def test_reduce_mean(self):
"""Test invoking ReduceMean in eager mode."""
with context.eager_mode():
input = np.random.rand(5, 10).astype(np.float32)
result = layers.ReduceMean(axis=1)(input)
assert result.shape == (5, )
assert np.allclose(result, np.mean(input, axis=1))
def _create_loss(self):
"""Create the loss function."""
prob = layers.ReduceSum(self.output * self._labels, axis=2)
mask = layers.ReduceSum(self._labels, axis=2)
log_prob = layers.Log(prob + 1e-20) * mask
loss = -layers.ReduceMean(
layers.ReduceSum(log_prob, axis=1), name='cross_entropy_loss')
if self._variational:
mean_sq = self._embedding_mean * self._embedding_mean
stddev_sq = self._embedding_stddev * self._embedding_stddev
kl = mean_sq + stddev_sq - layers.Log(stddev_sq + 1e-20) - 1
anneal_steps = self._annealing_final_step - self._annealing_start_step
if anneal_steps > 0:
current_step = tf.to_float(
self.get_global_step()) - self._annealing_start_step
anneal_frac = tf.maximum(0.0, current_step) / anneal_steps
kl_scale = layers.TensorWrapper(
tf.minimum(1.0, anneal_frac * anneal_frac), name='kl_scale')
else:
kl_scale = 1.0
loss += 0.5 * kl_scale * layers.ReduceMean(kl)
return loss
def create_generator_loss(self, discrim_output):
"""Create the loss function for the generator.
The default implementation is appropriate for most cases. Subclasses can
override this if the need to customize it.
Parameters
----------
discrim_output: Layer
the output from the discriminator on a batch of generated data. This is
its estimate of the probability that each sample is training data.
Returns
-------
A Layer object that outputs the loss function to use for optimizing the
generator.
"""
return -layers.ReduceMean(layers.Log(discrim_output + 1e-10))
def test_tensorboard(self):
"""Test creating an Estimator from a TensorGraph that logs information to TensorBoard."""
n_samples = 10
n_features = 3
n_tasks = 2
# Create a dataset and an input function for processing it.
np.random.seed(123)
X = np.random.rand(n_samples, n_features)
y = np.zeros((n_samples, n_tasks))
dataset = dc.data.NumpyDataset(X, y)
def input_fn(epochs):
x, y, weights = dataset.make_iterator(batch_size=n_samples,
epochs=epochs).get_next()
return {'x': x, 'weights': weights}, y
# Create a TensorGraph model.
model = dc.models.TensorGraph()
features = layers.Feature(shape=(None, n_features))
dense = layers.Dense(out_channels=n_tasks, in_layers=features)
dense.set_summary('histogram')
model.add_output(dense)
labels = layers.Label(shape=(None, n_tasks))
loss = layers.ReduceMean(layers.L2Loss(in_layers=[labels, dense]))
model.set_loss(loss)
# Create an estimator from it.
x_col = tf.feature_column.numeric_column('x', shape=(n_features, ))
estimator = model.make_estimator(feature_columns=[x_col])
# Train the model.
estimator.train(input_fn=lambda: input_fn(100))
model = dc.models.TensorGraph(model_dir='rnai')
features = layers.Feature(shape=(None, 21, 4))
labels = layers.Label(shape=(None, 1))
prev = features
for i in range(2):
prev = layers.Conv1D(filters=10,
kernel_size=10,
activation=tf.nn.relu,
padding='same',
in_layers=prev)
prev = layers.Dropout(dropout_prob=0.3, in_layers=prev)
output = layers.Dense(out_channels=1,
activation_fn=tf.sigmoid,
in_layers=layers.Flatten(prev))
model.add_output(output)
loss = layers.ReduceMean(layers.L2Loss(in_layers=[labels, output]))
model.set_loss(loss)
# Load the data.
train = dc.data.DiskDataset('train_siRNA')
valid = dc.data.DiskDataset('valid_siRNA')
