From 60544b5226b11e5ee457b81d806af8613548be6b Mon Sep 17 00:00:00 2001 From: ceci3 Date: Fri, 15 May 2020 11:15:01 +0800 Subject: [PATCH] fix something wrong about rlnas (#269) (#280) * fix * fix * update * update --- demo/nas/README.md | 32 ++- demo/nas/parl_nas_mobilenetv2.py | 244 ++++++++++++++++++ docs/zh_cn/quick_start/nas_tutorial.md | 18 +- docs/zh_cn/tutorials/sanas_darts_space.md | 24 +- paddleslim/common/__init__.py | 2 +- paddleslim/common/client.py | 2 +- paddleslim/common/rl_controller/base_env.py | 29 +++ .../rl_controller/ddpg/ddpg_controller.py | 22 +- .../rl_controller/lstm/lstm_controller.py | 28 +- paddleslim/common/server.py | 2 + 10 files changed, 355 insertions(+), 48 deletions(-) create mode 100644 demo/nas/parl_nas_mobilenetv2.py create mode 100644 paddleslim/common/rl_controller/base_env.py diff --git a/demo/nas/README.md b/demo/nas/README.md index f7ff5e32..b3ee0d18 100644 --- a/demo/nas/README.md +++ b/demo/nas/README.md @@ -1,6 +1,6 @@ -# 网络结构搜索示例 +# SANAS网络结构搜索示例 -本示例介绍如何使用网络结构搜索接口，搜索到一个更小或者精度更高的模型，该文档仅介绍paddleslim中SANAS的使用及如何利用SANAS得到模型结构，完整示例代码请参考sa_nas_mobilenetv2.py或者block_sa_nas_mobilenetv2.py。 +本示例介绍如何使用网络结构搜索接口，搜索到一个更小或者精度更高的模型，该示例介绍paddleslim中SANAS的使用及如何利用SANAS得到模型结构，完整示例代码请参考sa_nas_mobilenetv2.py或者block_sa_nas_mobilenetv2.py。 ## 数据准备 本示例默认使用cifar10数据，cifar10数据会根据调用的paddle接口自动下载，无需额外准备。 @@ -8,7 +8,7 @@ ## 接口介绍 请参考神经网络搜索API文档。 -本示例为在MobileNetV2的搜索空间上搜索FLOPs更小的模型。 +本示例为利用SANAS在MobileNetV2的搜索空间上搜索FLOPs更小的模型。 ## 1 搜索空间配置 默认搜索空间为`MobileNetV2`，详细的搜索空间配置请参考搜索空间配置文档。 @@ -24,3 +24,29 @@ CUDA_VISIBLE_DEVICES=0 python sa_nas_mobilenetv2.py ```shell CUDA_VISIBLE_DEVICES=0 python block_sa_nas_mobilenetv2.py ``` + +# RLNAS网络结构搜索示例 + +本示例介绍如何使用RLNAS接口进行网络结构搜索，该示例介绍paddleslim中RLNAS的使用，完整示例代码请参考rl_nas_mobilenetv2.py或者parl_nas_mobilenetv2.py。 + +## 数据准备 +本示例默认使用cifar10数据，cifar10数据会根据调用的paddle接口自动下载，无需额外准备。 + +## 接口介绍 +请参考神经网络搜索API文档。 + +示例为利用SANAS在MobileNetV2的搜索空间上搜索精度更高的模型。 +## 1 搜索空间配置 +默认搜索空间为`MobileNetV2`，详细的搜索空间配置请参考搜索空间配置文档。 + +## 2 启动训练 + +### 2.1 启动基于MobileNetV2初始模型结构构造搜索空间，强化学习算法为lstm的搜索实验 +```shell +CUDA_VISIBLE_DEVICES=0 python rl_nas_mobilenetv2.py +``` + +### 2.2 启动基于MobileNetV2初始模型结构构造搜索空间，强化学习算法为ddpg的搜索实验 +```shell +CUDA_VISIBLE_DEVICES=0 python parl_nas_mobilenetv2.py +``` diff --git a/demo/nas/parl_nas_mobilenetv2.py b/demo/nas/parl_nas_mobilenetv2.py new file mode 100644 index 00000000..1d698742 --- /dev/null +++ b/demo/nas/parl_nas_mobilenetv2.py @@ -0,0 +1,244 @@ +import sys +sys.path.append('..') +import numpy as np +import argparse +import ast +import time +import argparse +import ast +import logging +import paddle +import paddle.fluid as fluid +from paddle.fluid.param_attr import ParamAttr +from paddleslim.nas import RLNAS +from paddleslim.common import get_logger +from optimizer import create_optimizer +import imagenet_reader + +_logger = get_logger(__name__, level=logging.INFO) + + +def create_data_loader(image_shape): + data_shape = [None] + image_shape + data = fluid.data(name='data', shape=data_shape, dtype='float32') + label = fluid.data(name='label', shape=[None, 1], dtype='int64') + data_loader = fluid.io.DataLoader.from_generator( + feed_list=[data, label], + capacity=1024, + use_double_buffer=True, + iterable=True) + return data_loader, data, label + + +def build_program(main_program, + startup_program, + image_shape, + archs, + args, + is_test=False): + with fluid.program_guard(main_program, startup_program): + with fluid.unique_name.guard(): + data_loader, data, label = create_data_loader(image_shape) + output = archs(data) + output = fluid.layers.fc(input=output, size=args.class_dim) + + softmax_out = fluid.layers.softmax(input=output, use_cudnn=False) + cost = fluid.layers.cross_entropy(input=softmax_out, label=label) + avg_cost = fluid.layers.mean(cost) + acc_top1 = fluid.layers.accuracy( + input=softmax_out, label=label, k=1) + acc_top5 = fluid.layers.accuracy( + input=softmax_out, label=label, k=5) + + if is_test == False: + optimizer = create_optimizer(args) + optimizer.minimize(avg_cost) + return data_loader, avg_cost, acc_top1, acc_top5 + + +def search_mobilenetv2(config, args, image_size, is_server=True): + if is_server: + ### start a server and a client + rl_nas = RLNAS( + key='ddpg', + configs=config, + is_sync=False, + obs_dim=26, ### step + length_of_token + server_addr=(args.server_address, args.port)) + else: + ### start a client + rl_nas = RLNAS( + key='ddpg', + configs=config, + is_sync=False, + obs_dim=26, + server_addr=(args.server_address, args.port), + is_server=False) + + image_shape = [3, image_size, image_size] + for step in range(args.search_steps): + if step == 0: + action_prev = [1. for _ in rl_nas.range_tables] + else: + action_prev = rl_nas.tokens[0] + obs = [step] + obs.extend(action_prev) + archs = rl_nas.next_archs(obs=obs)[0][0] + + train_program = fluid.Program() + test_program = fluid.Program() + startup_program = fluid.Program() + train_loader, avg_cost, acc_top1, acc_top5 = build_program( + train_program, startup_program, image_shape, archs, args) + + test_loader, test_avg_cost, test_acc_top1, test_acc_top5 = build_program( + test_program, + startup_program, + image_shape, + archs, + args, + is_test=True) + test_program = test_program.clone(for_test=True) + + place = fluid.CUDAPlace(0) if args.use_gpu else fluid.CPUPlace() + exe = fluid.Executor(place) + exe.run(startup_program) + + if args.data == 'cifar10': + train_reader = paddle.batch( + paddle.reader.shuffle( + paddle.dataset.cifar.train10(cycle=False), buf_size=1024), + batch_size=args.batch_size, + drop_last=True) + + test_reader = paddle.batch( + paddle.dataset.cifar.test10(cycle=False), + batch_size=args.batch_size, + drop_last=False) + elif args.data == 'imagenet': + train_reader = paddle.batch( + imagenet_reader.train(), + batch_size=args.batch_size, + drop_last=True) + test_reader = paddle.batch( + imagenet_reader.val(), + batch_size=args.batch_size, + drop_last=False) + + train_loader.set_sample_list_generator( + train_reader, + places=fluid.cuda_places() if args.use_gpu else fluid.cpu_places()) + test_loader.set_sample_list_generator(test_reader, places=place) + + build_strategy = fluid.BuildStrategy() + train_compiled_program = fluid.CompiledProgram( + train_program).with_data_parallel( + loss_name=avg_cost.name, build_strategy=build_strategy) + for epoch_id in range(args.retain_epoch): + for batch_id, data in enumerate(train_loader()): + fetches = [avg_cost.name] + s_time = time.time() + outs = exe.run(train_compiled_program, + feed=data, + fetch_list=fetches)[0] + batch_time = time.time() - s_time + if batch_id % 10 == 0: + _logger.info( + 'TRAIN: steps: {}, epoch: {}, batch: {}, cost: {}, batch_time: {}ms'. + format(step, epoch_id, batch_id, outs[0], batch_time)) + + reward = [] + for batch_id, data in enumerate(test_loader()): + test_fetches = [ + test_avg_cost.name, test_acc_top1.name, test_acc_top5.name + ] + batch_reward = exe.run(test_program, + feed=data, + fetch_list=test_fetches) + reward_avg = np.mean(np.array(batch_reward), axis=1) + reward.append(reward_avg) + + _logger.info( + 