EmotiW2018—NLPR Teams Code (Acc: 60.33%)

EmotiW2018 focuses on audio-video emotion classification tasks, which contains seven basic emotion categories: Angry, Disgust, Fear, Happy, Sad, Surprise and Neutral. We adopt additional databases: FER+ database, MEC2017 database and SFEW2015 database. (Note: The MEC2017 database can only be avaliable if you participate in the MEC2017 challenges. The SFEW2015 database can be avalible if you ask organizers for the download link.) Since lack of permission, we only show our final solution in the EmotiW2018. Any information related to the challenge database can no be found.

Finally, we got 60.33% accuracy in EmotiW2018. Our rank: 7/32. Our result is just slightly lower than the Top1's. Please notice that there are only 653 test samples. Compared with Top1 solution, we only 1.5% lower than Top1, which means that our difference is just between 10 samples. Therefore, our solution is still very competitive.

Through sharing our solution, we hope that we can promote the development of emotion recognition.

Our Paper is avaliable in paper

1. Setup

PyTorch
Tensorflow: extract C3D, Inception features, sound features
sklearn: training linearSVC, RBFSVC, RF classifiers
cyvlfeat: realize Fisher Vectors (Change Frame-level features into Video-level features)
openface: face extration (which is based on dlib)
intraface: face extraction (which is appplied under vs2010 + opencv)
Seetaface: extract indentification information.
ffmpeg: split audio from video, extract frames from video
fire: available command line input
tqdm: show progress

2. Files

Audio_OpenSmile/: Extract opensmile featrues for wav files
- Audio/: save origin wav files
- opensmile-2.3.0/: opensmile source path download_link
- featureExtract.py: Extract audio features
baiduASR/: Recognize the content for each audio and extract textual features
C3D-tensorflow-master/: Extract C3D features
devicehive-audio-analysis-master/: Extract youtube-8M audio features
EmotiW2017/: save samples and features for EmotiW2018
EmotiW2018/: the main fold
- checkpoints/: save fintuned models (like DenseNet mdoels, VGG models)
- data/: for data loader, Fisher Vector and other functions
- Database/: store FER+ databases
- models/: image-level emotion classification models
- utils/: extract faces
- config.py: the configuration file
- main.py: fintuning pretrained models on FER+ database and SFEW database
- feature_extract.py: extract DenseNet and VGG bottleneck features
- classifier.py and run_train.py: train classifiers for different features
- classifier_lp.py and loupe.py: train 'NetVLAD', 'NetRVLAD', 'SoftDBoW', 'NetFV' classifiers
- predict.py: gain predictions for samples and save them.
- fusion_prediction.py and gain_cnn_para.py: gain fusion parameters and submission
MEC2017/: save samples and features for MEC2017 (same as EmotiW2017/)
Point_Pose_Indenfy/: Extract head pose, landmark points and Identification features.
Traditional_video_descriptor/: Extract DSIFT, HOG, HOG_LBP, LBP features. Since these codes are not written by me. I have no ability to open them. You can find extraction methods in the web.
youtube-8m-master/: Extract image-level Inception features of youtube-8M
Multi-model Emotion Recognition.pptx: conclusion of pervious EmotiW challenges.
EmotiW2018_LeaderBoard.pdf: leaderboard of EmotiW2018.

3. Data Preporcess

In this section, we show our face recognition methods.

Source codes are in EmotiW2018/utils

3.1 label Discription

Since MEC2017 and EmotiW2018 both are discrete emotion classification problems. Their labels have overlapping parts. Their 'label<->index map':

	MEC		EmotiW/SFEW
angry	0	Angry	0
disgust	1	Disgust	1
	2	Fear	2
happy	3	Happy	3
sad	4	Sad	4
surprise	5	Surprise	5
neutral	6	Neutral	6
anxious	7		7
worried	8		8

3.2 Face Extraction

Source codes are in EmotiW2018/utils

In this section, we change Video into Origin_Faces and Faces.

