-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain_all_models_a2a.py
176 lines (149 loc) · 6.53 KB
/
train_all_models_a2a.py
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
import os, json, glob
import numpy as np
import keras
from keras.layers import *
from keras.models import Model
#from data_generator import *
from keras import metrics
os.environ['CUDA_VISIBLE_DEVICES'] = "1"
import pandas as pd
import tensorflow as tf
from keras import callbacks
from keras.regularizers import l2
from architectures import *
ads_fps = json.load( open('all_ads_fps.json', 'r') )
def _read_av_data_and_splice(ad_id, db_name_): #, vid_dir = './jwt_ads/c3d', aud_dir='./jwt_ads/vggish'):
'''
CHECK:
1. here fps changes for each ad_id, so I have a cvpr_ads_fps: global
dict to look up ad_id for fps
2. vid_dir, aud_dir are predefine: I find this easier to deal with
input:
ad_id: filename indicator: in the directory vid_dir, aud_dir
output:
vid_slice; shape(?,11,512): 5 features context both ways
aud_slice: shape(?, 7, 128): 3 features context both ways
'''
f_id = ad_id.decode()
db_name = db_name_.decode()
#print(f_id)
#f_loc = (np.argwhere(filenames==f_id)[0,0]).astype(np.int32)
vid_file = os.path.join(db_name+'_ads', 'c3d', "%s.npz" % (f_id))
aud_file = os.path.join(db_name+'_ads', 'vggish', "%s.npy" % (f_id))
# squashing all into a GAP
vid = np.mean(np.squeeze(np.load(vid_file)['conv']), (1,2,3))
aud = np.load(aud_file)
fps = ads_fps[f_id]['fps']
# video duration per segment in frames
vT = 16.0
# audio duration per segment in secs
aT = 0.96
# window length and skip length [0...10], [3,...]
vid_length=11
vid_skip=3
aud_length=7
# get the indices for video frames with leaving out vid_skip windows
# when I do [::vid_skip] i don't account for corner case at end so may lose
# one sample window
v_ix = [range(i,i+vid_length) for i in range(vid.shape[0]-vid_length+1)]
v_ix = v_ix[::vid_skip]
#print(v_ix)
# work with the middle frame
a_map = [int((i[5]*vT)/(fps*aT)) for i in v_ix]
# since the aud length is 7 we need 3 fwd and bwd context
# done manually maybe do this elegantly?
a_ix = np.array([range(i-3, i+4) for i in a_map])
# take care of corner cases while mapping usually off by one at edges
a_ix[a_ix<0] = 0
a_ix[a_ix>len(aud)-1] = len(aud)-1
vid_slice = np.array([vid[i] for i in v_ix]).astype(np.float32)
aud_slice = np.array([aud[i] for i in a_ix.tolist()]).astype(np.float32)
# uncomment this to convince yourself that the suffle works as expected!
#f_ = (np.zeros_like(aud_slice)+f_loc).astype(np.int32)
return vid_slice, aud_slice#, f_
def _convert_splices_to_tensors(vid_slice_, aud_slice_):#, f_loc_):
'''
just convert the np arrays to tensors. important to explicitly mention
dtype and shape. Need to merge this with the above function but not today
'''
vid_slice = tf.reshape(tf.convert_to_tensor(vid_slice_, tf.float32),
[-1,11*512])
aud_slice = tf.reshape(tf.convert_to_tensor(aud_slice_, tf.float32),
[-1,7*128])
#f_loc = tf.reshape(tf.convert_to_tensor(f_loc_, tf.int32), [-1,7,128])
return vid_slice, aud_slice#, f_loc
def ae_input_fn(db_name = 'jwt', batch_size=100, n_epochs=10, n_threads=16):
if db_name == "jwt":
df = pd.read_pickle('jwt_ads_data_N9744.pkl')
filenames = np.array([str(i) for i in df['ad_id']])
elif db_name == "cvpr":
df = pd.read_pickle('cvpr_ads_data_labels_N2720.pkl')
filenames = np.array([str(i) for i in df['files']])
dataset = (tf.data.Dataset.from_tensor_slices(filenames)
.shuffle(buffer_size=10*len(filenames)) # shuffle filenames
.repeat(-1) # the n_epochs makes sure you do sampling with rep
)
#format: https://www.tensorflow.org/api_docs/python/tf/py_func
dataset = dataset.map(
lambda ad_id: tuple(tf.py_func(_read_av_data_and_splice, [ad_id,db_name],
[tf.float32, tf.float32] )),
num_parallel_calls=n_threads)
dataset = dataset.map(_convert_splices_to_tensors,
num_parallel_calls=20) # cpu-parallel
# unbatch the slices produce - shuffle - batch - prefetch
dataset = ( dataset.apply(tf.contrib.data.unbatch())
.shuffle(buffer_size=100*batch_size) # make a big buffer to shuffle well
.batch(batch_size)
.prefetch(1)
)
# make a iter and generate!
data_iter = dataset.make_one_shot_iterator()
vid,aud = data_iter.get_next()
#features = {'aud':aud, 'vid':vid}
#target=[]
return vid, aud #features, target
train_dataset = ae_input_fn(db_name="jwt")
test_dataset = ae_input_fn(db_name="cvpr")
model_256 = a2a_256()
model_128 = a2a_128()
model_64 = a2a_64()
model_256.compile( loss='mean_squared_error',optimizer='rmsprop',
metrics=[metrics.mse, metrics.mae])
model_128.compile( loss='mean_squared_error',optimizer='rmsprop',
metrics=[metrics.mse, metrics.mae])
model_64.compile( loss='mean_squared_error',optimizer='rmsprop',
metrics=[metrics.mse, metrics.mae])
n_epochs = 30
batch_size = 100
n_train_steps = 495531//batch_size
n_test_steps = 668
sess = tf.InteractiveSession()
all_test_loss = []
for ep_ix in range(n_epochs):
print(ep_ix, '-------------------------------------------')
for b_ix in range(n_train_steps):
_, aud = sess.run(train_dataset)
if not b_ix%1000: verbosity = 2
else: verbosity = 0
model_256.fit(aud, aud, batch_size=100, epochs=1,
verbose=verbosity)
model_128.fit(aud, aud, batch_size=100, epochs=1,
verbose=verbosity)
model_64.fit(aud, aud, batch_size=100, epochs=1,
verbose=verbosity)
# now test
test_loss_list = []
for b_test_ix in range(n_test_steps):
_, aud_test = sess.run(test_dataset)
test_loss = model_256.test_on_batch(aud_test, aud_test)
test_loss_list.append(test_loss)
test_loss = model_128.test_on_batch(aud_test, aud_test)
test_loss_list.append(test_loss)
test_loss = model_64.test_on_batch(aud_test, aud_test)
test_loss_list.append(test_loss)
#print(ep_ix, test_loss)
all_test_loss.append(test_loss_list)
print(ep_ix, ' mean loss: ',np.mean(test_loss_list), np.std(test_loss_list))
if not ep_ix%10: model_256.save('a2a_256_ep_%d.h5' % ep_ix )
if not ep_ix%10: model_128.save('a2a_128_ep_%d.h5' % ep_ix )
if not ep_ix%10: model_64.save('a2a_64_ep_%d.h5' % ep_ix )