Related
I have a huge TFRecord file with more than 4M entries. It is a very unbalanced dataset containing many more entries of some labels and few others - compare to the whole dataset. I want to filter a limited number of entries of some of these labels in order to have a balanced dataset. Below, you can see my attempt, but it takes more than 24 hours to filter 1k from each label (33 different labels).
import tensorflow as tf
tf.compat.as_str(
bytes_or_text='str', encoding='utf-8'
)
try:
tpu = tf.distribute.cluster_resolver.TPUClusterResolver.connect()
print("Device:", tpu.master())
strategy = tf.distribute.TPUStrategy(tpu)
except:
strategy = tf.distribute.get_strategy()
print("Number of replicas:", strategy.num_replicas_in_sync)
ignore_order = tf.data.Options()
ignore_order.experimental_deterministic = False
dataset = tf.data.TFRecordDataset('/test.tfrecord')
dataset = dataset.with_options(ignore_order)
features, feature_lists = detect_schema(dataset)
#Decodings TFRecord serialized data
def decode_data(serialized):
X, y = tf.io.parse_single_sequence_example(
serialized,
context_features=features,
sequence_features=feature_lists)
return X['title'], y['subject']
dataset = dataset.map(lambda x: tf.py_function(func=decode_data, inp=[x], Tout=(tf.string, tf.string)))
#Filtering and concatenating the samples
def balanced_dataset(dataset, labels_list, sample_size=1000):
datasets_list = []
for label in labels_list:
#Filtering the chosen labels
locals()[label] = dataset.filter(lambda x, y: tf.greater(tf.reduce_sum(tf.cast(tf.equal(tf.constant(label, dtype=tf.int64), y), tf.float32)), tf.constant(0.)))
#appending a limited sample
datasets_list.append(locals()[label].take(sample_size))
concat_dataset = datasets_list[0]
#concatenating the datasets
for dset in datasets_list[1:]:
concat_dataset = concat_dataset.concatenate(dset)
return concat_dataset
balanced_data = balanced_dataset(tabledataset, labels_list=list(decod_dic.values()), sample_size=1000)
One way to solve this is by using group_by_window method where the window_size would be the sample size of each class (in your case 1k).
ds = ds.group_by_window(
# Use label as key
key_func=lambda _, l: l,
# Convert each window to a sample_size
reduce_func=lambda _, window: window.batch(sample_size),
# Use window size as sample_size
window_size=sample_size)
This will form batches of single classes of size sample_size. But there is one problem, there will be multiple batches of same class, but you just need one of the batches in each class.
To solve the above problem, we need to add a count for each of the batches and then filter out count==0, which will fetch the first batch of all the classes.
Lets define an example:
labels = np.array(sum([[label]*repeat for label, repeat in zip([0, 1, 2], [100, 200, 15])], []))
features = np.arange(len(labels))
np.unique(labels, return_counts=True)
#(array([0, 1, 2]), array([100, 200, 15]))
# There are 3 labels chosen for simplicity and each of their counts are shown along.
sample_size = 15 # we choose to pick sample of 15 from each class
We create a dataset from the above inputs,
ds = tf.data.Dataset.from_tensor_slices((features, labels))
In the above window function we modify the reduce_func to make the counter, so the batch will have 3 elements (X_batch, y_batch, label_counter) :
def reduce_func(x, y):
#class_count[y] += 1
z = table.lookup(x)
table.insert(x, z+1)
return y.batch(sample_size).map(lambda a,b: (a, b, z))
# Group by window
ds = tf.data.Dataset.from_tensor_slices((features, labels))
ds = ds.group_by_window(
# Use label as key
key_func=lambda _, l: l,
# Convert each window to a sample_size
reduce_func=reduce_func,
# Use window size as sample_size
window_size=sample_size)
The counter logic in reduce_func is implemented as a table lookup where the counter needs to be updated and read from a lookup table. Its initialized as shown below:
n_classes = 3
keys = tf.range(0,n_classes, dtype=tf.int64)
vals = tf.zeros_like(keys, dtype=tf.int64)
table = tf.lookup.experimental.MutableHashTable(key_dtype=tf.int64,
value_dtype=tf.int64,
default_value=-1)
table.insert(keys, vals)
Now we filter out the batch where the count==0 and remove the count element to form (X, y) batch pairs:
ds = ds.filter(lambda x, y, count: count==0)
ds = ds.map(lambda x, y, count: (x, y))
Output,
for x, y in ds:
print(x.numpy(), y.numpy())
[ 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14] [0 0 0 0 0 0 0 0 0 0 0 0 0 0 0]
[100 101 102 103 104 105 106 107 108 109 110 111 112 113 114] [1 1 1 1 1 1 1 1 1 1 1 1 1 1 1]
[300 301 302 303 304 305 306 307 308 309 310 311 312 313 314] [2 2 2 2 2 2 2 2 2 2 2 2 2 2 2]
I am trying to implement a Bayesian network and solve a regression problem using PYMC3. In my model, I have a fair coin as the parent node. If the parent node is H, the child node selects the normal distribution N(5,0.2); if T, the child selects N(0,0.5). Here is an illustration of my network.
