DataParallel basics in PyTorch - Model Pipeline Trace

Model Pipeline - DataParallel basics

This pipeline shows how PyTorch's DataParallel helps train a model faster by splitting data across multiple GPUs. It copies the model to each GPU, processes parts of the data in parallel, then combines the results.

Data Flow - 5 Stages

1Input Data

1000 rows x 10 features→Original dataset loaded for training→1000 rows x 10 features

[[0.5, 1.2, ..., 0.3], [0.7, 0.8, ..., 0.1], ...]

↓

2Batch Split for GPUs

100 rows x 10 features (batch size)→Batch split evenly across 2 GPUs→2 batches of 50 rows x 10 features each

[GPU0 batch: [[0.5, 1.2, ...], ... 50 rows], GPU1 batch: [[0.7, 0.8, ...], ... 50 rows]]

↓

3Model Replication

Model with parameters (weights)→Model copied to each GPU→2 model replicas on GPU0 and GPU1

Same model architecture on both GPUs

↓

4Parallel Forward Pass

2 batches of 50 rows x 10 features→Each GPU processes its batch through its model copy→2 outputs of 50 rows x 1 prediction each

[GPU0 output: [0.8, 0.3, ...], GPU1 output: [0.5, 0.9, ...]]

↓

5Gather Outputs

2 outputs of 50 rows x 1 prediction→Combine outputs from GPUs into one batch→100 rows x 1 prediction

[0.8, 0.3, ..., 0.5, 0.9, ...]

Training Trace - Epoch by Epoch

Loss
0.7 |****
0.6 |*** 
0.5 |**  
0.4 |**  
0.3 |*   
0.2 |*   
    +---------
     1 2 3 4 5 Epochs

Epoch	Loss ↓	Accuracy ↑	Observation
1	0.65	0.60	Initial training with DataParallel, loss starts high
2	0.48	0.72	Loss decreases as model learns, accuracy improves
3	0.35	0.81	Continued improvement, model converging
4	0.28	0.86	Loss lowers further, accuracy nearing good performance
5	0.22	0.90	Training stabilizes with good accuracy

Prediction Trace - 4 Layers

Layer 1: Input batch split

Layer 2: Forward pass on GPU0

Layer 3: Forward pass on GPU1

Layer 4: Gather outputs

Model Quiz - 3 Questions

What is the main benefit of using DataParallel in PyTorch?

A It reduces the size of the model

B It increases the number of model parameters

C It splits data across GPUs to speed up training

D It changes the model architecture automatically

Key Insight

Using DataParallel allows the same model to run on multiple GPUs simultaneously by splitting input data batches. This speeds up training without changing the model itself. Loss decreases and accuracy improves as training progresses, showing effective learning.

Practice

1. What is the main purpose of torch.nn.DataParallel in PyTorch?

easy

A. To convert a model to run only on CPU

B. To run a model on multiple GPUs by splitting input data automatically

C. To reduce the size of the model parameters

D. To change the model architecture dynamically during training

Solution

1. Step 1: Understand DataParallel's role

   DataParallel helps use multiple GPUs by splitting input data and running model copies in parallel.

2. Step 2: Compare options

   Only To run a model on multiple GPUs by splitting input data automatically describes this behavior correctly; others describe unrelated tasks.

3. Final Answer:

   To run a model on multiple GPUs by splitting input data automatically -> Option B

4. Quick Check:

   DataParallel = multi-GPU input splitting [OK]

Hint: DataParallel splits input across GPUs automatically [OK]

Common Mistakes:

• Thinking DataParallel changes model architecture
• Confusing DataParallel with CPU-only mode
• Believing it reduces model size

2. Which of the following is the correct way to wrap a PyTorch model model with DataParallel?

easy

A. model = torch.DataParallel(model)

B. model = torch.nn.parallel.DataParallel(model)

C. model = torch.nn.DataParallel(model)

D. model = torch.parallel.DataParallel(model)

Solution

1. Step 1: Recall correct import path

   DataParallel is in torch.nn, so correct usage is torch.nn.DataParallel.

2. Step 2: Check options

   model = torch.nn.DataParallel(model) matches the correct usage; others are incorrect or incomplete.

3. Final Answer:

   model = torch.nn.DataParallel(model) -> Option C

4. Quick Check:

   Correct import path = torch.nn.DataParallel [OK]

Hint: DataParallel is in torch.nn module [OK]

Common Mistakes:

• Using torch.DataParallel instead of torch.nn.DataParallel
• Confusing torch.nn.parallel with torch.nn
• Confusing with other torch submodules

3. Given the code below, what will be the output shape of output if input shape is (8, 3, 32, 32) and 2 GPUs are used?

import torch
import torch.nn as nn

model = nn.Linear(32*32*3, 10)
model = nn.parallel.DataParallel(model)
input = torch.randn(8, 3, 32, 32)
input = input.view(8, -1)
output = model(input)
print(output.shape)

medium

A. torch.Size([8, 10])

B. torch.Size([4, 10])

C. torch.Size([16, 10])

D. RuntimeError due to shape mismatch

Solution

1. Step 1: Understand input and model

   Input is reshaped to (8, 3072) and model outputs 10 features per input.

2. Step 2: DataParallel splits batch across 2 GPUs

   Batch of 8 splits into 2 batches of 4, processed in parallel, outputs combined back to (8, 10).

3. Final Answer:

   torch.Size([8, 10]) -> Option A

4. Quick Check:

   Output shape = batch size x output features = (8, 10) [OK]

Hint: Output batch size stays same after DataParallel [OK]

Common Mistakes:

• Thinking output batch size halves with GPUs
• Confusing input reshape with output shape
• Expecting runtime error due to DataParallel

4. What is the likely cause of this error when using DataParallel?

RuntimeError: Expected tensor for argument #1 'input' to have scalar type Float but got Double

Code snippet:

model = torch.nn.parallel.DataParallel(model)
input = torch.randn(8, 3, 32, 32, dtype=torch.double)
output = model(input)

medium

A. Input batch size is too large for GPUs

B. DataParallel requires input to be on CPU only

C. Model is not wrapped correctly with DataParallel

D. Input tensor dtype does not match model parameters dtype

Solution

1. Step 1: Analyze error message

   Error says input tensor is Double but model expects Float tensors.

2. Step 2: Check input and model dtypes

   Model parameters default to Float; input is Double, causing mismatch and error.

3. Final Answer:

   Input tensor dtype does not match model parameters dtype -> Option D

4. Quick Check:

   Dtype mismatch causes runtime error [OK]

Hint: Ensure input and model dtypes match before DataParallel [OK]

Common Mistakes:

• Assuming DataParallel only works on CPU
• Not checking tensor data types
• Thinking batch size causes dtype errors

5. You want to speed up training on 4 GPUs using DataParallel. Your model has a batch normalization layer. What should you be careful about to avoid incorrect training behavior?

hard

A. BatchNorm layers behave differently in DataParallel; use SyncBatchNorm or move to DistributedDataParallel

B. DataParallel automatically fixes BatchNorm issues, no action needed

C. Remove BatchNorm layers before using DataParallel

D. Increase batch size to avoid BatchNorm problems

Solution

1. Step 1: Understand BatchNorm behavior in DataParallel

   BatchNorm computes stats per device, which can cause inconsistent normalization across GPUs.

2. Step 2: Recommended solutions

   Use SyncBatchNorm to sync stats or switch to DistributedDataParallel for correct BatchNorm behavior.

3. Final Answer:

   BatchNorm layers behave differently in DataParallel; use SyncBatchNorm or move to DistributedDataParallel -> Option A

4. Quick Check:

   BatchNorm needs syncing or DDP for multi-GPU [OK]

Hint: BatchNorm needs syncing or DDP on multiple GPUs [OK]

Common Mistakes:

• Assuming DataParallel handles BatchNorm correctly by default
• Removing BatchNorm layers unnecessarily
• Only increasing batch size without syncing stats