call(target, output)
+        w_1 = tf.cast(self.w_1, output.dtype)
+        l_1 = tf.math.multiply(l_1, w_1)
+
+        # l_2: L1 loss
+        l_2 = tf.norm(y_pred[..., 1:] - y_true[..., 1:], ord=1)
+
+        # l_2: L1 loss + L2 regularization
+        # l_2 = tf.norm(y_pred[..., 1:] - y_true[..., 1:], ord=1) + \
+        #       1e-3 * tf.norm(y_pred[..., 1:] - y_true[..., 1:], ord=2)
+
+        l_2 = tf.math.multiply(l_2, target)
+        l_2 = tf.math.divide(l_2, tf.math.reduce_sum(target))
+        w_2 = tf.cast(self.w_2, output.dtype)
+        l_2 = tf.math.multiply(l_2, w_2)
+
+        return l_1 + l_2
+
+
+class LabelAccuracy(tf.keras.metrics.Metric):
+    def __init__(self, name="label_accuracy", threshold=0.5, **kwargs):
+        super(LabelAccuracy, self).__init__(name=name, **kwargs)
+        self.total = self.add_weight(name="total", initializer="zeros")
+        self.count = self.add_weight(name="count", initializer="zeros")
+        self.threshold = float(threshold)
+
+    def update_state(self, y_true, y_pred, sample_weight=None):
+        output = y_pred[..., 0]
+        #         target = tf.cast(y_true[..., 0], output.dtype)
+        target = tf.cast(y_true[..., 0], tf.bool)
+
+        threshold = tf.cast(0.5, output.dtype)
+        output = tf.cast(output > threshold, tf.bool)
+
+        #         values = tf.cast(tf.math.equal(target, output), output.dtype)
+        values = tf.cast(tf.math.equal(target, output), tf.float32)
+        ones = tf.cast(tf.math.equal(target, target), tf.float32)
+
+        if sample_weight is not None:
+            sample_weight = tf.cast(sample_weight, self.dtype)
+            sample_weight = tf.broadcast_weights(sample_weight, values)
+            values = tf.multiply(values, sample_weight)
+
+        self.count.assign_add(tf.math.reduce_sum(values, axis=-1))
+        self.total.assign_add(tf.math.reduce_sum(ones, axis=-1))
+
+    def result(self):
+        return tf.math.divide(self.count, self.total)
+
+
+class PositionRootMeanSquarredError(tf.keras.metrics.Metric):
+    def __init__(self, name="position_rmse", scaling_factor=1, **kwargs):
+        super(PositionRootMeanSquarredError, self).__init__(name=name, **kwargs)
+        self.total = self.add_weight(name="total", initializer="zeros")
+        self.rmse = self.add_weight(name="rmse", initializer="zeros")
+        self.scaling_factor = float(scaling_factor)
+
+    def update_state(self, y_true, y_pred, sample_weight=None):
+        output = y_pred[..., 1:]
+        target = tf.cast(y_true[..., 1:], output.dtype)
+        label = tf.cast(y_true[..., 0], output.dtype)
+
+        rmse = tf.math.reduce_mean(
+            tf.math.sqrt(tf.math.squared_difference(output, target)), axis=-1
+        )
+        # only take positive examples into account:
+        rmse = tf.math.multiply(rmse, label)
+
+        self.rmse.assign_add(tf.math.reduce_sum(rmse, axis=-1))
+        # only count the positive examples:
+        self.total.assign_add(tf.math.reduce_sum(label, axis=-1))
+
+    def result(self):
+        sf = tf.constant(self.scaling_factor, dtype=self.rmse.dtype.base_dtype)
+        return tf.math.multiply(sf, tf.math.divide(self.rmse, self.total))
+
+
+def train_and_eval(
+    train_dataset,
+    val_dataset,
+    test_dataset,
+    steps_per_epoch_train,
+    steps_per_epoch_val,
+    epochs,
+    class_weight,
+    model_name: str = "tails",
+    tag="20210101",
+    w_1: float = 1.2,
+    w_2: float = 1,
+    class_threshold: float = 0.5,
+    scaling_factor=256,
+    input_shape=(256, 256, 3),
+    weights: str = None,
+    save_model=False,
+    verbose=False,
+    **kwargs,
+):
+    classifier = tails.models.DNN(name=model_name)
+
+    tails_loss = TailsLoss(name="loss", w_1=w_1, w_2=w_2)
+    label_accuracy = LabelAccuracy(threshold=class_threshold)
+    # convert position RMSE to pixels
+    position_rmse = PositionRootMeanSquarredError(scaling_factor=scaling_factor)
+
+    learning_rate = kwargs.get("learning_rate", 3e-4)
+    patience = kwargs.get("patience", 30)
+
+    classifier.setup(
+        input_shape=input_shape,
+        n_output_neurons=3,
+        architecture="tails",
+        loss=tails_loss,
+        optimizer="adam",
+        lr=learning_rate,  # epsilon=1e-3, beta_1=0.7,
+        metrics=[label_accuracy, position_rmse],
+        patience=patience,
+        monitor="val_position_rmse",
+        restore_best_weights=True,
+        callbacks=("early_stopping", "tensorboard"),
+        tag=tag,
+        logdir="logs",
+    )
+
+    # pre-load weights?
+    if weights is not None:
+        classifier.model.load_weights(weights)
+
+    classifier.train(
+        train_dataset,
+        val_dataset,
+        steps_per_epoch_train,
+        steps_per_epoch_val,
+        epochs=epochs,
+        class_weight=class_weight,
+        verbose=True,
+    )
+
+    # evaluate
+    stats = classifier.evaluate(test_dataset)
+    if verbose:
+        log(stats)
+
+    if save_model:
+        classifier.model.save_weights(f"{model_name}-{tag}")
+
+
+if __name__ == "__main__":
+    fire.Fire(train_and_eval)