Merge in HW 3

deepul/hw3_helper.py

@@ -88,10 +88,108 @@ def q1_save_results(part, dset_id, fn):
                     fname=f'results/q1_{part}_dset{dset_id}_sample_without_noise.png')
 
 
+def visualize_colored_shapes():
+    data_dir = get_data_dir(3)
+    train_data, test_data = load_pickled_data(join(data_dir, 'shapes_colored.pkl'))
+    idxs = np.random.choice(len(train_data), replace=False, size=(100,))
+    images = train_data[idxs]
+    show_samples(images, title='Colored Shapes Samples')
+
+
+def visualize_svhn():
+    data_dir = get_data_dir(3)
+    train_data, test_data = load_pickled_data(join(data_dir, 'svhn.pkl'))
+    idxs = np.random.choice(len(train_data), replace=False, size=(100,))
+    images = train_data[idxs]
+    show_samples(images, title='SVHN Samples')
+
+
+def visualize_cifar10():
+    data_dir = get_data_dir(3)
+    train_data, test_data = load_pickled_data(join(data_dir, 'cifar10.pkl'))
+    idxs = np.random.choice(len(train_data), replace=False, size=(100,))
+    images = train_data[idxs]
+    show_samples(images, title='CIFAR10 Samples')
+
+
+def q2_save_results(part, dset_id, fn):
+    assert part in ['a', 'b'] and dset_id in [1, 2]
+    data_dir = get_data_dir(3)
+    if dset_id == 1:
+        train_data, test_data = load_pickled_data(join(data_dir, 'svhn.pkl'))
+    else:
+        train_data, test_data = load_pickled_data(join(data_dir, 'cifar10.pkl'))
 
+    train_losses, test_losses, samples, reconstructions, interpolations = fn(train_data, test_data, dset_id)
+    samples, reconstructions, interpolations = samples.astype('float32'), reconstructions.astype('float32'), interpolations.astype('float32')
+    print(f'Final -ELBO: {test_losses[-1, 0]:.4f}, Recon Loss: {test_losses[-1, 1]:.4f}, '
+          f'KL Loss: {test_losses[-1, 2]:.4f}')
+    plot_vae_training_plot(train_losses, test_losses, f'Q2({part}) Dataset {dset_id} Train Plot',
+                           f'results/q2_{part}_dset{dset_id}_train_plot.png')
+    show_samples(samples, title=f'Q2({part}) Dataset {dset_id} Samples',
+                 fname=f'results/q2_{part}_dset{dset_id}_samples.png')
+    show_samples(reconstructions, title=f'Q2({part}) Dataset {dset_id} Reconstructions',
+                 fname=f'results/q2_{part}_dset{dset_id}_reconstructions.png')
+    show_samples(interpolations, title=f'Q2({part}) Dataset {dset_id} Interpolations',
+                 fname=f'results/q2_{part}_dset{dset_id}_interpolations.png')
 
