In [1]:
import os
import numpy as np
import matplotlib.pyplot as plt
import pandas as pd
import seaborn as sns
import torch
import torch.nn as nn
from torch.utils.data import DataLoader
from torchvision import transforms, datasets
import torch.nn.functional as F
from sklearn.metrics import confusion_matrix
from tqdm import tqdm
Download dataset¶
In [2]:
ROOT_DATA_DIR = "FashionMNISTDir"
train_data = datasets.FashionMNIST(
root = ROOT_DATA_DIR,
train = True,
download = True,
transform = transforms.ToTensor()
)
test_data = datasets.FashionMNIST(
root = ROOT_DATA_DIR,
train = False, ## <<< Test data
download = True,
transform = transforms.ToTensor()
)
In [3]:
train_data.data.shape
Out[3]:
torch.Size([60000, 28, 28])
In [4]:
test_data.data.shape
Out[4]:
torch.Size([10000, 28, 28])
In [5]:
label_map = {
0: 'T-shirt/top',
1: 'Trouser',
2: 'Pullover',
3:' Dress',
4: 'Coat',
5: 'Sandal',
6: 'Shirt',
7: 'Sneaker',
8: 'Bag',
9: 'Ankle boot',
}
Visualize one sample¶
In [6]:
def view_sample_img(data, index, label_map):
plt.imshow(data.data[index], cmap="gray")
plt.title(f"data label: {label_map[data.targets[index].item()]}")
plt.axis("off")
In [7]:
view_sample_img(train_data, index=1, label_map=label_map)
Create the dataloader¶
In [8]:
BATCH_SIZE = 64
train_data_loader = DataLoader(
dataset = train_data,
batch_size = BATCH_SIZE,
shuffle = True
)
test_data_loader = DataLoader(
dataset = test_data,
batch_size = BATCH_SIZE,
shuffle = True
)
In [9]:
for data, label in test_data_loader:
print(data.shape)
print(label.shape)
break
torch.Size([64, 1, 28, 28]) torch.Size([64])
In [10]:
device = "cuda" if torch.cuda.is_available() else "cpu"
device
Out[10]:
'cuda'
In [11]:
class CNN(nn.Module):
def __init__(self, in_, out_):
super(CNN, self).__init__()
self.conv_pool_01 = nn.Sequential(
nn.Conv2d(in_channels=in_, out_channels=8, kernel_size=5, stride=1, padding=0),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
)
self.conv_pool_02 = nn.Sequential(
nn.Conv2d(in_channels=8, out_channels=16, kernel_size=5, stride=1, padding=0),
nn.ReLU(),
nn.MaxPool2d(kernel_size=2, stride=2)
)
self.Flatten = nn.Flatten()
self.FC_01 = nn.Linear(in_features=16*4*4, out_features=128)
self.FC_02 = nn.Linear(in_features=128, out_features=64)
self.FC_03 = nn.Linear(in_features=64, out_features=out_)
def forward(self, x):
x = self.conv_pool_01(x)
x = self.conv_pool_02(x)
x = self.Flatten(x)
x = self.FC_01(x)
x = self.FC_02(x)
x = self.FC_03(x)
return x
In [12]:
model = CNN(1, 10)
model.to(device)
Out[12]:
CNN(
(conv_pool_01): Sequential(
(0): Conv2d(1, 8, kernel_size=(5, 5), stride=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(conv_pool_02): Sequential(
(0): Conv2d(8, 16, kernel_size=(5, 5), stride=(1, 1))
(1): ReLU()
(2): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)
)
(Flatten): Flatten(start_dim=1, end_dim=-1)
(FC_01): Linear(in_features=256, out_features=128, bias=True)
(FC_02): Linear(in_features=128, out_features=64, bias=True)
(FC_03): Linear(in_features=64, out_features=10, bias=True)
)
Count no. of trainable params¶
In [13]:
def count_params(model):
model_params = {"Modules": list(), "Parameters": list()}
total = 0