To simulate this network, I generated a sample dataset and tried doing Bayesian regression using the code below. Currently, the model does regression only for the child node as if the parent node does not exist. I would greatly appreciate it if anyone can let me know how to implement the conditional probability P(D|C). Ultimately, I am interested in finding the probability distribution for mu1 and mu2. Thank you!
# Generate data for coin flip P(C) and store in c1
theta_real = 0.5 # unkown value in a real experiment
n_sample = 10
c1 = bernoulli.rvs(p=theta_real, size=n_sample)
# Generate data for normal distribution P(D|C) and store in d1
np.random.seed(123)
mu1 = 0
sigma1 = 0.5
mu2 = 5
sigma2 = 0.2
d1 = []
for index, item in enumerate(c1):
if item == 0:
d1.extend(normal(mu1, sigma1, 1))
else:
d1.extend(normal(mu2, sigma2, 1))
# I start building PYMC3 model here
c1_tensor = theano.shared(np.array(c1))
d1_tensor = theano.shared(np.array(d1))
with pm.Model() as model:
# define prior for c1. I am not sure how to do this.
#c1_present = pm.Categorical('c1',observed=c1_tensor)
# how do I incorporate P(D | C)
mu_prior = pm.Normal('mu', mu=2, sd=2, shape=1)
sigma_prior = pm.HalfNormal('sigma', sd=2, shape=1)
y_likelihood = pm.Normal('y', mu=mu_prior, sd=sigma_prior, observed=d1_tensor)
You could use the Dirichlet distribution as a prior for the coin toss and NormalMixture as the prior of the two Gaussians. In the following snippet I changed the fairness of the coin and increased the number of coin tosses, but you could adjust these in any way want:
import numpy as np
import pymc3 as pm
from scipy.stats import bernoulli
# Generate data for coin flip P(C) and store in c1
theta_real = 0.2 # unkown value in a real experiment
n_sample = 2000
c1 = bernoulli.rvs(p=theta_real, size=n_sample)
# Generate data for normal distribution P(D|C) and store in d1
np.random.seed(123)
mu1 = 0
sigma1 = 0.5
mu2 = 5
sigma2 = 0.2
d1 = []
for index, item in enumerate(c1):
if item == 0:
d1.extend(np.random.normal(mu1, sigma1, 1))
else:
d1.extend(np.random.normal(mu2, sigma2, 1))
with pm.Model() as model:
w = pm.Dirichlet('p', a=np.ones(2))
mu = pm.Normal('mu', 0, 20, shape=2)
sigma = np.array([0.5,0.2])
pm.NormalMixture('like',w=w,mu=mu,sigma=sigma,observed=np.array(d1))
trace = pm.sample()
pm.summary(trace)
This will give you the following:
mean sd mc_error hpd_2.5 hpd_97.5 n_eff Rhat
mu__0 4.981222 0.023900 0.000491 4.935044 5.027420 2643.052184 0.999637
mu__1 -0.007660 0.004946 0.000095 -0.017388 0.001576 2481.146286 1.000312
p__0 0.213976 0.009393 0.000167 0.195602 0.231803 2245.905021 0.999302
p__1 0.786024 0.009393 0.000167 0.768197 0.804398 2245.905021 0.999302
The parameters are recovered nicely as you can also see from the traceplots:
The above implementation will give you the posterior of theta_real, mu1 and mu2 but I could not get convergence when I added sigma1 and sigma2 as parameters to be estimated by the data (even though the prior was quite narrow):
with pm.Model() as model:
w = pm.Dirichlet('p', a=np.ones(2))
mu = pm.Normal('mu', 0, 20, shape=2)
sigma = pm.HalfNormal('sigma', sd=2, shape=2)
pm.NormalMixture('like',w=w,mu=mu,sigma=sigma,observed=np.array(d1))
trace = pm.sample()
print(pm.summary(trace))
Auto-assigning NUTS sampler...
