TensorBoard: TF Dev Summit Tutorial

labels_1024.tsv

7
2
1
0
4
1
4
9
5
9
0
6
9
0
1
5
9
7
3
4
9
6
6
5
4
0
7
4
0
1
3
1
3
4
7
2
7
1
2
1
1
7
4
2
3
5
1
2
4
4
6
3
5
5
6
0
4
1
9
5
7
8
9
3
7
4
6
4
3
0
7
0
2
9
1
7
3
2
9
7
7
6
2
7
8
4
7
3
6
1
3
6
9
3
1
4
1
7
6
9
6
0
5
4
9
9
2
1
9
4
8
7
3
9
7
4
4
4
9
2
5
4
7
6
7
9
0
5
8
5
6
6
5
7
8
1
0
1
6
4
6
7
3
1
7
1
8
2
0
2
9
9
5
5
1
5
6
0
3
4
4
6
5
4
6
5
4
5
1
4
4
7
2
3
2
7
1
8
1
8
1
8
5
0
8
9
2
5
0
1
1
1
0
9
0
3
1
6
4
2
3
6
1
1
1
3
9
5
2
9
4
5
9
3
9
0
3
6
5
5
7
2
2
7
1
2
8
4
1
7
3
3
8
8
7
9
2
2
4
1
5
9
8
7
2
3
0
4
4
2
4
1
9
5
7
7
2
8
2
6
8
5
7
7
9
1
8
1
8
0
3
0
1
9
9
4
1
8
2
1
2
9
7
5
9
2
6
4
1
5
8
2
9
2
0
4
0
0
2
8
4
7
1
2
4
0
2
7
4
3
3
0
0
3
1
9
6
5
2
5
9
2
9
3
0
4
2
0
7
1
1
2
1
5
3
3
9
7
8
6
5
6
1
3
8
1
0
5
1
3
1
5
5
6
1
8
5
1
7
9
4
6
2
2
5
0
6
5
6
3
7
2
0
8
8
5
4
1
1
4
0
3
3
7
6
1
6
2
1
9
2
8
6
1
9
5
2
5
4
4
2
8
3
8
2
4
5
0
3
1
7
7
5
7
9
7
1
9
2
1
4
2
9
2
0
4
9
1
4
8
1
8
4
5
9
8
8
3
7
6
0
0
3
0
2
6
6
4
9
3
3
3
2
3
9
1
2
6
8
0
5
6
6
6
3
8
8
2
7
5
8
9
6
1
8
4
1
2
5
9
1
9
7
5
4
0
8
9
9
1
0
5
2
3
7
8
9
4
0
6
3
9
5
2
1
3
1
3
6
5
7
4
2
2
6
3
2
6
5
4
8
9
7
1
3
0
3
8
3
1
9
3
4
4
6
4
2
1
8
2
5
4
8
8
4
0
0
2
3
2
7
7
0
8
7
4
4
7
9
6
9
0
9
8
0
4
6
0
6
3
5
4
8
3
3
9
3
3
3
7
8
0
8
2
1
7
0
6
5
4
3
8
0
9
6
3
8
0
9
9
6
8
6
8
5
7
8
6
0
2
4
0
2
2
3
1
9
7
5
1
0
8
4
6
2
6
7
9
3
2
9
8
2
2
9
2
7
3
5
9
1
8
0
2
0
5
2
1
3
7
6
7
1
2
5
8
0
3
7
2
4
0
9
1
8
6
7
7
4
3
4
9
1
9
5
1
7
3
9
7
6
9
1
3
7
8
3
3
6
7
2
8
5
8
5
1
1
4
4
3
1
0
7
7
0
7
9
4
4
8
5
5
4
0
8
2
1
0
8
4
5
0
4
0
6
1
7
3
2
6
7
2
6
9
3
1
4
6
2
5
4
2
0
6
2
1
7
3
4
1
0
5
4
3
1
1
7
4
9
9
4
8
4
0
2
4
5
1
1
6
4
7
1
9
4
2
4
1
5
5
3
8
3
1
4
5
6
8
9
4
1
5
3
8
0
3
2
5
1
2
8
3
4
4
0
8
8
3
3
1
7
3
5
9
6
3
2
6
1
3
6
0
7
2
1
7
1
4
2
4
2
1
7
9
6
1
1
2
4
8
1
7
7
4
8
0
7
3
1
3
1
0
7
7
0
3
5
5
2
7
6
6
9
2
8
3
5
2
2
5
6
0
8
2
9
2
8
8
8
8
7
4
9
3
0
6
6
3
2
1
3
2
2
9
3
0
0
5
7
8
1
4
4
6
0
2
9
1
4
7
4
7
3
9
8
8
4
7
1
2
1
2
2
3
2
3
2
3
9
1
7
4
0
3
5
5
8
6
3
2
6
7
6
6
3
2
7
8
1
1
7
5
6
4
9
5
1
3
3
4
7
8
9
1
1
6
9
1
4
4
5
4
0
6
2
2
3
1
5
1
2
0
3
8
1
2
6
7
1
6
2
3
9
0
1
2
2
0
8
9
9
0
2
5
1
9
7
8
1
0
4
1
7
9
6
4
2
6
8
1
3
7
5
4

mnist.py

# Copyright 2017 Google, Inc. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
import os
import tensorflow as tf
import urllib

