experiments/STT-tensorflow
1
0
Fork 0
Code Issues Pull Requests Projects Releases Wiki Activity

Merge pull request #35891 from ROCmSoftwarePlatform:google_upstream_resnet50

Browse Source 
PiperOrigin-RevId: 291010290
Change-Id: I59831d2a93bad687473db9899f4ee7ffd13078bd
This commit is contained in:
TensorFlower Gardener 2020-01-22 12:50:38 -08:00
commit 82273f00d5
4 changed files with 11 additions and 9 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

8
tensorflow/python/eager/benchmarks/resnet50/hvp_test.py
View File

@ -35,7 +35,7 @@ def _forward_over_back_hvp(model, images, labels, vector):
model.trainable_variables, vector) as acc: model.trainable_variables, vector) as acc:
with tf.GradientTape() as grad_tape: with tf.GradientTape() as grad_tape:
logits = model(images, training=True) logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy( loss = tf.compat.v1.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels) logits=logits, onehot_labels=labels)
grads = grad_tape.gradient(loss, model.trainable_variables) grads = grad_tape.gradient(loss, model.trainable_variables)
return acc.jvp(grads) return acc.jvp(grads)
@ -47,7 +47,7 @@ def _back_over_forward_hvp(model, images, labels, vector):
with forwardprop.ForwardAccumulator( with forwardprop.ForwardAccumulator(
model.trainable_variables, vector) as acc: model.trainable_variables, vector) as acc:
logits = model(images, training=True) logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy( loss = tf.compat.v1.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels) logits=logits, onehot_labels=labels)
return grad_tape.gradient(acc.jvp(loss), model.trainable_variables) return grad_tape.gradient(acc.jvp(loss), model.trainable_variables)
@ -55,7 +55,7 @@ def _back_over_forward_hvp(model, images, labels, vector):
def _tf_gradients_forward_over_back_hvp(model, images, labels, vector): def _tf_gradients_forward_over_back_hvp(model, images, labels, vector):
with tf.GradientTape() as grad_tape: with tf.GradientTape() as grad_tape:
logits = model(images, training=True) logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy( loss = tf.compat.v1.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels) logits=logits, onehot_labels=labels)
variables = model.trainable_variables variables = model.trainable_variables
grads = grad_tape.gradient(loss, variables) grads = grad_tape.gradient(loss, variables)
@ -68,7 +68,7 @@ def _back_over_back_hvp(model, images, labels, vector):
with tf.GradientTape() as outer_tape: with tf.GradientTape() as outer_tape:
with tf.GradientTape() as inner_tape: with tf.GradientTape() as inner_tape:
logits = model(images, training=True) logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy( loss = tf.compat.v1.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels) logits=logits, onehot_labels=labels)
grads = inner_tape.gradient(loss, model.trainable_variables) grads = inner_tape.gradient(loss, model.trainable_variables)
return outer_tape.gradient( return outer_tape.gradient(

2
tensorflow/python/eager/benchmarks/resnet50/resnet50_graph_test.py
View File

@ -109,7 +109,7 @@ class ResNet50Benchmarks(tf.test.Benchmark):
model = resnet50.ResNet50(data_format()) model = resnet50.ResNet50(data_format())
logits = model(images, training=True) logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy( loss = tf.compat.v1.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels) logits=logits, onehot_labels=labels)
optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0) optimizer = tf.train.GradientDescentOptimizer(learning_rate=1.0)
train_op = optimizer.minimize(loss) train_op = optimizer.minimize(loss)

2
tensorflow/python/eager/benchmarks/resnet50/resnet50_test.py
View File

@ -36,7 +36,7 @@ from tensorflow.python.eager.benchmarks.resnet50 import resnet50_test_util
def compute_gradients(model, images, labels, num_replicas=1): def compute_gradients(model, images, labels, num_replicas=1):
with tf.GradientTape() as grad_tape: with tf.GradientTape() as grad_tape:
logits = model(images, training=True) logits = model(images, training=True)
loss = tf.losses.softmax_cross_entropy( loss = tf.compat.v1.losses.softmax_cross_entropy(
logits=logits, onehot_labels=labels) logits=logits, onehot_labels=labels)
tf.compat.v2.summary.write('loss', loss) tf.compat.v2.summary.write('loss', loss)
if num_replicas != 1: if num_replicas != 1:

8
tensorflow/python/eager/benchmarks/resnet50/resnet50_test_util.py
View File

@ -35,9 +35,11 @@ def random_batch(batch_size, data_format):
shape = (batch_size,) + shape shape = (batch_size,) + shape
num_classes = 1000 num_classes = 1000
images = tf.random_uniform(shape) images = tf.random.uniform(shape)
labels = tf.random_uniform( labels = tf.random.uniform([batch_size],
[batch_size], minval=0, maxval=num_classes, dtype=tf.int32) minval=0,
maxval=num_classes,
dtype=tf.int32)
one_hot = tf.one_hot(labels, num_classes) one_hot = tf.one_hot(labels, num_classes)
return images, one_hot return images, one_hot