for name, parameters in model.named_parameters():
if not parameters.requires_grad:
continue
param = parameters.numel()
model_params["Modules"].append(name)
model_params["Parameters"].append(param)
total += param
df = pd.DataFrame(model_params)
df = df.style.set_caption(f"Total trainable parameters: {total}")
return df
count_params(model)
Out[13]:
| Modules | Parameters | |
|---|---|---|
| 0 | conv_pool_01.0.weight | 200 |
| 1 | conv_pool_01.0.bias | 8 |
| 2 | conv_pool_02.0.weight | 3200 |
| 3 | conv_pool_02.0.bias | 16 |
| 4 | FC_01.weight | 32768 |
| 5 | FC_01.bias | 128 |
| 6 | FC_02.weight | 8192 |
| 7 | FC_02.bias | 64 |
| 8 | FC_03.weight | 640 |
| 9 | FC_03.bias | 10 |
In [14]:
next(model.parameters()).is_cuda
Out[14]:
True
Traning loop¶
In [15]:
learning_rate = 0.001
num_epochs = 20
In [16]:
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
In [17]:
n_total_steps = len(train_data_loader)
n_total_steps
Out[17]:
938
In [18]:
60000/BATCH_SIZE
Out[18]:
937.5
In [19]:
for epoch in range(num_epochs):
with tqdm(train_data_loader) as tqdm_epoch:
for images, labels in tqdm_epoch:
tqdm_epoch.set_description(f"Epoch {epoch + 1}/{num_epochs}")
images = images.to(device)
labels = labels.to(device)
# forward pass
outputs = model(images)
loss = criterion(outputs, labels)
# backward prop
optimizer.zero_grad()
loss.backward()
optimizer.step()
tqdm_epoch.set_postfix(loss=loss.item())
Epoch 1/20: 100%|██████████| 938/938 [00:20<00:00, 46.28it/s, loss=0.405] Epoch 2/20: 100%|██████████| 938/938 [00:14<00:00, 62.66it/s, loss=0.461] Epoch 3/20: 100%|██████████| 938/938 [00:15<00:00, 58.97it/s, loss=0.279] Epoch 4/20: 100%|██████████| 938/938 [00:15<00:00, 59.80it/s, loss=0.144] Epoch 5/20: 100%|██████████| 938/938 [00:15<00:00, 61.09it/s, loss=0.336] Epoch 6/20: 100%|██████████| 938/938 [00:15<00:00, 60.48it/s, loss=0.443] Epoch 7/20: 100%|██████████| 938/938 [00:15<00:00, 60.35it/s, loss=0.334] Epoch 8/20: 100%|██████████| 938/938 [00:15<00:00, 60.15it/s, loss=0.413] Epoch 9/20: 100%|██████████| 938/938 [00:15<00:00, 59.85it/s, loss=0.209] Epoch 10/20: 100%|██████████| 938/938 [00:15<00:00, 60.09it/s, loss=0.228] Epoch 11/20: 100%|██████████| 938/938 [00:15<00:00, 58.82it/s, loss=0.212] Epoch 12/20: 100%|██████████| 938/938 [00:15<00:00, 59.74it/s, loss=0.0203] Epoch 13/20: 100%|██████████| 938/938 [00:15<00:00, 59.96it/s, loss=0.437] Epoch 14/20: 100%|██████████| 938/938 [00:15<00:00, 60.11it/s, loss=0.38] Epoch 15/20: 100%|██████████| 938/938 [00:15<00:00, 60.20it/s, loss=0.292] Epoch 16/20: 100%|██████████| 938/938 [00:15<00:00, 59.15it/s, loss=0.177] Epoch 17/20: 100%|██████████| 938/938 [00:16<00:00, 58.12it/s, loss=0.381] Epoch 18/20: 100%|██████████| 938/938 [00:16<00:00, 57.91it/s, loss=0.173] Epoch 19/20: 100%|██████████| 938/938 [00:16<00:00, 57.57it/s, loss=0.276] Epoch 20/20: 100%|██████████| 938/938 [00:16<00:00, 57.84it/s, loss=0.233]
In [20]:
os.getcwd()
Out[20]:
'd:\\oneNeuron\\Pytorch\\Pytorch-basics\\codebase'
In [21]:
## save trained model -
os.makedirs("06_03_session_dir", exist_ok=True)
modle_file = os.path.join("06_03_session_dir", 'CNN_model.pth')
torch.save(model, modle_file)
In [ ]: