最近做实验要用到CMU-MOSI数据集，网上搜到的教程很少，经过一天时间的探索，最终成功安装配置数据集，这篇文章完整地整理一下该数据集的下载与使用方法。

配置环境：

window10，anaconda

1. 需要下载的内容

步骤1：下载官方github的SDK包：CMU-MultiComp-Lab/CMU-MultimodalSDK (github.com)

步骤2：解压的路径需要保存

2.anaconda环境配置

官方github的readme中写了需要配置环境，但该命令是基于linux系统，windows系统需要按照以下步骤设置。

步骤1：在anaconda的虚拟环境 路径下的Libsite-packages，创建一个文本文档，命名为’mypkpath‘，在该文档中添加上一步的SDK路径，保存之后将文件后缀改为“pth”。

 步骤2：用下面的命令安装mmsdk的依赖包：

pip install h5py validators tqdm numpy argparse requests

步骤3：因为win10和Linux的文件路径分隔符不同，所以这里要修改mmsdk里的文件：

修改文件1：路径如下：

****CMU-MultimodalSDK-mainmmsdkmmdatasdkcomputational_sequencedownload_ops.py

在文档最后修改为：(read和URL中间加一个'_') 

 修改文件2：和上一个文件同级的不同文件：computational_sequence.py

****CMU-MultimodalSDK-mainmmsdkmmdatasdkcomputational_sequencecomputational_sequence.py

文档第68行左右，把分隔符从'os.sep'改为'/'：

步骤4：启动conda的虚拟环境，导入mmsdk包验证是否安装成功，没报错就安装成功

from mmsdk import mmdatasdk as md

3. 下载数据集

步骤一：下载数据集，DATA_PATH是将下载的csd文件存放的目录，

from mmsdk import mmdatasdk as md
DATASET = md.cmu_mosi
DATA_PATH = './data'
try:
	md.mmdataset(DATASET.highlevel,DATA_PATH)
except:
	print('have been downloaded')

步骤2：加载csd文件并对齐

from mmsdk import mmdatasdk as md
DATASET = md.cmu_mosi
DATA_PATH = './data'

data_files = os.listdir(DATA_PATH)
print('
'.join(data_files))

visual_field = 'CMU_MOSI_Visual_Facet_41'
acoustic_field = 'CMU_MOSI_COVAREP'
text_field = 'CMU_MOSI_ModifiedTimestampedWords'

features = [
    text_field,
    visual_field,
    acoustic_field
]

recipe = {feat: os.path.join(DATA_PATH, feat) + '.csd' for feat in features}
dataset = md.mmdataset(recipe)
print(list(dataset[text_field].keys())[55])
print('done!')

步骤3：对齐csd文件并保存对齐之后的结果。保存的csd文件会存储到'./deployed'文件夹中，下一次使用数据时直接从deployed文件中加载数据就ok，这样就不用重复下载和对齐。

def avg(intervals: np.array, features: np.array) -> np.array:
    try:
        return np.average(features, axis=0)
    except:
        return features

# first we align to words with averaging, collapse_function receives a list of functions
dataset.align(text_field, collapse_functions=[avg])
label_field = 'CMU_MOSI_Opinion_Labels'

# we add and align to lables to obtain labeled segments
# this time we don't apply collapse functions so that the temporal sequences are preserved
label_recipe = {label_field: os.path.join(DATA_PATH, label_field + '.csd')}
dataset.add_computational_sequences(label_recipe, destination=None)
dataset.align(label_field, replace=True)
print(list(dataset[text_field].keys())[55])
print('done!')


