feat(mge/module):add param pack

GitOrigin-RevId: 9cf1dbe44d

python_module/megengine/core/tensor_nn.py

@@ -25,5 +25,14 @@ class Parameter(Tensor):
 
     def __init__(self, value, *, dtype=None, device=None, requires_grad=True):
         # pylint: disable=super-init-not-called
-        t = tensor(value, dtype=dtype, device=device, requires_grad=requires_grad)
+        if isinstance(value, Tensor):
+            t = value
+        else:
+            t = tensor(value, dtype=dtype, device=device, requires_grad=requires_grad)
         self.__dict__.update(t.__dict__)
+    
+    @property
+    def shape(self):
+        r"""Return shape of parameter.
+        """
+        return self._symvar.imm_shape

python_module/megengine/module/__init__.py

@@ -16,3 +16,4 @@ from .linear import Linear
 from .module import Module
 from .pooling import AvgPool2d, MaxPool2d
 from .sequential import Sequential
+from .parampack import ParamPack

python_module/megengine/module/module.py

@@ -168,6 +168,29 @@ class Module(metaclass=ABCMeta):
         """
         yield from self._flatten(predicate=_is_buffer, recursive=recursive)
 
+    def replace_param(self,
+                      params: dict,
+                      start_pos: int,
+                      seen: Optional[Set[int]] = None):
+        offset = 0
+        if seen is None:
+            seen = set([id(self)])
+        module_dict = vars(self)
+        for key in sorted(module_dict):
+            hash_id = id(module_dict[key])
+            if hash_id in seen:
+                continue
+            seen.add(hash_id)
+            if isinstance(module_dict[key], Parameter):
+                if start_pos + offset in params:
+                    assert module_dict[key].shape == params[start_pos +
+                                                            offset].shape
+                    module_dict[key] = params[start_pos + offset]
+                offset += 1
+            if isinstance(module_dict[key], Module):
+                offset += module_dict[key].replace_param(params, start_pos + offset, seen)
+        return offset
+
     def named_buffers(
         self, prefix: str = "", recursive: bool = True
     ) -> Iterable[Tuple[str, Buffer]]:

python_module/megengine/module/parampack.py

@@ -0,0 +1,117 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import collections
+from typing import Iterable, Optional
+import numpy as np
+
+from ..core import Parameter, Tensor
+from .module import Module
+from .._internal.opr import param_pack_split
+
+
+class ParamPack(Module):
+    def __init__(self,
+                 model: Module,
+                 nr_ignore_first:int = 8,
+                 max_size_per_group: int = 10,
+                 max_nr_params_per_group: int = 100):
+        super().__init__()
+        self._model = model
+        self._nr_ignore_first = nr_ignore_first
+        self._max_size_per_group = max_size_per_group
+        self._max_nr_params_per_group = max_nr_params_per_group
+        self._grouped_params = []
+        self._packed_params = []
+
+        params = model.parameters()
+        self._pack_params(params)
+
+    def parameters(self, requires_grad: Optional[bool] = None) -> Iterable[Parameter]:
+        for param in self._packed_params:
+            if requires_grad is None or param.requires_grad == requires_grad:
+                yield param
+
+    def _pack_params(self, params: Iterable[Parameter]):
+        groups = collections.defaultdict(list)
+        ignored = 0
+        param_id = 0
+        for param in params:
+            if self._nr_ignore_first > ignored:
+                ignored += 1
+                self._grouped_params.append([{'tensor': param, 'id': param_id}])
+                self._packed_params.append(param)
+            else:
+                key = (param.dtype, param.device, param.requires_grad)
+                groups[key].append({'tensor': param, 'id': param_id})
+            param_id += 1
+        for (dtype, device, requires_grad) in groups.keys():
+            dtype_sz = np.dtype(dtype).itemsize
+            align = device.mem_align
+            if align < dtype_sz:
+                align = 1
+            else:
+                assert align % dtype_sz == 0
+                align //= dtype_sz
+
+            group = groups[(dtype, device, requires_grad)]
+            while group:
+                aligned_pos = []
+                offset = 0
+                params = []
+                idx = 0
+                while idx < len(group):
+                    param = group[idx]
+                    assert param['tensor'].device == device
+                    padding = (align - (offset & (align - 1))) & (align - 1)
+                    offset += padding
+                    aligned_pos.append(offset)
+                    params.append(param)
+                    offset += int(np.prod(param['tensor'].shape))
+                    idx += 1
+
+                    if (offset * dtype_sz >=
+                            self._max_size_per_group * 1024 * 1024
+                            or idx >= self._max_nr_params_per_group):
+                        break
+                group = group[idx:]
+                if idx == 1:
+                    # ignore param packs with only one item
+                    self._packed_params.append(params[0])
+                    self._grouped_params.append(params)
+                    continue
+
+                packed_value = np.zeros((offset, ), dtype=dtype)
+                for param, pos in zip(params, aligned_pos):
+                    val = param['tensor'].numpy()
+                    packed_value[pos:pos + val.size] = val.flatten()
+                new_param = Parameter(value=packed_value,
+                                      device=device,
+                                      dtype=dtype,
+                                      requires_grad=requires_grad)
+                self._packed_params.append(new_param)
+                self._grouped_params.append(params)
+
+    def forward(self, *args, **kwargs):
+        replace_param = dict()
+        for i in range(len(self._packed_params)):
+            packed_param = self._packed_params[i]
+            grouped_params = self._grouped_params[i]
+            if len(grouped_params) == 1:
+                continue
+            split = param_pack_split(packed_param._symvar,
+                                     [i['tensor'].shape for i in grouped_params])
+            split = [
+                Parameter(Tensor(i, requires_grad=packed_param.requires_grad))
+                for i in split
+            ]
+            for j in range(len(split)):
+                replace_param[grouped_params[j]['id']] = split[j]
+        self._model.replace_param(replace_param, 0)
+
+        return self._model.forward(*args, **kwargs)