Initializing NUTS using jitter+adapt_diag...
Multiprocess sampling (4 chains in 4 jobs)
NUTS: [sigma, mu, p]
Sampling 4 chains: 100%|██████████| 4000/4000 [00:10<00:00, 395.57draws/s]
The acceptance probability does not match the target. It is 0.883057127209148, but should be close to 0.8. Try to increase the number of tuning steps.
The gelman-rubin statistic is larger than 1.4 for some parameters. The sampler did not converge.
The estimated number of effective samples is smaller than 200 for some parameters.
mean sd mc_error ... hpd_97.5 n_eff Rhat
mu__0 1.244021 2.165433 0.216540 ... 5.005507 2.002049 212.596596
mu__1 3.743879 2.165122 0.216510 ... 5.012067 2.002040 235.750129
p__0 0.643069 0.248630 0.024846 ... 0.803369 2.004185 30.966189
p__1 0.356931 0.248630 0.024846 ... 0.798632 2.004185 30.966189
sigma__0 0.416207 0.125435 0.012517 ... 0.504110 2.009031 17.333177
sigma__1 0.271763 0.125539 0.012533 ... 0.497208 2.007779 19.217223
[6 rows x 7 columns]
Based on that you most likely will need to reparametrize if you also wanted to estimate the two standard deviations from this data.
This answer is to supplement #balleveryday's answer, which suggests the Gaussian Mixture Model, but had some trouble getting the symmetry breaking to work. Admittedly, the symmetry breaking in the official example is done in the context of Metropolis-Hastings sampling, whereas I think NUTS might be a little more sensitive to encountering impossible values (not sure). Here's what worked for me:
import numpy as np
import pymc3 as pm
from scipy.stats import bernoulli
import theano.tensor as tt
# everything should reproduce
np.random.seed(123)
n_sample = 2000
# Generate data for coin flip P(C) and store in c1
theta_real = 0.2 # unknown value in a real experiment
c1 = bernoulli.rvs(p=theta_real, size=n_sample)
# Generate data for normal distribution P(D|C) and store in d1
mu1, mu2 = 0, 5
sigma1, sigma2 = 0.5, 0.2
d1 = np.empty_like(c1, dtype=np.float64)
d1[c1 == 0] = np.random.normal(mu1, sigma1, np.sum(c1 == 0))
d1[c1 == 1] = np.random.normal(mu2, sigma2, np.sum(c1 == 1))
with pm.Model() as gmm_asym:
# mixture vector
w = pm.Dirichlet('p', a=np.ones(2))
# Gaussian parameters (testval helps start off ordered)
mu = pm.Normal('mu', 0, 20, shape=2, testval=[-10, 10])
sigma = pm.HalfNormal('sigma', sd=2, shape=2)
# break symmetry, forcing mu[0] < mu[1]
order_means_potential = pm.Potential('order_means_potential',
tt.switch(mu[1] - mu[0] < 0, -np.inf, 0))
# observed
pm.NormalMixture('like', w=w, mu=mu, sigma=sigma, observed=d1)
# reproducible sampling
tr_gmm_asym = pm.sample(tune=2000, target_accept=0.9, random_seed=20191121)
This produces samples with the statistics
mean sd mc_error hpd_2.5 hpd_97.5 n_eff Rhat
mu__0 0.004549 0.011975 0.000226 -0.017398 0.029375 2425.487301 0.999916
mu__1 5.007663 0.008993 0.000166 4.989247 5.024692 2181.134002 0.999563
p__0 0.789983 0.009091 0.000188 0.773059 0.808062 2417.356539 0.999788
p__1 0.210017 0.009091 0.000188 0.191938 0.226941 2417.356539 0.999788
sigma__0 0.497322 0.009103 0.000186 0.480394 0.515867 2227.397854 0.999358
sigma__1 0.191310 0.006633 0.000141 0.178924 0.204859 2286.817037 0.999614
and the traces
I have been experimenting with tensorflow Datasets but I cannot figure out how to efficiently create RLE-masks.
