注意力 image caption

---

tf2 checkpoints bert

对应：
因为我这里用的是tf2.4.0

gs_folder_bert = "gs://cloud-tpu-checkpoints/bert/keras_bert/uncased_L-12_H-768_A-12"
tf.io.gfile.listdir(gs_folder_bert)

tf2 checkpoints bert

history = model.fit(train_images,
                        train_labels,
                        batch_size=1024,
                        epochs=max_epochs,
                        validation_data=(test_images, test_labels))

Downloading data from https://storage.googleapis.com/tensorflow/keras-applications/inception_v3/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5
87916544/87910968 [==============================] - 1s 0us/step

extend usage

#!/usr/bin/python

aList = [123, 'xyz', 'zara', 'abc', 123];
bList = [2009, 'manni'];
aList.extend(bList)

print "Extended List : ", aList ;

Extended List : [123, ‘xyz’, ‘zara’, ‘abc’, 123, 2009, ‘manni’]

In this example, you extract the features from the lower convolutional layer of InceptionV3 giving us a vector of shape (8, 8, 2048).

inceptionv3

Some common arguments:

split=: Which split to read (e.g. ‘train’, [‘train’, ‘test’], ‘train[80%:]’,…). See our split API guide.
shuffle_files=: Control whether to shuffle the files between each epoch (TFDS store big datasets in multiple smaller files).
data_dir=: Location where the dataset is saved ( defaults to ~/tensorflow_datasets/)
with_info=True: Returns the tfds.core.DatasetInfo containing dataset metadata
download=False: Disable download

手动下载tfds
手动下载tfds数据集

pil img show

from PIL import Image
img=Image.open('d:/dog.png')
img.show()

			import math
            x1 = math.ceil(len_result ** 0.5)
            ax = fig.add_subplot(x1, x1, l + 1)

打印模型的输出 形状 对于注意力的计算不是非常清楚 ，今天要搞懂。
特别是计算注意力分数：
tanh 相加 的原因

    class BahdanauAttention(tf.keras.Model):
        def __init__(self, units):
            super(BahdanauAttention, self).__init__()
            self.W1 = tf.keras.layers.Dense(units)
            self.W2 = tf.keras.layers.Dense(units)
            self.V = tf.keras.layers.Dense(1)
            self.isprint = True

        def call(self, features, hidden):
            # features(CNN_encoder output) shape == (batch_size, 64, embedding_dim)
            # hidden shape == (batch_size, hidden_size)
            # features.shape  (64, 64, 256)
            # hidden.shape (64, 512)
            # hidden_with_time_axis (64, 1, 512)
            # hidden_with_time_axis shape == (batch_size, 1, hidden_size)
            hidden_with_time_axis = tf.expand_dims(hidden, 1)

            # attention_hidden_layer shape == (batch_size, 64, units)
            # attention_hidden_layer.shape,(64, 64, 512)

            attention_hidden_layer = (tf.nn.tanh(self.W1(features) +
                                                 self.W2(hidden_with_time_axis)))

            if self.isprint:
                print(f'self.W1(features).shape,{self.W1(features).shape}')
                print(f'self.W2(hidden_with_time_axis).shape,{self.W2(hidden_with_time_axis).shape}')
                print(f'attention_hidden_layer.shape,{attention_hidden_layer.shape}')

            # score shape == (batch_size, 64, 1)
            # This gives you an unnormalized score for each image feature.
            # score shape == (64, 64, 1)
            score = self.V(attention_hidden_layer)

            # attention_weights shape == (batch_size, 64, 1)
            # attention_weights.shape, (64, 64, 1)
            # features.shape  (64, 64, 256)
            attention_weights = tf.nn.softmax(score, axis=1)

            # context_vector shape after sum == (batch_size, hidden_size)
            context_vector = attention_weights * features
            context_vector = tf.reduce_sum(context_vector, axis=1)

            self.isprint = False
            return context_vector, attention_weights

features.shape (64, 64, 256)
hidden.shape (64, 512)
target.shape:
(64, 51)
dec_input.shape
(64, 1)
img_tensor.shape
(64, 64, 2048)
features.shape (64, 64, 256)
context_vector.shape, (64, 256)
attention_weights.shape, (64, 64, 1)
self.embedding(x), (64, 1, 256)
tf.concat([tf.expand_dims(context_vector, 1), x], axis=-1), (64, 1, 512)
output.shape, (64, 1, 512)
state.shape, (64, 512)
self.fc1(output), (64, 1, 512)
tf.reshape(x, (-1, x.shape[2])), (64, 512)
self.fc2(x), (64, 5001)
target[:, i].shape,(64,)

return x, state, attention_weights

# predictions (64, 5001)
                # hidden  (64, 512)
                loss += loss_function(target[:, i], predictions)