pytorch-research（火炬研究）

PyTorch - Advanced Research & Engineering

Research-grade PyTorch requires moving beyond nn.Sequential. You need to control how gradients flow, how weights are initialized, and how computation is distributed across multiple GPUs. This guide covers the "internals" of the framework.

When to Use

• Implementing custom layers with non-standard mathematical derivatives.
• Debugging vanishing or exploding gradients using Hooks.
• Scaling models to multiple GPUs (Distributed Data Parallel).
• Fine-tuning model performance using the PyTorch Profiler.
• Creating complex learning rate schedules (Cyclic, OneCycle).
• Deploying models for high-performance inference (TorchScript, FX).
• Researching Weight Initialization and Normalization techniques.

Reference Documentation

• Autograd Mechanics: https://pytorch.org/docs/stable/notes/autograd.html
• Distributed Training: https://pytorch.org/docs/stable/distributed.html
• Profiler: https://pytorch.org/tutorials/recipes/recipes/profiler_recipe.html
• Search patterns: torch.autograd.Function, register_forward_hook, DistributedDataParallel, torch.nn.init

Core Principles

Beyond the Computational Graph

PyTorch is a "define-by-run" framework, but for research, you often need to intervene in the backward pass or inspect intermediate tensors without breaking the graph.

The Life of a Gradient

Understanding that gradients are accumulated in .grad attributes and that backward() consumes the graph unless retain_graph=True is specified.

Memory vs. Speed

In research, you often trade memory (activations) for speed (recomputation) using techniques like checkpointing.

Quick Reference

Standard Imports

import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.distributed as dist
from torch.nn.parallel import DistributedDataParallel as DDP
from torch.utils.data.distributed import DistributedSampler

Basic Pattern - Custom Autograd Function

class MySignFunction(torch.autograd.Function):
    @staticmethod
    def forward(ctx, input):
        # Save input for backward pass
        ctx.save_for_backward(input)
        return torch.sign(input)

    @staticmethod
    def backward(ctx, grad_output):
        # Straight-through estimator (STE) logic
        input, = ctx.saved_tensors
        grad_input = grad_output.clone()
        # Custom logic: gradients pass through as if it were an identity
        return grad_input

# Usage
my_sign = MySignFunction.apply

Critical Rules

✅ DO

• Use register_full_backward_hook - To inspect or modify gradients as they flow through a specific module.
• Initialize Weights Explicitly - Use torch.nn.init (Xavier, Kaiming) inside a model.apply(fn) loop.
• Use DistributedDataParallel (DDP) - Instead of DataParallel (DP). DDP is faster and handles multi-process communication correctly.
• Profile Before Optimizing - Use torch.profiler to find which operator (e.g., a slow view() or cat()) is actually slowing down the model.
• Use torch.cuda.empty_cache() sparingly - It doesn't free physical memory to the OS, but it fragments the PyTorch memory manager. Only use it in long-running loops if needed.

❌ DON'T

• Don't use inplace=True in custom layers - This often breaks Autograd's ability to compute gradients correctly.
• Don't use item() inside the loop - Calling .item() on a GPU tensor forces a CPU-GPU sync, which kills performance.
• Don't forget to set shuffle=False for DistributedSampler - Let the sampler handle the shuffling logic in a multi-GPU environment.
• Avoid Global Variables - PyTorch models should be self-contained for easy serialization and deployment.

Advanced Custom Layers

Hooks for Debugging and Feature Extraction

def print_grad_norm(module, grad_input, grad_output):
    print(f"Module: {module.__class__.__name__}, Grad Norm: {grad_output[0].norm().item()}")

# Attach to a specific layer
model.fc1.register_full_backward_hook(print_grad_norm)

# Extract activations (Forward Hook)
activations = {}
def get_activation(name):
    def hook(model, input, output):
        activations[name] = output.detach()
    return hook

model.conv1.register_forward_hook(get_activation('conv1'))

Advanced Training Patterns

Distributed Data Parallel (DDP) Skeleton

import torch.multiprocessing as mp

def setup(rank, world_size):
    dist.init_process_group("nccl", rank=rank, world_size=world_size)

def train(rank, world_size):
    setup(rank, world_size)

    model = MyModel().to(rank)
    ddp_model = DDP(model, device_ids=[rank])

    # Use DistributedSampler to ensure each GPU sees different data
    sampler = DistributedSampler(dataset, num_replicas=world_size, rank=rank)
    loader = DataLoader(dataset, sampler=sampler, batch_size=32)

    optimizer = torch.optim.Adam(ddp_model.parameters(), lr=0.001)
    # ... training loop ...

    dist.destroy_process_group()

# mp.spawn(train, args=(world_size,), nprocs=world_size)

Performance & Profiling

Using the Profiler

from torch.profiler import profile, record_function, ProfilerActivity

with profile(activities=[ProfilerActivity.CPU, ProfilerActivity.CUDA],
             record_shapes=True) as prof:
    with record_function("model_inference"):
        model(inputs)

print(prof.key_averages().table(sort_by="cuda_time_total", row_limit=10))

Gradient Checkpointing (Memory Saving)

If you have a very deep model and limited memory, trade computation for space.

from torch.utils.checkpoint import checkpoint

class DeepModel(nn.Module):
    def forward(self, x):
        # Instead of storing all activations, recompute them during backward
        x = checkpoint(self.heavy_layer_1, x)
        x = checkpoint(self.heavy_layer_2, x)
        return x

Practical Workflows

1. Custom Weight Initialization

def init_weights(m):
    if isinstance(m, nn.Linear):
        # Kaiming initialization for ReLU networks
        torch.nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
        if m.bias is not None:
            torch.nn.init.constant_(m.bias, 0)
    elif isinstance(m, nn.Conv2d):
        torch.nn.init.xavier_uniform_(m.weight)

model.apply(init_weights)

2. Gradient Clipping (Stability)

# Inside training loop
loss.backward()

# Clip to prevent exploding gradients (standard in RNNs/Transformers)
torch.nn.utils.clip_grad_norm_(model.parameters(), max_norm=1.0)

optimizer.step()

3. Dynamic Learning Rate (OneCycleLR)

optimizer = torch.optim.SGD(model.parameters(), lr=0.1)
scheduler = torch.optim.lr_scheduler.OneCycleLR(optimizer, max_lr=0.1,
                                                steps_per_epoch=len(train_loader),
                                                epochs=10)

for epoch in range(10):
    for batch in train_loader:
        train_batch()
        scheduler.step() # Step every batch for OneCycle

Common Pitfalls and Solutions

In-place Modification Error

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation.

# ❌ Problem: x += 1 (breaks backward pass)
# ✅ Solution: y = x + 1 (creates a new tensor)

CUDA Out of Memory (OOM) Strategies

• Batch Size: Reduce it.
• Gradient Accumulation: Compute loss for small batches, but only step() every N steps.
• Empty Cache: Use torch.cuda.empty_cache() between independent evaluations.
• Mixed Precision: Use torch.cuda.amp (saves 50% memory).

Silent Failure: zero_grad() position

If you call zero_grad() after backward() but before step(), your model will never learn.

# ✅ Correct order:
optimizer.zero_grad()
loss.backward()
optimizer.step()

Research PyTorch is about mastery over the mathematical engine. By leveraging custom gradients, hooks, and distributed infrastructure, you can move from training standard models to inventing the next generation of scientific AI.