FYI, I am using data from the Airbus Ship Detection Challenge in Kaggle: https://www.kaggle.com/c/airbus-ship-detection/data
I know my RLE-decoding function works (borrowed) from one of the kernels:
def rle_decode(mask_rle, shape=(768, 768)):
'''
mask_rle: run-length as string formated (start length)
shape: (height,width) of array to return
Returns numpy array, 1 - mask, 0 - background
'''
if not isinstance(mask_rle, str):
img = np.zeros(shape[0]*shape[1], dtype=np.uint8)
return img.reshape(shape).T
s = mask_rle.split()
starts, lengths = [np.asarray(x, dtype=int) for x in (s[0:][::2], s[1:][::2])]
starts -= 1
ends = starts + lengths
img = np.zeros(shape[0]*shape[1], dtype=np.uint8)
for lo, hi in zip(starts, ends):
img[lo:hi] = 1
return img.reshape(shape).T
.... BUT it does not seem to play nicely with the pipeline:
list_ds = tf.data.Dataset.list_files(train_paths_abs)
ds = list_ds.map(parse_img)
With the following parse function, everything works fine:
def parse_img(file_path,new_size=[128,128]):
img_content = tf.io.read_file(file_path)
img = tf.image.decode_jpeg(img_content)
img = tf.image.convert_image_dtype(img, tf.float32)
img = tf.image.resize(img,new_size)
return img
But things go rogue if I include the mask:
def parse_img(file_path,new_size=[128,128]):
# Image
img_content = tf.io.read_file(file_path)
img = tf.image.decode_jpeg(img_content)
img = tf.image.convert_image_dtype(img, tf.float32)
img = tf.image.resize(img,new_size)
# Mask
file_id = tf.strings.split(file_path,'/')[-1]
objects = [rle_decode(m) for m in df2[df.ImageId==file_id]]
mask = np.sum(objects,axis=0)
mask = np.expand_dims(mask,3) # Force mask to have 3 channels, necessary for resize step
mask = tf.image.convert_image_dtype(mask, tf.int8)
mask = tf.clip_by_value(mask,0,1)
mask = tf.image.resize(mask,new_size)
mask = tf.squeeze(mask) # squeeze back
mask = tf.image.convert_image_dtype(mask, tf.int8)
return img, mask
Although my parse_img function works fine (I have checked it on a sample, it takes 271 µs ± 67.9 µs per run); the list_ds.map step takes forever (>5 minutes) before hanging.
I can't figure out what's wrong and it drives me crazy!
Any idea?
You can rewrite the function rle_decode with tensorflow like this (here I do not do the final transposition to keep it more general, but you can do it later):
import tensorflow as tf
def rle_decode_tf(mask_rle, shape):
shape = tf.convert_to_tensor(shape, tf.int64)
size = tf.math.reduce_prod(shape)
# Split string
s = tf.strings.split(mask_rle)
s = tf.strings.to_number(s, tf.int64)
# Get starts and lengths
starts = s[::2] - 1
lens = s[1::2]
# Make ones to be scattered
total_ones = tf.reduce_sum(lens)
ones = tf.ones([total_ones], tf.uint8)
# Make scattering indices
r = tf.range(total_ones)
lens_cum = tf.math.cumsum(lens)
s = tf.searchsorted(lens_cum, r, 'right')
idx = r + tf.gather(starts - tf.pad(lens_cum[:-1], [(1, 0)]), s)
# Scatter ones into flattened mask
mask_flat = tf.scatter_nd(tf.expand_dims(idx, 1), ones, [size])
# Reshape into mask
return tf.reshape(mask_flat, shape)
A small test (TensorFlow 2.0):
mask_rle = '1 2 4 3 9 4 15 5'
shape = [4, 6]
# Original NumPy function
print(rle_decode(mask_rle, shape))
# [[1 0 0 1]
# [1 0 0 0]
# [0 1 1 0]
# [1 1 1 0]
# [1 1 1 0]
# [1 1 1 0]]
# TensorFlow function (transposing is done out of the function)
tf.print(tf.transpose(rle_decode_tf(mask_rle, shape)))
# [[1 0 0 1]
# [1 0 0 0]
# [0 1 1 0]
# [1 1 1 0]
# [1 1 1 0]
# [1 1 1 0]]
Given a tensorflow dataset
Train_dataset = tf.data.Dataset.from_tensor_slices((Train_Image_Filenames,Train_Image_Labels))
Train_dataset = Train_dataset.map(Parse_JPEG_Augmented)
...
