File size: 12,577 Bytes
b202543 7cb3d00 b202543 7cb3d00 b202543 7cb3d00 b202543 7cb3d00 b202543 7cb3d00 b202543 1035ec8 b202543 38b500c b202543 74808e7 b202543 94858c4 b202543 b055bcd b202543 78c42e7 b202543 145010b b202543 145010b b202543 145010b b202543 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 |
# Facial Recognition with Emotion / Sentiment Detector
# This is a custom, hard-coded version of darknet with
# YOLOv3 implementation for openimages database. This
# was written to test viability of implementing YOLO
# for face detection followed by emotion / sentiment
# analysis.
#
# Configuration, weights and data are hardcoded.
# This version takes any images, detects faces,
# and then runs emotion / sentiment analysis
#
# Author : Saikiran Tharimena
# Co-Authors: Kjetil Marinius Sjulsen, Juan Carlos Calvet Lopez
# Project : Emotion / Sentiment Detection from news images
# Date : 12 September 2022
# Version : v0.1
#
# (C) Schibsted ASA
# Libraries
import torch
from utils import *
import gradio as gr
from numpy import array
from darknet import Darknet
from torch.autograd import Variable
from torch.cuda import is_available as check_cuda
from PIL.ImageOps import grayscale
from fastai.vision.all import PILImage, load_learner
################## DARKNET ##################
# Parameters
batch_size = 1
confidence = 0.25
nms_thresh = 0.30
run_cuda = False
# CFG Files
cfg = 'cfg/yolov3-openimages.cfg'
clsnames= 'cfg/openimages.names'
weights = 'cfg/yolov3-openimages.weights'
# Load classes
classes = load_classes(clsnames)
num_classes = len(classes)
# Set up the neural network
print('Load Network')
model = Darknet(cfg)
print('Load Weights')
model.load_weights(weights)
print('Successfully loaded Network')
# Check CUDA
if run_cuda:
CUDA = check_cuda()
else:
CUDA = False
# Input dimension
inp_dim = int(model.net_info["height"])
# put the model on GPU
if CUDA:
model.cuda()
# Set the model in evaluation mode
model.eval()
def get_detections(x):
c1 = [int(y) for y in x[1:3]]
c2 = [int(y) for y in x[3:5]]
det_class = int(x[-1])
label = "{0}".format(classes[det_class])
return (label, tuple(c1 + c2))
# face detector
def detector(image):
# Just lazy to update this
imlist = [image]
loaded_ims = [image]
im_batches = list(map(prep_image, loaded_ims, [inp_dim for x in range(len(imlist))]))
im_dim_list = [(x.shape[1], x.shape[0]) for x in loaded_ims]
im_dim_list = torch.FloatTensor(im_dim_list).repeat(1,2)
leftover = 0
if (len(im_dim_list) % batch_size):
leftover = 1
if batch_size != 1:
num_batches = len(imlist) // batch_size + leftover
im_batches = [torch.cat((im_batches[i*batch_size : min((i + 1)*batch_size,
len(im_batches))])) for i in range(num_batches)]
write = 0
if CUDA:
im_dim_list = im_dim_list.cuda()
for i, batch in enumerate(im_batches):
# load the image
if CUDA:
batch = batch.cuda()
with torch.no_grad():
prediction = model(Variable(batch), CUDA)
prediction = write_results(prediction, confidence, num_classes, nms_conf = nms_thresh)
if type(prediction) == int:
for im_num, image in enumerate(imlist[i*batch_size: min((i + 1)*batch_size, len(imlist))]):
im_id = i*batch_size + im_num
continue
prediction[:,0] += i*batch_size # transform the atribute from index in batch to index in imlist
if not write: # If we have't initialised output
output = prediction
write = 1
else:
output = torch.cat((output, prediction))
for im_num, image in enumerate(imlist[i*batch_size: min((i + 1)*batch_size, len(imlist))]):
im_id = i * batch_size + im_num
objs = [classes[int(x[-1])] for x in output if int(x[0]) == im_id]
if CUDA:
torch.cuda.synchronize()
try:
output
except NameError:
return loaded_ims[0], []
im_dim_list = torch.index_select(im_dim_list, 0, output[:,0].long())
scaling_factor = torch.min(608/im_dim_list,1)[0].view(-1,1)
output[:, [1,3]] -= (inp_dim - scaling_factor*im_dim_list[:,0].view(-1,1))/2
output[:, [2,4]] -= (inp_dim - scaling_factor*im_dim_list[:,1].view(-1,1))/2
output[:, 1:5] /= scaling_factor
for i in range(output.shape[0]):
output[i, [1,3]] = torch.clamp(output[i, [1,3]], 0.0, im_dim_list[i,0])
output[i, [2,4]] = torch.clamp(output[i, [2,4]], 0.0, im_dim_list[i,1])
detections = list(map(get_detections, output))
if CUDA:
torch.cuda.empty_cache()
return loaded_ims[0], detections
#############################################
# Emotion
learn_emotion = load_learner('models/emotions_vgg19.pkl')
learn_emotion_labels = learn_emotion.dls.vocab
# Sentiment
learn_sentiment = load_learner('models/sentiment_vgg19.pkl')
learn_sentiment_labels = learn_sentiment.dls.vocab
def crop_images(img, bbox):
"Here image should be an image object from PILImage.create"
# Coordinates of face in cv2 format
xmin, ymin, xmax, ymax = bbox[1]
# resize and crop face
return img.crop((xmin, ymin, xmax, ymax))
def detect_person_face(img, detections):
'''This function is called from within detect face.
