華為開源自研AI框架昇思MindSpore應用案例：基于MindSpore框架實現PWCNet光流估計

如果你對MindSpore感興趣，可以關注昇思MindSpore社區

在這里插入圖片描述

1 環境準備

1.進入ModelArts官網
云平臺幫助用戶快速創建和部署模型，管理全周期AI工作流，選擇下面的云平臺以開始使用昇思MindSpore，可以在昇思教程中進入ModelArts官網

創建notebook，點擊【打開】啟動，進入ModelArts調試環境頁面。

注意選擇西南-貴陽一，mindspore_2.3.0

在這里插入圖片描述

等待環境搭建完成

在這里插入圖片描述

下載案例notebook文件

基于MindSpore框架實現PWCNet光流估計：https://github.com/mindspore-courses/applications/blob/master/pwc_net/pwc_net.ipynb

選擇ModelArts Upload Files上傳.ipynb文件

在這里插入圖片描述

進入昇思MindSpore官網，點擊上方的安裝獲取安裝命令

在這里插入圖片描述

MindSpore版本升級，鏡像自帶的MindSpore版本為2.3，該活動要求在MindSpore2.4.0版本體驗，所以需要進行MindSpore版本升級。
在這里插入圖片描述

命令如下：

export no_proxy='a.test.com,127.0.0.1,2.2.2.2'
pip install https://ms-release.obs.cn-north-4.myhuaweicloud.com/2.4.0/MindSpore/unified/aarch64/mindspore-2.4.0-cp39-cp39-linux_aarch64.whl --trusted-host ms-release.obs.cn-north-4.myhuaweicloud.com -i https://pypi.tuna.tsinghua.edu.cn/simple

在這里插入圖片描述

回到Notebook中，在第一塊代碼前加命令

pip install --upgrade pippip install mindvisionpip install download

2 案例實現

import os
import logging
os.environ['CUDA_VISIBLE_DEVICES'] = '0'
# logging.basicConfig(level=logging.ERROR)
logging.disable(logging.WARNING)

%matplotlib inline
import mindspore as mm.set_context(mode=m.GRAPH_MODE, device_target="GPU") # 訓練時使用靜態圖
# m.set_context(mode=m.PYNATIVE_MODE, device_target="GPU") # 設置為動態圖方便debug

將MindSpore設置為圖執行模式，并設置為使用GPU進行訓練。

train_data_path = r"data/MPI-Sintel-complete/training"
val_data_path = r"data/MPI-Sintel-complete/training"pretrained_path = r"pretrained_model/pwcnet-mindspore.ckpt"batch_size = 4
lr = 0.0001
num_parallel_workers = 4
lr_milestones = '6,10,12,16'
lr_gamma = 0.5
max_epoch = 20
loss_scale = 1024
warmup_epochs = 1

設置數據集路徑，設置訓練參數，包括batch_size、epoch_size、learning_rate等。

import mindspore.dataset.vision as Vfrom src.dataset_utils import RandomGammaaugmentation_list = [V.ToPIL(),V.RandomColorAdjust(brightness=0.5, contrast=0.5, saturation=0.5, hue=0.5),V.ToTensor(),RandomGamma(min_gamma=0.7, max_gamma=1.5, clip_image=True),
]

設置數據增強方法，包括使用隨機顏色變換和隨機Gamma變換。

from black import out
from src.dataset import getFlyingChairsTrainData, getSintelValDatadl_train, len_dl_train, dataset = getSintelValData(root=train_data_path,split="train",augmentations=augmentation_list,batch_size=batch_size,num_parallel_workers=num_parallel_workers,
)
dl_val, len_dl_val, val_dataset = getSintelValData(root=val_data_path,split="train",augmentations=augmentation_list,batch_size=batch_size,num_parallel_workers=num_parallel_workers,
)
train_len = dl_train.get_dataset_size()
dl_train = dl_train.repeat(max_epoch)
print(f"The dataset size of dl_train: {dl_train.get_dataset_size()}")
print(f"The dataset size of dl_val: {dl_val.get_dataset_size()}")dict_datasets = next(dl_train.create_dict_iterator())
print(dict_datasets.keys())
print(dict_datasets["im1"].shape)
print(dict_datasets["im2"].shape)
print(dict_datasets["flo"].shape)
print(type(dict_datasets["flo"]))
print(dict_datasets["flo"].max(), dict_datasets["flo"].min())
print(dict_datasets["flo"].max() * 0.05, dict_datasets["flo"].min() * 0.05)
dl_train = dl_train.create_tuple_iterator(output_numpy=False, do_copy=False)
dl_val = dl_val.create_tuple_iterator(output_numpy=False, do_copy=False)