# Train the model, tracking its performance on the training and validation datasets.
metric = dc.metrics.Metric(dc.metrics.pearsonr, mode='regression')
for i in range(20):
model.fit(train, nb_epoch=10)
print(model.evaluate(train, [metric])['pearsonr'][0])
print(model.evaluate(valid, [metric])['pearsonr'][0])
def create_discriminator_loss(self, discrim_output_train, discrim_output_gen):
gradient_penalty = GradientPenaltyLayer(discrim_output_train, self)
return gradient_penalty + layers.ReduceMean(discrim_output_train -
discrim_output_gen)
def create_generator_loss(self, discrim_output): return layers.ReduceMean(discrim_output)
conv2d_1 = layers.Conv2D(num_outputs=32,
activation_fn=tf.nn.relu,
in_layers=make_image)
conv2d_2 = layers.Conv2D(num_outputs=64,
activation_fn=tf.nn.relu,
in_layers=conv2d_1)
flatten = layers.Flatten(in_layers=conv2d_2)
dense1 = layers.Dense(out_channels=1024,
activation_fn=tf.nn.relu,
in_layers=flatten)
dense2 = layers.Dense(out_channels=10, activation_fn=None, in_layers=dense1)
# Computes the loss for every sample
smce = layers.SoftMaxCrossEntropy(in_layers=[label, dense2])
# Average all the losses
loss = layers.ReduceMean(in_layers=smce)
model.set_loss(loss)
# Convert the output from logits to probs
output = layers.SoftMax(in_layers=dense2)
model.add_output(output)
model.fit(train_dataset, nb_epoch=1) # nb_epoch=10
# Use as metric accuracy: the fraction of labels that are correctly predicted
metric = dc.metrics.Metric(dc.metrics.accuracy_score)
train_scores = model.evaluate(train_dataset, [metric])
test_scores = model.evaluate(test_dataset, [metric])
def create_model():
"""
Create our own MNIST model from scratch
:return:
:rtype:
"""
mnist = input_data.read_data_sets("MNIST_DATA/", one_hot=True)
# the layers from deepchem are the building blocks of what we will use to make our deep learning architecture
# now we wrap our dataset into a NumpyDataset
train_dataset = dc.data.NumpyDataset(mnist.train.images,
mnist.train.labels)
test_dataset = dc.data.NumpyDataset(mnist.test.images, mnist.test.labels)
# we will create a model that will take an input, add multiple layers, where each layer takes input from the
# previous layers.
model = dc.models.TensorGraph(model_dir='mnist')
# 784 corresponds to an image of size 28 X 28
# 10 corresponds to the fact that there are 10 possible digits (0-9)
# the None indicates that we can accept any size input (e.g. an empty array or 500 items each with 784 features)
# our data is also categorical so we must one hot encode, set single array element to 1 and the rest to 0
feature = layers.Feature(shape=(None, 784))
labels = layers.Label(shape=(None, 10))
# in order to apply convolutional layers to our input, we convert flat vector of 785 to 28X28
# in_layers means it takes our feature layer as an input
make_image = layers.Reshape(shape=(None, 28, 28), in_layers=feature)
# now that we have reshaped the input, we pass to convolution layers
conv2d_1 = layers.Conv2D(num_outputs=32,
activation_fn=tf.nn.relu,
in_layers=make_image)
conv2d_2 = layers.Conv2D(num_outputs=64,
activation_fn=tf.nn.relu,
in_layers=conv2d_1)
# we want to end by applying fully connected (Dense) layers to the outputs of our convolutional layer
# but first, we must flatten the layer from a 2d matrix to a 1d vector
flatten = layers.Flatten(in_layers=conv2d_2)
dense1 = layers.Dense(out_channels=1024,
activation_fn=tf.nn.relu,
in_layers=flatten)
# note that this is final layer so out_channels of 10 represents the 10 outputs and no activation_fn
dense2 = layers.Dense(out_channels=10,
activation_fn=None,
in_layers=dense1)
# next we want to connect this output to a loss function, so we can train the output
# compute the value of loss function for every sample then average of all samples to get final loss (ReduceMean)
smce = layers.SoftMaxCrossEntropy(in_layers=[labels, dense2])
loss = layers.ReduceMean(in_layers=smce)
model.set_loss(loss)
# for MNIST we want the probability that a given sample represents one of the 10 digits
# we can achieve this using a softmax function to get the probabilities, then cross entropy to get the labels
output = layers.SoftMax(in_layers=dense2)
model.add_output(output)
# if our model takes long to train, reduce nb_epoch to 1
model.fit(train_dataset, nb_epoch=10)
# our metric is accuracy, the fraction of labels that are accurately predicted
metric = dc.metrics.Metric(dc.metrics.accuracy_score)
train_scores = model.evaluate(train_dataset, [metric])
test_scores = model.evaluate(test_dataset, [metric])
print('train_scores', train_scores)
print('test_scores', test_scores)
