def get_config(opts):
"""Builds ipu_options"""
profile = opts.report
config = utils.create_ipu_config(profiling=profile,
profile_execution=profile,
report_every_nth_execution=1)
if opts.device_id == -1:
config = utils.auto_select_ipus(config, opts.shards * opts.replicas)
else:
config = utils.select_ipus(config, [opts.device_id])
if opts.convolution_options:
config = utils.set_convolution_options(
config, json.loads(opts.convolution_options))
if opts.matmul_options:
config = utils.set_matmul_options(config,
json.loads(opts.matmul_options))
if opts.enable_half_partials:
config = utils.set_matmul_options(config, {"partialsType": 'half'})
config = utils.set_convolution_options(config,
{"partialsType": 'half'})
return config
def testCrossReplicaAndStatefulGradientAccumulate(self):
with self.session() as sess:
dtype = np.float32
def my_net(y):
def cond(i, y):
del y
return i < 10
def body(i, y):
cr = gen_popops_ops.ipu_cross_replica_sum(
array_ops.ones_like(y))
ga = gen_poputil_ops.ipu_stateful_gradient_accumulate(
cr, num_mini_batches=5)
y = y + ga
i = i + 1
return (i, y)
i = 0
return control_flow_ops.while_loop(cond, body, (i, y))
with ops.device('cpu'):
y = array_ops.placeholder(dtype, [1])
opts = utils.create_ipu_config()
opts = utils.auto_select_ipus(opts, num_ipus=2)
utils.configure_ipu_system(opts)
with ops.device("/device:IPU:0"):
r = xla.compile(my_net, inputs=[y])
y = sess.run(r, {y: [10]})
self.assertEqual(y[0], 10)
self.assertAllEqual(y[1], [30])
def testCborReport(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out = math_ops.add(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=True,
profile_execution=True,
use_poplar_text_report=False,
use_poplar_cbor_report=True)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
sess.run(out, fd)
rep = sess.run(report, fd)
evts = utils.extract_all_events(rep)
self.assertEqual(len(evts), 4) # engine, begin, end, execute
self.assertEqual(evts[1].compile_end.compilation_report[0],
bytes(bytearray([217]))[0])
self.assertEqual(evts[3].execute.execution_report[0],
bytes(bytearray([217]))[0])
def testIpuModelDeviceWithMultipleReport(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out1 = pa + pb
out2 = pa - pb
with ops.device('cpu'):
with ops.control_dependencies([out1, out2]):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=True,
profile_execution=True)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
result = sess.run(out1, fd)
self.assertAllClose(result, [[1., 2.], [6., 8.]])
result, rep = sess.run([out2, report], fd)
self.assertAllClose(result, [[1., 0.], [-2., -2.]])
# 2x engine, 2x compile_begin, 2x compile_end, 2x load engine
self.assertEqual(len(rep), 8)
def testPrefixPathWithTranspose(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
z = array_ops.placeholder(np.float32, shape=[4, 4, 2, 1])
with variable_scope.variable_scope("vs", use_resource=True):
y = layers.Conv2D(
2,
1,
use_bias=True,
kernel_initializer=init_ops.ones_initializer())(x)
res = array_ops.transpose(y, [1, 2, 3, 0]) + z
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
sess.run(variables.global_variables_initializer())
result = sess.run(
res, {
x: np.reshape(np.arange(32), [1, 4, 4, 2]),
z: np.ones([4, 4, 2, 1])
})
self.assertAllClose(result, [[[[2.], [2.]], [[6.], [6.]],
[[10.], [10.]], [[14.], [14.]]],
[[[18.], [18.]], [[22.], [22.]],
[[26.], [26.]], [[30.], [30.]]],
[[[34.], [34.]], [[38.], [38.]],
[[42.], [42.]], [[46.], [46.]]],
[[[50.], [50.]], [[54.], [54.]],
[[58.], [58.]], [[62.], [62.]]]])