'TEST: step: {}, batch: {}, avg_cost: {}, acc_top1: {}, acc_top5: {}'. + format(step, batch_id, batch_reward[0], batch_reward[1], + batch_reward[2])) + + finally_reward = np.mean(np.array(reward), axis=0) + _logger.info( + 'FINAL TEST: avg_cost: {}, acc_top1: {}, acc_top5: {}'.format( + finally_reward[0], finally_reward[1], finally_reward[2])) + + obs = np.expand_dims(obs, axis=0).astype('float32') + actions = rl_nas.tokens + obs_next = [step + 1] + obs_next.extend(actions[0]) + obs_next = np.expand_dims(obs_next, axis=0).astype('float32') + + if step == args.search_steps - 1: + terminal = np.expand_dims([True], axis=0).astype(np.bool) + else: + terminal = np.expand_dims([False], axis=0).astype(np.bool) + rl_nas.reward( + np.expand_dims( + np.float32(finally_reward[1]), axis=0), + obs=obs, + actions=actions.astype('float32'), + obs_next=obs_next, + terminal=terminal) + + if step == 2: + sys.exit(0) + + +if __name__ == '__main__': + + parser = argparse.ArgumentParser( + description='RL NAS MobileNetV2 cifar10 argparase') + parser.add_argument( + '--use_gpu', + type=ast.literal_eval, + default=True, + help='Whether to use GPU in train/test model.') + parser.add_argument( + '--batch_size', type=int, default=256, help='batch size.') + parser.add_argument( + '--class_dim', type=int, default=10, help='classify number.') + parser.add_argument( + '--data', + type=str, + default='cifar10', + choices=['cifar10', 'imagenet'], + help='server address.') + parser.add_argument( + '--is_server', + type=ast.literal_eval, + default=True, + help='Whether to start a server.') + parser.add_argument( + '--search_steps', + type=int, + default=100, + help='controller server number.') + parser.add_argument( + '--server_address', type=str, default="", help='server ip.') + parser.add_argument('--port', type=int, default=8881, help='server port') + parser.add_argument( + '--retain_epoch', type=int, default=5, help='epoch for each token.') + parser.add_argument('--lr', type=float, default=0.1, help='learning rate.') + args = parser.parse_args() + print(args) + + if args.data == 'cifar10': + image_size = 32 + block_num = 3 + elif args.data == 'imagenet': + image_size = 224 + block_num = 6 + else: + raise NotImplementedError( + 'data must in [cifar10, imagenet], but received: {}'.format( + args.data)) + + config = [('MobileNetV2Space')] + + search_mobilenetv2(config, args, image_size, is_server=args.is_server) diff --git a/docs/zh_cn/quick_start/nas_tutorial.md b/docs/zh_cn/quick_start/nas_tutorial.md index 94216ba5..e68cb7fe 100644 --- a/docs/zh_cn/quick_start/nas_tutorial.md +++ b/docs/zh_cn/quick_start/nas_tutorial.md @@ -9,12 +9,12 @@ 4. 定义输入数据函数 5. 定义训练函数 6. 定义评估函数 -7. 启动搜索实验 - 7.1 获取模型结构 - 7.2 构造program - 7.3 定义输入数据 - 7.4 训练模型 - 7.5 评估模型 +7. 启动搜索实验 + 7.1 获取模型结构 + 7.2 构造program + 7.3 定义输入数据 + 7.4 训练模型 + 7.5 评估模型 7.6 回传当前模型的得分 8. 完整示例 @@ -53,7 +53,7 @@ def build_program(archs): acc_top1 = fluid.layers.accuracy(input=softmax_out, label=label, k=1) acc_top5 = fluid.layers.accuracy(input=softmax_out, label=label, k=5) test_program = fluid.default_main_program().clone(for_test=True) - + optimizer = fluid.optimizer.Adam(learning_rate=0.1) optimizer.minimize(avg_cost) @@ -77,7 +77,7 @@ def input_data(inputs): ## 5. 定义训练函数 根据训练program和训练数据进行训练。 ```python -def start_train(program, data_reader, data_feeder): +def start_train(program, data_reader, data_feeder): outputs = [avg_cost.name, acc_top1.name, acc_top5.name] for data in data_reader(): batch_reward = exe.run(program, feed=data_feeder.feed(data), fetch_list = outputs) @@ -146,7 +146,7 @@ for step in range(3): current_flops = slim.analysis.flops(train_program) if current_flops > 321208544: continue - + for epoch in range(7): start_train(train_program, train_reader, train_feeder) diff --git a/docs/zh_cn/tutorials/sanas_darts_space.md b/docs/zh_cn/tutorials/sanas_darts_space.md index d934aa23..939ded2b 100644 --- a/docs/zh_cn/tutorials/sanas_darts_space.md +++ b/docs/zh_cn/tutorials/sanas_darts_space.md @@ -15,13 +15,13 @@ 6. 定义造program的函数 7. 定义训练函数 8. 定义预测函数 -9. 启动搜索 - 9.1 获取下一个模型结构 - 9.2 构造相应的训练和预测program - 9.3 添加搜索限制 - 9.4 定义环境 - 9.5 定义输入数据 - 9.6 启动训练和评估 +9. 启动搜索 + 9.1 获取下一个模型结构 + 9.2 构造相应的训练和预测program + 9.3 添加搜索限制 + 9.4 定义环境 + 9.5 定义输入数据 + 9.6 启动训练和评估 9.7 回传当前模型的得分reward 10. 利用demo下的脚本启动搜索 11. 利用demo下的脚本启动最终实验 @@ -33,7 +33,7 @@ 按照通道数来区分DARTS_model中block的话，则DARTS_model中共有3个block，第一个block仅包含6个normal cell，之后的两个block每个block都包含和一个reduction cell和6个normal cell，共有20个cell。在构造搜索空间的时候我们定义每个cell中的所有卷积操作都使用相同的通道数，共有20位token。 -完整的搜索空间可以参考[基于DARTS_model的搜索空间](../../../paddleslim/nas/search_space/darts_space.py) +完整的搜索空间可以参考[基于DARTS_model的搜索空间](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/paddleslim/nas/search_space/darts_space.py) ### 2. 引入依赖包并定义全局变量 ```python @@ -232,9 +232,9 @@ exe.run(startup_program) ``` #### 9.5 定义输入数据 -由于本示例中对cifar10中的图片进行了一些额外的预处理操作，和[快速开始](../quick_start/nas_tutorial.md)示例中的reader不同，所以需要自定义cifar10的reader，不能直接调用paddle中封装好的`paddle.dataset.cifar10`的reader。自定义cifar10的reader文件位于[demo/nas](../../../demo/nas/darts_cifar10_reader.py)中。 +由于本示例中对cifar10中的图片进行了一些额外的预处理操作，和[快速开始](https://paddlepaddle.github.io/PaddleSlim/quick_start/nas_tutorial.html)示例中的reader不同，所以需要自定义cifar10的reader，不能直接调用paddle中封装好的`paddle.dataset.cifar10`的reader。自定义cifar10的reader文件位于[demo/nas](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/demo/nas/darts_cifar10_reader.py)中。 -**注意：**本示例为了简化代码直接调用`paddle.dataset.cifar10`定义训练数据和预测数据，实际训练需要使用自定义cifar10的reader。 +**注意：**本示例为了简化代码直接调用`paddle.dataset.cifar10`定义训练数据和预测数据，实际训练需要使用自定义cifar10文件中的reader。 ```python train_reader = paddle.batch(paddle.reader.shuffle(paddle.dataset.cifar.train10(cycle=False), buf_size=1024), batch_size=BATCH_SIZE, drop_last=True) test_reader = paddle.batch(paddle.dataset.cifar.test10(cycle=False), batch_size=BATCH_SIZE, drop_last=False) @@ -261,14 +261,14 @@ sa_nas.reward(float(valid_top1_list[-1] + valid_top1_list[-2]) / 2) ### 10. 利用demo下的脚本启动搜索 -搜索文件位于: [darts_sanas_demo](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/nas/sanas_darts_nas.py)，搜索过程中限制模型参数量为不大于3.77M。 +搜索文件位于: [darts_sanas_demo](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/demo/nas/sanas_darts_space.py)，搜索过程中限制模型参数量为不大于3.77M。 ```python cd demo/nas/ python darts_nas.py ``` ### 11. 利用demo下的脚本启动最终实验 -最终实验文件位于: [darts_sanas_demo](https://github.com/PaddlePaddle/PaddleSlim/tree/develop/demo/nas/sanas_darts_nas.py)，最终实验需要训练600epoch。以下示例输入token为`[5, 5, 0, 5, 5, 10, 7, 7, 5, 7, 7, 11, 10, 12, 10, 0, 5, 3, 10, 8]`。 +最终实验文件位于: [darts_sanas_demo](https://github.com/PaddlePaddle/PaddleSlim/blob/develop/demo/nas/sanas_darts_space.py)，最终实验需要训练600epoch。以下示例输入token为`[5, 5, 0, 5, 5, 10, 7, 7, 5, 7, 7, 11, 10, 12, 10, 0, 5, 3, 10, 8]`。 ```python cd demo/nas/ python darts_nas.py --token 5 5 0 5 5 10 7 7 5 7 7 11 10 12 10 0 5 3 10 8 --retain_epoch 600 diff --git a/paddleslim/common/__init__.py b/paddleslim/common/__init__.py index 9a21bb77..4d1eb1e6 100644 --- a/paddleslim/common/__init__.py +++ b/paddleslim/common/__init__.py @@ -11,7 +11,7 @@ # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. -from .controller import EvolutionaryController +from .controller import EvolutionaryController, RLBaseController from .sa_controller import SAController from .log_helper import get_logger from .controller_server import ControllerServer diff --git a/paddleslim/common/client.py b/paddleslim/common/client.py index f5fded05..a0feacf1 100644 --- a/paddleslim/common/client.py +++ b/paddleslim/common/client.py @@ -99,7 +99,7 @@ class Client(object): assert self._params_dict != None, "Please call next_token to get token first, then call update" current_params_dict = self._controller.update( rewards, self._params_dict, **kwargs) - params_grad = compute_grad(self._params_dict, current_params_dict) + params_grad = compute_grad(current_params_dict, self._params_dict) _logger.debug("Client: update weight {}".format(self._client_name)) self._client_socket.send_multipart([ pickle.dumps(ConnectMessage.UPDATE_WEIGHT), diff --git a/paddleslim/common/rl_controller/base_env.py b/paddleslim/common/rl_controller/base_env.py new file mode 100644 index 00000000..cd20dc3f --- /dev/null +++ b/paddleslim/common/rl_controller/base_env.py @@ -0,0 +1,29 @@ +# Copyright (c) 2020 PaddlePaddle Authors. All Rights Reserve. +# +# Licensed under the Apache License, Version 2.0 (the "License"); +# you may not use this file except in compliance with the License. +# You may obtain a copy of the License at +# +# http://www.apache.org/licenses/LICENSE-2.0 +# +# Unless required by applicable law or agreed to in writing, software +# distributed under the License is distributed on an "AS IS" BASIS, +# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. +# See the License for the specific language governing permissions and +# limitations under the License. +"""Base environment used in reinforcement learning""" + +import numpy as np + +__all__ = ['BaseEnv'] + + +class BaseEnv: + def reset(self): + raise NotImplementedError('Abstract method.') + + def step(self): + raise NotImplementedError('Abstract method.') + + def _build_state_embedding(self): + raise NotImplementedError('Abstract method.') diff --git a/paddleslim/common/rl_controller/ddpg/ddpg_controller.py b/paddleslim/common/rl_controller/ddpg/ddpg_controller.py index d25f556f..50216adb 100644 --- a/paddleslim/common/rl_controller/ddpg/ddpg_controller.py +++ b/paddleslim/common/rl_controller/ddpg/ddpg_controller.py @@ -41,24 +41,24 @@ class DDPGAgent(parl.Agent): self.learn_program = fluid.Program() with fluid.program_guard(self.pred_program): - obs = layers.data( - name='obs', shape=[self.obs_dim], dtype='float32') + obs = fluid.data( + name='obs', shape=[None, self.obs_dim], dtype='float32') self.pred_act = self.alg.predict(obs) with fluid.program_guard(self.learn_program): - obs = layers.data( - name='obs', shape=[self.obs_dim], dtype='float32') - act = layers.data( - name='act', shape=[self.act_dim], dtype='float32') - reward = layers.data(name='reward', shape=[], dtype='float32') - next_obs = layers.data( - name='next_obs', shape=[self.obs_dim], dtype='float32') - terminal = layers.data(name='terminal', shape=[], dtype='bool') + obs = fluid.data( + name='obs', shape=[None, self.obs_dim], dtype='float32') + act = fluid.data( + name='act', shape=[None, self.act_dim], dtype='float32') + reward = fluid.data(name='reward', shape=[None], dtype='float32') + next_obs = fluid.data( + name='next_obs', shape=[None, self.obs_dim], dtype='float32') + terminal = fluid.data( + name='terminal', shape=[None, 1], dtype='bool') _, self.critic_cost = self.alg.learn(obs, act, reward, next_obs, terminal) def predict(self, obs): - obs = np.expand_dims(obs, axis=0) act = self.fluid_executor.run(self.pred_program, feed={'obs': obs}, fetch_list=[self.pred_act])[0] diff --git a/paddleslim/common/rl_controller/lstm/lstm_controller.py b/paddleslim/common/rl_controller/lstm/lstm_controller.py index 20f5c632..0e32be6d 100644 --- a/paddleslim/common/rl_controller/lstm/lstm_controller.py +++ b/paddleslim/common/rl_controller/lstm/lstm_controller.py @@ -62,6 +62,7 @@ class LSTM(RLBaseController): self.lstm_num_layers = kwargs.get('lstm_num_layers') or 1 self.hidden_size = kwargs.get('hidden_size') or 100 self.temperature = kwargs.get('temperature') or None + self.controller_lr = kwargs.get('controller_lr') or 1e-4 self.tanh_constant = kwargs.get('tanh_constant') or None self.decay = kwargs.get('decay') or 0.99 self.weight_entropy = kwargs.get('weight_entropy') or None @@ -91,12 +92,6 @@ class LSTM(RLBaseController): return logits, output, new_states def _create_parameter(self): - self.emb_w = fluid.layers.create_parameter( - name='emb_w', - shape=(self.max_range_table, self.hidden_size), - dtype='float32', - default_initializer=uniform_initializer(1.0)) - self.g_emb = fluid.layers.create_parameter( name='emb_g', shape=(self.controller_batch_size, self.hidden_size), @@ -133,11 +128,16 @@ class LSTM(RLBaseController): axes=[1], starts=[idx], ends=[idx + 1]) + action = fluid.layers.squeeze(action, axes=[1]) action.stop_gradient = True else: action = fluid.layers.sampling_id(probs) actions.append(action) - log_prob = fluid.layers.cross_entropy(probs, action) + log_prob = fluid.layers.softmax_with_cross_entropy( + logits, + fluid.layers.reshape( + action, shape=[fluid.layers.shape(action), 1]), + axis=1) sample_log_probs.append(log_prob) entropy = log_prob * fluid.layers.exp(-1 * log_prob) @@ -145,10 +145,14 @@ class LSTM(RLBaseController): entropies.append(entropy) action_emb = fluid.layers.cast(action, dtype=np.int64) - inputs = fluid.layers.gather(self.emb_w, action_emb) + inputs = fluid.embedding( + action_emb, + size=(self.max_range_table, self.hidden_size), + param_attr=fluid.ParamAttr( + name='emb_w', initializer=uniform_initializer(1.0))) - sample_log_probs = fluid.layers.stack(sample_log_probs) - self.sample_log_probs = fluid.layers.reduce_sum(sample_log_probs) + self.sample_log_probs = fluid.layers.concat( + sample_log_probs, axis=0) entropies = fluid.layers.stack(entropies) self.sample_entropies = fluid.layers.reduce_sum(entropies) @@ -196,7 +200,9 @@ class LSTM(RLBaseController): self.loss = self.sample_log_probs * (self.rewards - self.baseline) fluid.clip.set_gradient_clip( clip=fluid.clip.GradientClipByGlobalNorm(clip_norm=5.0)) - optimizer = fluid.optimizer.Adam(learning_rate=1e-3) + lr = fluid.layers.exponential_decay( + self.controller_lr, decay_steps=1000, decay_rate=0.8) + optimizer = fluid.optimizer.Adam(learning_rate=lr) optimizer.minimize(self.loss) def _create_input(self, is_test=True, actual_rewards=None): diff --git a/paddleslim/common/server.py b/paddleslim/common/server.py index dc7253e5..abd34e29 100644 --- a/paddleslim/common/server.py +++ b/paddleslim/common/server.py @@ -161,6 +161,8 @@ class Server(object): if len(self._client) == len( self._client_dict.items()): self._done = True + self._params_dict = sum_params_dict + del sum_params_dict self._server_socket.send_multipart([ pickle.dumps(ConnectMessage.WAIT), -- GitLab