use resnet50 to train a cifar10 classifier

.gitignore

@@ -1,2 +1,4 @@
 data
 *.pt
+.DS_Store
+__pycache__

README.md

@@ -1,6 +1,6 @@
 # Abstract
 
-Attempt to use [DVC](https://dvc.ai/), a data versioning tool, to track model training with PyTorch, including data, trained model file, and used parameters. The data will be recorded and pushed to my private DVC remote via webdav🎁
+Attempt to use [DVC](https://dvc.ai/), a data versioning tool, to track image classification model training with PyTorch, including data, trained model file, and used parameters. The data will be recorded and pushed to my private DVC remote via webdav🎁
 
 # Requirements
 
@@ -11,6 +11,8 @@ Attempt to use [DVC](https://dvc.ai/), a data versioning tool, to track model tr
 * **env**
   * **pt.yaml**
     * conda env yaml to run this repo
+* **utils**
+  * house pre-built functions
 
 # Files
 

dvc.lock

@@ -0,0 +1,48 @@
+schema: '2.0'
+stages:
+  prepare:
+    cmd: python prepare.py
+    deps:
+    - path: prepare.py
+      hash: md5
+      md5: a1c07d1d5caf6e5288560a189415785c
+      size: 2979
+    params:
+      params.yaml:
+        prepare:
+          data_dir: data/raw
+          save_dir: data/processed
+          train_valid_split:
+          - 0.7
+          - 0.3
+          random_seed: 0
+    outs:
+    - path: data/processed
+      hash: md5
+      md5: f4bf62ffa725ca9144b7852a283dc1da.dir
+      size: 295118798
+      nfiles: 60000
+  train:
+    cmd: python train.py
+    deps:
+    - path: data/processed
+      hash: md5
+      md5: f4bf62ffa725ca9144b7852a283dc1da.dir
+      size: 295118798
+      nfiles: 60000
+    - path: train.py
+      hash: md5
+      md5: b797ccf2fe61952bbf6d83fa51b0b11f
+      size: 3407
+    params:
+      params.yaml:
+        train:
+          data_dir: data/processed
+          epochs: 5
+          batch_size: 256
+          learning_rate: 5e-05
+    outs:
+    - path: model.pt
+      hash: md5
+      md5: 8ead2a7cd52d70b359d3cdc3df5e43e3
+      size: 102592994

dvc.yaml

@@ -2,27 +2,42 @@ stages:
   prepare:
     cmd: python prepare.py
     deps:
-      - prepare.py
+    - prepare.py
     params:
-      - prepare
+    - prepare
     outs:
-      - data/processed
+    - data/processed
   train:
     cmd: python train.py
     deps:
-      - data/processed
-      - train.py
+    - data/processed
+    - train.py
     params:
-      - train
+    - train
     outs:
-      - model.pt
+    - model.pt
   evaluate:
     cmd: python evaluate.py
     deps:
-      - data/processed
-      - evaluate.py
-      - model.pt
+    - data/processed
+    - evaluate.py
+    - model.pt
     params:
-      - evaluate
+    - evaluate
     outs:
-      - eval
+    - eval
+params:
+- dvclive/train/params.yaml
+- dvclive/eval/params.yaml
+artifacts:
+  resnet50:
+    path: model.pt
+    type: model
+metrics:
+- dvclive/train/metrics.json
+- dvclive/eval/metrics.json
+plots:
+- dvclive/train/plots/metrics:
+    x: step
+- dvclive/eval/plots/metrics:
+    x: step

dvclive/eval/metrics.json

@@ -0,0 +1,3 @@
+{
+    "test_acc": 0.7336928844451904
+}

dvclive/eval/params.yaml

@@ -0,0 +1 @@
+data_dir: data/processed

dvclive/eval/plots/metrics/test_acc.tsv

@@ -0,0 +1,2 @@
+step	test_acc
+0	0.7336928844451904

dvclive/eval/report.md

@@ -0,0 +1,15 @@
+# DVC Report
+
+params.yaml
+
+| data_dir       |
+|----------------|
+| data/processed |
+
+metrics.json
+
+|   test_acc |
+|------------|
+|   0.733693 |
+
+![static/test_acc](static/test_acc.png)

dvclive/train/metrics.json

@@ -0,0 +1,7 @@
+{
+    "train_loss": 2.2422571182250977,
+    "train_acc": 0.7347080707550049,
+    "valid_loss": 2.3184266090393066,
+    "valid_acc": 0.7381500005722046,
+    "step": 4
+}

dvclive/train/params.yaml

@@ -0,0 +1,5 @@
+data_dir: data/processed
+epochs: 5
+batch_size: 256
+learning_rate: 5e-05
+epoch: 4