 
-
-
-
-
+def q3_save_results(dset_id, fn):
+    assert dset_id in [1, 2]
+    data_dir = get_data_dir(3)
+    if dset_id == 1:
+        train_data, test_data = load_pickled_data(join(data_dir, 'svhn.pkl'))
+    else:
+        train_data, test_data = load_pickled_data(join(data_dir, 'cifar10.pkl'))
+
+    vqvae_train_losses, vqvae_test_losses, pixelcnn_train_losses, pixelcnn_test_losses, samples, reconstructions = fn(train_data, test_data, dset_id)
+    samples, reconstructions = samples.astype('float32'), reconstructions.astype('float32')
+    print(f'VQ-VAE Final Test Loss: {vqvae_test_losses[-1]:.4f}')
+    print(f'PixelCNN Prior Final Test Loss: {pixelcnn_test_losses[-1]:.4f}')
+    save_training_plot(vqvae_train_losses, vqvae_test_losses,f'Q3 Dataset {dset_id} VQ-VAE Train Plot',
+                       f'results/q3_dset{dset_id}_vqvae_train_plot.png')
+    save_training_plot(pixelcnn_train_losses, pixelcnn_test_losses,f'Q3 Dataset {dset_id} PixelCNN Prior Train Plot',
+                       f'results/q3_dset{dset_id}_pixelcnn_train_plot.png')
+    show_samples(samples, title=f'Q3 Dataset {dset_id} Samples',
+                 fname=f'results/q3_dset{dset_id}_samples.png')
+    show_samples(reconstructions, title=f'Q3 Dataset {dset_id} Reconstructions',
+                 fname=f'results/q3_dset{dset_id}_reconstructions.png')
+
+
+def q4_a_save_results(dset_id, fn):
+    assert dset_id in [1, 2]
+    data_dir = get_data_dir(3)
+    if dset_id == 1:
+        train_data, test_data = load_pickled_data(join(data_dir, 'svhn.pkl'))
+    else:
+        train_data, test_data = load_pickled_data(join(data_dir, 'cifar10.pkl'))
+
+    vqvae_train_losses, vqvae_test_losses, pixelcnn_train_losses, pixelcnn_test_losses, samples, reconstructions = fn(train_data, test_data, dset_id)
+    samples, reconstructions = samples.astype('float32'), reconstructions.astype('float32')
+    print(f'VQ-VAE Final Test Loss: {vqvae_test_losses[-1]:.4f}')
+    print(f'PixelCNN Prior Final Test Loss: {pixelcnn_test_losses[-1]:.4f}')
+    save_training_plot(vqvae_train_losses, vqvae_test_losses,f'Q4(a) Dataset {dset_id} VQ-VAE Train Plot',
+                       f'results/q4_a_dset{dset_id}_vqvae_train_plot.png')
+    save_training_plot(pixelcnn_train_losses, pixelcnn_test_losses,f'Q4(a) Dataset {dset_id} PixelCNN Prior Train Plot',
+                       f'results/q4_a_dset{dset_id}_pixelcnn_train_plot.png')
+    show_samples(samples, title=f'Q4(a) Dataset {dset_id} Samples',
+                 fname=f'results/q4_a_dset{dset_id}_samples.png')
+    show_samples(reconstructions, title=f'Q4(a) Dataset {dset_id} Reconstructions',
+                 fname=f'results/q4_a_dset{dset_id}_reconstructions.png')
+
+
+def q4_b_save_results(fn):
+    part = 'b'
+    data_dir = get_data_dir(3)
+    train_data, test_data = load_pickled_data(join(data_dir, 'mnist.pkl'))
+
+    train_losses, test_losses, samples, reconstructions = fn(train_data, test_data)
+    samples, reconstructions = samples.astype('float32') * 255, reconstructions.astype('float32') * 255
+    print(f'Final -ELBO: {test_losses[-1, 0]:.4f}, Recon Loss: {test_losses[-1, 1]:.4f}, '
+          f'KL Loss: {test_losses[-1, 2]:.4f}')
+    plot_vae_training_plot(train_losses, test_losses, f'Q4({part}) Train Plot',
+                           f'results/q4_{part}_train_plot.png')
+    show_samples(samples, title=f'Q4({part}) Samples',
+                 fname=f'results/q4_{part}_samples.png')
+    show_samples(reconstructions, title=f'Q4({part}) Reconstructions',
+                 fname=f'results/q4_{part}_reconstructions.png')

deepul/utils.py

@@ -99,4 +99,9 @@ def load_pickled_data(fname, include_labels=False):
 
 
 def get_data_dir(hw_number):
-    return join('deepul', 'homeworks', f'hw{hw_number}', 'data')
+    return join('deepul', 'homeworks', f'hw{hw_number}', 'data')
+
+
+def quantize(images, n_bits):
+    images = np.floor(images / 256. * 2 ** n_bits)
+    return images.astype('uint8')

homeworks/hw3/README.md

@@ -0,0 +1,15 @@
+# Homework 3: Latent Variable Models
+
+**Due March 10, 11:59pm on Gradescope**
+
+In this homework, you will implement and train a variety of different latent variable models models. The homework consists of two components: 
+* `hw3_notebook.ipynb` : Contains all of the coding questions, and will automatically generate and display results for you after completing each question. 
+You will submit the notebook to Gradescope after completing the homework. 
+Open it on Colab by clicking on the file, and then "Open in Colab" at the top. 
+**Submit a PDF version of the notebook to the code (Print -> Preview -> Save) on Gradescope in the assignment with (code)**
+* `hw3_latex` :  Contains LaTeX files and figures needed to generate your PDF submission to Gradescope. Copy the images saved from the notebook into the `figures/` folder and fill out the test losses.
+**Submit the Latex PDF in the assignment with (PDF)**
+
+You can open the notebook in Google Colab to get access to a free GPU, or you can link Colab to a local runtime to run it on your own GPU.  
+
+**Note that for this homework, the data files are not in the git repo. If you would like to download them, you can use this [link](https://drive.google.com/open?id=1lWjGICwgzgcBDejo9S5g69hLAf0O3lGF)