查看數據集的訓練集和測試集的數量。同時查看數據集中RGB圖片和光流圖片的分辨率大小。

import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import flow_visfig = matplotlib.pyplot.gcf()
fig.set_size_inches(18.5, 10.5)
ax = plt.subplot(131)
ax.imshow(np.transpose(dict_datasets["im1"][0].asnumpy(), (1, 2, 0)))
ax.set_title("Image 1")
ax.set_axis_off()
ax = plt.subplot(132)
ax.imshow(np.transpose(dict_datasets["im2"][0].asnumpy(), (1, 2, 0)))
ax.set_title("Image 2")
ax.set_axis_off()
ax = plt.subplot(133)
ax.imshow(flow_vis.flow_to_color(np.transpose(dict_datasets["flo"][0].asnumpy(), (1, 2, 0)))
)
ax.set_axis_off()
ax.set_title("Optical Flow")

在這里插入圖片描述

使用flow_vis和matplotlib庫分別將光流圖片與RGB圖片可視化。

# from src.pwc_net import PWCNet
# from src.loss import PyramidEPE, MultiStepLR# from mindspore.nn import Adam# network = PWCNet()
# criterion = PyramidEPE()# optimizer = Adam(params=net.trainable_params(), learning_rate=lr, loss_scale=loss_scale)

from collections import Counter
import numpy as npclass _WarmUp():"""Basic class for warm up"""def __init__(self, warmup_init_lr):self.warmup_init_lr = warmup_init_lrdef get_lr(self):# Get learning rate during warmupraise NotImplementedErrorclass _LRScheduler():"""Basic class for learning rate scheduler"""def __init__(self, lr, max_epoch, steps_per_epoch):self.base_lr = lrself.steps_per_epoch = steps_per_epochself.total_steps = int(max_epoch * steps_per_epoch)def get_lr(self):# Compute learning rate using chainable form of the schedulerraise NotImplementedErrorclass _LinearWarmUp(_WarmUp):"""Class for linear warm up"""def __init__(self, lr, warmup_epochs, steps_per_epoch, warmup_init_lr=0):self.base_lr = lrself.warmup_init_lr = warmup_init_lrself.warmup_steps = int(warmup_epochs * steps_per_epoch)super(_LinearWarmUp, self).__init__(warmup_init_lr)def get_warmup_steps(self):return self.warmup_stepsdef get_lr(self, current_step):lr_inc = (float(self.base_lr) - float(self.warmup_init_lr)) / float(self.warmup_steps)lr = float(self.warmup_init_lr) + lr_inc * current_stepreturn lrclass MultiStepLR(_LRScheduler):"""Multi-step learning rate schedulerDecays the learning rate by gamma once the number of epoch reaches one of the milestones.Args:lr (float): Initial learning rate which is the lower boundary in the cycle.milestones (list): List of epoch indices. Must be increasing.gamma (float): Multiplicative factor of learning rate decay.steps_per_epoch (int): The number of steps per epoch to train for.max_epoch (int): The number of epochs to train for.warmup_epochs (int, optional): The number of epochs to Warmup. Default: 0Outputs:numpy.ndarray, shape=(1, steps_per_epoch*max_epoch)Example:>>> # Assuming optimizer uses lr = 0.05 for all groups>>> # lr = 0.05     if epoch < 30>>> # lr = 0.005    if 30 <= epoch < 80>>> # lr = 0.0005   if epoch >= 80>>> scheduler = MultiStepLR(lr=0.1, milestones=[30,80], gamma=0.1, steps_per_epoch=5000, max_epoch=90)>>> lr = scheduler.get_lr()"""def __init__(self, lr, milestones, gamma, steps_per_epoch, max_epoch, warmup_epochs=0):self.milestones = Counter(milestones)self.gamma = gammaself.warmup = _LinearWarmUp(lr, warmup_epochs, steps_per_epoch)super(MultiStepLR, self).__init__(lr, max_epoch, steps_per_epoch)def get_lr(self):warmup_steps = self.warmup.get_warmup_steps()lr_each_step = []current_lr = self.base_lrfor i in range(self.total_steps):if i < warmup_steps:lr = self.warmup.get_lr(i+1)else:cur_ep = i // self.steps_per_epochif i % self.steps_per_epoch == 0 and cur_ep in self.milestones:current_lr = current_lr * self.gammalr = current_lrlr_each_step.append(lr)return np.array(lr_each_step).astype(np.float32)