I would like to stratify my batches to deal with class imbalance. I found tf.contrib.training.stratified_sample and thought I could use it in the following way:
Train_dataset_iter = Train_dataset.make_one_shot_iterator()
Train_dataset_Image_Batch,Train_dataset_Label_Batch = Train_dataset_iter.get_next()
Train_Stratified_Images,Train_Stratified_Labels = tf.contrib.training.stratified_sample(Train_dataset_Image_Batch,Train_dataset_Label_Batch,[1/Classes]*Classes,Batch_Size)
But it gives the following error and I'm not sure that this would allow me to keep the performance benefits of tensorflow dataset as I may have then have to pass Train_Stratified_Images and Train_Stratified_Labels via feed_dict ?
File "/xxx/xxx/anaconda3/lib/python3.6/site-packages/tensorflow/contrib/training/python/training/sampling_ops.py", line 192, in stratified_sample
with ops.name_scope(name, 'stratified_sample', list(tensors) + [labels]):
File "/xxx/xxx/anaconda3/lib/python3.6/site-packages/tensorflow/python/framework/ops.py", line 459, in __iter__
"Tensor objects are only iterable when eager execution is "
TypeError: Tensor objects are only iterable when eager execution is enabled. To iterate over this tensor use tf.map_fn.
What would be the "best practice" way of using dataset with stratified batches?
Here is below a simple example to demonstrate the usage of sample_from_datasets (thanks #Agade for the idea).
import math
import tensorflow as tf
import numpy as np
def print_dataset(name, dataset):
elems = np.array([v.numpy() for v in dataset])
print("Dataset {} contains {} elements :".format(name, len(elems)))
print(elems)
def combine_datasets_balanced(dataset_smaller, size_smaller, dataset_bigger, size_bigger, batch_size):
ds_smaller_repeated = dataset_smaller.repeat(count=int(math.ceil(size_bigger / size_smaller)))
# we repeat the smaller dataset so that the 2 datasets are about the same size
balanced_dataset = tf.data.experimental.sample_from_datasets([ds_smaller_repeated, dataset_bigger], weights=[0.5, 0.5])
# each element in the resulting dataset is randomly drawn (without replacement) from dataset even with proba 0.5 or from odd with proba 0.5
balanced_dataset = balanced_dataset.take(2 * size_bigger).batch(batch_size)
return balanced_dataset
N, M = 3, 10
even = tf.data.Dataset.range(0, 2 * N, 2).repeat(count=int(math.ceil(M / N)))
odd = tf.data.Dataset.range(1, 2 * M, 2)
even_odd = combine_datasets_balanced(even, N, odd, M, 2)
print_dataset("even", even)
print_dataset("odd", odd)
print_dataset("even_odd_all", even_odd)
Output :
Dataset even contains 12 elements : # 12 = 4 x N (because of .repeat)
[0 2 4 0 2 4 0 2 4 0 2 4]
Dataset odd contains 10 elements :
[ 1 3 5 7 9 11 13 15 17 19]
Dataset even_odd contains 10 elements : # 10 = 2 x M / 2 (2xM because of .take(2 * M) and /2 because of .batch(2))
[[ 0 2]
[ 1 4]
[ 0 2]
[ 3 4]
[ 0 2]
[ 4 0]
[ 5 2]
[ 7 4]
[ 0 9]
[ 2 11]]
I am relatively new to machine learning as well as tensorflow. I would like to train the data so that predictions with 2 targets and multiple classes could be made. Is this something that can be done? I was able to implement the algorithm for 1 target but don't know how I need to do it for a second target as well.
