Skip to content

3D Models

Source code in k3im/cait_3d.py 
 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
def CAiT3DModel(
    image_size,
    image_patch_size,
    frames,
    frame_patch_size,
    num_classes,
    dim,
    depth,
    cls_depth,
    heads,
    mlp_dim,
    channels=3,
    dim_head=64,
):
    image_height, image_width = pair(image_size)
    patch_height, patch_width = pair(image_patch_size)

    assert (
        image_height % patch_height == 0 and image_width % patch_width == 0
    ), "Image dimensions must be divisible by the patch size."
    assert (
        frames % frame_patch_size == 0
    ), "Frames must be divisible by the frame patch size"

    nf, nh, nw = (
        frames // frame_patch_size,
        image_height // patch_height,
        image_width // patch_width,
    )
    patch_dim = channels * patch_height * patch_width * frame_patch_size

    i_p = layers.Input((frames, image_height, image_width, channels))
    tubelets = layers.Reshape(
        (frame_patch_size, nf, patch_height, nh, patch_width, nw, channels)
    )(i_p)
    tubelets = ops.transpose(tubelets, (0, 2, 4, 6, 1, 3, 5, 7))
    tubelets = layers.Reshape((nf, nh, nw, -1))(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Dense(dim)(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Reshape((-1, dim))(tubelets)
    tubelets = Transformer(dim, depth, heads, dim_head, mlp_dim)(tubelets)

    _, cls_token = CLS_Token(dim)(tubelets)
    cls_token = Transformer(dim, cls_depth, heads, dim_head, mlp_dim)(
        cls_token, context=tubelets
    )
    cls_token = ops.squeeze(cls_token, axis=1)
    o_p = layers.Dense(num_classes)(cls_token)

    return keras.Model(inputs=i_p, outputs=o_p)
Source code in k3im/cct_3d.py 
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
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
def CCT3DModel(
    input_shape,
    num_heads,
    projection_dim,
    kernel_size,
    stride,
    padding,
    transformer_units,
    stochastic_depth_rate,
    transformer_layers,
    num_classes,
    positional_emb=False,
):
    inputs = layers.Input(input_shape)

    # Encode patches.

    cct_tokenizer = CCTTokenizer3D(
        kernel_size,
        stride,
        padding,
        n_output_channels=[64, projection_dim],
        n_conv_layers=2,
    )
    encoded_patches = cct_tokenizer(inputs)

    # Apply positional embedding.
    if positional_emb:
        sequence_length = encoded_patches.shape[1]
        encoded_patches += PositionEmbedding(sequence_length=sequence_length)(
            encoded_patches
        )

    # Calculate Stochastic Depth probabilities.
    dpr = [x for x in np.linspace(0, stochastic_depth_rate, transformer_layers)]

    # Create multiple layers of the Transformer block.
    for i in range(transformer_layers):
        # Layer normalization 1.
        x1 = layers.LayerNormalization(epsilon=1e-5)(encoded_patches)

        # Create a multi-head attention layer.
        attention_output = layers.MultiHeadAttention(
            num_heads=num_heads, key_dim=projection_dim, dropout=0.1
        )(x1, x1)

        # Skip connection 1.
        attention_output = StochasticDepth(dpr[i])(attention_output)
        x2 = layers.Add()([attention_output, encoded_patches])

        # Layer normalization 2.
        x3 = layers.LayerNormalization(epsilon=1e-5)(x2)

        # MLP.
        x3 = mlp(x3, hidden_units=transformer_units, dropout_rate=0.1)

        # Skip connection 2.
        x3 = StochasticDepth(dpr[i])(x3)
        encoded_patches = layers.Add()([x3, x2])

    # Apply sequence pooling.
    representation = layers.LayerNormalization(epsilon=1e-5)(encoded_patches)
    weighted_representation = SequencePooling()(representation)

    # Classify outputs.
    logits = layers.Dense(num_classes)(weighted_representation)
    # Create the Keras model.
    model = keras.Model(inputs=inputs, outputs=logits)
    return model

ConvMixer-256/8: https://openreview.net/pdf?id=TVHS5Y4dNvM. The hyperparameter values are taken from the paper.

Source code in k3im/convmixer_3d.py 
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
def ConvMixer3DModel(
    image_size=28,
    num_frames=28,
    filters=256,
    depth=8,
    kernel_size=5,
    kernel_depth=5,
    patch_size=2,
    patch_depth=2,
    num_classes=10,
    num_channels=3,
):
    """ConvMixer-256/8: https://openreview.net/pdf?id=TVHS5Y4dNvM.
    The hyperparameter values are taken from the paper.
    """

    inputs = keras.Input((num_frames, image_size, image_size, num_channels))
    x = layers.Rescaling(scale=1.0 / 255)(inputs)

    kernel_size = (kernel_depth,) + pair(kernel_size)
    patch_size = (patch_depth,) + pair(patch_size)
    # Extract patch embeddings.
    x = conv_stem(x, filters, patch_size)

    # ConvMixer blocks.
    for _ in range(depth):
        x = conv_mixer_block(x, filters, kernel_size)

    # Classification block.
    x = layers.GlobalAvgPool3D()(x)
    outputs = layers.Dense(num_classes)(x)

    return keras.Model(inputs, outputs)
Source code in k3im/eanet3d.py 
 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
def EANet3DModel(
    image_size,
    image_patch_size,
    frames,
    frame_patch_size,
    num_classes,
    dim,
    depth,
    heads,
    mlp_dim,
    channels=3,
    dim_coefficient=4,
    projection_dropout=0.0,
    attention_dropout=0,
):
    image_height, image_width = pair(image_size)
    patch_height, patch_width = pair(image_patch_size)

    assert (
        image_height % patch_height == 0 and image_width % patch_width == 0
    ), "Image dimensions must be divisible by the patch size."
    assert (
        frames % frame_patch_size == 0
    ), "Frames must be divisible by the frame patch size"

    nf, nh, nw = (
        frames // frame_patch_size,
        image_height // patch_height,
        image_width // patch_width,
    )
    patch_dim = channels * patch_height * patch_width * frame_patch_size

    i_p = layers.Input((frames, image_height, image_width, channels))
    tubelets = layers.Reshape(
        (frame_patch_size, nf, patch_height, nh, patch_width, nw, channels)
    )(i_p)
    tubelets = ops.transpose(tubelets, (0, 2, 4, 6, 1, 3, 5, 7))
    tubelets = layers.Reshape((nf, nh, nw, -1))(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Dense(dim)(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Reshape((-1, dim))(tubelets)
    tubelets = Transformer(
        dim,
        depth,
        heads,
        mlp_dim,
        dim_coefficient=dim_coefficient,
        projection_dropout=projection_dropout,
        attention_dropout=attention_dropout,
    )(tubelets)

    tubelets = layers.GlobalAveragePooling1D(name="avg_pool")(tubelets)
    o_p = layers.Dense(num_classes)(tubelets)

    return keras.Model(inputs=i_p, outputs=o_p)
Source code in k3im/fnet_3d.py 
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
def FNet3DModel(
    image_size,
    image_patch_size,
    frames,
    frame_patch_size,
    num_classes,
    dim,
    depth,
    hidden_units,
    dropout_rate,
    channels=3,
):
    image_height, image_width = pair(image_size)
    patch_height, patch_width = pair(image_patch_size)

    assert (
        image_height % patch_height == 0 and image_width % patch_width == 0
    ), "Image dimensions must be divisible by the patch size."
    assert (
        frames % frame_patch_size == 0
    ), "Frames must be divisible by the frame patch size"

    nf, nh, nw = (
        frames // frame_patch_size,
        image_height // patch_height,
        image_width // patch_width,
    )
    patch_dim = channels * patch_height * patch_width * frame_patch_size

