Inputs and outputs | Dagnam.AI Docs

The Input node, the Dataset node, augmentations, and every field on the Output node that controls training.

Inputs and outputs Every architecture needs exactly one kind of data source an Input node or one or more Dataset nodes, never both and an Output node, which carries the training configuration for the whole design. This page covers both, plus the augmentations you can attach to a data source. The Input node The Input node is the entry point for data you are not linking from a dataset. Its Input Type decides which shape fields it shows: Input Type Shape fields defaults ------------------- --------------------------------------------------------------------------------------------------------------------------------------------------------- Image default Image Size Preset ImageNet, CIFAR-10/100, MNIST or Custom , Height 224, Width 224, Channels 3 Audio Sample Rate 16000 8000 to 48000 , Duration seconds 4, Audio Channels 1 1 to 2 Sequence Sequence Preset Token IDs, BERT Base, GPT-2 Small, Short Text or Custom; default Custom, with Sequence Length 256 and Feature Dimension 128 Video Frame Count 16, Height 224, Width 224, Channels 3, Frames Per Second 30 Tabular Column Count 10, Feature Types Numeric Numeric, Categorical, Mixed Feature Feature Size Preset Small, Standard, Large or Custom; default Standard, 128 features ; Custom shows Number of Features Every Input Type also shows Normalization None, Z-Score, or Rescale; default None and Data Loading Workers default 4, 0 to 16 . Augmentations A data source can turn on augmentations for its domain. Each one is off by default; turning it on reveals its own parameters. Deterministic transforms Normalize can also be applied to validation and test, not just training; random ones stay training-only. Augmentation Applies to Key parameters defaults ----------------------- -------------------------- ---------------------------------------------------------------------------------- Random Crop Image Size inherits the upstream shape by default , Padding 4 Horizontal Flip Image Probability 0.5 Normalize Image Mean, Std per channel Color Jitter Image Brightness 0.2, Contrast 0.2, Saturation 0.2, Hue 0.05 Random Resized Crop Image Size inherits the upstream shape by default , Scale 0.08 to 1, Ratio 0.75 to 1.33 MixUp Image, classification only Alpha 0.2, Probability 1 CutMix Image, classification only Alpha 1, Probability 1 RandAugment Image Ops 2, Magnitude 9 SpecAugment Audio Frequency Mask 15, Time Mask 35, Masks 2 Time Stretch Audio Rate 0.9 to 1.1 Noise Injection Audio SNR dB 10 to 30 The Dataset node Drag a dataset card from the Data tab filters for All, System, User, and by task; an Upload Dataset action; click a card to preview it onto the canvas. Its "Dataset Configuration" panel shows: - Role chips for Training , Validation , and Testing : which splits this node covers. - A split slider for each covered role, as a percentage. - Adopt Dataset Shape , shown when the dataset's real column count does not match the configured shape. - Refresh Metadata , shown once a dataset is linked. - The same Input Type and shape fields as the Input node. - A summary of which columns feed the model as input and which is the target, and how many rows land in each split, once the dataset's metadata is available. One data source kind per design A design can have at most one Input node, and an Input node cannot be mixed with Dataset nodes; use one Input node, or one or more Dataset nodes. Switching from an Input node to a Dataset node works the same way as deleting the last one: add the Dataset node first, then delete the Input node see Deleting and the last Output or data source rule /docs/studio/overview deleting-and-the-last-output-or-data-source-rule . Training from Studio's Train button uses the dataset and splits of the Dataset node with the Training role, or of the first Dataset node when none has it. The Output node The Output node carries the whole training configuration and adds no layer of its own, so the last layer you connect to it produces the model's output. Set the final activation with Final Activation on the Output node: a layer connected directly to the Output node has its Activation fixed to Linear, with the note "Final activation is controlled by the Output node." - Output Type : Classification default , Regression, Segmentation, Language Modeling, Generation. - Number of Classes default 10, 1 to 10000 : for Classification and Segmentation. - Vocabulary Size default 10000 and Sequence Length default 128 : for Language Modeling. - Optimizer : Adam default , SGD, RMSProp, AdamW, Adagrad. SGD and RMSProp add Optimizer Momentum default 0 , and AdamW, SGD and RMSProp add Weight Decay default 0 here as well as under Regularization. - Learning Rate : default 0.001. - Epochs : default 10 1 to 1000 . - Batch Size : default 32 1 to 2048 . - Loss Function and Metrics a multiselect : the options depend on the Output Type, as in the table below. Label Smoothing default 0, up to 0.2 appears alongside CrossEntropyLoss for Classification. - Final Activation : None, Softmax the default on a new Output node , Sigmoid, Argmax, Beam Search, or Top-K Sampling. Output Type Loss Function options Metrics options ----------------- ---------------------------------------------------------- ----------------------------------------- Classification CrossEntropyLoss, BCELoss, BCEWithLogitsLoss, NLLLoss accuracy, precision, recall, f1, auc Regression MSELoss, L1Loss, HuberLoss, SmoothL1Loss mse, mae, rmse, r2, mape Segmentation CrossEntropyLoss, BCEWithLogitsLoss, DiceLoss, DiceBCELoss dice score, iou, pixel accuracy, mean iou Language Modeling CrossEntropyLoss perplexity, token accuracy Generation MSELoss, L1Loss, BCELoss, KLDivLoss fid, inception score, ssim, psnr, lpips Changing Output Type resets Loss Function and Metrics to the first option for the new type, and Final Activation to Softmax for Classification and Segmentation or None for every other type. Generating code and training from Studio both use these fields, including the loss and compute settings, as configured here. Compute - Mode : CPU default , Single GPU, or Multi-GPU. - Precision : FP32 default , plus BF16 on CPU, or FP16 and BF16 on a GPU mode. - Gradient Accumulation : shown once Mode leaves CPU; default 1. - GPUs per Node default 2 and Nodes default 1 : Multi-GPU only. - Advanced Multi-GPU : a distributed Backend Auto, plus a framework-specific list , and PyTorch-only settings that depend on the backend: Find Unused Parameters , CPU Offload , Activation Checkpointing , and a ZeRO Stage default 2 . Multi-GPU training strategies need access beyond the Free plan; CPU and single GPU are available on Free. Backends your plan does not include are disabled in the Backend list, and a few are not on any public plan; ask through Support /support . Training on Dagnam's cloud GPUs also needs cloud GPU access, available on request during the beta through Support. See Account /docs/getting-started/account for plans in words. Learning rate schedule LR Scheduler : None default , Reduce on Plateau, Cosine Annealing, Warmup Decay, or OneCycle, with Warmup Epochs default 0, up to 100 once a scheduler is chosen. Regularization and optimization Regularization is off by default Enable Regularization . Its fields, Weight Decay default 0 , Gradient Clipping default 0 , and Early Stopping Patience default 10 , are always listed but stay dimmed until you turn it on; turning it off resets them to those defaults. Optimization always shows Gradient Accumulation Steps default 1 and Gradient Checkpointing off by default ; Compile Model torch.compile is PyTorch only and off by default. Checkpointing and logging Checkpointing: Save Frequency Epochs default every epoch , Max Checkpoints default 5, capped by the number of epochs , Save Best Only on by default , and Save Optimizer State on by default . Logging: Log Frequency Steps default every 10 steps and Log Level INFO by default; DEBUG, INFO, WARNING, or ERROR .