我在TF训练了一个模型，我想要想象它的重量。下面是我用来训练模型的代码。它看起来像很多代码，但它非常直接。我只是构建网络，定义占位符并使用optimize（）函数训练它。
# Create the neural network
def conv_net(x, reuse=tf.AUTO_REUSE):
# Define a scope for reusing the variables
with tf.variable_scope('ConvNet', reuse=tf.AUTO_REUSE):
# Reshape to match picture format [Height x Width x Channel]
# Tensor input become 4-D: [Batch Size, Height, Width, Channel]
#x = tf.reshape(x, shape=[-1, height, width, channels])
# -- FIRST BLOCK --
# Two convolutional layers with 32 filters, size of 5, padding = 'same'
# One max-pool layer 2X2
# One dropout layer
conv1 = tf.layers.conv2d(x, 32, 5, activation=tf.nn.relu, padding='SAME', name='conv1')
conv1 = tf.layers.batch_normalization(conv1,training=training)
conv2 = tf.layers.conv2d(conv1, 32, 5, activation=tf.nn.relu, padding='SAME', name='conv2')
conv2 = tf.layers.batch_normalization(conv2, training=training)
conv2 = tf.layers.max_pooling2d(conv2, 2, 2)
conv2 = tf.layers.dropout(conv2, rate=dropout)
# -- SECOND BLOCK --
# Two convolutional layers with 32 filters, size of 3, padding = 'same'
# One max-pool layer 2X2
# One dropout layer with 1.5 dropout rate
conv3 = tf.layers.conv2d(conv2, 32, 3, activation=tf.nn.relu, padding='SAME', name='conv3')
conv3 = tf.layers.batch_normalization(conv3, training=training)
conv4 = tf.layers.conv2d(conv3, 32, 3, activation=tf.nn.relu, padding='SAME', name='conv4')
conv4 = tf.layers.batch_normalization(conv4, training=training)
conv4 = tf.layers.max_pooling2d(conv4, 2, 2)
conv4 = tf.layers.dropout(conv4, rate=dropout*1.25)
# Flatten the data to a 1-D vector for fully connected layers
fc1 = tf.contrib.layers.flatten(conv4)
# -- THIRD BLOCK --
# Two fully connected layers with 1024 neurons
# 2x dropout rate on each layer
fc1 = tf.layers.dense(fc1, 1024)
fc1 = tf.layers.batch_normalization(fc1, training=training)
fc1 = tf.layers.dropout(fc1, rate=dropout*2)
fc2 = tf.layers.dense(fc1, 1024)
fc2 = tf.layers.batch_normalization(fc2, training=training)
fc2 = tf.layers.dropout(fc2, rate=dropout*2)
# Output layer
out = tf.layers.dense(fc2, num_classes)
return out
#### Misc stuff tu run train part ####
# Predicted classes
logits_out = conv_net(images_placeholder, reuse=tf.AUTO_REUSE)
y_pred = tf.nn.softmax(logits_out, name='y_pred')
y_pred_cls = tf.argmax(y_pred, axis=1)
y_true_cls = tf.argmax(labels_placeholder, axis=1)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(logits=logits_out, labels=labels_placeholder)
cost = tf.reduce_mean(cross_entropy)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
correct_prediction = tf.equal(y_pred_cls, y_true_cls)
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
#Optimizer
def optimize(epochs):
start_time = time.time()
for i in range(epochs):
indices = np.random.choice(trainX.shape[0], batch_size)
x_batch, y_true_batch = trainX[indices], trainY[indices]
batch_norm_update = tf.get_collection(tf.GraphKeys.UPDATE_OPS)
session.run([optimizer, batch_norm_update], feed_dict={images_placeholder:x_batch,
labels_placeholder: y_true_batch,
training: True})
saver.save(session, 'C:\\Users\\ggiorcelli\\Tensorflow Workbooks\\animals_model')
if (i+1) % 10 == 0:
train_accuracy = session.run(accuracy, feed_dict={images_placeholder: x_batch,
labels_placeholder: y_true_batch,
training: True})
indices_test = np.random.choice(testX.shape[0], batch_size)
images_test = testX[indices_test]
labels_test = testY[indices_test]
test_accuracy = session.run(accuracy, feed_dict={images_placeholder: images_test,
labels_placeholder: labels_test,
training: True})
print('Step {:3d}: training accuracy = {:0.3f}, test accuracy = {:0.3f}'.format(i+1, train_accuracy, test_accuracy))
end_time = time.time()
time_dif = end_time - start_time
print("Time usage: " + str(timedelta(seconds=int(round(time_dif)))))
#Train
with tf.Session() as session:
init = tf.global_variables_initializer()
session.run(init, {training:True})
optimize(epochs=200)
使用此代码，我能够成功训练模型。 现在，我想获得卷积层的权重。为此，我使用此代码段：
with tf.variable_scope('conv1', reuse=True):
w1 = tf.get_variable('weights')
x_image = tf.reshape(w1, [5, 5, 1, 32])
img = tf.summary.image('img',x_image)
然而，使用上面的代码我只得到一个我无法想象的图像对象。我甚至不确定这是否是获得权重的正确方法。我已经看到其他人已经问过这个问题，但大多数答案都会导致Tensorboard，我想这样做是在iPython shell中。
以下是我希望看到的一个例子。