python_module/megengine/optimizer/optimizer.py

@@ -168,6 +168,8 @@ class Optimizer(metaclass=ABCMeta):
 
         cg = get_default_graph()
         grads = grad_func(loss, params, use_virtual_grad=not cg.is_eager())
+        if not isinstance(grads, list):
+            grads = [grads]
         assert len(grads) == len(params)
 
         for param, grad in zip(params, grads):

python_module/test/integration/test_parampack.py

@@ -0,0 +1,207 @@
+# -*- coding: utf-8 -*-
+# MegEngine is Licensed under the Apache License, Version 2.0 (the "License")
+#
+# Copyright (c) 2014-2020 Megvii Inc. All rights reserved.
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT ARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+import itertools
+
+import numpy as np
+import pytest
+
+import megengine as mge
+from megengine.core import tensor
+from megengine.functional import cross_entropy_with_softmax, tanh
+from megengine.jit import trace
+from megengine.module import Linear, Module, ParamPack
+from megengine.optimizer import SGD
+
+batch_size = 64
+data_shape = (batch_size, 2)
+label_shape = (batch_size,)
+
+
+def minibatch_generator():
+    while True:
+        inp_data = np.zeros((batch_size, 2))
+        label = np.zeros(batch_size, dtype=np.int32)
+        for i in range(batch_size):
+            # [x0, x1], sampled from U[-1, 1]
+            inp_data[i, :] = np.random.rand(2) * 2 - 1
+            label[i] = 0 if np.prod(inp_data[i]) < 0 else 1
+        yield inp_data.astype(np.float32), label.astype(np.int32)
+
+
+def calculate_precision(data: np.ndarray, pred: np.ndarray) -> float:
+    """ Calculate precision for given data and prediction.
+
+    :type data: [[x, y], ...]
+    :param data: Input data
+    :type pred: [[x_pred, y_pred], ...]
+    :param pred: Network output data
+    """
+    correct = 0
+    assert len(data) == len(pred)
+    for inp_data, pred_output in zip(data, pred):
+        label = 0 if np.prod(inp_data) < 0 else 1
+        pred_label = np.argmax(pred_output)
+        if pred_label == label:
+            correct += 1
+    return float(correct) / len(data)
+
+
+class XORNet(Module):
+    def __init__(self):
+        self.mid_layers = 14
+        self.num_class = 2
+        super().__init__()
+
+        self.fc0 = Linear(self.num_class, self.mid_layers, bias=True)
+        self.fc1 = Linear(self.mid_layers, self.mid_layers, bias=True)
+
+        self.fc2 = Linear(self.mid_layers, self.num_class, bias=True)
+
+    def forward(self, x):
+        x = self.fc0(x)
+        x = tanh(x)
+        x = self.fc1(x)
+        x = tanh(x)
+        x = self.fc2(x)
+        return x
+
+
+@pytest.mark.slow
+def test_static_graph_parampack():
+    net = XORNet()
+    net = ParamPack(net,
+                    nr_ignore_first=0,
+                    max_size_per_group=10,
+                    max_nr_params_per_group=100)
+    opt = SGD(
+        net.parameters(requires_grad=True), lr=0.01, momentum=0.9, weight_decay=5e-4
+    )
+
+    @trace(symbolic=True)
+    def train(data, label):
+        pred = net(data)
+        opt.zero_grad()
+        loss = cross_entropy_with_softmax(pred, label)
+        opt.backward(loss)
+        return loss
+
+    @trace(symbolic=True)
+    def infer(data):