###保存
deploy_files={x:x for x in dataset.computational_sequences.keys()}
dataset.deploy("./deployed",deploy_files)
aligned_cmumosi_highlevel=md.mmdataset('./deployed')

4. 一次小小的完整训练

from mmsdk import mmdatasdk
import os, re
import numpy as np
from torch.utils.data import DataLoader
from collections import defaultdict
import torch
import torch.nn as nn
from tqdm import tqdm_notebook
from torch.optim import Adam, SGD
from sklearn.metrics import accuracy_score
import torch, os
from torch.nn.utils.rnn import pad_sequence, pack_padded_sequence, pad_packed_sequence
from torch.utils.data import DataLoader, Dataset
from tqdm import tqdm_notebook

word2id = defaultdict(lambda: len(word2id))
UNK = word2id['<unk>']
PAD = word2id['<pad>']
# def return_unk():
#     return UNK
# word2id.default_factory = return_unk
def splittraindevtest(train_split, dev_split, test_split, dataset, features):
    EPS = 0
    text_field, visual_field, acoustic_field, label_field = features
    # construct a word2id mapping that automatically takes increment when new words are encountered
    # place holders for the final train/dev/test dataset
    train = []
    dev = []
    test = []
    # define a regular expression to extract the video ID out of the keys
    pattern = re.compile('(.*)[.*]')
    num_drop = 0  # a counter to count how many data points went into some processing issues
    for segment in dataset[label_field].keys():

        # get the video ID and the features out of the aligned dataset
        vid = re.search(pattern, segment).group(1)
        label = dataset[label_field][segment]['features']
        _words = dataset[text_field][segment]['features']
        _visual = dataset[visual_field][segment]['features']
        _acoustic = dataset[acoustic_field][segment]['features']

        # if the sequences are not same length after alignment, there must be some problem with some modalities
        # we should drop it or inspect the data again
        if not _words.shape[0] == _visual.shape[0] == _acoustic.shape[0]:
            print(
                f"Encountered datapoint {vid} with text shape {_words.shape}, visual shape {_visual.shape}, acoustic shape {_acoustic.shape}")
            num_drop += 1
            continue

        # remove nan values
        label = np.nan_to_num(label)
        _visual = np.nan_to_num(_visual)
        _acoustic = np.nan_to_num(_acoustic)

        # remove speech pause tokens - this is in general helpful
        # we should remove speech pauses and corresponding visual/acoustic features together
        # otherwise modalities would no longer be aligned
        words = []
        visual = []
        acoustic = []
        for i, word in enumerate(_words):
            if word[0] != b'sp':
                words.append(word2id[word[0].decode('utf-8')])  # SDK stores strings as bytes, decode into strings here
                visual.append(_visual[i, :])
                acoustic.append(_acoustic[i, :])

        words = np.asarray(words)
        visual = np.asarray(visual)
        acoustic = np.asarray(acoustic)

        # z-normalization per instance and remove nan/infs
        visual = np.nan_to_num((visual - visual.mean(0, keepdims=True)) / (EPS + np.std(visual, axis=0, keepdims=True)))
        acoustic = np.nan_to_num(
            (acoustic - acoustic.mean(0, keepdims=True)) / (EPS + np.std(acoustic, axis=0, keepdims=True)))

        if vid in train_split:
            train.append(((words, visual, acoustic), label, segment))
        elif vid in dev_split:
            dev.append(((words, visual, acoustic), label, segment))
        elif vid in test_split:
            test.append(((words, visual, acoustic), label, segment))
        else:
            print(f"Found video that doesn't belong to any splits: {vid}")

    def return_unk():
        return UNK

    word2id.default_factory = return_unk
    print(f"Total number of {num_drop} datapoints have been dropped.")
    return train, dev, test

def multi_collate(batch):
    '''
    Collate functions assume batch = [Dataset[i] for i in index_set]
    '''
    # for later use we sort the batch in descending order of length
    batch = sorted(batch, key=lambda x: x[0][0].shape[0], reverse=True)

    # get the data out of the batch - use pad sequence util functions from PyTorch to pad things
    labels = torch.cat([torch.from_numpy(sample[1]) for sample in batch], dim=0)
    sentences = pad_sequence([torch.LongTensor(sample[0][0]) for sample in batch], padding_value=PAD)
    visual = pad_sequence([torch.FloatTensor(sample[0][1]) for sample in batch])
    acoustic = pad_sequence([torch.FloatTensor(sample[0][2]) for sample in batch])

