Semisupervised Deep Features of Time-Frequency Maps for Multimodal Emotion Recognition

Table 4

Structure of Inception-v3.

Types	Patch size/stride (or remarks)	Input size
Convolution	3 × 3/2	299 × 299 × 3
Convolution	3 × 3/1	149 × 149 × 32
Convolution padded	3 × 3/1	147 × 147 × 32
Maximum pooling	3 × 3/2	147 × 147 × 64
Convolution	3 × 3/1	73 × 73 × 64
Convolution	3 × 3/2	71 × 71 × 80
Convolution	3 × 3/1	35 × 35 × 192
3 × inception	As in Figure 3(a)	35 × 35 × 288
5 × inception	As in Figure 3(b)	17 × 17 × 768
2 × inception	As in Figure 3(c)	8 × 8 × 1280
Maximum pooling	8 × 8	8 × 8 × 2048
Linear	Logits (unnormalized log-probabilities)	8 × 8 × 2048
Softmax	Classifier	8 × 8 ×