Origin_Faces: the origin frames in the video.
Faces: the aligned faces of each videos.

We utilize Intraface toolkit in the face extraction. Source code is based on vs2010+opencv. We provide our modified version. (link: https://pan.baidu.com/s/10TaL2tBSD-w4c_hApUbvIg password: zf7n)

## step 0: change video into frames
python Frame_Extraction.py all_video_to_frames --VIDEO_PATH='Video' --ORIGIN_FACES_PATH='Origin_Faces'

## step 1: recognize faces through intraface (realized under vs2010, C++, Window 10)
## source code is in 'IntraFace.rar' (change Origin_Faces into Faces)

## step 2: del empty files(which means faces in the video is false detected)
python MEC_EmotiW_data_clear.py data_item_clear --pic_root='Faces'

## step 3: find empty files and track faces through openface
### step 3.1: init bb in the first frame
python Track_Video.py all_video_to_bb --frames_root='Origin_Faces' --save_root='IMAGE' --bb_txt_save_path='IMAGE/bb.txt'
### step 3.2: mannual correct bb in the first frame (mannual change the bb.txt)
### step 3.3: show whether bb is correct
python Track_Video.py all_video_show_bb --image_root='IMAGE' --bb_txt_save_path='IMAGE/bb.txt'
### step 3.4: track bb
python Track_Video.py track_frames_from_video --origin_faces_root='Origin_Faces' --aligned_faces_root='Aligned_openface' --txt_save_path='IMAGE/bb.txt'

3.3 Data Format Normalization

MEC2017 path: MEC2017

data format：

Faces: Extract faces based on openface
train_eval_test_MEC2017_filter.npz: save emotion labels [pic_path='2_004_03_022_F003_e', label=1]
train_eval_test_MEC2017_filter_bottleneck.npz: save frame path for features extraction
data_format.py: generate two .npz files

EmoitW2018 path: EmotiW2017/For_EmotiW2018

data format：

Faces: Extracted based on openface
train_eval_test_EmotiW2017.npz: save emotion labels [pic_path='020205600', label=1]
- train_datas(773), eval_datas(383), test_datas(653)
train_eval_test_EmotiW2017.npz_bottleneck.npz: save frame path for features extraction
data_format.py: generate two .npz files

4. Image Feature Extractor

To extract frame features, we need to train frame-level emotion classifiers. And then we can extract frame features through bottleneck layers.

Training frame-level emotion classifiers
- Fintuning on FER+ database
- Fintuning on FER+ database at first, and then fintuning on SFEW database
Bottleneck features extraction: visual features are listed behind.
- EmotiW_DenseNet_FER_Fintuning.npz (frame-level features): features_1 ~ features_4
- EmotiW_VGG_FER_Fintuning.npz (frame-level features): features_1 ~ features_9
- EmotiW_C3D_features_all.npz (segment-level features): features_1
- EmotiW_Inception_features_Faces.npz (frame-level features): features_1
- EmotiW_Inception_features_Origin.npz (frame-level features): features_1
- EmotiW_traditional_descriptor.npz (video-level features): HOG, LBP, HOGLBP, DSIFT
- EmotiW_EmotiW2017_LBPTOP.npz (video-level features): LBPTOP (不是每个视频都有LBPTOP)
- EmotiW_Pose.npz (frame-level features): poses, posesDev
- EmotiW_Identify.npz (frame-level features): identifies

4.1 Training frame-level emotion classifiers

4.1.1 Fintuning on FER+ Database

Source code is in EmotiW2018/main.py, parameters are following behind:

--decay_every: change according to batch-size and training_sample_numbers. Must follow the following equation: decay_every*batch_size<training_sample_numbers. Because we save model and adjust learning rate each opt.decay_every time
--plot-every: visualize one batch result every opt.plot_every time
--lr and --lr2 : learning rate
--lr_decay: if scores don't imporve in the val dataset, we will reducelr into lr*lr_decay
--cuda: -1(CPU), >=0(GPU)
--num_classes: output number of the model
--max_epoch: max training epoch
--train_eval_test_path: generated through Database/datas_norm_format.py

FER+ database can be downloaded from link. You can also download our format-normlized FER+ data from (link: https://pan.baidu.com/s/1cmhrAmZmqHpOh0aL8lDTiw password: 4zo9). And put FER+.zip into the EmotiW2018/Database/.