An example dataset:
DayOfYear Temperature Flow Visibility
316 8 1 4
285 -1 1 4
326 8 2 5
323 -1 0 3
10 7 3 6
62 8 0 3
56 8 1 4
347 7 2 5
363 7 0 3
77 7 3 6
1 7 1 4
308 -1 2 5
364 7 3 6
If I train (DayOfYear Temperature Flow) I can predict the Visibility quite well. But I need to predict Flow as well somehow. I am pretty sure that Flow will influence Visibility so I am not sure how to go with that.
This is the implementation that I have
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
import os
import urllib
import numpy as np
import tensorflow as tf
# Data sets
TRAINING = "/ml_baetterich_learn.csv"
TEST = "/ml_baetterich_test.csv"
VALIDATION = "/ml_baetterich_validation.csv"
def main():
# Load datasets.
training_set = tf.contrib.learn.datasets.base.load_csv_without_header(
filename=TRAINING,
target_dtype=np.int,
features_dtype=np.int,
target_column=-1)
test_set = tf.contrib.learn.datasets.base.load_csv_without_header(
filename=TEST,
target_dtype=np.int,
features_dtype=np.int,
target_column=-1)
validation_set = tf.contrib.learn.datasets.base.load_csv_without_header(
filename=VALIDATION,
target_dtype=np.int,
features_dtype=np.int,
target_column=-1)
# Specify that all features have real-value data
feature_columns = [tf.contrib.layers.real_valued_column("", dimension=3)]
# Build 3 layer DNN with 10, 20, 10 units respectively.
classifier = tf.contrib.learn.DNNClassifier(feature_columns=feature_columns,
hidden_units=[10, 20, 10],
n_classes=9,
model_dir="/tmp/iris_model")
# Define the training inputs
def get_train_inputs():
x = tf.constant(training_set.data)
y = tf.constant(training_set.target)
return x, y
# Fit model.
classifier.fit(input_fn=get_train_inputs, steps=4000)
# Define the test inputs
def get_test_inputs():
x = tf.constant(test_set.data)
y = tf.constant(test_set.target)
return x, y
# Define the test inputs
def get_validation_inputs():
x = tf.constant(validation_set.data)
y = tf.constant(validation_set.target)
return x, y
# Evaluate accuracy.
accuracy_test_score = classifier.evaluate(input_fn=get_test_inputs,
steps=1)["accuracy"]
accuracy_validation_score = classifier.evaluate(input_fn=get_validation_inputs,
steps=1)["accuracy"]
print ("\nValidation Accuracy: {0:0.2f}\nTest Accuracy: {1:0.2f}\n".format(accuracy_validation_score,accuracy_test_score))
# Classify two new flower samples.
def new_samples():
return np.array(
[[327,8,3],
[47,8,0]], dtype=np.float32)
predictions = list(classifier.predict_classes(input_fn=new_samples))
print(
"New Samples, Class Predictions: {}\n"
.format(predictions))
if __name__ == "__main__":
main()
Option 1: multi-headed model
You could use a multi-headed DNNEstimator model. This treats Flow and Visibility as two separate softmax classification targets, each with their own set of classes. I had to modify the load_csv_without_header helper function to support multiple targets (which could be cleaner, but is not the point here - feel free to ignore its details).