def get_config(report_n=1):
"""Builds ipu_options"""
config = utils.create_ipu_config(profiling=False, use_poplar_text_report=False, report_every_nth_execution=report_n)
config = utils.auto_select_ipus(config, [1])
return config
def testIoTilesAreExcludedFromShard(self):
def my_net(a, b):
with ipu_shard(0):
aa = math_ops.matmul(a, a, transpose_b=True, name="aa")
with ipu_shard(1):
bb = math_ops.matmul(b, b, transpose_b=True, name="bb")
return aa, bb
input_a = array_ops.placeholder(np.float32, [1216, 1])
input_b = array_ops.placeholder(np.float32, [1216, 1])
with ops.device("/device:IPU:0"):
compiled_net = ipu_compiler.compile(my_net, inputs=[input_a, input_b])
num_io_tiles = 128
cfg = ipu_utils.create_ipu_config(profiling=True)
cfg = ipu_utils.set_gcl_options(cfg, num_io_tiles=num_io_tiles)
cfg = ipu_utils.auto_select_ipus(cfg, num_ipus=2)
ipu_utils.configure_ipu_system(cfg)
with session.Session() as sess:
report = ReportJSON(self, sess, configure_device=False)
report.reset()
sess.run(compiled_net, {
input_a: np.ones(input_a.shape),
input_b: np.ones(input_b.shape)
})
report.parse_log()
num_compute_tiles = report.get_num_tiles_per_ipu() - num_io_tiles
for t in report.get_tensor_map().all_tensors():
self.assertLessEqual(len(t.tiles), num_compute_tiles)
def testPrefixPathWithReshape(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
z = array_ops.placeholder(np.float32, shape=[32])
with variable_scope.variable_scope("vs", use_resource=True):
y = layers.Conv2D(
2,
1,
use_bias=True,
kernel_initializer=init_ops.ones_initializer())(x)
res = gen_array_ops.reshape(y, [32]) + z
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
sess.run(variables.global_variables_initializer())
result = sess.run(res, {
x: np.reshape(np.arange(32), [1, 4, 4, 2]),
z: np.ones([32])
})
# Confirmed with values on the CPU.
self.assertAllClose(result, [
2., 2., 6., 6., 10., 10., 14., 14., 18., 18., 22., 22., 26.,
26., 30., 30., 34., 34., 38., 38., 42., 42., 46., 46., 50.,
50., 54., 54., 58., 58., 62., 62.
])
def testStatefulGradientAccumulate(self):
with self.session() as sess:
dtype = np.float32
def my_net(y):
def cond(i, x, y):
del x
del y
return i < 10
def body(i, x, y):
x = x + gen_poputil_ops.ipu_stateful_gradient_accumulate(
array_ops.ones_like(x),
num_mini_batches=5,
verify_usage=False)
y = y + array_ops.ones_like(x)
i = i + 1
return (i, x, y)
i = 0
return control_flow_ops.while_loop(cond, body, (i, y, y))
with ops.device('cpu'):
y = array_ops.placeholder(dtype, [1])
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
with ops.device("/device:IPU:0"):
r = xla.compile(my_net, inputs=[y])
y = sess.run(r, {y: [10]})
self.assertEqual(y[0], 10)
self.assertAllEqual(y[1], [20])
self.assertAllEqual(y[2], [20])
def testStatefulGradientAccumulateInvalidUse(self):
with self.session() as sess:
dtype = np.float32
def my_net(y):
def cond(i, x, y):
del x
del y
return i < 10
def body(i, x, y):
x = x + gen_poputil_ops.ipu_stateful_gradient_accumulate(
array_ops.ones_like(x), num_mini_batches=5)
y = y + array_ops.ones_like(x)
i = i + 1
return (i, x, y)
i = 0
return control_flow_ops.while_loop(cond, body, (i, y, y))
with ops.device('cpu'):
y = array_ops.placeholder(dtype, [1])
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
with ops.device("/device:IPU:0"):
r = xla.compile(my_net, inputs=[y])
with self.assertRaisesRegex(
errors.FailedPreconditionError,
"The .*IpuStatefulGradientAccumulate op"):
sess.run(r, {y: [10]})
def testTrainReplicated(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_model_fn(features, labels, mode): # pylint: disable=unused-argument
self.assertEqual(model_fn_lib.ModeKeys.TRAIN, mode)
loss = ipu.ops.cross_replica_ops.cross_replica_sum(features,
name="loss")
train_op = array_ops.identity(loss)
return model_fn_lib.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
def my_input_fn():
dataset = tu.create_dual_increasing_dataset(10,
data_shape=[1],
label_shape=[1])
dataset = dataset.batch(batch_size=1, drop_remainder=True)
return dataset
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, 4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=2, num_replicas=4,
ipu_options=ipu_options),
log_step_count_steps=1,
save_summary_steps=1)
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
session_run_counter = _SessionRunCounter()
num_steps = 6
estimator.train(input_fn=my_input_fn,
steps=num_steps,
hooks=[session_run_counter])
self.assertEqual(