# DenseNet
python main.py main --model='DenseNet' --model_path=None --loss='crossentropyloss' --num_classes=8 --train_eval_test_path='Database/train_eval_test.npz' --pic_root='Database' --plot_every=100 --batch-size=128  --lr=0.001 --lr2=0 --lr_decay=0.5  --decay-every=234 --max_epoch=30 --cuda=-1

# VGG (add bn on VGG is better)
nohup python -u main.py main --model='VGG' --model_path=None --loss='crossentropyloss' --num_classes=8 --train_eval_test_path='Database/train_eval_test.npz' --pic_root='Database' --plot_every=100 --batch-size=128 --lr=0.01 --lr2=0 --lr_decay=0.5 --decay-every=234 --max_epoch=20 --cuda=-1

4.1.2 Fintuning on SFEW2015

In this section, we need to change --num_classes to 7. Because SFEW2015 has seven emotion labels. Therefore, the output number of the model is 7. And we fintune on FER+ pretrained models.

python main.py main --model='DenseNet' --model_path='checkpoints/DenseNet_80.9635722679' --loss='crossentropyloss' --num_classes=7 --train_eval_test_path='../SFEW2015/train_eval_test_SFEW2015_filter.npz' --pic_root='../SFEW2015/Faces' --plot_every=1 --batch-size=128  --lr=0.0001 --lr2=0 --lr_decay=0.5  --decay-every=1 --max_epoch=20 --cuda=-1

python main.py main --model='VGG' --model_path='checkpoints/VGG_80.0528789659' --loss='crossentropyloss' --num_classes=7 --train_eval_test_path='../SFEW2015/train_eval_test_SFEW2015_filter.npz' --pic_root='../SFEW2015/Faces' --plot_every=1 --batch-size=128  --lr=0.0001 --lr2=0 --lr_decay=0.5  --decay-every=1 --max_epoch=20 --cuda=-1

4.1.3 Scores

Pretrained DenseNet and VGG models can be found in link: https://pan.baidu.com/s/1ZWEGy_BctbcUq0z359s9Jg password: x7z0. Pretraied models should be put into the EmotiW2018/checkpoints

model	score on val
DenseNet, FER+	0.8096
VGG, FER+	0.8005
DenseNet, SFEW2015	0.8081
VGG, SFEW2015	0.7980

4.2 Bottleneck features extraction

In this section, we extract bottleneck features of DenseNet and VGG.

Source code is in EmotiW2018/feature_extract.py.

--model: 'DenseNet' or 'VGG'
--model_path: model path for feature extraction
--train_eval_test_path: generated from '../EmotiW2017/For_EmotiW2018/data_format.py' and 'MEC2017/data_format.py'
--pic_root: aligned Face path
--save_: save path for features
--cuda: -1(CPU), 0~(GPU)

