Merge pull request #5 from dvgodoy/ch4

Ch4

README.md

@@ -15,11 +15,12 @@ You can easily **load the notebooks directly from GitHub** using Colab and run t
 
 You can go through the chapters already using the links below:
 
-- [Chapter 0](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter00.ipynb)
-- [Chapter 1](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter01.ipynb)
-- [Chapter 2](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter02.ipynb)
-- [Chapter 2.1](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter02.1.ipynb)
-- [Chapter 3](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter03.ipynb)
+- [Chapter 0 - Visualizing Gradient Descent](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter00.ipynb)
+- [Chapter 1 - A Simple Regression Problem](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter01.ipynb)
+- [Chapter 2 - Rethinking the Training Loop](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter02.ipynb)
+- [Chapter 2.1 - Going Classy](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter02.1.ipynb)
+- [Chapter 3 - A Simple Classification Problem](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter03.ipynb)
+- [Chapter 4 - Classifying Images](https://colab.research.google.com/github/dvgodoy/PyTorchStepByStep/blob/master/Chapter04.ipynb)
 
 ### Binder
 

config.py

@@ -17,14 +17,15 @@ def tensorboard_cleanup():
     2: ['plots', 'data_generation', 'data_preparation', 'model_configuration', 'model_training'],
     21: ['plots', 'data_generation', 'data_preparation', 'model_configuration', 'stepbystep'],
     3: ['plots', 'stepbystep'],
+    4: ['plots', 'stepbystep', 'data_generation'],
 }
 FILENAMES = {
     0: ['chapter0.py'],
     1: ['chapter1.py'],
     2: ['chapter2.py', 'simple_linear_regression.py', 'v0.py', 'v0.py', 'v0.py'],
     21: ['chapter2_1.py', 'simple_linear_regression.py', 'v2.py', '', 'v0.py'],
     3: ['chapter3.py', 'v0.py'],
-}
+    4: ['chapter4.py', 'v0.py', 'image_classification.py'],}
 
 try:
     host = os.environ['BINDER_SERVICE_HOST']
@@ -40,8 +41,8 @@ def tensorboard_cleanup():
 
 IS_LOCAL = (not IS_BINDER) and (not IS_COLAB)
 
-def download_to_colab(chapter):    
-    base_url = 'https://raw.githubusercontent.com/dvgodoy/PyTorchStepByStep/master/'
+def download_to_colab(chapter, branch='master'):    
+    base_url = 'https://raw.githubusercontent.com/dvgodoy/PyTorchStepByStep/{}/'.format(branch)
 
     folders = FOLDERS[chapter]
     filenames = FILENAMES[chapter]
@@ -53,10 +54,10 @@ def download_to_colab(chapter):
                 raise
 
         if len(filename):
-	        path = os.path.join(folder, filename)
-	        url = '{}{}'.format(base_url, path)
-	        r = requests.get(url, allow_redirects=True)
-	        open(path, 'wb').write(r.content)
+            path = os.path.join(folder, filename)
+            url = '{}{}'.format(base_url, path)
+            r = requests.get(url, allow_redirects=True)
+            open(path, 'wb').write(r.content)
 
     try:
         os.mkdir('runs')
@@ -75,18 +76,18 @@ def download_to_colab(chapter):
     </script>
     ''')
 
-def config_chapter0():
+def config_chapter0(branch='master'):
     if IS_COLAB:
         print('Downloading files from GitHub repo to Colab...')
-        download_to_colab(0)
+        download_to_colab(0, branch)
         print('Finished!')
 
-def config_chapter1():
+def config_chapter1(branch='master'):
     if IS_COLAB:
         print('Installing torchviz...')
         subprocess.run([sys.executable, '-m', 'pip', 'install', 'torchviz'])
         print('Downloading files from GitHub repo to Colab...')
-        download_to_colab(1)
+        download_to_colab(1, branch)
         print('Creating folders...')
         folders = ['data_preparation', 'model_configuration', 'model_training']
 
@@ -99,20 +100,26 @@ def config_chapter1():
                     raise
         print('Finished!')
 