If only a person is detected, then this will crop
image and then try to detect face again.'''
faces = []
# Loop through people
for detection in detections:
# Get cropped image of person
temp = crop_images(img, detection)
# run detector again
_, detect = detector(array(temp)[...,:3])
# check for human faces
human_face = [idx for idx, val in enumerate(detect) if val[0] == 'Human face']
if len(human_face) == 0:
continue
# Force it to take only 1 face per person
# crop face and append to list
faces.append(crop_images(temp, detect[human_face[0]]))
return faces
def detect_face(img):
_, detections = detector(array(img)[...,:3])
# check for human faces
human_face = [idx for idx, val in enumerate(detections) if val[0] == 'Human face']
if len(human_face) == 0:
human_face = [idx for idx, val in enumerate(detections) if val[0] == 'Person']
if len(human_face) == 0:
return []
else:
# Only get human face detections
faces = detect_person_face(img, [detections[idx] for idx in human_face])
else:
# Only get human face detections
faces = []
for idx in human_face:
faces.append(crop_images(img, detections[idx]))
return faces
# Predict
def predict(img):
img = PILImage.create(img)
# Detect faces
faces = detect_face(img)
output = []
if len(faces) == 0:
img = img.resize((48, 48))
pred_emotion, pred_emotion_idx, probs_emotion = learn_emotion.predict(array(grayscale(img)))
pred_sentiment, pred_sentiment_idx, probs_sentiment = learn_sentiment.predict(array(grayscale(img)))
emotions = {learn_emotion_labels[i]: float(probs_emotion[i]) for i in range(len(learn_emotion_labels))}
sentiments = {learn_sentiment_labels[i]: float(probs_sentiment[i]) for i in range(len(learn_sentiment_labels))}
output = [img, emotions, sentiments, img, emotions, sentiments, img, emotions, sentiments]
else: # Max 3 for now
for face in faces[:3]:
img = face.resize((48, 48))
pred_emotion, pred_emotion_idx, probs_emotion = learn_emotion.predict(array(grayscale(img)))
pred_sentiment, pred_sentiment_idx, probs_sentiment = learn_sentiment.predict(array(grayscale(img)))
emotions = {learn_emotion_labels[i]: float(probs_emotion[i]) for i in range(len(learn_emotion_labels))}
sentiments = {learn_sentiment_labels[i]: float(probs_sentiment[i]) for i in range(len(learn_sentiment_labels))}
output.append(img)
output.append(emotions)
output.append(sentiments)
temp = output[-3:]
while len(output) < 9:
output = output + temp
return output
# Gradio
title = 'Face Recognition with Emotion and Sentiment Detector'
description = gr.Markdown(
"""Ever wondered what a person might be feeling looking at their picture?
Well, now you can! Try this fun app. Just upload a facial image in JPG or
PNG format. Voila! you can now see what they might have felt when the picture
was taken.
This is an updated version of Facial Expression Classifier:
https://huggingface.co/spaces/schibsted/facial_expression_classifier
""").value
article = gr.Markdown(
"""**DISCLAIMER:** This model does not reveal the actual emotional state of a person. Use and
interpret results at your own risk! It was built as a demo for AI course. Samples images
were downloaded from VG & AftenPosten news webpages. Copyrights belong to respective
brands. All rights reserved.
**PREMISE:** The idea is to determine an overall sentiment of a news site on a daily basis
based on the pictures. We are restricting pictures to only include close-up facial
images.
**DATA:** FER2013 dataset consists of 48x48 pixel grayscale images of faces. There are 28,709
images in the training set and 3,589 images in the test set. However, for this demo all
pictures were combined into a single dataset and 80:20 split was used for training. Images
are assigned one of the 7 emotions: Angry, Disgust, Fear, Happy, Sad, Surprise, and Neutral.
In addition to these 7 classes, images were re-classified into 3 sentiment categories based
on emotions:
Positive (Happy, Surprise)
Negative (Angry, Disgust, Fear, Sad)
Neutral (Neutral)
FER2013 (preliminary version) dataset can be downloaded at:
https://www.kaggle.com/c/challenges-in-representation-learning-facial-expression-recognition-challenge/data
**EMOTION / SENTIMENT MODEL:** VGG19 was used as the base model and trained on FER2013 dataset. Model was trained
using PyTorch and FastAI. Two models were trained, one for detecting emotion and the other
for detecting sentiment. Although, this could have been done with just one model, here two
models were trained for the demo.
**FACE DETECTOR:** Darknet with YOLOv3 architecture was used for face detection. Reach out to me for full details.
In short, any image is first sent through darknet. If face is detected, then it is passed through emotion/sentiment
model for each face in the picture. If a person is detected rather than a face, the image is cropped and run through
face detector again. If a face is detected, then it is passed through emotion/sentiment model. In case face is not
detected in an image, then the entire image is evaluated to generate some score. This is done because, I couldn't
figure out how to pipe None/blank output to Gradio.Interface(). There maybe option through Gradio.Blocks() but was
too lazy to go through that at this stage. In addition, the output is restricted to only 3 faces in a picture.
""").value
enable_queue=True
examples = ['happy1.jpg', 'happy2.jpg', 'angry1.png', 'angry2.jpg', 'neutral1.jpg', 'neutral2.jpg']
gr.Interface(fn = predict,
inputs = gr.Image(),
outputs = [gr.Image(shape=(48, 48), label='Person 1'),
gr.Label(label='Emotion - Person 1'),
gr.Label(label='Sentiment - Person 1'),
gr.Image(shape=(48, 48), label='Person 2'),
gr.Label(label='Emotion - Person 2'),
gr.Label(label='Sentiment - Person 2'),
gr.Image(shape=(48, 48), label='Person 3'),
gr.Label(label='Emotion - Person 3'),
gr.Label(label='Sentiment - Person 3'),], #gr.Label(),
title = title,
examples = examples,
description = description,
article=article,
allow_flagging='never').launch(enable_queue=enable_queue) |