session_run_counter.num_session_runs,
num_steps // config.ipu_run_config.iterations_per_loop)
model_dir = estimator.model_dir
events_file = glob.glob(model_dir + "/*tfevents*")
assert len(events_file) == 1
events_file = events_file[0]
loss_output = list()
for e in summary_iterator.summary_iterator(events_file):
for v in e.summary.value:
if "loss" in v.tag:
loss_output.append(v.simple_value)
# loss is averaged across iterations per loop
self.assertEqual(loss_output, [14.0, 16.0, 18.0])
def testNumUniqueDevicesBelowNumShardsRange(self):
def model_fn_with_zero_stages(mode):
def optimizer_function():
pass
return IPUPipelineEstimatorSpec(mode,
computational_stages=[],
gradient_accumulation_count=1,
device_mapping=[0, 1, 0],
optimizer_function=optimizer_function)
def my_input_fn():
return dataset_ops.Dataset.from_tensor_slices(([0], [0]))
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, num_ipus=4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
num_shards=4, iterations_per_loop=1, ipu_options=ipu_options))
estimator = IPUPipelineEstimator(model_fn=model_fn_with_zero_stages,
config=config)
with self.assertRaisesRegex(
ValueError, r"This pipeline requires 2 devices, but "
"`IPURunConfig.num_shards` was set to 4"):
estimator.train(input_fn=my_input_fn, steps=1)
def testPrefixPathWithElementwiseInPath(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
x = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
z = array_ops.placeholder(np.float32, shape=[1, 4, 4, 2])
s = array_ops.placeholder(np.float32, shape=[])
with variable_scope.variable_scope("vs", use_resource=True):
y = layers.Conv2D(
2,
1,
use_bias=True,
kernel_initializer=init_ops.ones_initializer())(x)
res = y + z * s
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
sess.run(variables.global_variables_initializer())
result = sess.run(
res, {
x: np.reshape(np.arange(32), [1, 4, 4, 2]),
z: np.reshape(np.arange(32), [1, 4, 4, 2]),
s: 2.0
})
# Confirmed with values on the CPU.
self.assertAllClose(
result,
[[[[1., 3.], [9., 11.], [17., 19.], [25., 27.]],
[[33., 35.], [41., 43.], [49., 51.], [57., 59.]],
[[65., 67.], [73., 75.], [81., 83.], [89., 91.]],
[[97., 99.], [105., 107.], [113., 115.], [121., 123.]]]])
def testIpuEventsWithoutPoplarReporting(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out = math_ops.add(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=False,
enable_ipu_events=True)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
sess.run(out, fd)
rep = sess.run(report, fd)
evts = utils.extract_all_events(rep)
self.assertEqual(len(evts),
3) # compile begin, compile end, execute
for e in evts:
if e.type == IpuTraceEvent.COMPILE_END:
self.assertFalse(e.compile_end.compilation_report)
if e.type == IpuTraceEvent.EXECUTE:
self.assertFalse(e.execute.execution_report)
sess.close()
def testSendScalar(self, dtype):
with self.session() as sess:
def device_fn(x):
return gen_sendrecv_ops.ipu_send_to_host(
x,
tensor_name="test_tensor",
send_device="/device:IPU:0",
send_device_incarnation=0,
recv_device="/device:CPU:0")
inputs = array_ops.placeholder(dtype=dtype, shape=())
with ipu_scope("/device:IPU:0"):
send_op = ipu_compiler.compile(device_fn, inputs=[inputs])
with ops.device("/device:CPU:0"):
recv_op = gen_sendrecv_ops.ipu_recv_at_host(
T=dtype,
tensor_name="test_tensor",
send_device="/device:IPU:0",
send_device_incarnation=0,
recv_device="/device:CPU:0")
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
sent, received = sess.run([send_op, recv_op],
feed_dict={inputs: 1})
self.assertIsNone(sent) # Send op has no output
self.assertEqual(dtype, received.dtype)
self.assertEqual(0, len(received.shape))
self.assertEqual(1, received)
def test_ipu_horovod_strategy(self):
hvd_size = hvd.size()
hvd_rank = hvd.rank()
strategy = IPUHorovodStrategy()
self.assertEqual(strategy.num_replicas_in_sync, hvd_size)
cfg = ipu_utils.create_ipu_config()
cfg = ipu_utils.auto_select_ipus(cfg, num_ipus=1)
ipu_utils.configure_ipu_system(cfg)
with strategy.scope():
def per_replica_fn():
w = variable_scope.get_variable(name="w",
initializer=hvd_rank + 1.0)
self.assertEqual("/replica:0/task:0/device:IPU:0", w.device)
return w * w
per_replica_val = strategy.experimental_run_v2(per_replica_fn)
strategy_sum = strategy.reduce(ReduceOp.SUM, per_replica_val)
strategy_mean = strategy.reduce(ReduceOp.MEAN, per_replica_val)
with session.Session() as sess:
sess.run(variables.global_variables_initializer())