    # lengths are useful later in using RNNs
    lengths = torch.LongTensor([sample[0][0].shape[0] for sample in batch])
    return sentences, visual, acoustic, labels, lengths


class LFLSTM(nn.Module):
    def __init__(self, input_sizes, hidden_sizes, fc1_size, output_size, dropout_rate):
        super(LFLSTM, self).__init__()
        self.input_size = input_sizes
        self.hidden_size = hidden_sizes
        self.fc1_size = fc1_size
        self.output_size = output_size
        self.dropout_rate = dropout_rate

        # defining modules - two layer bidirectional LSTM with layer norm in between
        self.embed = nn.Embedding(len(word2id), input_sizes[0])
        self.trnn1 = nn.LSTM(input_sizes[0], hidden_sizes[0], bidirectional=True)
        self.trnn2 = nn.LSTM(2 * hidden_sizes[0], hidden_sizes[0], bidirectional=True)

        self.vrnn1 = nn.LSTM(input_sizes[1], hidden_sizes[1], bidirectional=True)
        self.vrnn2 = nn.LSTM(2 * hidden_sizes[1], hidden_sizes[1], bidirectional=True)

        self.arnn1 = nn.LSTM(input_sizes[2], hidden_sizes[2], bidirectional=True)
        self.arnn2 = nn.LSTM(2 * hidden_sizes[2], hidden_sizes[2], bidirectional=True)

        self.fc1 = nn.Linear(sum(hidden_sizes) * 4, fc1_size)
        self.fc2 = nn.Linear(fc1_size, output_size)
        self.relu = nn.ReLU()
        self.dropout = nn.Dropout(dropout_rate)
        self.tlayer_norm = nn.LayerNorm((hidden_sizes[0] * 2,))
        self.vlayer_norm = nn.LayerNorm((hidden_sizes[1] * 2,))
        self.alayer_norm = nn.LayerNorm((hidden_sizes[2] * 2,))
        self.bn = nn.BatchNorm1d(sum(hidden_sizes) * 4)

    def extract_features(self, sequence, lengths, rnn1, rnn2, layer_norm):
        packed_sequence = pack_padded_sequence(sequence, lengths.cpu())
        packed_h1, (final_h1, _) = rnn1(packed_sequence)
        padded_h1, _ = pad_packed_sequence(packed_h1)
        normed_h1 = layer_norm(padded_h1)
        packed_normed_h1 = pack_padded_sequence(normed_h1, lengths.cpu())
        _, (final_h2, _) = rnn2(packed_normed_h1)
        return final_h1, final_h2

    def fusion(self, sentences, visual, acoustic, lengths):
        batch_size = lengths.size(0)
        sentences = self.embed(sentences)

        # extract features from text modality
        final_h1t, final_h2t = self.extract_features(sentences, lengths, self.trnn1, self.trnn2, self.tlayer_norm)

        # extract features from visual modality
        final_h1v, final_h2v = self.extract_features(visual, lengths, self.vrnn1, self.vrnn2, self.vlayer_norm)

        # extract features from acoustic modality
        final_h1a, final_h2a = self.extract_features(acoustic, lengths, self.arnn1, self.arnn2, self.alayer_norm)

        # simple late fusion -- concatenation + normalization
        h = torch.cat((final_h1t, final_h2t, final_h1v, final_h2v, final_h1a, final_h2a),
                      dim=2).permute(1, 0, 2).contiguous().view(batch_size, -1)
        return self.bn(h)

