Files

481 lines
18 KiB
Python

import pickle
from collections import namedtuple
import numpy as np
import pytest
from einops import EinopsError, rearrange, reduce
from einops.tests import FLOAT_REDUCTIONS as REDUCTIONS
from einops.tests import collect_test_backends, is_backend_tested
__author__ = "Alex Rogozhnikov"
testcase = namedtuple("testcase", ["pattern", "axes_lengths", "input_shape", "wrong_shapes"])
rearrangement_patterns = [
testcase(
"b c h w -> b (c h w)",
dict(c=20),
(10, 20, 30, 40),
[(), (10,), (10, 10, 10), (10, 21, 30, 40), [1, 20, 1, 1, 1]],
),
testcase(
"b c (h1 h2) (w1 w2) -> b (c h2 w2) h1 w1",
dict(h2=2, w2=2),
(10, 20, 30, 40),
[(), (1, 1, 1, 1), (1, 10, 3), ()],
),
testcase(
"b ... c -> c b ...",
dict(b=10),
(10, 20, 30),
[(), (10,), (5, 10)],
),
]
def test_rearrange_imperative():
for backend in collect_test_backends(symbolic=False, layers=True):
print("Test layer for ", backend.framework_name)
for pattern, axes_lengths, input_shape, wrong_shapes in rearrangement_patterns:
x = np.arange(np.prod(input_shape), dtype="float32").reshape(input_shape)
result_numpy = rearrange(x, pattern, **axes_lengths)
layer = backend.layers().Rearrange(pattern, **axes_lengths)
for shape in wrong_shapes:
try:
layer(backend.from_numpy(np.zeros(shape, dtype="float32")))
except BaseException:
pass
else:
raise AssertionError("Failure expected")
# simple pickling / unpickling
layer2 = pickle.loads(pickle.dumps(layer))
result1 = backend.to_numpy(layer(backend.from_numpy(x)))
result2 = backend.to_numpy(layer2(backend.from_numpy(x)))
assert np.allclose(result_numpy, result1)
assert np.allclose(result1, result2)
just_sum = backend.layers().Reduce("...->", reduction="sum")
variable = backend.from_numpy(x)
result = just_sum(layer(variable))
result.backward()
assert np.allclose(backend.to_numpy(variable.grad), 1)
def test_rearrange_symbolic():
for backend in collect_test_backends(symbolic=True, layers=True):
print("Test layer for ", backend.framework_name)
for pattern, axes_lengths, input_shape, _wrong_shapes in rearrangement_patterns:
x = np.arange(np.prod(input_shape), dtype="float32").reshape(input_shape)
result_numpy = rearrange(x, pattern, **axes_lengths)
layer = backend.layers().Rearrange(pattern, **axes_lengths)
input_shape_of_nones = [None] * len(input_shape)
shapes = [input_shape, input_shape_of_nones]
for shape in shapes:
symbol = backend.create_symbol(shape)
eval_inputs = [(symbol, x)]
result_symbol1 = layer(symbol)
result1 = backend.eval_symbol(result_symbol1, eval_inputs)
assert np.allclose(result_numpy, result1)
layer2 = pickle.loads(pickle.dumps(layer))
result_symbol2 = layer2(symbol)
result2 = backend.eval_symbol(result_symbol2, eval_inputs)
assert np.allclose(result1, result2)
# now testing back-propagation
just_sum = backend.layers().Reduce("...->", reduction="sum")
result_sum1 = backend.eval_symbol(just_sum(result_symbol1), eval_inputs)
result_sum2 = np.sum(x)
assert np.allclose(result_sum1, result_sum2)
reduction_patterns = [
*rearrangement_patterns,
testcase("b c h w -> b ()", dict(b=10), (10, 20, 30, 40), [(10,), (10, 20, 30)]),
testcase("b c (h1 h2) (w1 w2) -> b c h1 w1", dict(h1=15, h2=2, w2=2), (10, 20, 30, 40), [(10, 20, 31, 40)]),
testcase("b ... c -> b", dict(b=10), (10, 20, 30, 40), [(10,), (11, 10)]),
]
def test_reduce_imperative():
for backend in collect_test_backends(symbolic=False, layers=True):
print("Test layer for ", backend.framework_name)
for reduction in REDUCTIONS:
for pattern, axes_lengths, input_shape, wrong_shapes in reduction_patterns:
print(backend, reduction, pattern, axes_lengths, input_shape, wrong_shapes)
x = np.arange(1, 1 + np.prod(input_shape), dtype="float32").reshape(input_shape)
x /= x.mean()
result_numpy = reduce(x, pattern, reduction, **axes_lengths)
layer = backend.layers().Reduce(pattern, reduction, **axes_lengths)