# All workers should have the initial value from the first worker.
self.assertEqual([1.0], sess.run(variables.global_variables()))
self.assertEqual(1.0 * hvd_size, strategy_sum.eval())
self.assertEqual(1.0, strategy_mean.eval())
def testVectorInputOutput(self):
with self.session() as sess:
def device_fn(x):
with ipu_scope("/device:IPU:0"):
x = x + x
with outside_compilation_scope():
# Use float64 which is not supported on IPU
x = math_ops.cast(x, dtype=dtypes.float64)
c = constant_op.constant(2.0,
dtype=dtypes.float64,
shape=(2, ))
x += c
x = math_ops.cast(x, dtype=dtypes.float32)
x = x + 2.0
return x
inputs = array_ops.placeholder(dtype=dtypes.float32, shape=(2, ))
[device_out] = ipu_compiler.compile(device_fn, inputs=[inputs])
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
result = sess.run(device_out, feed_dict={inputs: [1.0, 2.0]})
self.assertEqual((2, ), result.shape)
self.assertAllEqual([6.0, 8.0], result)
def testSentTensorIsUsedAfterReceive(self):
with self.session() as sess:
def device_fn(x):
with ipu_scope("/device:IPU:0"):
x *= x # 4
with outside_compilation_scope():
y = x + 1.0 # 5
# Use `x` after receiving `y` and make sure that we still have the correct
# value of `x` (i.e. it is not overwritten by the receive, in which case
# we would get 25).
z = x * y # 20
return z
inputs = array_ops.placeholder(dtype=dtypes.float32, shape=())
[out] = ipu_compiler.compile(device_fn, inputs=[inputs])
opts = utils.create_ipu_config()
utils.configure_ipu_system(opts)
res = sess.run(out, feed_dict={inputs: 2.0})
self.assertEqual(20.0, res)
def testTwoInputsTwoOutputs(self):
with self.session() as sess:
def device_fn(x1, x2):
with ipu_scope("/device:IPU:0"):
x1 *= x1
x2 *= x2
with outside_compilation_scope():
x1 += 1.0
x2 += 2.0
x1 *= 1.0
x2 *= 2.0
return x1, x2
input1 = array_ops.placeholder(dtype=dtypes.float32, shape=())
input2 = array_ops.placeholder(dtype=dtypes.float32, shape=())
out1, out2 = ipu_compiler.compile(device_fn,
inputs=[input1, input2])
opts = utils.create_ipu_config()
opts = utils.set_optimization_options(opts,
max_send_recv_cluster_size=8)
utils.configure_ipu_system(opts)
res1, res2 = sess.run([out1, out2],
feed_dict={
input1: 1.0,
input2: 2.0
})
self.assertEqual(2.0, res1)
self.assertEqual(12.0, res2)
def testReportEveryNthExecution_Every1(self):
with self.session() as sess:
with ops.device("/device:IPU:0"):
pa = array_ops.placeholder(np.float32, [2, 2], name="a")
pb = array_ops.placeholder(np.float32, [2, 2], name="b")
out = math_ops.add(pa, pb)
with ops.device('cpu'):
report = gen_ipu_ops.ipu_event_trace()
opts = utils.create_ipu_config(profiling=True,
profile_execution=True,
report_every_nth_execution=1,
use_poplar_text_report=False)
utils.configure_ipu_system(opts)
fd = {pa: [[1., 1.], [2., 3.]], pb: [[0., 1.], [4., 5.]]}
sess.run(report, fd)
sess.run(out, fd)
sess.run(out, fd)
sess.run(out, fd)
sess.run(out, fd)
sess.run(out, fd)
rep = sess.run(report, fd)
r = tu.ReportJSON(self)
types = r.parse_events(rep)
self.assertEqual(types[IpuTraceEvent.EXECUTE], 5)
self.assertEqual(len(r.get_execution_reports()), 5,
"Every execution should have generated a report")
def testPipelineIterationsNotMultiple(self):
dataset = tu.create_single_increasing_dataset(5, shape=[4, 4, 2])
dataset = dataset.batch(batch_size=2, drop_remainder=True)
def dataset_parser(value):
a = value
b = (value + 10.) / 2.0
return {"a": a, "b": b}
