tensorboard可视化

  1 import os
  2 import datetime
  3 import tensorflow as tf
  4 import matplotlib.pyplot as plt
  5 import io
  6 from tensorflow.keras import datasets, layers, Sequential, optimizers
  7 
  8 os.environ[‘TF_CPP_MIN_LOG_LEVEL‘] = ‘2‘
  9 tf.random.set_seed(2345)
 10 
 11 conv_layers = [
 12     # unit1
 13     layers.Conv2D(6, kernel_size=[5, 5], strides=1, padding=‘valid‘, activation=tf.nn.sigmoid),
 14     layers.MaxPool2D(pool_size=[2, 2], padding=‘same‘, strides=2),
 15 
 16     # unit2
 17     layers.Conv2D(16, kernel_size=[5, 5], strides=1, padding=‘valid‘, activation=tf.nn.sigmoid),
 18     layers.MaxPool2D(pool_size=[2, 2], padding=‘same‘, strides=2),
 19 ]
 20 
 21 
 22 def preprocess(x, y):
 23     x = tf.cast(x, dtype=tf.float32) / 255
 24     # x = tf.reshape(x, [32, 32])
 25     y = tf.cast(y, dtype=tf.int32)
 26     # y = tf.one_hot(y, depth=10)
 27     return x, y
 28 
 29 
 30 def plot_to_image(figure):
 31     """Converts the matplotlib plot specified by ‘figure‘ to a PNG image and
 32     returns it. The supplied figure is closed and inaccessible after this call."""
 33     # Save the plot to a PNG in memory.
 34     buf = io.BytesIO()
 35     plt.savefig(buf, format=‘png‘)
 36     # Closing the figure prevents it from being displayed directly inside
 37     # the notebook.
 38     plt.close(figure)
 39     buf.seek(0)
 40     # Convert PNG buffer to TF image
 41     image = tf.image.decode_png(buf.getvalue(), channels=4)
 42     # Add the batch dimension
 43     image = tf.expand_dims(image, 0)
 44     return image
 45 
 46 
 47 def image_grid(images):
 48     """Return a 5x5 grid of the MNIST images as a matplotlib figure."""
 49     # Create a figure to contain the plot.
 50     figure = plt.figure(figsize=(10, 10))
 51     for i in range(25):
 52         # Start next subplot.
 53         plt.subplot(5, 5, i + 1, title=‘name‘)
 54         plt.xticks([])
 55         plt.yticks([])
 56         plt.grid(False)
 57         plt.imshow(images[i], cmap=plt.cm.binary)
 58 
 59     return figure
 60 
 61 
 62 def main():
 63     (x_train, y_train), (x_test, y_test) = datasets.cifar10.load_data()
 64     print("x_train.shape:", x_train.shape, "y_train.shape:", y_train.shape)
 65     print("x_test.shape:", x_test.shape, ‘y_test.shape:‘, y_test.shape)
 66     y_test = tf.squeeze(y_test, axis=1)
 67     y_train = tf.squeeze(y_train, axis=1)
 68     train_db = tf.data.Dataset.from_tensor_slices((x_train, y_train))
 69     train_db = train_db.shuffle(1000).map(preprocess).batch(256)
 70 
 71     test_db = tf.data.Dataset.from_tensor_slices((x_test, y_test))
 72     test_db = test_db.map(preprocess).batch(256)
 73 
 74     sample_train = next(iter(train_db))
 75     sample_test = next(iter(test_db))
 76     print("sample_train[0].shape:", sample_train[0].shape, "sample_train[1].shape:", sample_train[1].shape)
 77     print("sample_test[0].shape:", sample_test[0].shape, "sample_test[1].shape:", sample_test[1].shape)
 78 
 79     conv_net = Sequential(conv_layers)
 80     fc_net = Sequential([
 81         layers.Dense(120, activation=tf.nn.tanh),
 82         layers.Dense(84, activation=tf.nn.tanh),
 83         layers.Dense(10, activation=None),
 84     ])
 85     conv_net.build(input_shape=[None, 32, 32, 3])
 86     fc_net.build(input_shape=[None, 400])
 87 
 88     optimizer = optimizers.Adam(lr=1e-4)
 89     variables = conv_net.trainable_variables + fc_net.trainable_variables
 90 
 91     current_time = datetime.datetime.now().strftime("%Y%m%d-%H%M%S")
 92     log_dir = ‘logs/‘ + current_time
 93     summary_writer = tf.summary.create_file_writer(log_dir)
 94     sample_img = next(iter(train_db))[0]
 95     # get first image instance
 96     sample_img = sample_img[0]
 97     sample_img = tf.reshape(sample_img, [1, 32, 32, 3])
 98     with summary_writer.as_default():
 99         tf.summary.image("Training sample:", sample_img, step=0)
100 
101     for epoch in range(2000):
102         for step, (x, y) in enumerate(train_db):
103             with tf.GradientTape() as tape:
104                 # [b, 32, 32, 3]->[b, 5,5,16]
105                 out = conv_net(x)
106                 # [b,5,5,16]->[b,400]
107                 out = tf.reshape(out, [-1, 400])
108                 # [b, 16, 16, 5]->[b, 10]
109                 logits = fc_net(out)
110                 y_onehot = tf.one_hot(y, depth=10)
111                 loss = tf.losses.categorical_crossentropy(y_onehot, logits, from_logits=True)
112                 loss = tf.reduce_mean(loss)
113             grads = tape.gradient(loss, variables)
114             optimizer.apply_gradients(zip(grads, variables))
115 
116             if step % 100 == 0:
117                 print(epoch, step, ‘loss:‘, float(loss))
118         with summary_writer.as_default():
119             tf.summary.scalar(‘train-loss‘, float(loss), step=epoch)
120         total_num = 0
121         total_correct = 0
122         for x, y in test_db:
123             out = conv_net(x)
124             out = tf.reshape(out, [-1, 400])
125             logits = fc_net(out)
126             prob = tf.nn.softmax(logits, axis=1)
127             pred = tf.argmax(prob, axis=1)
128             pred = tf.cast(pred, dtype=tf.int32)
129             # y = tf.cast(y, dtype=tf.int32)
130             correct = tf.cast(tf.equal(pred, y), dtype=tf.int32)
131             correct = tf.reduce_sum(correct)
132 
133             total_num += x.shape[0]
134             total_correct += int(correct)
135         acc = total_correct / total_num
136         val_images = x[:25]
137         val_images = tf.reshape(val_images, [-1, 32, 32, 3])
138         with summary_writer.as_default():
139             tf.summary.scalar(‘test-acc‘, float(acc), step=epoch)
140             tf.summary.image("val-onebyone-images:", val_images, max_outputs=25, step=epoch)
141 
142             val_images = tf.reshape(val_images, [-1, 32, 32])
143             figure = image_grid(val_images)
144             tf.summary.image(‘val-images:‘, plot_to_image(figure), step=epoch)
145         print(epoch, ‘acc:‘, acc)
146 
147 
148 if __name__ == ‘__main__‘:
149     main()