import numpy as np
from keras.applications.inception_v3 import InceptionV3
from keras.models import Sequential, load_model, Model
from keras.layers import Input, average, concatenate, GlobalAveragePooling2D
from keras.layers import TimeDistributed, GlobalAveragePooling1D
from keras.layers.core import Dense, Dropout, Activation, Flatten
from keras.layers.convolutional import Conv2D, MaxPooling2D
from keras.optimizers import SGD, Adam
from keras.layers.normalization import BatchNormalization
class ResearchModels():
def __init__(self, nb_classes, n_snip, opt_flow_len, image_shape = (224, 224), saved_model=None, saved_temporal_weights=None, saved_spatial_weights=None):
"""
`nb_classes` = the number of classes to predict
`opt_flow_len` = the length of optical flow frames
`image_shape` = shape of image frame
`saved_model` = the path to a saved Keras model to load
"""
self.nb_classes = nb_classes
self.n_snip = n_snip
self.opt_flow_len = opt_flow_len
self.load_model = load_model
self.saved_model = saved_model
self.saved_temporal_weights = saved_temporal_weights
self.saved_spatial_weights = saved_spatial_weights
self.input_shape_spatial = (image_shape[0], image_shape[1], 3)
self.input_shape_temporal = (image_shape[0], image_shape[1], opt_flow_len * 2)
self.input_shape_spatial_multi = (self.n_snip, image_shape[0], image_shape[1], 3)
self.input_shape_temporal_multi = (self.n_snip, image_shape[0], image_shape[1], opt_flow_len * 2)
# Set the metrics. Only use top k if there's a need.
metrics = ['accuracy']
if self.nb_classes >= 10:
metrics.append('top_k_categorical_accuracy')
# Load model
# If saved fuse model exists, directly load
if self.saved_model is not None:
print("\nLoading model %s" % self.saved_model)
self.model = load_model(self.saved_model)
# Otherwise build the model and load weights for both streams
else:
print("\nLoading the two-stream model...")
self.model = self.two_stream_fuse()
optimizer = Adam()
# optimizer = SGD(lr=0.01, momentum=0.9, nesterov=True)
self.model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=metrics)
# Two-stream fused model
def two_stream_fuse(self):
# spatial stream (frozen)
cnn_spatial_multi = self.cnn_spatial_multi()
# temporal stream (frozen)
cnn_temporal_multi = self.cnn_temporal_multi()
# fused by taking average
outputs = average([cnn_spatial_multi.output, cnn_temporal_multi.output])
model = Model([cnn_spatial_multi.input, cnn_temporal_multi.input], outputs)
return model
# CNN model for the temporal stream with multiple inputs
def cnn_spatial_multi(self):
# spatial stream (frozen)
cnn_spatial = self.cnn_spatial()
if self.saved_spatial_weights is None:
print("[ERROR] No saved_spatial_weights weights file!")
else:
cnn_spatial.load_weights(self.saved_spatial_weights)
for layer in cnn_spatial.layers:
layer.trainable = False
# building inputs and output
model = Sequential()
model.add(TimeDistributed((cnn_spatial), input_shape=self.input_shape_spatial_multi))
model.add(GlobalAveragePooling1D())
return model
# CNN model for the temporal stream with multiple inputs
def cnn_temporal_multi(self):
# spatial stream (frozen)
cnn_temporal = self.cnn_temporal()
if self.saved_temporal_weights is None:
print("[ERROR] No saved_temporal_weights weights file!")
else:
cnn_temporal.load_weights(self.saved_temporal_weights)
for layer in cnn_temporal.layers:
layer.trainable = False
# building inputs and output
model = Sequential()
model.add(TimeDistributed((cnn_temporal), input_shape=self.input_shape_temporal_multi))
model.add(GlobalAveragePooling1D())
return model
# CNN model for the spatial stream
def cnn_spatial(self):
base_model = InceptionV3(weights='imagenet', include_top=False)
# add a global spatial average pooling layer
x = base_model.output
x = GlobalAveragePooling2D()(x)
# let's add a fully-connected layer
x = Dense(1024, activation='relu')(x)
# and a logistic layer
predictions = Dense(self.nb_classes, activation='softmax')(x)
model = Model(inputs=base_model.input, outputs=predictions)
return model
# CNN model for the temporal stream
def cnn_temporal(self):
#model
model = Sequential()
#conv1
model.add(Conv2D(96, (7, 7), strides=2, padding='same', input_shape=self.input_shape_temporal))
model.add(BatchNormalization())
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#conv2
model.add(Conv2D(256, (5, 5), strides=2, padding='same'))
model.add(Activation('relu'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#conv3
model.add(Conv2D(512, (3, 3), strides=1, activation='relu', padding='same'))
#conv4
model.add(Conv2D(512, (3, 3), strides=1, activation='relu', padding='same'))
#conv5
model.add(Conv2D(512, (3, 3), strides=1, activation='relu', padding='same'))
model.add(MaxPooling2D(pool_size=(2, 2)))
#full6
model.add(Flatten())
model.add(Dense(4096, activation='relu'))
model.add(Dropout(0.9))
#full7
model.add(Dense(2048, activation='relu'))
model.add(Dropout(0.9))
#softmax
model.add(Dense(self.nb_classes, activation='softmax'))
return model