dataset = dataset.map(dataset_parser)
infeed_queue = ipu_infeed_queue.IPUInfeedQueue(dataset, "__feed1")
outfeed_queue = ipu_outfeed_queue.IPUOutfeedQueue("__feed1")
def stage1(c, **kwargs):
with variable_scope.variable_scope("vs", use_resource=True):
y = layers.Conv2D(
2,
1,
use_bias=True,
kernel_initializer=init_ops.ones_initializer(),
name='conv1')(kwargs["a"])
return y + kwargs["b"], c
def stage2(x, c):
return math_ops.reduce_sum(x) + c
def stage3(x):
return x
def my_net(c):
return pipelining_ops.pipeline(
[stage1, stage2, stage3],
10,
inputs=[c],
infeed_queue=infeed_queue,
outfeed_queue=outfeed_queue,
pipeline_schedule=pipelining_ops.PipelineSchedule.Grouped)
with ops.device('cpu'):
c = array_ops.placeholder(np.float32, shape=[])
with tu.ipu_session() as sess:
with ops.device("/device:IPU:0"):
r = ipu_compiler.compile(my_net, inputs=[c])
cfg = utils.create_ipu_config(profiling=True,
profile_execution=True)
cfg = utils.auto_select_ipus(cfg, 4)
utils.configure_ipu_system(cfg)
utils.move_variable_initialization_to_cpu()
sess.run(variables.global_variables_initializer())
sess.run(infeed_queue.initializer)
with self.assertRaisesRegex(
errors.FailedPreconditionError,
'The pipeline depth of the pipeline must be a multiple of 3'
):
sess.run(r, {c: 10.01})
def _gradient_accumulation_loop(test_wrapper,
fwd_fn,
inputs_fn,
input_values,
repeat_count,
num_batches_to_accumulate,
dataset_fn,
optimizer,
num_iterations=None):
g = ops.Graph()
if num_iterations is None:
num_iterations = repeat_count * num_batches_to_accumulate
with g.as_default(), test_wrapper.test_session(graph=g) as session:
dataset = dataset_fn()
inputs = inputs_fn()
infeed_queue = ipu_infeed_queue.IPUInfeedQueue(dataset, next_feed_id())
outfeed_queue = ipu_outfeed_queue.IPUOutfeedQueue(next_feed_id())
with variable_scope.variable_scope("ipu", use_resource=True, reuse=False):
def model(*args):
loss = fwd_fn(*functional_ops._convert_to_list(args)) # pylint: disable=W0212
enqueue_op = outfeed_queue.enqueue(loss)
opt = gradient_accumulation_optimizer.GradientAccumulationOptimizerV2(
optimizer, num_batches_to_accumulate)
outs = list(args[:len(args) - infeed_queue.number_of_tuple_elements])
outs.append(enqueue_op)
outs.append(opt.minimize(loss))
return outs
def my_net(*args):
return loops.repeat(num_iterations,
model,
inputs=args,
infeed_queue=infeed_queue)
with ops.device("/device:IPU:0"):
loop_ret = ipu_compiler.compile(my_net, inputs=inputs)
outfeed_op = outfeed_queue.dequeue()
profiling = utils.running_on_ipu_model()
cfg = utils.create_ipu_config(profiling=profiling,
profile_execution=profiling)
cfg = utils.set_ipu_model_options(cfg,
compile_ipu_code=True,
tiles_per_ipu=128)
cfg = utils.auto_select_ipus(cfg, 1)
utils.configure_ipu_system(cfg)
utils.move_variable_initialization_to_cpu()
session.run(variables.global_variables_initializer())
session.run(infeed_queue.initializer)
session.run(loop_ret, feed_dict=dict(zip(inputs, input_values)))
return session.run(outfeed_op)
def testTrainWithAutomaticSharding(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_model_fn(features, labels, mode):
self.assertEqual(model_fn_lib.ModeKeys.TRAIN, mode)
with variable_scope.variable_scope("vs", use_resource=True):
predictions = layers.Dense(units=1)(features)
loss = losses.mean_squared_error(labels=labels,
predictions=predictions)
sharded_optimizer_obj = sharded_optimizer.ShardedOptimizer(
gradient_descent.GradientDescentOptimizer(0.1))