4.2.1 EmotiW: Bottleneck features extraction

######################## DenseNet ##############################
nohup python -u feature_extract.py main_DenseNet --model='DenseNet' --num_classes=8 --model_path='checkpoints/DenseNet_80.9635722679' --train_eval_test_path='../EmotiW2017/For_EmotiW2018/train_eval_test_EmotiW2017_bottleneck_Faces.npz' --pic_root='../EmotiW2017/For_EmotiW2018/Faces' --type_='test' --save_='EmotiW_DenseNet_FER_Fintuning.npz' --cuda=-1

nohup python -u feature_extract.py main_DenseNet --model='DenseNet' --num_classes=7 --model_path='checkpoints/DenseNet_80.8080808081_fintuning_SFEW' --train_eval_test_path='../EmotiW2017/For_EmotiW2018/train_eval_test_EmotiW2017_bottleneck_Faces.npz' --pic_root='../EmotiW2017/For_EmotiW2018/Faces' --type_='test' --save_='EmotiW_DenseNet_SFEW_Fintuning.npz' --cuda=-1

######################## VGG ##############################
nohup python -u feature_extract.py main_VGG --model='VGG' --num_classes=8 --model_path='checkpoints/VGG_80.0528789659' --train_eval_test_path='../EmotiW2017/For_EmotiW2018/train_eval_test_EmotiW2017_bottleneck_Faces.npz' --pic_root='../EmotiW2017/For_EmotiW2018/Faces' --type_='test' --save_='EmotiW_VGG_FER_Fintuning.npz' --cuda=-1

nohup python -u feature_extract.py main_VGG --model='VGG' --num_classes=7 --model_path='checkpoints/VGG_79.797979798_fintuning_SFEW' --train_eval_test_path='../EmotiW2017/For_EmotiW2018/train_eval_test_EmotiW2017_bottleneck_Faces.npz' --pic_root='../EmotiW2017/For_EmotiW2018/Faces' --type_='test' --save_='EmotiW_VGG_SFEW_Fintuning.npz' --cuda=-1

4.2.2 MEC: Bottleneck features extraction

######################## DenseNet ##############################
nohup python -u feature_extract.py main_DenseNet --model='DenseNet' --num_classes=8 --model_path='checkpoints/DenseNet_80.9635722679' --train_eval_test_path='../MEC2017/train_eval_test_MEC2017_filter_bottleneck.npz' --pic_root='../MEC2017/Faces' --type_='test' --save_='MEC_DenseNet_FER_Fintuning.npz' --cuda=-1

nohup python -u feature_extract.py main_DenseNet --model='DenseNet' --num_classes=7 --model_path='checkpoints/DenseNet_80.8080808081_fintuning_SFEW' --train_eval_test_path='../MEC2017/train_eval_test_MEC2017_filter_bottleneck.npz' --pic_root='../MEC2017/Faces' --type_='test' --save_='MEC_DenseNet_SFEW_Fintuning.npz' --cuda=-1

######################## VGG ##############################
nohup python -u feature_extract.py main_VGG --model='VGG' --num_classes=8 --model_path='checkpoints/VGG_80.0528789659' --train_eval_test_path='../MEC2017/train_eval_test_MEC2017_filter_bottleneck.npz' --pic_root='../MEC2017/Faces' --type_='test' --save_='MEC_VGG_FER_Fintuning.npz' --cuda=-1

nohup python -u feature_extract.py main_VGG --model='VGG' --num_classes=7 --model_path='checkpoints/VGG_79.797979798_fintuning_SFEW' --train_eval_test_path='../MEC2017/train_eval_test_MEC2017_filter_bottleneck.npz' --pic_root='../MEC2017/Faces' --type_='test' --save_='MEC_VGG_SFEW_Fintuning.npz' --cuda=-1

4.3. C3D feature extraction

Source codes are in C3D-tensorflow-master/C3D-tensorflow-master/.

This section extract C3D features, which follows the link. models/sports1m_finetuning_ucf101.model can also be download through link. Or you can find in link: https://pan.baidu.com/s/1P7z54BZOLLhoXGjz4NZJbA password: 4sru.