import numpy as np
import tensorflow as tf
from tensorflow.python.platform import gfile
import csv
import collections
num_flow_classes = 4
num_visib_classes = 7
Dataset = collections.namedtuple('Dataset', ['data', 'target'])
def load_csv_without_header(fn, target_dtype, features_dtype, target_columns):
with gfile.Open(fn) as csv_file:
data_file = csv.reader(csv_file)
data = []
targets = {
target_cols: []
for target_cols in target_columns.keys()
}
for row in data_file:
cols = sorted(target_columns.items(), key=lambda tup: tup[1], reverse=True)
for target_col_name, target_col_i in cols:
targets[target_col_name].append(row.pop(target_col_i))
data.append(np.asarray(row, dtype=features_dtype))
targets = {
target_col_name: np.array(val, dtype=target_dtype)
for target_col_name, val in targets.items()
}
data = np.array(data)
return Dataset(data=data, target=targets)
feature_columns = [
tf.contrib.layers.real_valued_column("", dimension=1),
tf.contrib.layers.real_valued_column("", dimension=2),
]
head = tf.contrib.learn.multi_head([
tf.contrib.learn.multi_class_head(
num_flow_classes, label_name="Flow", head_name="Flow"),
tf.contrib.learn.multi_class_head(
num_visib_classes, label_name="Visibility", head_name="Visibility"),
])
classifier = tf.contrib.learn.DNNEstimator(
feature_columns=feature_columns,
hidden_units=[10, 20, 10],
model_dir="iris_model",
head=head,
)
def get_input_fn(filename):
def input_fn():
dataset = load_csv_without_header(
fn=filename,
target_dtype=np.int,
features_dtype=np.int,
target_columns={"Flow": 2, "Visibility": 3}
)
x = tf.constant(dataset.data)
y = {k: tf.constant(v) for k, v in dataset.target.items()}
return x, y
return input_fn
classifier.fit(input_fn=get_input_fn("tmp_train.csv"), steps=4000)
res = classifier.evaluate(input_fn=get_input_fn("tmp_test.csv"), steps=1)
print("Validation:", res)
Option 2: multi-labeled head
If you keep your CSV data separated by commas, and keep the last column for all the classes a row might have (separated by some token such as space), you can use the following code:
import numpy as np
import tensorflow as tf
all_classes = ["0", "1", "2", "3", "4", "5", "6"]
def k_hot(classes_col, all_classes, delimiter=' '):
table = tf.contrib.lookup.index_table_from_tensor(
mapping=tf.constant(all_classes)
)
classes = tf.string_split(classes_col, delimiter)
ids = table.lookup(classes)
num_items = tf.cast(tf.shape(ids)[0], tf.int64)
num_entries = tf.shape(ids.indices)[0]
y = tf.SparseTensor(
indices=tf.stack([ids.indices[:, 0], ids.values], axis=1),
values=tf.ones(shape=(num_entries,), dtype=tf.int32),
dense_shape=(num_items, len(all_classes)),
)
y = tf.sparse_tensor_to_dense(y, validate_indices=False)
return y
def feature_engineering_fn(features, labels):
labels = k_hot(labels, all_classes)
return features, labels
feature_columns = [
tf.contrib.layers.real_valued_column("", dimension=1), # DayOfYear
tf.contrib.layers.real_valued_column("", dimension=2), # Temperature
]
classifier = tf.contrib.learn.DNNEstimator(
feature_columns=feature_columns,
hidden_units=[10, 20, 10],
model_dir="iris_model",
head=tf.contrib.learn.multi_label_head(n_classes=len(all_classes)),
feature_engineering_fn=feature_engineering_fn,
)
def get_input_fn(filename):
def input_fn():
dataset = tf.contrib.learn.datasets.base.load_csv_without_header(
filename=filename,
target_dtype="S100", # strings of length up to 100 characters
features_dtype=np.int,
target_column=-1
)
x = tf.constant(dataset.data)
y = tf.constant(dataset.target)
return x, y
return input_fn
classifier.fit(input_fn=get_input_fn("tmp_train.csv"), steps=4000)
res = classifier.evaluate(input_fn=get_input_fn("tmp_test.csv"), steps=1)
print("Validation:", res)
We are using DNNEstimator with a multi_label_head, which uses sigmoid crossentropy rather than softmax crossentropy as a loss function. This means that each of the output units/logits are passed through the sigmoid function, which gives the likelihood of the data point belonging to that class, i.e. the classes are computed independently and are not mutually exclusive as they are with softmax crossentropy. This means that you could have between 0 and len(all_classes) classes set for each row in the training set and final predictions.
Also notice that the classes are represented as strings (and k_hot makes the conversion to token indices), so that you could use arbitrary class identifiers such as category UUIDs in e-commerce settings. If the categories in the 3rd and 4th column are different (Flow ID 1 != Visibility ID 1), you could prepend the column name to each class ID, e.g.
316,8,flow1 visibility4
285,-1,flow1 visibility4
326,8,flow2 visibility5
For a description of how k_hot works, see my other SO answer. I decided to use k_hot as a separate function (rather than define it directly in feature_engineering_fn because it's a distinct piece of functionality, and probably TensorFlow will soon have a similar utility function.
Note that if you're now using the first two columns to predict the last two columns, your accuraccy will certainly go down, as the last two columns are highly correlated and using one of them will give you a lot of information about the other. Actually, your code was using only the 3rd column, which was kind of a cheat anyway if the goal is to predict the 3rd and 4th columns.