train_op = sharded_optimizer_obj.minimize(loss)
return model_fn_lib.EstimatorSpec(mode=mode,
loss=loss,
train_op=train_op)
def my_input_fn():
dataset = dataset_ops.Dataset.from_tensor_slices(
_create_regression_dataset(num_samples=1000, num_features=5))
dataset = dataset.batch(batch_size=2, drop_remainder=True).repeat()
return dataset
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, 4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=2,
num_shards=4,
autosharding=True,
ipu_options=ipu_options),
log_step_count_steps=1,
save_summary_steps=1)
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
estimator.train(input_fn=my_input_fn, steps=10)
model_dir = estimator.model_dir
events_file = glob.glob(model_dir + "/*tfevents*")
assert len(events_file) == 1
events_file = events_file[0]
loss_output = list()
for e in summary_iterator.summary_iterator(events_file):
for v in e.summary.value:
if "loss" in v.tag:
loss_output.append(v.simple_value)
self.assertTrue(loss_output[0] > loss_output[-1])
def get_ipu_config(fp_exceptions=True,
stochastic_rounding=True,
xla_recompute=False,
available_memory_proportion=None,
disable_graph_outlining=False,
num_ipus_required=0,
max_cross_replica_sum_buffer_size=0,
scheduler_selection='',
compile_only=False,
partials_type="half"):
"""Builds ipu_options"""
config = utils.create_ipu_config(
max_report_size=3001819596000,
merge_infeed_io_copies=True,
always_rearrange_copies_on_the_host=False,
selection_order=utils.SelectionOrder.AUTO,
disable_graph_outlining=disable_graph_outlining,
max_cross_replica_sum_buffer_size=max_cross_replica_sum_buffer_size,
scheduler_selection=scheduler_selection)
config = utils.auto_select_ipus(config, num_ipus_required)
config = utils.set_matmul_options(config, clear_pass_type=True)
if available_memory_proportion is not None:
config = utils.set_convolution_options(
config, {
"availableMemoryProportion": str(available_memory_proportion),
"partialsType": partials_type
})
config = utils.set_matmul_options(
config, {
"availableMemoryProportion": str(available_memory_proportion),
"partialsType": partials_type
})
config = utils.set_norm_options(config, use_stable_statistics=True)
config = utils.set_recomputation_options(config,
allow_recompute=xla_recompute)
if compile_only:
config = utils.set_ipu_connection_type(
config,
utils.DeviceConnectionType.NEVER,
ipu_version=2,
enable_remote_buffers=True)
config = utils.set_floating_point_behaviour_options(
config,
inv=fp_exceptions,
div0=fp_exceptions,
oflo=fp_exceptions,
esr=stochastic_rounding,
nanoo=fp_exceptions)
return config
def configureIPU(self,
serialization_folder=None,
offline_compilation=True):
opts = utils.create_ipu_config()
if offline_compilation:
opts = utils.set_ipu_connection_type(
opts, utils.DeviceConnectionType.NEVER, 1)
if serialization_folder:
opts = utils.set_serialization_options(opts, serialization_folder)
utils.configure_ipu_system(opts)
def _configureIPU(self, serialization_folder, verification_options=None):
opts = utils.create_ipu_config()
opts = utils.set_ipu_connection_type(opts,
utils.DeviceConnectionType.NEVER,
1)
opts = utils.set_serialization_options(opts, serialization_folder)
if verification_options:
opts = utils.set_transfer_options(opts, True)
opts = utils.set_verification_options(opts, verification_options)
utils.configure_ipu_system(opts)
def testInput():
config = utils.create_ipu_config()