    i_p = layers.Input((frames, image_height, image_width, channels))
    tubelets = layers.Reshape(
        (frame_patch_size, nf, patch_height, nh, patch_width, nw, channels)
    )(i_p)
    tubelets = ops.transpose(tubelets, (0, 2, 4, 6, 1, 3, 5, 7))
    tubelets = layers.Reshape((nf, nh, nw, -1))(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Dense(dim)(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Reshape((-1, dim))(tubelets)
    num_patches = ops.shape(tubelets)[1]
    for _ in range(depth):
        tubelets = FNetLayer(hidden_units, dropout_rate)(tubelets)
    tubelets = layers.GlobalAveragePooling1D(name="avg_pool")(tubelets)
    o_p = layers.Dense(num_classes)(tubelets)

    return keras.Model(inputs=i_p, outputs=o_p)
Source code in k3im/gmlp_3d.py 
 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
def gMLP3DModel(
    image_size,
    image_patch_size,
    frames,
    frame_patch_size,
    num_classes,
    dim,
    depth,
    hidden_units,
    dropout_rate,
    channels=3,
):
    image_height, image_width = pair(image_size)
    patch_height, patch_width = pair(image_patch_size)

    assert (
        image_height % patch_height == 0 and image_width % patch_width == 0
    ), "Image dimensions must be divisible by the patch size."
    assert (
        frames % frame_patch_size == 0
    ), "Frames must be divisible by the frame patch size"

    nf, nh, nw = (
        frames // frame_patch_size,
        image_height // patch_height,
        image_width // patch_width,
    )
    patch_dim = channels * patch_height * patch_width * frame_patch_size

    i_p = layers.Input((frames, image_height, image_width, channels))
    tubelets = layers.Reshape(
        (frame_patch_size, nf, patch_height, nh, patch_width, nw, channels)
    )(i_p)
    tubelets = ops.transpose(tubelets, (0, 2, 4, 6, 1, 3, 5, 7))
    tubelets = layers.Reshape((nf, nh, nw, -1))(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Dense(dim)(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Reshape((-1, dim))(tubelets)
    num_patches = ops.shape(tubelets)[1]
    for _ in range(depth):
        tubelets = gMLPLayer(num_patches, hidden_units, dropout_rate)(tubelets)
    tubelets = layers.GlobalAveragePooling1D(name="avg_pool")(tubelets)
    o_p = layers.Dense(num_classes)(tubelets)

    return keras.Model(inputs=i_p, outputs=o_p)
Source code in k3im/mlp_mixer_3d.py 
 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
def MLPMixer3DModel(
    image_size,
    image_patch_size,
    frames,
    frame_patch_size,
    num_classes,
    dim,
    depth,
    hidden_units,
    dropout_rate,
    channels=3,
):
    image_height, image_width = pair(image_size)
    patch_height, patch_width = pair(image_patch_size)

    assert (
        image_height % patch_height == 0 and image_width % patch_width == 0
    ), "Image dimensions must be divisible by the patch size."
    assert (
        frames % frame_patch_size == 0
    ), "Frames must be divisible by the frame patch size"

    nf, nh, nw = (
        frames // frame_patch_size,
        image_height // patch_height,
        image_width // patch_width,
    )
    patch_dim = channels * patch_height * patch_width * frame_patch_size

    i_p = layers.Input((frames, image_height, image_width, channels))
    tubelets = layers.Reshape(
        (frame_patch_size, nf, patch_height, nh, patch_width, nw, channels)
    )(i_p)
    tubelets = ops.transpose(tubelets, (0, 2, 4, 6, 1, 3, 5, 7))
    tubelets = layers.Reshape((nf, nh, nw, -1))(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Dense(dim)(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Reshape((-1, dim))(tubelets)
    num_patches = ops.shape(tubelets)[1]
    for _ in range(depth):
        tubelets = MLPMixerLayer(num_patches, hidden_units, dropout_rate)(tubelets)
    tubelets = layers.GlobalAveragePooling1D(name="avg_pool")(tubelets)
    o_p = layers.Dense(num_classes)(tubelets)