for shape in wrong_shapes:
try:
layer(backend.from_numpy(np.zeros(shape, dtype="float32")))
except BaseException:
pass
else:
raise AssertionError("Failure expected")
# simple pickling / unpickling
layer2 = pickle.loads(pickle.dumps(layer))
result1 = backend.to_numpy(layer(backend.from_numpy(x)))
result2 = backend.to_numpy(layer2(backend.from_numpy(x)))
assert np.allclose(result_numpy, result1)
assert np.allclose(result1, result2)
just_sum = backend.layers().Reduce("...->", reduction="sum")
variable = backend.from_numpy(x)
result = just_sum(layer(variable))
result.backward()
grad = backend.to_numpy(variable.grad)
if reduction == "sum":
assert np.allclose(grad, 1)
if reduction == "mean":
assert np.allclose(grad, grad.min())
if reduction in ["max", "min"]:
assert np.all(np.isin(grad, [0, 1]))
assert np.sum(grad) > 0.5
def test_reduce_symbolic():
for backend in collect_test_backends(symbolic=True, layers=True):
print("Test layer for ", backend.framework_name)
for reduction in REDUCTIONS:
for pattern, axes_lengths, input_shape, _wrong_shapes in reduction_patterns:
x = np.arange(1, 1 + np.prod(input_shape), dtype="float32").reshape(input_shape)
x /= x.mean()
result_numpy = reduce(x, pattern, reduction, **axes_lengths)
layer = backend.layers().Reduce(pattern, reduction, **axes_lengths)
input_shape_of_nones = [None] * len(input_shape)
shapes = [input_shape, input_shape_of_nones]
for shape in shapes:
symbol = backend.create_symbol(shape)
eval_inputs = [(symbol, x)]
result_symbol1 = layer(symbol)
result1 = backend.eval_symbol(result_symbol1, eval_inputs)
assert np.allclose(result_numpy, result1)
layer2 = pickle.loads(pickle.dumps(layer))
result_symbol2 = layer2(symbol)
result2 = backend.eval_symbol(result_symbol2, eval_inputs)
assert np.allclose(result1, result2)
def create_torch_model(use_reduce=False, add_scripted_layer=False):
if not is_backend_tested("torch"):
pytest.skip()
else:
import torch.jit
from torch.nn import Conv2d, Linear, MaxPool2d, ReLU, Sequential
from einops.layers.torch import EinMix, Rearrange, Reduce
return Sequential(
Conv2d(3, 6, kernel_size=(5, 5)),
Reduce("b c (h h2) (w w2) -> b c h w", "max", h2=2, w2=2) if use_reduce else MaxPool2d(kernel_size=2),
Conv2d(6, 16, kernel_size=(5, 5)),
Reduce("b c (h h2) (w w2) -> b c h w", "max", h2=2, w2=2),
torch.jit.script(Rearrange("b c h w -> b (c h w)"))
if add_scripted_layer
else Rearrange("b c h w -> b (c h w)"),
Linear(16 * 5 * 5, 120),
ReLU(),
Linear(120, 84),
ReLU(),
EinMix("b c1 -> (b c2)", weight_shape="c1 c2", bias_shape="c2", c1=84, c2=84),
EinMix("(b c2) -> b c3", weight_shape="c2 c3", bias_shape="c3", c2=84, c3=84),
Linear(84, 10),
)
def test_torch_layer():
if not is_backend_tested("torch"):
pytest.skip()
else:
# checked that torch present
import torch
import torch.jit
model1 = create_torch_model(use_reduce=True)
model2 = create_torch_model(use_reduce=False)
input = torch.randn([10, 3, 32, 32])
# random models have different predictions
assert not torch.allclose(model1(input), model2(input))
model2.load_state_dict(pickle.loads(pickle.dumps(model1.state_dict())))
assert torch.allclose(model1(input), model2(input))
# tracing (freezing)
model3 = torch.jit.trace(model2, example_inputs=input)
torch.testing.assert_close(model1(input), model3(input), atol=1e-3, rtol=1e-3)
torch.testing.assert_close(model1(input + 1), model3(input + 1), atol=1e-3, rtol=1e-3)
model4 = torch.jit.trace(model2, example_inputs=input)
torch.testing.assert_close(model1(input), model4(input), atol=1e-3, rtol=1e-3)
torch.testing.assert_close(model1(input + 1), model4(input + 1), atol=1e-3, rtol=1e-3)
def test_torch_layers_scripting():
if not is_backend_tested("torch"):
pytest.skip()
else:
import torch
for script_layer in [False, True]:
model1 = create_torch_model(use_reduce=True, add_scripted_layer=script_layer)
model2 = torch.jit.script(model1)
input = torch.randn([10, 3, 32, 32])
torch.testing.assert_close(model1(input), model2(input), atol=1e-3, rtol=1e-3)
def test_keras_layer():
rng = np.random.default_rng()
if not is_backend_tested("tensorflow"):
pytest.skip()
else:
import tensorflow as tf
if tf.__version__ < "2.16.":
# current implementation of layers follows new TF interface
pytest.skip()
from tensorflow.keras.layers import Conv2D as Conv2d
from tensorflow.keras.layers import Dense as Linear
from tensorflow.keras.layers import ReLU
from tensorflow.keras.models import Sequential
from einops.layers.keras import EinMix, Rearrange, Reduce, keras_custom_objects
def create_keras_model():
return Sequential(
[
Conv2d(6, kernel_size=5, input_shape=[32, 32, 3]),
Reduce("b c (h h2) (w w2) -> b c h w", "max", h2=2, w2=2),
Conv2d(16, kernel_size=5),
Reduce("b c (h h2) (w w2) -> b c h w", "max", h2=2, w2=2),
Rearrange("b c h w -> b (c h w)"),
Linear(120),
ReLU(),
Linear(84),
ReLU(),
EinMix("b c1 -> (b c2)", weight_shape="c1 c2", bias_shape="c2", c1=84, c2=84),
EinMix("(b c2) -> b c3", weight_shape="c2 c3", bias_shape="c3", c2=84, c3=84),
Linear(10),
]
)
model1 = create_keras_model()
model2 = create_keras_model()
input = rng.normal(size=[10, 32, 32, 3]).astype("float32")
# two randomly init models should provide different outputs
assert not np.allclose(model1.predict_on_batch(input), model2.predict_on_batch(input))
# get some temp filename
tmp_model_filename = "/tmp/einops_tf_model.h5"
# save arch + weights
print("temp_path_keras1", tmp_model_filename)
tf.keras.models.save_model(model1, tmp_model_filename)
model3 = tf.keras.models.load_model(tmp_model_filename, custom_objects=keras_custom_objects)
np.testing.assert_allclose(model1.predict_on_batch(input), model3.predict_on_batch(input))
weight_filename = "/tmp/einops_tf_model.weights.h5"
# save arch as json
model4 = tf.keras.models.model_from_json(model1.to_json(), custom_objects=keras_custom_objects)
model1.save_weights(weight_filename)
model4.load_weights(weight_filename)
model2.load_weights(weight_filename)
# check that differently-inialized model receives same weights
np.testing.assert_allclose(model1.predict_on_batch(input), model2.predict_on_batch(input))
# ulimate test
# save-load architecture, and then load weights - should return same result
np.testing.assert_allclose(model1.predict_on_batch(input), model4.predict_on_batch(input))
def test_flax_layers():
"""
One-off simple tests for Flax layers.
Unfortunately, Flax layers have a different interface from other layers.
"""
if not is_backend_tested("jax"):
pytest.skip()
else:
import flax
import jax
import jax.numpy as jnp
from flax import linen as nn
from einops.layers.flax import EinMix, Rearrange, Reduce
class NN(nn.Module):
@nn.compact
def __call__(self, x):
x = EinMix(
"b (h h2) (w w2) c -> b h w c_out", "h2 w2 c c_out", "c_out", sizes=dict(h2=2, w2=3, c=4, c_out=5)
)(x)
x = Rearrange("b h w c -> b (w h c)", sizes=dict(c=5))(x)
x = Reduce("b hwc -> b", "mean", dict(hwc=2 * 3 * 5))(x)
return x
model = NN()
fixed_input = jnp.ones([10, 2 * 2, 3 * 3, 4])
params = model.init(jax.random.PRNGKey(0), fixed_input)
def eval_at_point(params):
return jnp.linalg.norm(model.apply(params, fixed_input))
vandg = jax.value_and_grad(eval_at_point)
value0 = eval_at_point(params)
value1, grad1 = vandg(params)
assert jnp.allclose(value0, value1)
if jax.__version__ < "0.6.0":
tree_map = jax.tree_map
else:
tree_map = jax.tree.map
params2 = tree_map(lambda x1, x2: x1 - x2 * 0.001, params, grad1)
value2 = eval_at_point(params2)
assert value0 >= value2, (value0, value2)
# check serialization
fbytes = flax.serialization.to_bytes(params)
_loaded = flax.serialization.from_bytes(params, fbytes)
def test_einmix_decomposition():
"""
Testing that einmix correctly decomposes into smaller transformations.
