Read CSV → Dataset

What this node actually is

ComfyDL is Cynthia-lxx's attempt to make ComfyUI do deep learning instead of diffusion - the pack's own tagline is "DeepLearning is just a few clicks away," and it's built on the d2l (Dive into Deep Learning) codebase, so the nodes map onto textbook recipes: linear regression, Squared Loss, mini-batch SGD, CNNs, eventually BERT. It ships ~207 nodes across 35 categories, 109 of them its own.

Read CSV → Dataset is where you come in. It takes a path to a .csv and hands you the pack's DATASET object - a single teal socket that carries a feature tensor, an optional label tensor, the column names, and some metadata, passed by reference between nodes. Every other data node in the pack speaks that type: the adapters, the trainers, the writers, the preview. So this node is the door: read a CSV, and you're inside.

Reach for it when your data lives anywhere sane - a pandas pickle you exported, a Google Sheets download, a metrics dump - and you want to train on it inside the graph rather than in a notebook.

How it works

It's a thin, honest wrapper over pandas. The node calls pd.read_csv(file, sep=delimiter, header=0 if header else None, low_memory=False) - that last flag is there to stop pandas silently upcasting dtypes on mixed files - then hands the DataFrame to CdlDataset.from_dataframe. That function pops the target column off the frame and turns it into the labels tensor, converts every remaining column to float32/float64 through torch.from_numpy (a zero-copy view when the dtype already matches), and keeps the column names as feature_names. So the pipeline is: file → DataFrame → features + labels.

One thing to internalise: everything that isn't the target becomes a float feature. There's no "ignore this column" switch.

Inputs and outputs

Required: file (default cdl_dataset.csv) and target - a column name, matched exactly, and empty means "unlabelled dataset" (fine for writing, useless for training). Optional: delimiter (default ,), dtype (float32/float64, default float32), and header (default True). Output is a single dataset.

In practice you set two things: the absolute path, and the name of the column you're predicting. Then wire dataset into Dataset → Preview to eyeball it, Dataset → Tensors to feed the trainers, or straight into Linear Regression Train.

Install

Search the pack title ComfyDL in ComfyUI Manager, or do it by hand:

cd ComfyUI/custom_nodes
git clone https://github.com/Cynthia-lxx/ComfyDL
pip install -r ./ComfyDL/requirements.txt

Restart ComfyUI. The console prints [ComfyDL] Loaded <n> deep learning nodes. if it worked. Note the pack's requirements.txt is only matplotlib, IPython and matplotlib-inline - pandas is not in it, even though every reader and writer imports it. Usually it's already in your venv from other packs; if not, pip install pandas. If you don't have ComfyUI at all, the author also ships ComfyDL_UI, a ComfyUI fork with the pack built in.

Where people get burned

Relative paths. The file widget is a plain string handed to pandas, so cdl_dataset.csv resolves against ComfyUI's process working directory - your ComfyUI folder - not the folder you launched from in your head. Use a full path and this class of bug disappears.

A stray text column. An ID column, an ISO date, a category name: any of those makes the float conversion blow up with a numpy error. Drop them in the CSV, or split the file.

target typos. The name is matched literally; a mismatch raises target column not found in DataFrame. With header = False the columns have no names at all, so target can never match - leave it empty in that case.

NaNs. Nothing in the chain drops them. A blank cell survives as NaN, becomes NaN in the tensor, and your loss goes NaN a few steps into training. Clean the CSV first.

Multi-character delimiters. delimiter is a single-character separator here (;, |, tab are all fine). Semicolon-separated European exports work great; \s+ does not - that's what Read Text is for.

Finally, remember the pack is young: the author archived the shipped example workflows because a node refactor made them stale, so there are no templates to crib from. You're wiring this by hand, which is exactly why the path and target fields matter so much.