    def forward(self, sentences, visual, acoustic, lengths):
        batch_size = lengths.size(0)
        h = self.fusion(sentences, visual, acoustic, lengths)
        h = self.fc1(h)
        h = self.dropout(h)
        h = self.relu(h)
        o = self.fc2(h)
        return o
def load_emb(w2i, path_to_embedding, embedding_size=300, embedding_vocab=2196017, init_emb=None):
    if init_emb is None:
        emb_mat = np.random.randn(len(w2i), embedding_size)
    else:
        emb_mat = init_emb
    f = open(path_to_embedding, 'r')
    found = 0
    for line in tqdm_notebook(f, total=embedding_vocab):
        content = line.strip().split()
        vector = np.asarray(list(map(lambda x: float(x), content[-300:])))
        word = ' '.join(content[:-300])
        if word in w2i:
            idx = w2i[word]
            emb_mat[idx, :] = vector
            found += 1
    print(f"Found {found} words in the embedding file.")
    return torch.tensor(emb_mat).float()
def run(train_loader, dev_loader, test_loader):
    torch.manual_seed(123)
    torch.cuda.manual_seed_all(123)

    CUDA = True  # torch.cuda.is_available()
    MAX_EPOCH = 5

    text_size = 300
    visual_size = 47
    acoustic_size = 74

    # define some model settings and hyper-parameters
    input_sizes = [text_size, visual_size, acoustic_size]
    hidden_sizes = [int(text_size * 1.5), int(visual_size * 1.5), int(acoustic_size * 1.5)]
    fc1_size = sum(hidden_sizes) // 2
    dropout = 0.25
    output_size = 1
    curr_patience = patience = 8
    num_trials = 3
    grad_clip_value = 1.0
    weight_decay = 0.1
    CACHE_PATH = r'D:Speechjupyter_notes_zyxCMU-MultimodalSDK-Tutorials-masterCMU-MultimodalSDK-Tutorials-masterdataembedding_and_mapping.pt'
    if os.path.exists(CACHE_PATH):
        pretrained_emb, word2id = torch.load(CACHE_PATH)

    else:
        pretrained_emb = None

    model = LFLSTM(input_sizes, hidden_sizes, fc1_size, output_size, dropout)
    if pretrained_emb is not None:
        model.embed.weight.data = pretrained_emb
    model.embed.requires_grad = False
    optimizer = Adam([param for param in model.parameters() if param.requires_grad], weight_decay=weight_decay)

    if CUDA:
        model.cuda()
    criterion = nn.L1Loss(reduction='sum')
    criterion_test = nn.L1Loss(reduction='sum')
    best_valid_loss = float('inf')
    lr_scheduler = torch.optim.lr_scheduler.StepLR(optimizer, step_size=1, gamma=0.1)
    lr_scheduler.step()  # for some reason it seems the StepLR needs to be stepped once first
    train_losses = []
    valid_losses = []
    for e in range(MAX_EPOCH):
        model.train()
        train_iter = tqdm_notebook(train_loader)
        train_loss = 0.0
        for batch in train_iter:
            model.zero_grad()
            t, v, a, y, l = batch
            batch_size = t.size(0)
            if CUDA:
                t = t.cuda()
                v = v.cuda()
                a = a.cuda()
                y = y.cuda()
                l = l.cuda()
            y_tilde = model(t, v, a, l)
            loss = criterion(y_tilde, y)
            loss.backward()
            torch.nn.utils.clip_grad_value_([param for param in model.parameters() if param.requires_grad], grad_clip_value)
            optimizer.step()
            train_iter.set_description(f"Epoch {e}/{MAX_EPOCH}, current batch loss: {round(loss.item() / batch_size, 4)}")
            train_loss += loss.item()
        train_loss = train_loss / len(train)
        train_losses.append(train_loss)
        print(f"Training loss: {round(train_loss, 4)}")

        model.eval()
        with torch.no_grad():
            valid_loss = 0.0
            for batch in dev_loader:
                model.zero_grad()
                t, v, a, y, l = batch
                if CUDA:
                    t = t.cuda()
                    v = v.cuda()
                    a = a.cuda()
                    y = y.cuda()
                    l = l.cuda()
                y_tilde = model(t, v, a, l)
                loss = criterion(y_tilde, y)
                valid_loss += loss.item()