    return keras.Model(inputs=i_p, outputs=o_p)
Source code in k3im/simple_vit_3d.py 
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
def SimpleViT3DModel(
    image_size,
    image_patch_size,
    frames,
    frame_patch_size,
    num_classes,
    dim,
    depth,
    heads,
    mlp_dim,
    channels=3,
    dim_head=64,
):
    image_height, image_width = pair(image_size)
    patch_height, patch_width = pair(image_patch_size)

    assert (
        image_height % patch_height == 0 and image_width % patch_width == 0
    ), "Image dimensions must be divisible by the patch size."
    assert (
        frames % frame_patch_size == 0
    ), "Frames must be divisible by the frame patch size"

    nf, nh, nw = (
        frames // frame_patch_size,
        image_height // patch_height,
        image_width // patch_width,
    )
    patch_dim = channels * patch_height * patch_width * frame_patch_size

    i_p = layers.Input((frames, image_height, image_width, channels))
    tubelets = layers.Reshape(
        (frame_patch_size, nf, patch_height, nh, patch_width, nw, channels)
    )(i_p)
    tubelets = ops.transpose(tubelets, (0, 2, 4, 6, 1, 3, 5, 7))
    tubelets = layers.Reshape((nf, nh, nw, -1))(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Dense(dim)(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    # pos_embedding = posemb_sincos_3d(tubelets)
    tubelets = layers.Reshape((-1, dim))(tubelets)
    # tubelets += pos_embedding
    tubelets = Transformer(dim, depth, heads, dim_head, mlp_dim)(tubelets)

    tubelets = layers.GlobalAveragePooling1D(name="avg_pool")(tubelets)
    o_p = layers.Dense(num_classes)(tubelets)

    return keras.Model(inputs=i_p, outputs=o_p)
Source code in k3im/vit_3d.py 
 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
def ViT3DModel(
    image_size,
    image_patch_size,
    frames,
    frame_patch_size,
    num_classes,
    dim,
    depth,
    heads,
    mlp_dim,
    pool,
    channels=3,
    dim_head=64,
):
    image_height, image_width = pair(image_size)
    patch_height, patch_width = pair(image_patch_size)

    assert (
        image_height % patch_height == 0 and image_width % patch_width == 0
    ), "Image dimensions must be divisible by the patch size."
    assert (
        frames % frame_patch_size == 0
    ), "Frames must be divisible by the frame patch size"

    assert pool in {
        "cls",
        "mean",
    }, "pool type must be either cls (cls token) or mean (mean pooling)"

    nf, nh, nw = (
        frames // frame_patch_size,
        image_height // patch_height,
        image_width // patch_width,
    )
    patch_dim = channels * patch_height * patch_width * frame_patch_size

    i_p = layers.Input((frames, image_height, image_width, channels))
    tubelets = layers.Reshape(
        (frame_patch_size, nf, patch_height, nh, patch_width, nw, channels)
    )(i_p)
    tubelets = ops.transpose(tubelets, (0, 2, 4, 6, 1, 3, 5, 7))
    tubelets = layers.Reshape((nf, nh, nw, -1))(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Dense(dim)(tubelets)
    tubelets = layers.LayerNormalization()(tubelets)
    tubelets = layers.Reshape((-1, dim))(tubelets)
    num_patches = ops.shape(tubelets)[1]
    tubelets = ClassTokenPositionEmb(num_patches, dim)(tubelets)
    tubelets = Transformer(dim, depth, heads, dim_head, mlp_dim)(tubelets)
    if pool == "mean":
        tubelets = layers.GlobalAveragePooling1D(name="avg_pool")(tubelets)
    else:
        tubelets = tubelets[:, -1]
    o_p = layers.Dense(num_classes)(tubelets)

    return keras.Model(inputs=i_p, outputs=o_p)