是否可以在 tensorboard 中可视化 keras 嵌入？

Question

以 tensorboard 兼容格式导出它的一些训练数据

但是不支持tensorboard中的embedding visualisation

有解决办法吗？

Answer 1

找到解决方案：

import os

import keras
import tensorflow

ROOT_DIR = '/tmp/tfboard'

os.makedirs(ROOT_DIR, exist_ok=True)


OUTPUT_MODEL_FILE_NAME = os.path.join(ROOT_DIR,'tf.ckpt')

# get the keras model
model = get_model()
# get the tensor name from the embedding layer
tensor_name = next(filter(lambda x: x.name == 'embedding', model.layers)).W.name

# the vocabulary
metadata_file_name = os.path.join(ROOT_DIR,tensor_name)

embedding_df = get_embedding()
embedding_df.to_csv(metadata_file_name, header=False, columns=[])

saver = tensorflow.train.Saver()
saver.save(keras.backend.get_session(), OUTPUT_MODEL_FILE_NAME)

summary_writer = tensorflow.train.SummaryWriter(ROOT_DIR)

config = tensorflow.contrib.tensorboard.plugins.projector.ProjectorConfig()
embedding = config.embeddings.add()
embedding.tensor_name = tensor_name
embedding.metadata_path = metadata_file_name
tensorflow.contrib.tensorboard.plugins.projector.visualize_embeddings(summary_writer, config)

Answer 2

存在具有此功能的拉取请求 - https://github.com/fchollet/keras/pull/5247 扩展回调以创建特定嵌入层的可视化。

Answer 3

现在可以直接使用 keras.callbacks.TensorBoard 回调：

from keras import callbacks

model.fit(x_train, y_train,
        batch_size=batch_size,
        epochs=10,
        callbacks=[
                   callbacks.TensorBoard(batch_size=batch_size,
                                         embeddings_freq=3,  # Store embeddings every 3 epochs (this can be time consuming)
                                         embeddings_layer_names=['fc1', 'fc2'],  # Embeddings are taken from layers with names fc1 and fc2
                                         embeddings_metadata='metadata.tsv',  # This file will describe the embeddings data (see below)
                                         embeddings_data=x_test),  # Data used for the embeddings
                   ],
        )


# Use this metadata.tsv file before you have a trained model:
with open("metadata.tsv", 'w') as f:
    f.write("label\tidx\n")
    f.write('\n'.join(["{}\t{}".format(class_names[int(y.argmax())], i)
                       for i, y in enumerate(y_test)]))


# After the model is trained, you can update the metadata file to include more information, such as the predicted labels and the mistakes:
y_pred = model.predict(x_test)
with open("metadata.tsv", 'w') as f:
    f.write("label\tidx\tpredicted\tcorrect\n")
    f.write('\n'.join(["{}\t{}\t{}\t{}".format(class_names[int(y.argmax())],
                                               i,
                                               class_names[int(y_pred[i].argmax())],
                                               class_names[int(y.argmax())]==class_names[int(y_pred[i].argmax())])
                       for i, y in enumerate(y_test)]))

注意： Tensorboard 通常会在 logs 目录中查找您的 metadata.tsv。如果找不到它，它会告诉你它正在寻找的路径，你可以将它复制到那里并刷新 tensorboard。

是否可以在 tensorboard 中可视化 keras 嵌入？

Is it possible to visualize keras embeddings in tensorboard?

python

tensorflow

keras

tensorboard