Architectures / Recommendation
🛒 DLRM
Meta's Deep Learning Recommendation Model — bottom MLP for dense, embedding for sparse, feature interaction, top MLP
From Naumov et al. (2019). Deep Learning Recommendation Model for Personalization and Recommendation Systems. This page is the graph, not the PDF: open it, edit it, verify it, export it.
Layers
14
Parameters
64.35M
Input
13
Output
1
Verifier
Clean
Every number on this page is computed from the graph by the same functions the app runs, not written by hand.
Open DLRM on the canvas
Free, no account needed
When to pick it
Pick when you have substantial dense + sparse features and need a production-validated CTR baseline. Bare-bones — needs feature engineering.
Structure
14 layers. Output shapes are propagated from the input shape, batch dimension excluded.
| Layer | Type | Parameters | Output shape | |
|---|---|---|---|---|
| 1 | Dense Features | Input | shape=[13] | 13 |
| 2 | Bottom MLP 1 | Linear | outFeatures=64, inFeatures=13 | 64 |
| 3 | ReLU | ReLU | 64 | |
| 4 | Bottom MLP 2 | Linear | outFeatures=64, inFeatures=64 | 64 |
| 5 | Sparse Features | Input | shape=[26] | 26 |
| 6 | EmbeddingBag | EmbeddingBag | vocabSize=1000000 | 26 × 64 |
| 7 | Feature Interaction | Feature Interaction | 415 | |
| 8 | Top MLP 1 | Linear | outFeatures=512, inFeatures=415 | 512 |
| 9 | ReLU | ReLU | 512 | |
| 10 | Top MLP 2 | Linear | outFeatures=256, inFeatures=512 | 256 |
| 11 | ReLU | ReLU | 256 | |
| 12 | CTR Head | Linear | outFeatures=1, inFeatures=256 | 1 |
| 13 | Sigmoid | Sigmoid | 1 | |
| 14 | P(click) | Output | 1 |
What the verifier says
The same 43 single-graph checks that run on every edit in the app, on this graph.
info"Sigmoid" feeds directly into Output. PyTorch's nn.CrossEntropyLoss already applies log-softmax internally, an explicit Softmax causes double-application and degrades training stability. Fix: Remove Softmax/Sigmoid for training. Restore it in a separate inference wrapper or ONNX export. (Sigmoid)
output-activation
output-activation
infoSigmoid saturates to [0,1] / [-1,1], and its gradient approaches zero for large inputs. In networks deeper than 5 layers, this halts learning in early layers. Fix: Use ReLU, GELU, or SiLU for hidden layers. Keep Sigmoid only at binary classification outputs; Tanh in specific contexts (GAN generators, LSTM gates). (Sigmoid)
vanishing-gradient
vanishing-gradient
info11 layers with no BatchNorm, LayerNorm, or GroupNorm. Without normalization, activations can explode or vanish across layers, causing slow or unstable training. Fix: Add BatchNorm after Conv2d (CV tasks), LayerNorm after attention/FFN (NLP/LLM), or GroupNorm for small batch sizes.
deep-no-norm
deep-no-norm
infoPyTorch initializes Linear/Conv with Kaiming (He) init, which is derived for ReLU-family activations. Feeding a saturating activation (sigmoid/tanh) from a He-initialized layer starts training in the saturated tails, shrinking early gradients. Fix: Initialize these layers with Xavier instead: nn.init.xavier_uniform_(w, gain=nn.init.calculate_gain("sigmoid"|"tanh")), or switch the activation to a ReLU-family one. (CTR Head)
init-activation-mismatch
init-activation-mismatch
The PyTorch it exports
Generated from the graph above. First 46 lines; the app exports the whole file, plus the training loop, the data contract and a deploy bundle.
# Architecture designed with Neurarch: https://neurarch.com
# PyTorch: compatible with Python 3.8+ and torch>=1.12
# Colab: pip install torch torchvision (usually pre-installed)
import torch
import torch.nn as nn
import torch.nn.functional as F
from typing import Tuple
class DLRM(nn.Module):
def __init__(self):
super().__init__()
self.linear_1 = nn.Linear(13, 64)
self.linear_2 = nn.Linear(64, 64)
self.embeddingBag_1 = nn.Embedding(1000000, 64)
self.linear_3 = nn.Linear(415, 512)
self.linear_4 = nn.Linear(512, 256)
self.linear_5 = nn.Linear(256, 1)
@staticmethod
def _dlrm_interact(dense, emb):
"""DLRM feature interaction: pairwise dots over [dense] + embedding rows,
concatenated back onto dense. Returns [B, D + C(N+1, 2)]."""
if emb.dim() == 2:
emb = emb.unsqueeze(0)
if dense.dim() == 1:
dense = dense.unsqueeze(0)
v = torch.cat([dense.unsqueeze(1), emb], dim=1) # [B, N+1, D]
z = torch.bmm(v, v.transpose(1, 2)) # [B, N+1, N+1]
i, j = torch.triu_indices(z.size(1), z.size(2), offset=1)
return torch.cat([dense, z[:, i, j]], dim=-1)
def forward(self, src, tgt=None):
# Dense Features shape: [13]
# Sparse Features shape: [26]
linear_ot_fc1 = self.linear_1(src)
relu__relu1 = F.relu(linear_ot_fc1)
linear_ot_fc2 = self.linear_2(relu__relu1)
embedding_bag_ed_bag = self.embeddingBag_1(tgt)
feature_interaction_teract = self._dlrm_interact(linear_ot_fc2, embedding_bag_ed_bag)
linear_op_fc1 = self.linear_3(feature_interaction_teract)
relu__relu1 = F.relu(linear_op_fc1)
linear_op_fc2 = self.linear_4(relu__relu1)
relu__relu2 = F.relu(linear_op_fc2)
linear_op_out = self.linear_5(relu__relu2)For agents
This architecture is machine-readable end to end. An agent can list the set, fetch this graph, edit it, and have the edit verified before any GPU time is spent.