SGF/dataloader.py at main · stddddd/SGF · GitHub

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
import torch
import torch.nn.functional as F
from torch.utils.data import Dataset
from torch.nn.utils.rnn import pad_sequence
from torch.utils.data.sampler import SubsetRandomSampler, WeightedRandomSampler
import pickle, pandas as pd
import numpy
import json
import yaml

meld_speakers = ['Rachel', 'Joey', 'Ross', 'Monica', 'Chandler', 'Phoebe']
# meld_speakers = ['Rachel', 'Joey', 'Monica']#, 'Ross', 'Chandler', 'Phoebe']
# meld_speakers = ['Rachel', 'Joey', 'Chandler']#, 'Ross', 'Monica', 'Phoebe']
iemocap_speakers = ['Ses01_M', 'Ses01_F', 'Ses02_M', 'Ses02_F', 'Ses03_M', 'Ses03_F', 'Ses04_M', 'Ses04_F']
# iemocap_speakers = ['Ses02_M', 'Ses02_F', 'Ses04_M', 'Ses04_F']
no_speakers = ['Ross', 'Monica', 'Phoebe']
# iemocap_speakers = ['M', 'F']

class IEMOCAPDataset(Dataset):

	def __init__(self, train=True):
		self.videoIDs, self.videoSpeakers, self.videoLabels, self.videoText,\
		self.videoAudio, self.videoVisual, self.videoSentence, self.trainVid,\
		self.testVid = pickle.load(open('./IEMOCAP_features/IEMOCAP_features_raw.pkl', 'rb'), encoding='latin1')
		#self.testVid = pickle.load(open('./IEMOCAP_features/IEMOCAP_features.pkl', 'rb'), encoding='latin1')

		_, _, self.roberta1, self.roberta2, self.roberta3, self.roberta4,\
		_, _, _, _ = pickle.load(open('./IEMOCAP_features/iemocap_features_roberta.pkl', 'rb'), encoding='latin1')
		'''
		label index mapping = {'hap':0, 'sad':1, 'neu':2, 'ang':3, 'exc':4, 'fru':5}
		'''
		self.keys = [x for x in (self.trainVid if train else self.testVid)]
		self.speakerNames = {}
		# self.keys = []
		# tag = 0
		for diaid in self.videoSpeakers.keys():
			# if diaid[:5] == 'Ses01' or diaid[:5] == 'Ses05':
			# 	if train:
			# 		self.keys.append(diaid)
			# elif diaid[:5] == 'Ses03':
			# 	if not train:
			# 		self.keys.append(diaid)
			# else:
			# 	if train == tag & 1:
			# 		self.keys.append(diaid)
			# 	tag += 1
			self.speakerNames[diaid] = []
			for sp in self.videoSpeakers[diaid]:
				speaker = diaid[:5] + '_' + sp
				if speaker in iemocap_speakers:
					self.speakerNames[diaid].append(iemocap_speakers.index(speaker))
				else:
					self.speakerNames[diaid].append(-1)

		self.len = len(self.keys)

	def __getitem__(self, index):
		vid = self.keys[index]
		return torch.FloatTensor(numpy.array(self.roberta1[vid])),\
			   torch.FloatTensor(numpy.array(self.roberta2[vid])),\
			   torch.FloatTensor(numpy.array(self.roberta3[vid])),\
			   torch.FloatTensor(numpy.array(self.roberta4[vid])),\
			   torch.FloatTensor(numpy.array(self.videoVisual[vid])),\
			   torch.FloatTensor(numpy.array(self.videoAudio[vid])),\
			   torch.FloatTensor(numpy.array([[1,0] if x=='M' else [0,1] for x in\
								  self.videoSpeakers[vid]])),\
			   torch.FloatTensor(numpy.array([1 if i != -1 else 0 for i in self.speakerNames[vid]])),\
			   torch.FloatTensor(numpy.array([1] * len(self.speakerNames[vid]))),\
			   torch.LongTensor(numpy.array(self.videoLabels[vid])),\
			   torch.LongTensor(numpy.array(self.speakerNames[vid])),\
			   vid

	def __len__(self):
		return self.len

	def collate_fn(self, data):
		dat = pd.DataFrame(data)
		return [pad_sequence(dat[i]) if i<7 else pad_sequence(dat[i], True) if i<10 else pad_sequence(dat[i], True, -1) if i<11 else dat[i].tolist() for i in dat]

class MELDDataset(Dataset):

	def __init__(self, path, train=True):
		self.videoIDs, self.videoSpeakers, self.videoLabels, self.videoText,\
		self.videoAudio, self.videoVisual, self.videoSentence, self.trainVid,\
		self.testVid, _ = pickle.load(open(path, 'rb'))
		self.keys = [x for x in (self.trainVid if train else self.testVid)]
		self.trainvis = pickle.load(open('./V/meld_train_vision_utt.pkl', 'rb'))['train']
		self.valvis = pickle.load(open('./V/meld_val_vision_utt.pkl', 'rb'))['val']
		self.testvis = pickle.load(open('./V/meld_test_vision_utt.pkl', 'rb'))['test']
		self.visdata = {}
		self.vismask = {}
		self.speakerNames = {}
		with open('MELD/utterance-ordered.json', 'r') as fl:
			self.utterance_ordered = json.load(fl)
		self.train_err = 'dia125_utt3'
		self.test_err = 'dia220_utt0'

		i, j = 0, 0
		if train:
			nkeys = []
			diaids = list(self.utterance_ordered['train'].keys())
			for diaid in diaids:
				flag = 1
				self.speakerNames[self.keys[i]], self.visdata[self.keys[i]], self.vismask[self.keys[i]] = [], [], []
				for uttid in self.utterance_ordered['train'][diaid]:
					self.visdata[self.keys[i]].append(numpy.pad(self.trainvis['vision'][j], pad_width=((0, 2), (0, 0)), mode='constant'))
					self.vismask[self.keys[i]].append(numpy.pad(self.trainvis['vision_utt_mask'][j], pad_width=(0, 2), mode='constant'))
					if uttid != self.train_err:
						j += 1
					with open('MELD/raw-texts/train/' + uttid + '.json', 'r') as fl:
						sp = json.load(fl)['Speaker']
						if sp in meld_speakers:
							self.speakerNames[self.keys[i]].append(meld_speakers.index(sp))
						else:
							self.speakerNames[self.keys[i]].append(-1)
						if sp in no_speakers:
							flag = 0
				if flag:
					nkeys.append(self.keys[i])
				i += 1
			j = 0
			diaids = list(self.utterance_ordered['val'].keys())
			for diaid in diaids:
				flag = 1
				self.speakerNames[self.keys[i]], self.visdata[self.keys[i]], self.vismask[self.keys[i]] = [], [], []
				for uttid in self.utterance_ordered['val'][diaid]:
					self.visdata[self.keys[i]].append(self.valvis['vision'][j])
					self.vismask[self.keys[i]].append(self.valvis['vision_utt_mask'][j])
					j += 1
					with open('MELD/raw-texts/val/' + uttid + '.json', 'r') as fl:
						sp = json.load(fl)['Speaker']
						if sp in meld_speakers:
							self.speakerNames[self.keys[i]].append(meld_speakers.index(sp))
						else:
							self.speakerNames[self.keys[i]].append(-1)
						if sp in no_speakers:
							flag = 0
				if flag:
					nkeys.append(self.keys[i])
				i += 1
			# self.keys = nkeys
		else:
			diaids = list(self.utterance_ordered['test'].keys())
			for diaid in diaids:
				self.speakerNames[self.keys[i]], self.visdata[self.keys[i]], self.vismask[self.keys[i]] = [], [], []
				for uttid in self.utterance_ordered['test'][diaid]:
					self.visdata[self.keys[i]].append(self.testvis['vision'][j])
					self.vismask[self.keys[i]].append(self.testvis['vision_utt_mask'][j])
					if uttid != self.test_err:
						j += 1
					with open('MELD/raw-texts/test/' + uttid + '.json', 'r') as fl:
						sp = json.load(fl)['Speaker']
						if sp in meld_speakers:
							self.speakerNames[self.keys[i]].append(meld_speakers.index(sp))
						else:
							self.speakerNames[self.keys[i]].append(-1)
				i += 1
		self.len = len(self.keys)
		cnt1, cnt2 = 0, 0
		for k in self.keys:
			for sp in self.speakerNames[k]:
				if sp == -1:
					cnt2 += 1
				else:
					cnt1 += 1
		print(train, self.len, cnt1, cnt2)
		_, _, _, self.roberta1, self.roberta2, self.roberta3, self.roberta4, \
			_, self.trainIds, self.testIds, self.validIds \
			= pickle.load(open("./MELD_features/meld_features_roberta.pkl", 'rb'), encoding='latin1')

	def __getitem__(self, index):
		vid = self.keys[index]
		return torch.FloatTensor(numpy.array(self.roberta1[vid])),\
			   torch.FloatTensor(numpy.array(self.roberta2[vid])),\
			   torch.FloatTensor(numpy.array(self.roberta3[vid])),\
			   torch.FloatTensor(numpy.array(self.roberta4[vid])),\
			   torch.FloatTensor(numpy.array(self.videoVisual[vid])),\
			   torch.FloatTensor(numpy.array(self.videoAudio[vid])),\
			   torch.FloatTensor(numpy.array(self.videoSpeakers[vid])),\
			   torch.FloatTensor(numpy.array(self.visdata[vid])),\
			   torch.FloatTensor(numpy.array(self.vismask[vid])),\
			   torch.FloatTensor(numpy.array([1 if i != -1 else 0 for i in self.speakerNames[vid]])),\
			   torch.FloatTensor(numpy.array([1] * len(self.speakerNames[vid]))),\
			   torch.LongTensor(numpy.array(self.videoLabels[vid])),\
			   torch.LongTensor(numpy.array(self.speakerNames[vid])),\
			   vid

	def __len__(self):
		return self.len

	def return_labels(self):
		return_label = []
		for key in self.keys:
			return_label+=self.videoLabels[key]
		return return_label

	def collate_fn(self, data):
		dat = pd.DataFrame(data)
		return [pad_sequence(dat[i]) if i<9 else pad_sequence(dat[i], True) if i<12 else pad_sequence(dat[i], True, -1) if i<13 else dat[i].tolist() for i in dat]