There are several important parameters in the predict_c3d_ucf101.py:

--pic_root: aligned face path
--test_list_file: 'a.txt' save 'video_name label' in each line. For example: '000534800 3'
--save_path: save features

# Extract C3D for EmotiW2018
python predict_c3d_ucf101.py --pic_root='../../EmotiW2017/For_EmotiW2018/Faces' --test_list_file='../../EmotiW2017/For_EmotiW2018/a.txt' --save_path='EmotiW_C3D_features_all.npz'

# Extract C3D for MEC2017
python predict_c3d_ucf101.py --pic_root='../../MEC2017/Faces' --test_list_file='../../MEC2017/a.txt' --save_path='MEC_C3D_features_all.npz'

4.4. Inception feature extraction(from Youtube8M)

Source code is in youtube-8m-master/youtube-8m-master/feature_extractor/

This section extractes Inception features for both aligned faces and original frames, which follows the link.

Models in yt8m/ can be found in link. Or for convience, you can download yt8m_pca.tgz from link: https://pan.baidu.com/s/1u3L4rGwu-ZfxS5aa4oUL_w password: nbsj. Download inception-2015-12-05.tgz from link: https://pan.baidu.com/s/1qcUDNbbOxlTlnPPzLJOtYw password: 2bow.

There are several important parameters in the feature_extractor_main.py:

--train_eval_test_path: show the frame path of Origin_Faces or Faces
--pic_root: show the frame root of Origin_Faces or Faces
--save_path: save path

# Extract 1024D inception features from aligned faces
python feature_extractor_main.py extract_from_faces --train_eval_test_path='../../../EmotiW2017/For_EmotiW2018/train_eval_test_EmotiW2017_bottleneck_Origin_Faces.npz' --pic_root='../../../EmotiW2017/For_EmotiW2018/Origin_Faces' --save_path='./EmotiW_Inception_features_Origin.npz'

# Extract 1024D inception features from origin frames
python feature_extractor_main.py extract_from_origin_faces --train_eval_test_path='./../../EmotiW2017/For_EmotiW2018/train_eval_test_EmotiW2017_bottleneck_Faces.npz' --pic_root='../../../EmotiW2017/For_EmotiW2018/Faces' --save_path='./EmotiW_Inception_features_Faces.npz'

4.5 Pose features extraction and Indentify features Extraction

This section extract Pose features and Indentify features.

Source codes are in Point_Pose_Indenfy/, They are all writen in C++.

'Pose features' include head pose and facial landmark points. These Features are extracted using Dlib toolkit under vs2015. Our simplfied version can be found in link: https://pan.baidu.com/s/1eR_PsRSZmUoAkKO_h--4og password: dvf3
'Identify features' : These Features are extracted using Seetaface toolkit under vs2013. Our simplfied version can be found in link: https://pan.baidu.com/s/1RTiHQABFd4GH_omtXm7KnA password: apdw.

4.6 Traditional Video Features Extraction

This section extract HOG, LBP, HOGLBP, DSIFT.

Source codes are in Traditional_video_descriptor/

## Dsift features
python dsift.py dsift_extractor --video_root='../EmotiW2017/For_EmotiW2018/Video' --face_root='../EmotiW2017/For_EmotiW2018/Faces' --save_path='./Dsift_features.npz'

## HOG features
python hog.py hog_extractor --video_root='../EmotiW2017/For_EmotiW2018/Video' --face_root='../EmotiW2017/For_EmotiW2018/Faces' --save_path='./Hog_features.npz'

## HOG_LBP features
python hog_lbp.py hoglbp_extractor --video_root='../EmotiW2017/For_EmotiW2018/Video' --face_root='../EmotiW2017/For_EmotiW2018/Faces' --save_path='./HogLBP_features.npz'

## LBP features
python LBP.py lbp_extractor --video_root='../EmotiW2017/For_EmotiW2018/Video' --face_root='../EmotiW2017/For_EmotiW2018/Faces' --save_path='./LBP_features.npz'

4.7 EmotiW_EmotiW2017_LBPTOP

LBPTOP features are provided by organizers.