config = utils.auto_select_ipus(config, 1)
config = utils.create_ipu_config(profiling=True,
use_poplar_text_report=True)
utils.configure_ipu_system(config)
# config = utils.set_convolution_options(config, {"partialsType": str('half')})
# config = utils.set_matmul_options(config, {"partialsType": str('half')})
gdv = tf.Graph()
with gdv.as_default():
g1 = tf.GraphDef()
# Load model with pywrap isntead? https://github.com/graphcore/examples/blob/master/applications/tensorflow/cnns/training/weight_avg.py#L33
with tf.gfile.GFile('model.pb', 'rb') as fid:
serialized_graph = fid.read()
g1.ParseFromString(serialized_graph)
tf.import_graph_def(g1, name='')
with tf.Session(graph=gdv) as sess:
inp_tensor = gdv.get_tensor_by_name('input:0')
out_tensor = gdv.get_tensor_by_name(
'InceptionV3/Predictions/Softmax:0')
image_np = getExamples()
#image_np = getSyntheticExamples()
np.set_printoptions(threshold=np.inf)
import time
tic = time.time()
# This is new and doesn't crash
# But doesn't seem to do anything either
with ipu_scope("/device:IPU:0"):
proba = sess.run(out_tensor, {inp_tensor: image_np})
print(proba)
toc = time.time()
duration = toc - tic
num_images = len(image_np)
print("Total time taken: {0} seconds".format(duration))
print("Number of examples: {0}".format(num_images))
print("Throughput: {0} im/s".format(num_images / duration))
def testReplicatedEvaluationOnHost(self):
if ipu_utils.running_on_ipu_model():
self.skipTest(
"Replicated top level graphs are not supported on the "
"IPU_MODEL target")
def my_input_fn():
features = [0, 0, 0, 1, 0, 0, 0, 1]
labels = [0, 1, 0, 1, 0, 1, 0, 1]
return dataset_ops.Dataset.from_tensor_slices(
(features, labels)).batch(2, drop_remainder=True)
def my_metrics_fn(features, labels):
labels64 = math_ops.cast(labels, np.int64)
return {
"accuracy": metrics_impl.accuracy(labels, features),
"precision": metrics_impl.precision(labels, features),
"recall": metrics_impl.recall(labels, features),
"recall_at_1": metrics_impl.recall_at_k(labels64,
features,
k=1),
"recall_at_2": metrics_impl.recall_at_k(labels64,
features,
k=2),
"mse": metrics_impl.mean_squared_error(labels, features),
"rmse": metrics_impl.root_mean_squared_error(labels, features),
}
def my_model_fn(features, labels, mode):
loss = math_ops.cast(replication_ops.replication_index(),
np.float32)
eval_metrics = (my_metrics_fn, [features, labels])
return ipu_estimator.IPUEstimatorSpec(mode,
loss=loss,
eval_metrics=eval_metrics)
ipu_options = ipu_utils.create_ipu_config()
ipu_options = ipu_utils.auto_select_ipus(ipu_options, num_ipus=4)
config = ipu_run_config.RunConfig(
ipu_run_config=ipu_run_config.IPURunConfig(
iterations_per_loop=1, num_replicas=4,
ipu_options=ipu_options))
estimator = ipu_estimator.IPUEstimator(model_fn=my_model_fn,
config=config)
scores = estimator.evaluate(my_input_fn, steps=1)
self.assertEqual(0.75, scores["accuracy"])
self.assertEqual(1.0, scores["precision"])
self.assertEqual(0.5, scores["recall"])
self.assertEqual(0.5, scores["recall_at_1"])
self.assertEqual(1.0, scores["recall_at_2"])
self.assertEqual(0.25, scores["mse"])
self.assertEqual(0.5, scores["rmse"])
self.assertEqual(1.5, scores[model_fn_lib.LOSS_METRIC_KEY])
def testResetSeed(self):
# The dataset for feeding the graphs
ds = dataset_ops.Dataset.from_tensors(
array_ops.constant(1.0, shape=[SIZE]))
ds = ds.map(lambda x: [x, x])
ds = ds.repeat()
# The host side queues
infeed_queue = ipu_infeed_queue.IPUInfeedQueue(