        valid_loss = valid_loss / len(dev)
        valid_losses.append(valid_loss)
        print(f"Validation loss: {round(valid_loss, 4)}")
        print(f"Current patience: {curr_patience}, current trial: {num_trials}.")
        if valid_loss <= best_valid_loss:
            best_valid_loss = valid_loss
            print("Found new best model on dev set!")
            torch.save(model.state_dict(), 'model.std')
            torch.save(optimizer.state_dict(), 'optim.std')
            curr_patience = patience
        else:
            curr_patience -= 1
            if curr_patience <= -1:
                print("Running out of patience, loading previous best model.")
                num_trials -= 1
                curr_patience = patience
                model.load_state_dict(torch.load('model.std'))
                optimizer.load_state_dict(torch.load('optim.std'))
                lr_scheduler.step()
                print(f"Current learning rate: {optimizer.state_dict()['param_groups'][0]['lr']}")

        if num_trials <= 0:
            print("Running out of patience, early stopping.")
            break

    model.load_state_dict(torch.load('model.std'))
    y_true = []
    y_pred = []
    model.eval()
    with torch.no_grad():
        test_loss = 0.0
        for batch in test_loader:
            model.zero_grad()
            t, v, a, y, l = batch
            if CUDA:
                t = t.cuda()
                v = v.cuda()
                a = a.cuda()
                y = y.cuda()
                l = l.cuda()
            y_tilde = model(t, v, a, l)
            loss = criterion_test(y_tilde, y)
            y_true.append(y_tilde.detach().cpu().numpy())
            y_pred.append(y.detach().cpu().numpy())
            test_loss += loss.item()
    print(f"Test set performance: {test_loss / len(test)}")
    y_true = np.concatenate(y_true, axis=0)
    y_pred = np.concatenate(y_pred, axis=0)

    y_true_bin = y_true >= 0
    y_pred_bin = y_pred >= 0
    bin_acc = accuracy_score(y_true_bin, y_pred_bin)
    print(f"Test set accuracy is {bin_acc}")

if __name__=='__main__':
    visual_field1 = 'CMU_MOSI_Visual_Facet_41'
    acoustic_field1 = 'CMU_MOSI_COVAREP'
    text_field1 = 'CMU_MOSI_ModifiedTimestampedWords'
    label_field1 = 'CMU_MOSI_Opinion_Labels'
    features1 = [
        text_field1,
        visual_field1,
        acoustic_field1,
        label_field1
    ]
    DATA_PATH1 = './mymydeployed'
    recipe1 = {feat: os.path.join(DATA_PATH1, feat) + '.csd' for feat in features1}
    dataset1 = mmdatasdk.mmdataset(recipe1)
    print(list(dataset1[text_field1].keys())[55])
    tensors = dataset1.get_tensors(seq_len=25, non_sequences=["Opinion Segment Labels"], direction=False,
                                   folds=[mmdatasdk.cmu_mosi.standard_folds.standard_train_fold, mmdatasdk.cmu_mosi.
                                   standard_folds.standard_valid_fold,
                                          mmdatasdk.cmu_mosi.standard_folds.standard_test_fold])

    fold_names = ["train", "valid", "test"]
    train, dev, test = splittraindevtest(mmdatasdk.cmu_mosi.standard_folds.standard_train_fold, mmdatasdk.cmu_mosi.
                                   standard_folds.standard_valid_fold, mmdatasdk.cmu_mosi.standard_folds.standard_test_fold,
                                         dataset1, features1)

    batch_sz = 56
    train_loader = DataLoader(train, shuffle=True, batch_size=batch_sz, collate_fn=multi_collate)
    dev_loader = DataLoader(dev, shuffle=False, batch_size=batch_sz * 3, collate_fn=multi_collate)
    test_loader = DataLoader(test, shuffle=False, batch_size=batch_sz * 3, collate_fn=multi_collate)
    run(train_loader, dev_loader, test_loader)