+        return net(data)
+
+    train_dataset = minibatch_generator()
+    losses = []
+
+    for data, label in itertools.islice(train_dataset, 2000):
+        loss = train(data, label)
+        loss = loss[0][0]
+        opt.step()
+        losses.append(loss.numpy())
+
+    assert np.mean(losses[-100:]) < 0.1, "Final training Loss must be low enough"
+
+    data, _ = next(train_dataset)
+    pred = infer(data).numpy()
+    assert calculate_precision(data, pred) > 0.95, "Test precision must be high enough"
+
+@pytest.mark.slow
+def test_dynamic_graph_parampack():
+    net = XORNet()
+    net = ParamPack(net,
+                    nr_ignore_first=0,
+                    max_size_per_group=10,
+                    max_nr_params_per_group=100)
+    opt = SGD(
+        net.parameters(requires_grad=True), lr=0.01, momentum=0.9, weight_decay=5e-4
+    )
+
+    @trace(symbolic=False)
+    def train(data, label):
+        pred = net(data)
+        opt.zero_grad()
+        loss = cross_entropy_with_softmax(pred, label)
+        opt.backward(loss)
+        return loss
+
+    @trace(symbolic=False)
+    def infer(data):
+        return net(data)
+
+    train_dataset = minibatch_generator()
+    losses = []
+
+    for data, label in itertools.islice(train_dataset, 2000):
+        loss = train(data, label)
+        loss = loss[0][0]
+        opt.step()
+        losses.append(loss.numpy())
+
+    assert np.mean(losses[-100:]) < 0.1, "Final training Loss must be low enough"
+
+    data, _ = next(train_dataset)
+    pred = infer(data).numpy()
+    assert calculate_precision(data, pred) > 0.95, "Test precision must be high enough"
+
+@pytest.mark.slow
+def test_correctness_parampack():
+    net1 = XORNet()
+    net2 = XORNet()
+    params1 = net1.parameters()
+    params2 = net2.parameters()
+    for param1, param2 in zip(params1, params2):
+        param1.set_value(param2.numpy())
+    net1 = ParamPack(net1,
+                     nr_ignore_first=0,
+                     max_size_per_group=10,
+                     max_nr_params_per_group=100)
+    opt1 = SGD(
+        net1.parameters(requires_grad=True), lr=0.01, momentum=0.9, weight_decay=5e-4
+    )
+
+    opt2 = SGD(
+        net2.parameters(requires_grad=True), lr=0.01, momentum=0.9, weight_decay=5e-4
+    )
+
+    @trace(symbolic=False)
+    def train1(data, label):
+        pred = net1(data)
+        opt1.zero_grad()
+        loss = cross_entropy_with_softmax(pred, label)
+        opt1.backward(loss)
+        return loss
+
+    @trace(symbolic=False)
+    def train2(data, label):
+        pred = net2(data)
+        opt2.zero_grad()
+        loss = cross_entropy_with_softmax(pred, label)
+        opt2.backward(loss)
+        return loss
+
+    @trace(symbolic=False)
+    def infer1(data):
+        return net1(data)
+
+    @trace(symbolic=False)
+    def infer2(data):
+        return net2(data)
+
+    train_dataset = minibatch_generator()
+
+    for data, label in itertools.islice(train_dataset, 2000):
+        train1(data, label)
+        opt1.step()
+
+        train2(data, label)
+        opt2.step()
+
+    data, _ = next(train_dataset)
+    pred1 = infer1(data).numpy()
+    pred2 = infer2(data).numpy()
+    assert np.allclose(pred1, pred2)