-def config_chapter2():
+def config_chapter2(branch='master'):
+    if IS_COLAB:
+        print('Downloading files from GitHub repo to Colab...')
+        download_to_colab(2, branch)
+        print('Finished!')
+
+def config_chapter2_1(branch='master'):
     if IS_COLAB:
         print('Downloading files from GitHub repo to Colab...')
-        download_to_colab(2)
+        download_to_colab(21, branch)
         print('Finished!')
 
-def config_chapter2_1():
+def config_chapter3(branch='master'):
     if IS_COLAB:
         print('Downloading files from GitHub repo to Colab...')
-        download_to_colab(21)
+        download_to_colab(3, branch)
         print('Finished!')
 
-def config_chapter3():
+def config_chapter4(branch='master'):
     if IS_COLAB:
         print('Downloading files from GitHub repo to Colab...')
-        download_to_colab(3)
-        print('Finished!')
+        download_to_colab(4, branch)
+        print('Finished!')

data_generation/image_classification.py

@@ -0,0 +1,56 @@
+import numpy as np
+
+
+def gen_img(start, target, fill=1, img_size=10):
+    # Generates empty image
+    img = np.zeros((img_size, img_size), dtype=np.float)
+
+    start_row, start_col = None, None
+
+    if start > 0:
+        start_row = start
+    else:
+        start_col = np.abs(start)
+
+    if target == 0:
+        if start_row is None:
+            img[:, start_col] = fill
+        else:
+            img[start_row, :] = fill
+    else:
+        if start_col == 0:
+            start_col = 1
+
+        if target == 1:
+            if start_row is not None:
+                up = (range(start_row, -1, -1), 
+                      range(0, start_row + 1))
+            else:
+                up = (range(img_size - 1, start_col - 1, -1), 
+                      range(start_col, img_size))
+            img[up] = fill
+        else:
+            if start_row is not None:
+                down = (range(start_row, img_size, 1), 
+                        range(0, img_size - start_row))
+            else:
+                down = (range(0, img_size - 1 - start_col + 1), 
+                        range(start_col, img_size))
+            img[down] = fill
+
+    return 255 * img.reshape(1, img_size, img_size)
+
+
+def generate_dataset(img_size=10, n_images=100, binary=True, seed=17):
+    np.random.seed(seed)
+
+    starts = np.random.randint(-(img_size - 1), img_size, size=(n_images,))
+    targets = np.random.randint(0, 3, size=(n_images,))
+
+    images = np.array([gen_img(s, t, img_size=img_size) 
+                       for s, t in zip(starts, targets)], dtype=np.uint8)
+
+    if binary:
+        targets = (targets > 0).astype(np.int)
+
+    return images, targets

helpers.py

@@ -47,4 +47,35 @@ def perform_val_step(x, y):
         # There is no need to compute Steps 3 and 4, since we don't update parameters during evaluation
         return loss.item()
 
-    return perform_val_step
+    return perform_val_step
+
+def index_splitter(n, splits, seed=13):
+    idx = torch.arange(n)
+    # Makes the split argument a tensor
+    splits_tensor = torch.as_tensor(splits)
+    # Finds the correct multiplier, so we don't have
+    # to worry about summing up to N (or one)
+    multiplier = n / splits_tensor.sum()    
+    splits_tensor = (multiplier * splits_tensor).long()
+    # If there is a difference, throws at the first split
+    # so random_split does not complain
+    diff = n - splits_tensor.sum()
+    splits_tensor[0] += diff
+    # Uses PyTorch random_split to split the indices
+    torch.manual_seed(seed)
+    return random_split(idx, splits_tensor)
+
+def make_balanced_sampler(y):
+    # Computes weights for compensating imbalanced classes
+    classes, counts = y.unique(return_counts=True)
+    weights = 1.0 / counts.float()
+    sample_weights = weights[y.squeeze().long()]
+    # Builds sampler with compute weights
+    generator = torch.Generator()
+    sampler = WeightedRandomSampler(
+        weights=sample_weights,
+        num_samples=len(sample_weights),
+        generator=generator,
+        replacement=True
+    )
+    return sampler