LOGDIR = '/tmp/mnist_tutorial/'
GIST_URL = 'https://gist.githubusercontent.com/dandelionmane/4f02ab8f1451e276fea1f165a20336f1/raw/dfb8ee95b010480d56a73f324aca480b3820c180'

### MNIST EMBEDDINGS ###
mnist = tf.contrib.learn.datasets.mnist.read_data_sets(train_dir=LOGDIR + 'data', one_hot=True)
### Get a sprite and labels file for the embedding projector ###
urllib.urlretrieve(GIST_URL + 'labels_1024.tsv', LOGDIR + 'labels_1024.tsv')
urllib.urlretrieve(GIST_URL + 'sprite_1024.png', LOGDIR + 'sprite_1024.png')


def conv_layer(input, size_in, size_out, name="conv"):
  with tf.name_scope(name):
    w = tf.Variable(tf.truncated_normal([5, 5, size_in, size_out], stddev=0.1), name="W")
    b = tf.Variable(tf.constant(0.1, shape=[size_out]), name="B")
    conv = tf.nn.conv2d(input, w, strides=[1, 1, 1, 1], padding="SAME")
    act = tf.nn.relu(conv + b)
    tf.summary.histogram("weights", w)
    tf.summary.histogram("biases", b)
    tf.summary.histogram("activations", act)
    return tf.nn.max_pool(act, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")


def fc_layer(input, size_in, size_out, name="fc"):
  with tf.name_scope(name):
    w = tf.Variable(tf.truncated_normal([size_in, size_out], stddev=0.1), name="W")
    b = tf.Variable(tf.constant(0.1, shape=[size_out]), name="B")
    act = tf.nn.relu(tf.matmul(input, w) + b)
    tf.summary.histogram("weights", w)
    tf.summary.histogram("biases", b)
    tf.summary.histogram("activations", act)
    return act


def mnist_model(learning_rate, use_two_conv, use_two_fc, hparam):
  tf.reset_default_graph()
  sess = tf.Session()

  # Setup placeholders, and reshape the data
  x = tf.placeholder(tf.float32, shape=[None, 784], name="x")
  x_image = tf.reshape(x, [-1, 28, 28, 1])
  tf.summary.image('input', x_image, 3)
  y = tf.placeholder(tf.float32, shape=[None, 10], name="labels")

  if use_two_conv:
    conv1 = conv_layer(x_image, 1, 32, "conv1")
    conv_out = conv_layer(conv1, 32, 64, "conv2")
  else:
    conv1 = conv_layer(x_image, 1, 64, "conv")
    conv_out = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")

  flattened = tf.reshape(conv_out, [-1, 7 * 7 * 64])


  if use_two_fc:
    fc1 = fc_layer(flattened, 7 * 7 * 64, 1024, "fc1")
    embedding_input = fc1
    embedding_size = 1024
    logits = fc_layer(fc1, 1024, 10, "fc2")
  else:
    embedding_input = flattened
    embedding_size = 7*7*64
    logits = fc_layer(flattened, 7*7*64, 10, "fc")

  with tf.name_scope("xent"):
    xent = tf.reduce_mean(
        tf.nn.softmax_cross_entropy_with_logits(
            logits=logits, labels=y), name="xent")
    tf.summary.scalar("xent", xent)

  with tf.name_scope("train"):
    train_step = tf.train.AdamOptimizer(learning_rate).minimize(xent)

  with tf.name_scope("accuracy"):
    correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))
    accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
    tf.summary.scalar("accuracy", accuracy)

  summ = tf.summary.merge_all()


  embedding = tf.Variable(tf.zeros([1024, embedding_size]), name="test_embedding")
  assignment = embedding.assign(embedding_input)
  saver = tf.train.Saver()

  sess.run(tf.global_variables_initializer())
  writer = tf.summary.FileWriter(LOGDIR + hparam)
  writer.add_graph(sess.graph)

  config = tf.contrib.tensorboard.plugins.projector.ProjectorConfig()
  embedding_config = config.embeddings.add()
  embedding_config.tensor_name = embedding.name
  embedding_config.sprite.image_path = LOGDIR + 'sprite_1024.png'
  embedding_config.metadata_path = LOGDIR + 'labels_1024.tsv'
  # Specify the width and height of a single thumbnail.
  embedding_config.sprite.single_image_dim.extend([28, 28])
  tf.contrib.tensorboard.plugins.projector.visualize_embeddings(writer, config)