5. Audio Feature Extraction

Original audio files are extracted from video by ffmpeg.

ffmpeg -i video_path audio_path

After we separate audio files from video, we extract different audio features sets.

EmotiW_audio.npz (frame-level features): mfcc, soundnet,IS10, Egemaps
EmotiW_EmotiW2017_features.npz (frame-level features): English, Chinese
EmotiW_EmotiW2017_features.npz (utterance-level features): Egmaps
EmotiW_IS091113.npz (utterance-level features): IS09 ,IS11, IS13
EmotiW_youtubeAudio.npz (segment-level features): logspec, emb, qemb

5.1 EmotiW_audio

This section includes mfcc, soundnet,IS10, Egemaps . They are all frame-level features.

soundnet: extract based on torch7. link
mfcc, IS10, Egemaps: extract based on opensmile

5.2 EmotiW_EmotiW2017_features

This section includesEnglish, Chinese, Egmaps.

English, Chinese: They are all frame-level features. These features are extracted from Chinese-ASR bottleneck layers and English-ASR bottleneck layers, respectively.
Egmaps: utterance-level features, extracted through opensmile

5.3 EmotiW_IS091113

This section includes IS09 ,IS11, IS13. They are all utterance-level features. They are extract in windows though OpenSmile.

opensmile-2.3.0.rar can be downloaded from [link](opensmile source path download_link), which should be put into the Audio_OpenSmile/.

Source codes are in Audio_OpenSmile/featureExtract.py.

# extract audio features
python featureExtract.py main

# save all features into .npz files
python featureExtract.py change_features_into_npz

5.4 EmotiW_youtubeAudio

This section includes logspec, emb, qemb. They are all segment-level features. Each segment is one second.

Source codes are in devicehive-audio-analysis-master/, which refers to link.

For convenience, models in devicehive-audio-analysis-master/models/ can be found in link: https://pan.baidu.com/s/1jreEsgjtYFONcOdIwBKEVQ password: kqw0

Main files is devicehive-audio-analysis-master/parse_file.py. Several parameters are listed:

--wav_root: wav dir contains all wav files.
--save_path: save path

# extract audio features
python devicehive-audio-analysis-master/parse_file.py  gain_features --wav_root='Audio_Dataset' --save_path='./EmotiW_youtubeAudio.npz'

6. Textual Feature Extraction

We utilize open-source toolkit, baidu ASR API, to recognize text in the audio.

After we recognize text in audio, we extract different audio features sets:

EmotiW_TFIDF.npz (utterance-level features): TFIDF
EmotiW_TFIDFNoEmpty.npz (utterance-level features): TFIDFNoEmpty（不是每句话都有）
EmotiW_txt_vec.npz (utterance-level features): word2vec（不是每句话都有）

6.1 ASR in Baidu API

Source codes are in baiduASR/baiduapi_my.py

To utilize baiduAPI, we need to register. Please follow link.

Then set your own APP_ID, API_KEY and SECRET_KEY in baiduASR/baiduapi_my.py.

--file_root: wav root path
--save_path: save path

python baiduapi_my.py asr_for_one_root --file_root='wav' --save_path='EmotiW_txt_pcm.npz'

Through this file, we change 'wav' to 'EmotiW_txt.npz'

6.2 Extract TFIDF

Source codes are in baiduASR/Text_Features.py

## preporcess on text. Add 'Word_Tabel' in 'EmotiW_txt.npz'
python Text_Features.py emotiW_extraction --asr_root='EmotiW_txt.npz'

## Extract TFIDF features (which will utilize word_Table)
python Text_Features.py extract_TFIDF

6.3 Extract Word2Vec Features

Source codes are in baiduASR/gain_word2vec.py. We utilize pre-trained English fasttext word vectors: wiki-news-300d-1M.vec can be download through link.

