task_predict_modality/src/methods/suzuki_mlp/config.vsh.yaml at 8fe5caea0ea537320fffd8bcfb04bf7410b603d5 · openproblems-bio/task_predict_modality · 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
__merge__: ../../api/comp_method.yaml
name: suzuki_mlp
label: Suzuki MLP
summary: Hierarchical encoder-decoder neural network with task-specific preprocessing and residual connections for cross-modal prediction.
description: |
  A hierarchical neural network encoder-decoder model based on Shuji Suzuki's 1st place solution
  in the Open Problems Multimodal Single-Cell Integration competition. The model uses task-specific
  preprocessing, SVD dimensionality reduction, and hierarchical MLP blocks with residual connections
  for learning cross-modal mappings.

  The original author's code was adapted by GitHub Copilot
  (using Claude Sonnet) to integrate with this repository's framework and standards.
links:
  documentation: https://www.kaggle.com/competitions/open-problems-multimodal/discussion/348468
  repository: https://github.com/shu65/open-problems-multimodal
info:
  preferred_normalization: log_cp10k
arguments:
  # Task configuration
  - name: "--task_type"
    type: "string"
    default: "auto"
    description: Task type - 'auto' for automatic detection, 'cite' for CITE-seq, 'multi' for multiome.

  # Preprocessing arguments
  - name: "--inputs_n_components"
    type: "integer"
    default: 128
    description: Number of SVD components for input modality dimensionality reduction.
  - name: "--targets_n_components"
    type: "integer"
    default: 128
    description: Number of SVD components for target modality dimensionality reduction.

  # Model architecture arguments
  - name: "--encoder_h_dim"
    type: "integer"
    default: 512
    description: Hidden dimension size for the encoder.
  - name: "--decoder_h_dim"
    type: "integer"
    default: 512
    description: Hidden dimension size for the decoder.
  - name: "--n_encoder_block"
    type: "integer"
    default: 3
    description: Number of encoder blocks.
  - name: "--n_decoder_block"
    type: "integer"
    default: 3
    description: Number of decoder blocks.
  - name: "--dropout_p"
    type: "double"
    default: 0.1
    description: Dropout probability.
  - name: "--activation"
    type: "string"
    default: "relu"
    description: Activation function (relu or gelu).
  - name: "--norm"
    type: "string"
    default: "layer_norm"
    description: Normalization type (layer_norm or batch_norm).
  - name: "--use_skip_connections"
    type: "boolean"
    default: true
    description: Whether to use skip connections in blocks.

  # Training arguments
  - name: "--learning_rate"
    type: "double"
    default: 1e-4
    description: Learning rate for training.
  - name: "--weight_decay"
    type: "double"
    default: 1e-6
    description: Weight decay for regularization.
  - name: "--epochs"
    type: "integer"
    default: 40
    description: Number of training epochs.
  - name: "--batch_size"
    type: "integer"
    default: 64
    description: Batch size for training.
  - name: "--use_residual_connections"
    type: "boolean"
    default: true
    description: Whether to use residual connections for multi task.

resources:
  - type: python_script
    path: script.py
  - path: utils.py
engines:
  - type: docker
    image: openproblems/base_pytorch_nvidia:1
runners:
  - type: executable
  - type: nextflow
    directives:
      label: [hightime, highmem, midcpu, gpu]