dvclive/train/plots/metrics/train_acc.tsv

@@ -0,0 +1,6 @@
+step	train_acc
+0	0.6712241768836975
+1	0.6976224184036255
+2	0.7157850861549377
+3	0.7277812957763672
+4	0.7347080707550049

dvclive/train/plots/metrics/train_loss.tsv

@@ -0,0 +1,6 @@
+step	train_loss
+0	3.0726168155670166
+1	2.7409346103668213
+2	2.5224294662475586
+3	2.364570140838623
+4	2.2422571182250977

dvclive/train/plots/metrics/valid_acc.tsv

@@ -0,0 +1,6 @@
+step	valid_acc
+0	0.6918894052505493
+1	0.7131190896034241
+2	0.7261338233947754
+3	0.7339118123054504
+4	0.7381500005722046

dvclive/train/plots/metrics/valid_loss.tsv

@@ -0,0 +1,6 @@
+step	valid_loss
+0	2.890321969985962
+1	2.669679880142212
+2	2.5183584690093994
+3	2.4061686992645264
+4	2.3184266090393066

dvclive/train/report.md

@@ -0,0 +1,21 @@
+# DVC Report
+
+params.yaml
+
+| data_dir       |   epochs |   batch_size |   learning_rate |   epoch |
+|----------------|----------|--------------|-----------------|---------|
+| data/processed |        5 |          256 |           5e-05 |       4 |
+
+metrics.json
+
+|   train_loss |   train_acc |   valid_loss |   valid_acc |   step |
+|--------------|-------------|--------------|-------------|--------|
+|      2.24226 |    0.734708 |      2.31843 |     0.73815 |      4 |
+
+![static/valid_loss](static/valid_loss.png)
+
+![static/train_acc](static/train_acc.png)
+
+![static/valid_acc](static/valid_acc.png)
+
+![static/train_loss](static/train_loss.png)

env/pt.yaml

@@ -21,9 +21,11 @@ dependencies:
   - billiard=4.1.0
   - boto3=1.34.9
   - botocore=1.34.9
+  - brotli=1.1.0
+  - brotli-bin=1.1.0
   - brotli-python=1.1.0
   - bzip2=1.0.8
-  - ca-certificates=2023.08.22
+  - ca-certificates=2023.11.17
   - cairo=1.18.0
   - celery=5.3.4
   - certifi=2023.11.17
@@ -35,7 +37,9 @@ dependencies:
   - click-repl=0.3.0
   - colorama=0.4.6
   - configobj=5.0.8
+  - contourpy=1.2.0
   - cryptography=41.0.7
+  - cycler=0.12.1
   - dav1d=1.2.1
   - decorator=5.1.1
   - dictdiffer=0.9.0
@@ -64,6 +68,7 @@ dependencies:
   - fontconfig=2.14.2
   - fonts-conda-ecosystem=1
   - fonts-conda-forge=1
+  - fonttools=4.47.0
   - freetype=2.12.1
   - fribidi=1.0.10
   - frozenlist=1.4.1
@@ -94,6 +99,7 @@ dependencies:
   - jinja2=3.1.2
   - jmespath=1.0.1
   - joblib=1.2.0
+  - kiwisolver=1.4.5
   - kombu=5.3.4
   - krb5=1.21.2
   - lame=3.100
@@ -101,6 +107,9 @@ dependencies:
   - lerc=4.0.0
   - libass=0.17.1
   - libblas=3.9.0
+  - libbrotlicommon=1.1.0
+  - libbrotlidec=1.1.0
+  - libbrotlienc=1.1.0
   - libcblas=3.9.0
   - libcxx=16.0.6
   - libdeflate=1.19
@@ -131,14 +140,18 @@ dependencies:
   - libxcb=1.15
   - libxml2=2.12.3
   - libzlib=1.2.13
+  - lightning-utilities=0.10.0
   - llvm-openmp=17.0.6
   - markdown-it-py=3.0.0
   - markupsafe=2.1.3
+  - matplotlib=3.8.2
+  - matplotlib-base=3.8.2
   - mdurl=0.1.0
   - mpc=1.3.1
   - mpfr=4.2.1
   - mpmath=1.3.0
   - multidict=6.0.4
+  - munkres=1.1.4
   - nanotime=0.5.2
   - ncurses=6.4
   - nettle=3.9.1
@@ -206,12 +219,15 @@ dependencies:
   - tk=8.6.13
   - tomlkit=0.12.3
   - torchaudio=2.1.2
+  - torchmetrics=1.2.1
   - torchvision=0.16.2
+  - tornado=6.3.3
   - tqdm=4.66.1
   - typer=0.9.0
   - typing-extensions=4.9.0
   - typing_extensions=4.9.0
   - tzdata=2023d
+  - unicodedata2=15.1.0
   - urllib3=1.26.18
   - vine=5.0.0
   - voluptuous=0.14.1

evaluate.py

@@ -3,9 +3,15 @@
 # author: deng
 # date  : 20231228
 
+from pathlib import Path
 import yaml
 
 import torch
+from torch.utils.data import DataLoader
+from torchmetrics.classification import MulticlassAccuracy
+from dvclive import Live
+
+from utils.dataset import ProcessedDataset
 
 
 def evaluate(params_path: str = 'params.yaml') -> None:
@@ -17,6 +23,30 @@ def evaluate(params_path: str = 'params.yaml') -> None:
 
     with open(params_path, 'r') as f:
         params = yaml.safe_load(f)
+    data_dir = Path(params['evaluate']['data_dir'])
+
+    test_dataset = ProcessedDataset(dataset_dir=data_dir.joinpath('test'))
+    test_dataloader = DataLoader(test_dataset, batch_size=64)
+
+    device = torch.device('mps')
+    net = torch.load('model.pt')
+    net.to(device)
+    net.eval()
+
+    metric = MulticlassAccuracy(num_classes=10, top_k=1, average='weighted')
+    metric.to(device)
+
+    with Live(dir='dvclive/eval', report='md') as live:
+        live.log_params(params['evaluate'])
+
+        for data in test_dataloader:
+            inputs, labels = data[0].to(device), data[1].to(device)
+            outputs = net(inputs)
+            _ = metric(outputs, labels)
+        test_acc = metric.compute()
+
+        print(f'test_acc:{test_acc}')
+        live.log_metric('test_acc', float(test_acc.cpu()))
 
 
 if __name__ == '__main__':

params.yaml

@@ -1,3 +1,12 @@
 prepare:
+  data_dir: data/raw
+  save_dir: data/processed
+  train_valid_split: [0.7, 0.3]
+  random_seed: 0
 train:
+  data_dir: data/processed
+  epochs: 5
+  batch_size: 256
+  learning_rate: 0.00005
 evaluate:
+  data_dir: data/processed

prepare.py

@@ -3,9 +3,13 @@
 # author: deng
 # date  : 20231228
 
+from pathlib import Path
+from shutil import rmtree
+import random
+import pickle
 import yaml
 
-import torch
+import numpy as np
 
 
 def prepare(params_path: str = 'params.yaml') -> None:
@@ -17,7 +21,75 @@ def prepare(params_path: str = 'params.yaml') -> None:
 
     with open(params_path, 'r') as f:
         params = yaml.safe_load(f)
+    data_dir = Path(params['prepare']['data_dir'])
+    save_dir = Path(params['prepare']['save_dir'])
+    train_valid_split = params['prepare']['train_valid_split']
+    random_seed = params['prepare']['random_seed']
+
+    train_dir = save_dir.joinpath('train')
+    valid_dir = save_dir.joinpath('valid')
+    test_dir = save_dir.joinpath('test')
+
+    if train_dir.is_dir():
+        rmtree(train_dir)
+    train_dir.mkdir()
+    if valid_dir.is_dir():
+        rmtree(valid_dir)
+    valid_dir.mkdir()
+    if test_dir.is_dir():
+        rmtree(test_dir)
+    test_dir.mkdir()
+
+    # Process training data
+    ids = list(range(50000))
+    random.Random(random_seed).shuffle(ids)
+    train_ids = ids[:int(50000 * train_valid_split[0])]
+    valid_ids = ids[int(50000 * train_valid_split[0]):]
+
+    current_id, train_count, valid_count = 0, 0, 0
+    cifar_10_dir = data_dir.joinpath('cifar-10-batches-py')
+    for data_path in cifar_10_dir.glob('data_batch_*'):
+        with open(data_path, 'rb') as f:
+            data = pickle.load(f, encoding='bytes')
+            for i, label in enumerate(data[b'labels']):
+                x = data[b'data'][i]
+                x = x.reshape(3, 32, 32)
+                x = x / 255
+                x = x.astype(np.float32)  # mps does not support float64
+                y = np.zeros(10, dtype=np.float32)
+                y[label] = 1.
+                if current_id in train_ids:
+                    npz_path = train_dir.joinpath(f'{train_count}.npz')
+                    train_ids.remove(current_id)
+                    train_count += 1
+                else:
+                    npz_path = valid_dir.joinpath(f'{valid_count}.npz')
+                    valid_ids.remove(current_id)
+                    valid_count += 1
+                np.savez_compressed(
+                    npz_path,
+                    x=x,
+                    y=y
+                )
+                current_id += 1
+
+    # Process testing data
+    data_path = cifar_10_dir.joinpath('test_batch')
+    with open(data_path, 'rb') as f:
+        data = pickle.load(f, encoding='bytes')
+        for i, label in enumerate(data[b'labels']):
+            x = data[b'data'][i]
+            x = x.reshape(3, 32, 32)
+            x = x / 255
+            x = x.astype(np.float32)
+            y = np.zeros(10, dtype=np.float32)
+            y[label] = 1.
+            np.savez_compressed(
+                save_dir.joinpath('test', f'{i}.npz'),
+                x=x,
+                y=y
+            )
 
 
 if __name__ == '__main__':
-    prepare()
+    prepare('params.yaml')

train.py

@@ -3,21 +3,97 @@
 # author: deng
 # date  : 20231228
 
+from pathlib import Path
 import yaml
 
 import torch
+from rich.progress import track
+from torch.utils.data import DataLoader
+from torchvision.models import resnet50
+from torchmetrics.classification import MulticlassAccuracy
 from dvclive import Live
 
+from utils.dataset import ProcessedDataset
+
 
 def train(params_path: str = 'params.yaml') -> None:
     """Train a simple model using Pytorch
 
     Args:
-        params_path (str, optional): path of config yaml. Defaults to 'params.yaml'.
+        params_path (str, optional): path of parameter yaml. Defaults to 'params.yaml'.
     """
 
     with open(params_path, 'r') as f:
         params = yaml.safe_load(f)
+    data_dir = Path(params['train']['data_dir'])
+    epochs = params['train']['epochs']
+    batch_size = params['train']['batch_size']
+    learning_rate = params['train']['learning_rate']
+
+    train_dataset = ProcessedDataset(dataset_dir=data_dir.joinpath('train'))
+    valid_dataset = ProcessedDataset(dataset_dir=data_dir.joinpath('valid'))
+
+    train_dataloader = DataLoader(train_dataset, batch_size=batch_size, num_workers=2)
+    valid_dataloader = DataLoader(valid_dataset, batch_size=batch_size, num_workers=2)
+
+    device = torch.device('mps')
+    net = torch.nn.Sequential(
+        resnet50(weights='IMAGENET1K_V1'),
+        torch.nn.Linear(1000, 10)
+    )
+    net.to(device)
+
+    criterion = torch.nn.CrossEntropyLoss()
+    optimizer = torch.optim.SGD(net.parameters(), lr=learning_rate)
+    metric = MulticlassAccuracy(num_classes=10, top_k=1, average='weighted')
+    metric.to(device)
+
+    with Live(dir='dvclive/train', report='md') as live:
+        live.log_params(params['train'])
+
+        for epoch in track(range(epochs), 'Epochs'):
+
+            net.train()
+            train_loss = 0.
+            for data in train_dataloader:
+                inputs, labels = data[0].to(device), data[1].to(device)
+                optimizer.zero_grad()
+                outputs = net(inputs)
+                loss = criterion(outputs, labels)
+                train_loss += loss
+                loss.backward()
+                optimizer.step()
+                _ = metric(outputs, labels)
+            train_loss /= len(train_dataloader)
+            train_acc = metric.compute()
+            metric.reset()
+
+            net.eval()
+            with torch.no_grad():
+                valid_loss = 0.
+                for data in valid_dataloader:
+                    inputs, labels = data[0].to(device), data[1].to(device)
+                    outputs = net(inputs)
+                    loss = criterion(outputs, labels)
+                    valid_loss += loss
+                    _ = metric(outputs, labels)
+                valid_loss /= len(valid_dataloader)
+                valid_acc = metric.compute()
+                metric.reset()
+
+            print(
+                f'Epoch {epoch} - train_loss:{train_loss} train_acc:{train_acc} '
+                f'valid_loss:{valid_loss} valid_acc:{valid_acc}'
+            )
+            live.log_param('epoch', epoch)
+            live.log_metric('train_loss', float(train_loss.cpu()))
+            live.log_metric('train_acc', float(train_acc.cpu()))
+            live.log_metric('valid_loss', float(valid_loss.cpu()))
+            live.log_metric('valid_acc', float(valid_acc.cpu()))
+            live.next_step()
+
+        torch.save(net, 'model.pt')
+        live.log_artifact('model.pt', type='model', name='resnet50')
 
 
 if __name__ == '__main__':

utils/dataset.py

@@ -0,0 +1,26 @@
+# prepare.py
+#
+# author: deng
+# date  : 20231229
+
+from pathlib import PosixPath
+
+import torch
+import numpy as np
+from torch.utils.data import Dataset
+
+
+class ProcessedDataset(Dataset):
+    """"Load processed data"""
+    def __init__(self, dataset_dir: PosixPath):
+        self.dataset_dir = dataset_dir
+        self.file_paths = list(self.dataset_dir.glob('*.npz'))
+
+    def __len__(self):
+        return len(self.file_paths)
+
+    def __getitem__(self, idx):
+        npz = np.load(self.file_paths[idx])
+        x = torch.from_numpy(npz['x'])
+        y = torch.from_numpy(npz['y'])
+        return x, y