初始化神經網絡、損失函數、優化器、模型和回調函數。

import mindspore.nn as nn
import mindspore.ops as ops
from mindspore.train.callback import LossMonitor, TimeMonitor
from mindspore import Model, load_checkpoint, load_param_into_netfrom src.pwc_net import BuildTrainNetwork, PWCNet
from src.loss import PyramidEPEclass CustomWithLossCell(nn.Cell):def __init__(self, network, criterion):super(CustomWithLossCell, self).__init__(auto_prefix=False)self.network = networkself.criterion = criteriondef construct(self, im1, im2, flow):out = self.network(im1, im2)loss = self.criterion(out, flow)return lossnetwork = PWCNet()
criterion = PyramidEPE()param_dict = load_checkpoint(pretrained_path)
param_dict_new = {}
for key, values in param_dict.items():if key.startswith('moment1.' or 'moment2' or 'global_step' or 'beta1_power' or 'beta2_power' or'learning_rate'):continueelif key.startswith('network.'):param_dict_new[key[8:]] = valueselse:param_dict_new[key] = values
load_param_into_net(network, param_dict_new)train_net = BuildTrainNetwork(network, criterion)# model = Model(
#     network=net_with_loss,
#     # loss_fn=criterion,
#     optimizer=optimizer,
#     eval_network=net_with_loss,
#     metrics={"loss"},
#     amp_level="O0",
# )

import mindspore.nn as nn
import mindspore.ops as ops
from mindspore.train.callback import LossMonitor, TimeMonitor
from mindspore import Model, load_checkpoint, load_param_into_netfrom src.pwc_net import BuildTrainNetwork, PWCNet
from src.loss import PyramidEPEclass CustomWithLossCell(nn.Cell):def __init__(self, network, criterion):super(CustomWithLossCell, self).__init__(auto_prefix=False)self.network = networkself.criterion = criteriondef construct(self, im1, im2, flow):out = self.network(im1, im2)loss = self.criterion(out, flow)return lossnetwork = PWCNet()
criterion = PyramidEPE()param_dict = load_checkpoint(pretrained_path)
param_dict_new = {}
for key, values in param_dict.items():if key.startswith('moment1.' or 'moment2' or 'global_step' or 'beta1_power' or 'beta2_power' or'learning_rate'):continueelif key.startswith('network.'):param_dict_new[key[8:]] = valueselse:param_dict_new[key] = values
load_param_into_net(network, param_dict_new)train_net = BuildTrainNetwork(network, criterion)# model = Model(
#     network=net_with_loss,
#     # loss_fn=criterion,
#     optimizer=optimizer,
#     eval_network=net_with_loss,
#     metrics={"loss"},
#     amp_level="O0",
# )

print('Start training...')
for i, data in enumerate(dl_train):# clean grad + adjust lr + put data into device + forward + backward + optimizer, return loss# print(data[0].shape, data[1].shape, data[2].shape)# print(data[0].max(), data[0].min(), data[1].max(), data[1].min(), data[2].max(), data[2].min())loss = train_net_step(data[0], data[1], data[2])# print(loss)loss_meter.update(loss.asnumpy())if i == 0:time_for_graph_compile = time.time() - create_network_startprint('graph compile time={:.2f}s'.format(time_for_graph_compile))if i % 10 == 0 and i > 0:t_now = time.time()epoch = int(i / train_len)print('epoch: [{}], iter: [{}], loss: [{:.4f}], time: [{:.2f}]s'.format(epoch, i, loss_meter.avg, t_now - t_end))t_end = t_nowloss_meter.reset()if i % train_len == 0  and i > 0:epoch_time_used = time.time() - t_epochepoch = int(i / train_len)fps = batch_size * train_len / epoch_time_usedprint('=================================================')print('epoch[{}], iter[{}], [{:.2f}] imgs/sec'.format(epoch, i, fps))t_epoch = time.time()validation_loss = 0sum_num = 0for _, val_data in enumerate(dl_val):network.set_train(False)val_output = network(val_data[0], val_data[1], training=False)val_loss = criterion(val_output, val_data[2], training=False)validation_loss += val_losssum_num += 1if (validation_loss / sum_num) < best_val_loss:best_val_loss = validation_loss / sum_numprint('validation EPE: {}, best validation EPE: {}'.format(validation_loss / sum_num, best_val_loss))

在這里插入圖片描述

import matplotlib
import matplotlib.pyplot as plt
import numpy as np
import flow_visout_flow = network(dict_datasets['im1'][3][None, ...], dict_datasets['im2'][3][None, ...], training=False)fig = matplotlib.pyplot.gcf()
fig.set_size_inches(18.5, 10.5)
ax = plt.subplot(141)
ax.imshow(np.transpose(dict_datasets["im1"][3].asnumpy(), (1, 2, 0)))
ax.set_title("Image 1")
ax.set_axis_off()
ax = plt.subplot(142)
ax.imshow(np.transpose(dict_datasets["im2"][3].asnumpy(), (1, 2, 0)))
ax.set_title("Image 2")
ax.set_axis_off()
ax = plt.subplot(143)
ax.imshow(flow_vis.flow_to_color(np.transpose(dict_datasets["flo"][3].asnumpy(), (1, 2, 0)))
)
ax.set_axis_off()
ax.set_title("Optical Flow")
ax = plt.subplot(144)
ax.imshow(flow_vis.flow_to_color(np.transpose(out_flow[0].asnumpy(), (1, 2, 0)))
)
ax.set_axis_off()
ax.set_title("Predicted Optical Flow")
# plt.show()