# change text to vector
python gain_word2vec.py change_txt_vec --features_path='EmotiW_txt.npz' --save_path='EmotiW_txt_vec.npz'

# del empty features (whose text is false recognized through ASR)
python gain_word2vec.py del_empty --features_path='EmotiW_txt_vec.npz'

7. Classifier

Source code is in EmotiW2018/classifier.py

--label_type: change label_root and classifier_save_root in config.py

--features_path：

As for frame level features, data_path['pic_path'] refers to frame path
As for video level features, data_path['pic_path'] refers to video name

--features_name:

visual features
- EmotiW_DenseNet_FER_Fintuning.npz (frame-level features): features_1 ~ features_4
- EmotiW_VGG_FER_Fintuning.npz (frame-level features): features_1 ~ features_9
- EmotiW_C3D_features_all.npz (segment-level features): features_1
- EmotiW_Inception_features_Faces.npz (frame-level features): features_1
- EmotiW_Inception_features_Origin.npz (frame-level features): features_1
- EmotiW_traditional_descriptor.npz (video-level features): HOG, LBP, HOGLBP, DSIFT
- EmotiW_EmotiW2017_LBPTOP.npz (video-level features): LBPTOP (不是每个视频都有LBPTOP)
- EmotiW_Pose.npz (frame-level features): poses, posesDev
- EmotiW_Identify.npz (frame-level features): identifies
auditory features:
- EmotiW_audio.npz (frame-level features): mfcc, soundnet,IS10, Egemaps
- EmotiW_EmotiW2017_features.npz (frame-level features): English, Chinese
- EmotiW_EmotiW2017_features.npz (utterance-level features): Egmaps
- EmotiW_IS091113.npz (utterance-level features): IS09 ,IS11, IS13
- EmotiW_youtubeAudio.npz (segment-level features): logspec, emb, qemb
textual features:
- EmotiW_TFIDF.npz (utterance-level features): TFIDF
- EmotiW_TFIDFNoEmpty.npz (utterance-level features): TFIDFNoEmpty（不是每句话都有）
- EmotiW_txt_vec.npz (utterance-level features): word2vec（不是每句话都有）

--temporal_：

Video: whether data_path['pic_path'] refers to frame path or video path
max, mean, FV_N_K, _None: methods for encoding frame-level features into vidoe-level features

--model:

linearSVC, RBFSVC, RF, LR, NN: ignore temporal information
LSTM: consider temporal information
'NetVLAD', 'NetRVLAD', 'SoftDBoW', 'NetFV': realized through classifier_lp.py. reference

# train all models through one file
sh run_train.sh 1 # 基于train_eval_test_EmotiW2017_filter.npz classifier_1

8. Prediction

In this section, we gain predictions for train_datas, eval_datas and test_datas, respectively. And results are saved as .pth file into classifier_1/train_result, classifier_1/eval_result and classifier_1/test_result , respectively.

Source code is in EmotiW2018/predict.py, which has two different calling methods.

# method 1: gain prediction for single model
python predict.py main --label_type=1 --model_path='classifier_1/EmotiW_VGG_FER_Fintuning_features_1_mean_RF_0.425267993874'

# method 2: gain all predictions of all models in 'classifier_1/'
python predict.py main_all_classifiers --label_type=1

9. Fusion

In this section, we gain fusion parameters and final submission.

Source code is in EmotiW2018/fusion_predict.py, which has following parameters:

--label_type: change label_root and classifier_save_root in config.py
--fusion_type: 'gain_para' or 'gain_submit'
--best_para_path: save path for 'gain_para' and call path for 'gain_submit'
--max_epoch: number of interations in 'gain_para'

# step 1: Gain fusion parameters
python fusion_predict.py main --label_type=1 --fusion_type='gain_para' --max_epoch=100 --best_para_path='1_best_para.npz'

# step 2: Gain submission files
python fusion_predict.py  main --label_type=1 --fusion_type='gain_submit' --best_para_path='1_best_para_58.81.npz'

HojaMuerta/EmotiW2018