MouseGrooming/Inference/TestInfer.py at master · KumarLabJax/MouseGrooming · 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
import keras
from keras.models import load_model
import imageio
import os
import numpy as np
from scipy.misc import imresize
import sys, getopt, re, argparse
import tensorflow as tf
from time import time

# Keras' definition converted to numpy...
def softmax(x, axis=-1):
	ndim = np.ndim(x)
	if ndim >= 2:
		e = np.exp(x - np.max(x, axis=axis, keepdims=True))
		s = np.sum(e, axis=axis, keepdims=True)
		return e / s
	else:
		raise ValueError('Cannot apply softmax to a tensor that is 1D')


def main(argv):
	parser = argparse.ArgumentParser(description='Inference 3DConv Models')
	parser.add_argument('--mov_name', help='Name of movie to process')

	args = parser.parse_args()
	arg_dict = args.__dict__

	# Helix
	net = load_model('3Dconv_Keras.h5')
	# Remove the softmax layer...
	net.pop()
	net.compile('adam','categorical_crossentropy') # rebuild to set the output correctly
	print(net.summary())

	if 'mov_name' in arg_dict.keys() and arg_dict['mov_name'] is not None:
		video_pattern = os.path.splitext(args.mov_name)[0]

	reader = imageio.get_reader(video_pattern+'.avi')
	writer = imageio.get_writer(video_pattern+'_out_keras_BN.avi', fps=30, codec='mpeg4', quality=10)
	im_iter = reader.iter_data()

	input_size = 112
	time_depth = 16
	frames = [np.zeros([input_size, input_size, 1]) for x in range(time_depth)]

	framenum = 0
	while True:
		try:
			start_time = time()
			frames[0:time_depth-1] = np.copy(frames[1:time_depth])
			frame = np.uint8(next(im_iter))
			frame = imresize(frame, (input_size, input_size, 3))
			frame = frame[:,:,0]
			frame = np.reshape(frame, [input_size, input_size, 1])
			frames[time_depth-1] = frame
			batch_input_single = np.reshape(frames,[time_depth, input_size, input_size, 1])
			batch_input = np.reshape([batch_input_single,
									np.flipud(batch_input_single),
									np.fliplr(batch_input_single),
									np.fliplr(np.flipud(batch_input_single)),
									np.transpose(batch_input_single,(0,2,1,3)),
									np.transpose(np.flipud(batch_input_single),(0,2,1,3)),
									np.transpose(np.fliplr(batch_input_single),(0,2,1,3)),
									np.transpose(np.fliplr(np.flipud(batch_input_single)),(0,2,1,3))], [8, time_depth, input_size, input_size, 1])
			if framenum % 1000 == 0:
				print('Batch ' + str(framenum))
				print('Batch Assembled in: ' + str(time()-start_time))
			start_time = time()
			results_nosoftmax = net.predict(batch_input, batch_size=8)
			if framenum % 1000 == 0:
				print('Batch Computed in: ' + str(time()-start_time))
			start_time = time()
			results = softmax(results_nosoftmax)
			#print(results)
			predictions = np.argmax(results, 1)
			#prob = np.exp(results)[:,1] / np.sum(np.exp(results), axis=1)
			new_frame = np.zeros([input_size, input_size, 3])
			new_frame[:,:,0] = np.squeeze(frame)
			new_frame[:,:,1] = np.squeeze(frame)
			new_frame[:,:,2] = np.squeeze(frame)
			new_frame[:,0:15,:] = 255
			for i in range(np.shape(results)[0]):
				if results[i][1] < 0.50000001: # Negative
					level = 112-int(results[i][1]*112)
					new_frame[level:111,(0+2*i):(1+2*i),0] = 200
					new_frame[level:111,(0+2*i):(1+2*i),1] = 0
					new_frame[level:111,(0+2*i):(1+2*i),2] = 0
				else: # Positive
					level = 112-int(results[i][1]*112)
					new_frame[level:111,(0+2*i):(1+2*i),1] = 200
					new_frame[level:111,(0+2*i):(1+2*i),0] = 0
					new_frame[level:111,(0+2*i):(1+2*i),2] = 0
			# Tick marks at 25, 50, 75%
			new_frame[int(112/2),0:5,:] = 0
			new_frame[int(3*112/4),0:5,:] = 0
			new_frame[int(112/4),0:5,:] = 0
			# Consensus predictions
			mean_pred = np.mean(results[:,1]) > 0.5
			vote_pred = np.sum(predictions) > 4 # only majority, not half
			maxpool_pred = np.argmax(np.diag(results_nosoftmax[np.argmax(results_nosoftmax,0)])) == 1
			if mean_pred:
				new_frame[1:6,20:25,1] = 200
				new_frame[1:6,20:25,0] = 0
				new_frame[1:6,20:25,2] = 0
			else:
				new_frame[1:6,20:25,0] = 200
				new_frame[1:6,20:25,1] = 0
				new_frame[1:6,20:25,2] = 0
			if vote_pred:
				new_frame[1:6,30:35,1] = 200
				new_frame[1:6,30:35,0] = 0
				new_frame[1:6,30:35,2] = 0
			else:
				new_frame[1:6,30:35,0] = 200
				new_frame[1:6,30:35,1] = 0
				new_frame[1:6,30:35,2] = 0
			if maxpool_pred:
				new_frame[1:6,40:45,1] = 200
				new_frame[1:6,40:45,0] = 0
				new_frame[1:6,40:45,2] = 0
			else:
				new_frame[1:6,40:45,0] = 200
				new_frame[1:6,40:45,1] = 0
				new_frame[1:6,40:45,2] = 0
			writer.append_data(new_frame.astype('u1'))
			if framenum % 1000 == 0:
				print('Batch Saved in: ' + str(time()-start_time))
			framenum = framenum + 1
		except StopIteration:
			break
	reader.close()
	writer.close()


if __name__ == '__main__':
	main(sys.argv[1:])