ds, feed_name="infeed", replication_factor=REPLICAS)
outfeed_queue = ipu_outfeed_queue.IPUOutfeedQueue(
feed_name="outfeed", replication_factor=REPLICAS)
# The device side
def body(x1, x2):
d1 = rand_ops.dropout(x1)
d2 = rand_ops.dropout(x2)
outfeed = outfeed_queue.enqueue({'d1': d1, 'd2': d2})
return outfeed
def my_net():
r = loops.repeat(REPEATS, body, [], infeed_queue)
return r
with scopes.ipu_scope('/device:IPU:0'):
res = ipu_compiler.compile(my_net, inputs=[])
# The outfeed dequeue has to happen after the outfeed enqueue
dequeue_outfeed = outfeed_queue.dequeue()
# Configure the hardware
config = utils.create_ipu_config(profiling=True)
config = utils.auto_select_ipus(config, REPLICAS)
config = utils.set_floating_point_behaviour_options(config)
utils.configure_ipu_system(config)
with session.Session() as sess:
res_all = set()
total = 0
sess.run(infeed_queue.initializer)
for _ in range(EXECS):
sess.run(res)
outfed_result = sess.run(dequeue_outfeed)
for r in np.array(list(outfed_result.values())).reshape(
[-1, SIZE]):
total += 1
res_all.add(r.tostring())
# 2 dropouts per replica * REPLICAS * REPEATS * EXECS
expected = 2 * REPLICAS * REPEATS * EXECS
self.assertEqual(total, expected)
self.assertEqual(len(res_all), expected)
def generic_train_graph(opts, is_training):
data_type = 'float32'
train_graph = tf.Graph()
with train_graph.as_default():
placeholders = {}
placeholders["learning_rate"] = tf.compat.v1.placeholder(data_type, shape=[])
uid_embedding, mid_embedding, cat_embedding = id_embedding(opts, is_training, seed)
if opts['use_synthetic_data']:
dataset_train = get_synthetic_dataset(opts)
else:
dataset_train = get_dataset_embed(opts, is_training=True)
infeed_train = ipu_infeed_queue.IPUInfeedQueue(dataset_train, feed_name = 'DIN_dataset_infeed_train', replication_factor = (opts['replicas']))
with ipu_scope('/device:IPU:0'):
def comp_fn():
def body(total_loss, total_aux_loss, total_accuracy, uids, mids, cats, mid_his, cat_his, mid_mask, target, seqlen):
prob, loss, aux_loss, accuracy, grad_op = graph_builder(opts, uid_embedding, mid_embedding, cat_embedding, placeholders['learning_rate'], uids, mids, cats, mid_his, cat_his, mid_mask, target, seqlen, use_negsampling=False)
with tf.control_dependencies([grad_op]):
return total_loss + loss, total_aux_loss + aux_loss, total_accuracy + accuracy
return loops.repeat(opts['batches_per_step'], body, [tf.constant(0, getattr(np, 'float32'))] * 3, infeed_train)
outputs_train = ipu_compiler.compile(comp_fn, [])
avg_loss, avg_aux_loss, avg_accuracy = [x / opts['batches_per_step'] for x in outputs_train]
outfeed = None
saver = tf.compat.v1.train.Saver()
utils.move_variable_initialization_to_cpu()
init = tf.compat.v1.global_variables_initializer()
if opts['use_ipu_model']:
os.environ["TF_POPLAR_FLAGS"] = "--use_ipu_model"
ipu_options = utils.create_ipu_config()
ipu_options = utils.set_optimization_options(ipu_options,
combine_embedding_lookups=True)
ipu_options = utils.set_recomputation_options(ipu_options, allow_recompute=True)
ipu_options = utils.auto_select_ipus(ipu_options, [opts['replicas']])
utils.configure_ipu_system(ipu_options)
if seed is not None:
utils.reset_ipu_seed(seed)
ops_train = [avg_loss, avg_aux_loss, avg_accuracy]
sess = tf.compat.v1.Session(graph=train_graph)
return GraphOps(sess,
init,
ops_train,
placeholders,
infeed_train,
outfeed,
saver), uid_embedding, mid_embedding, cat_embedding