"""
from einops.layers._einmix import _EinmixDebugger
mixin1 = _EinmixDebugger(
"a b c d e -> e d c b a",
weight_shape="d a b",
d=2, a=3, b=5,
) # fmt: off
assert mixin1.pre_reshape_pattern is None
assert mixin1.post_reshape_pattern is None
assert mixin1.einsum_pattern == "abcde,dab->edcba"
assert mixin1.saved_weight_shape == [2, 3, 5]
assert mixin1.saved_bias_shape is None
mixin2 = _EinmixDebugger(
"a b c d e -> e d c b a",
weight_shape="d a b",
bias_shape="a b c d e",
a=1, b=2, c=3, d=4, e=5,
) # fmt: off
assert mixin2.pre_reshape_pattern is None
assert mixin2.post_reshape_pattern is None
assert mixin2.einsum_pattern == "abcde,dab->edcba"
assert mixin2.saved_weight_shape == [4, 1, 2]
assert mixin2.saved_bias_shape == [5, 4, 3, 2, 1]
mixin3 = _EinmixDebugger(
"... -> ...",
weight_shape="",
bias_shape="",
) # fmt: off
assert mixin3.pre_reshape_pattern is None
assert mixin3.post_reshape_pattern is None
assert mixin3.einsum_pattern == "...,->..."
assert mixin3.saved_weight_shape == []
assert mixin3.saved_bias_shape == []
mixin4 = _EinmixDebugger(
"b a ... -> b c ...",
weight_shape="b a c",
a=1, b=2, c=3,
) # fmt: off
assert mixin4.pre_reshape_pattern is None
assert mixin4.post_reshape_pattern is None
assert mixin4.einsum_pattern == "ba...,bac->bc..."
assert mixin4.saved_weight_shape == [2, 1, 3]
assert mixin4.saved_bias_shape is None
mixin5 = _EinmixDebugger(
"(b a) ... -> b c (...)",
weight_shape="b a c",
a=1, b=2, c=3,
) # fmt: off
assert mixin5.pre_reshape_pattern == "(b a) ... -> b a ..."
assert mixin5.pre_reshape_lengths == dict(a=1, b=2)
assert mixin5.post_reshape_pattern == "b c ... -> b c (...)"
assert mixin5.einsum_pattern == "ba...,bac->bc..."
assert mixin5.saved_weight_shape == [2, 1, 3]
assert mixin5.saved_bias_shape is None
mixin6 = _EinmixDebugger(
"b ... (a c) -> b ... (a d)",
weight_shape="c d",
bias_shape="a d",
a=1, c=3, d=4,
) # fmt: off
assert mixin6.pre_reshape_pattern == "b ... (a c) -> b ... a c"
assert mixin6.pre_reshape_lengths == dict(a=1, c=3)
assert mixin6.post_reshape_pattern == "b ... a d -> b ... (a d)"
assert mixin6.einsum_pattern == "b...ac,cd->b...ad"
assert mixin6.saved_weight_shape == [3, 4]
assert mixin6.saved_bias_shape == [1, 1, 4] # (b) a d, ellipsis does not participate
mixin7 = _EinmixDebugger(
"a ... (b c) -> a (... d b)",
weight_shape="c d b",
bias_shape="d b",
b=2, c=3, d=4,
) # fmt: off
assert mixin7.pre_reshape_pattern == "a ... (b c) -> a ... b c"
assert mixin7.pre_reshape_lengths == dict(b=2, c=3)
assert mixin7.post_reshape_pattern == "a ... d b -> a (... d b)"
assert mixin7.einsum_pattern == "a...bc,cdb->a...db"
assert mixin7.saved_weight_shape == [3, 4, 2]
assert mixin7.saved_bias_shape == [1, 4, 2] # (a) d b, ellipsis does not participate
def test_einmix_restrictions():
"""
Testing different cases
"""
from einops.layers._einmix import _EinmixDebugger
with pytest.raises(EinopsError):
_EinmixDebugger(
"a b c d e -> e d c b a",
weight_shape="d a b",
d=2, a=3, # missing b
) # fmt: off
with pytest.raises(EinopsError):
_EinmixDebugger(
"a b c d e -> e d c b a",
weight_shape="w a b",
d=2, a=3, b=1 # missing d
) # fmt: off
with pytest.raises(EinopsError):
_EinmixDebugger(
"(...) a -> ... a",
weight_shape="a", a=1, # ellipsis on the left
) # fmt: off
with pytest.raises(EinopsError):
_EinmixDebugger(
"(...) a -> a ...",
weight_shape="a", a=1, # ellipsis on the right side after bias axis
bias_shape="a",
) # fmt: off