  for i in range(2001):
    batch = mnist.train.next_batch(100)
    if i % 5 == 0:
      [train_accuracy, s] = sess.run([accuracy, summ], feed_dict={x: batch[0], y: batch[1]})
      writer.add_summary(s, i)
    if i % 500 == 0:
      sess.run(assignment, feed_dict={x: mnist.test.images[:1024], y: mnist.test.labels[:1024]})
      saver.save(sess, os.path.join(LOGDIR, "model.ckpt"), i)
    sess.run(train_step, feed_dict={x: batch[0], y: batch[1]})

def make_hparam_string(learning_rate, use_two_fc, use_two_conv):
  conv_param = "conv=2" if use_two_conv else "conv=1"
  fc_param = "fc=2" if use_two_fc else "fc=1"
  return "lr_%.0E,%s,%s" % (learning_rate, conv_param, fc_param)

def main():
  # You can try adding some more learning rates
  for learning_rate in [1E-4]:

    # Include "False" as a value to try different model architectures
    for use_two_fc in [True]:
      for use_two_conv in [True]:
        # Construct a hyperparameter string for each one (example: "lr_1E-3,fc=2,conv=2)
        hparam = make_hparam_string(learning_rate, use_two_fc, use_two_conv)
        print('Starting run for %s' % hparam)

	    # Actually run with the new settings
        mnist_model(learning_rate, use_two_fc, use_two_conv, hparam)


if __name__ == '__main__':
  main()

slides.pdf

sprite_1024.png

I have one more questions. Isn't this implementation using the RELU also in the output layer before the softmax?
I think that is screwing up the training?

Great tutorial and show! Thanks.

When running the mnist.py with python3.5 I get the failure:

File "mnist.py", line 25, in
urllib.urlretrieve(GIST_URL + 'labels_1024.tsv', LOGDIR + 'labels_1024.tsv')
AttributeError: module 'urllib' has no attribute 'urlretrieve'

Problem solved.

1. I downloaded the code from the other source mentioned already above:
   https://github.com/mamcgrath/TensorBoard-TF-Dev-Summit-Tutorial

2. I corrected the cuda installation as described here:
   tensorflow/tensorflow#5968

I got the nice Tensorboard graphs and scalars running:
tensorboard --logdir /tmp/mnist_tutorial

so nobody solve the problem the pic sprite_1024.png is broken and we can't load the data in the first step....

embeddings visualizer is not working (the rest seem to be working fine). I got the file from the other location (one location has an empty file) mentioned in the thread above (~ 32kb in size?). But Tensorboard gets stuck "Fetching sprite image.."

Actually the sprite_1024.png from the location that @rafalfirlejczyk mentioned up above works. Thanks @rafalfirlejczyk !

I can see PCA and T-SNE views of the 1024 data points / labels. It would be very convenient if the code itself just generates the tsv and png files when it is writing out the tensor variables. Perhaps it does and I am just not seeing it? (I am new to this).

Save sprite_1024.png.

import numpy as np
import scipy.misc as misc

sprite_images = mnist.test.images[:1024]

x = None
res = None
for i in range(32):
    x = None
    for j in range(32):
        img = sprite_images[i*32 + j,:].reshape((28, 28))
        x = np.concatenate((x, img), axis=1) if x is not None else img
    res = np.concatenate((res, x), axis=0) if res is  not None else  x

misc.toimage(256 - res, channel_axis=0).save('sprite_1024.png')

Last fc layer should be without tf.relu function, because later we use softmax.

I've moved the tutorial (and added a few fixes) to a GitHub repository:
https://github.com/dandelionmane/tf-dev-summit-tensorboard-tutorial

Great job, after learning how to use tensorboard, I can easily to know the performance of the algorithm via web browser.

The slides file are broken for me too

Hi,

With this version of code I am getting very poor training results, not at all like in the video,

I have no idea why. It is in the default settings 2 conv, 2 fc, learning rate 1e-4 Adam. Different runs may land in very different training accuracy but more often a poor accuracy and never close to 1.

Ok, after reading the other comments, the problem is clear:
it is the ReLu + softmax activation on the output. The moved tutorial repository does not have this problem. Maybe you should take this one down.

This is an amazing TensorBoard example! Love it!

Beautiful relatable example for humans to comprehend the power of tensorboard. Switching to my own data use cases will be cool.

@shekhovt Yes, the same problem to me. At this time, one dropout layer between two fully-connected neural network would make results better. Have a try.

embedding visualisation is not working for me. I can see both label and sprite image files, but tensorboard is unable to load them, it just says loading forever... I have downloaded the files from https://github.com/dandelionmane/tf-dev-summit-tensorboard-tutorial
labels file does not have a header in the first line, it simply has label(digit) in each row. Could that be a problem?
I am able to see all other graphs without any issue... Any help appreciated