Improve data quality workflows - #17
Conversation
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
| function build_grm(dataset::GenomicDataset; method::Symbol = :vanraden) | ||
| X = dataset.genotype | ||
| n, p = size(X) | ||
| p == 0 && throw(ArgumentError("基因型矩阵为空, 无法构建 GRM")) | ||
| method == :vanraden || throw(ArgumentError("暂不支持的方法: $(method)")) | ||
|
|
||
| # 还原原始 0/1/2 计数以计算等位基因频率。 | ||
| Z = copy(X) | ||
| denom = 0.0 | ||
| @inbounds for j in 1:p | ||
| col = view(Z, :, j) | ||
| μ = mean(col) | ||
| col .-= μ | ||
| freq = (μ + 1.0) / 2 # 假设之前标准化后的均值约为 0 | ||
| denom += 2 * freq * (1 - freq) | ||
| end |
There was a problem hiding this comment.
The VanRaden GRM builder derives each locus frequency as freq = (μ + 1.0) / 2 after simply copying the genotype matrix. For raw 0/1/2 genotypes the correct estimator is mean(col) / 2; adding 1 forces frequencies ≥0.5 and makes common cases (e.g. μ ≈ 1 for p ≈ 0.5) yield freq = 1, driving 2*freq*(1−freq) to zero so the denominator collapses to eps(). When μ > 1 the term becomes negative. As a result the returned GRM is arbitrarily scaled or numerically unstable for any realistic dataset, breaking downstream models that expect a properly scaled relationship matrix.
Useful? React with 👍 / 👎.
| """ | ||
| build_transformer(seq_len, embed_dim, num_heads, num_layers; ff_dim=4*embed_dim, | ||
| dropout_rate=0.1) | ||
|
|
||
| 构建轻量 Transformer, 利用多头注意力捕捉远距离位点关联。 | ||
| """ | ||
| function build_transformer(seq_len::Integer, embed_dim::Integer, num_heads::Integer, num_layers::Integer; | ||
| ff_dim::Integer = 4 * embed_dim, dropout_rate::Real = 0.1) | ||
| encoder_blocks = Any[] | ||
| for _ in 1:num_layers | ||
| attn = MultiheadAttention(embed_dim, num_heads; dropout = dropout_rate) | ||
| ff = Chain(Dense(embed_dim, ff_dim, gelu), Dropout(dropout_rate), Dense(ff_dim, embed_dim)) | ||
| push!(encoder_blocks, TransformerEncoder(attn, ff, LayerNorm(embed_dim))) | ||
| end | ||
| pos_enc = reshape(positionalencoding(embed_dim, seq_len), (embed_dim, seq_len, 1)) | ||
| embedding = Dense(seq_len, embed_dim * seq_len, gelu) | ||
| return Chain( | ||
| x -> reshape(x, (seq_len, size(x, 2))), | ||
| embedding, | ||
| x -> reshape(x, (embed_dim, seq_len, size(x, 2))), | ||
| x -> x .+ pos_enc, | ||
| encoder_blocks..., |
There was a problem hiding this comment.
build_transformer captures pos_enc inside the anonymous layer x -> x .+ pos_enc. When the model is moved to GPU via train_deep_model! (which calls fmap(cu, model)), that captured array is not transferred, so with device = :gpu the forward pass attempts to add a CPU matrix to a CuArray and raises a MethodError. This makes transformer training/inference fail on GPU despite the API exposing GPU support.
Useful? React with 👍 / 👎.
1 participant
Summary
Testing
https://chatgpt.com/codex/tasks/task_e_68f98c6e9d388321897d42279e9d067d