\nA world class football coach may have a top secret playbook, mind blowing creativity and seemingly psychic intuition, skills which are learned through study, practice and experience. If these skills can be learned by a human, could they be learned by a machine?\n
\nThe subject of artificial intelligence undertaking jobs traditionally completed by humans is frequently discussed across a wide variety of platforms and industries. This notebook will not answer the question of whether a special teams coach can be replaced by an artificial intelligence, there is far more to being a special teams coach than what can be learned from this data. However, it will take a step on the road to answering that question by attempting to train a Generative Adversarial Network (GAN) to generate new and unique punt plays. If it can learn how 22 players and the ball line up and react to each others movements then perhaps, with further work, it could predict the optimum location of the field to aim a punt or the return route that will yield the most yards.\n
\n
Importing the libraries needed in the notebook and loading the raw data.
","metadata":{}},{"cell_type":"code","source":"import numpy as np\nimport pandas as pd\nimport matplotlib as mpl\nimport matplotlib.pyplot as plt\nimport seaborn as sns\nimport tensorflow as tf\nfrom tensorflow.keras import layers\nimport math\nimport time\nimport os","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2021-12-30T05:59:14.345364Z","iopub.execute_input":"2021-12-30T05:59:14.345707Z","iopub.status.idle":"2021-12-30T05:59:21.899364Z","shell.execute_reply.started":"2021-12-30T05:59:14.345569Z","shell.execute_reply":"2021-12-30T05:59:21.898171Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Read in the data\nroot_dir = '../input/nfl-big-data-bowl-2022/'\n\nli = []\n\ntracking_cols = ['gameId','playId','displayName','position','nflId','position','playDirection','x','y','team','frameId']\nfor filename in ['tracking2018.csv', 'tracking2019.csv', 'tracking2020.csv']:\n df = pd.read_csv(root_dir + filename, index_col=None, header=0, usecols=tracking_cols)\n li.append(df)\n\ntracking = pd.concat(li, axis=0, ignore_index=True)\n\nplay_cols = ['gameId','playId', 'possessionTeam', 'specialTeamsPlayType', 'specialTeamsResult']\nplays = pd.read_csv(root_dir + 'plays.csv', usecols=play_cols)\n\nprint(f\"Loaded in {len(plays)} plays and {len(tracking)} tracking records\")","metadata":{"_kg_hide-input":true,"execution":{"iopub.status.busy":"2021-12-30T05:59:21.901298Z","iopub.execute_input":"2021-12-30T05:59:21.901506Z","iopub.status.idle":"2021-12-30T06:01:22.493204Z","shell.execute_reply.started":"2021-12-30T05:59:21.901479Z","shell.execute_reply":"2021-12-30T06:01:22.492236Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n\nCreate a unique identifier by combining the playId, which is only unique within a game, and the gameId which is unique globally. This gives each play a global unique identifier.\n
\nDuring the period covered by the data the team known as the Raiders re-located from Oakland to Las Vegas. The team abbreviation used for the Raiders also changed from OAK to LV. To achieve consistency throughout the data all instances of the OAK team abbreviation will be renamed to LV.\n
\nThe tracking data for the ball does not hold a position or an NFL ID. To ensure the ball can be separated from players within the tracking data, these values will be populated with some arbitrary values.\n
\nSelect only the punt related data and clean up any irrelevant data.\n
","metadata":{}},{"cell_type":"code","source":"# combine gameid and playid to create a unique play id\ntracking['playUniqueId'] = tracking['gameId'].astype(str) + tracking['playId'].astype(str)\nplays['playUniqueId'] = plays['gameId'].astype(str) + plays['playId'].astype(str)\n\n# Rename all instances of OAK to LV\nplays.loc[plays['possessionTeam']=='OAK', 'possessionTeam'] = 'LV'\n\n#Fill in some gaps in the tracking data for the football\ntracking.loc[tracking['displayName']=='football', 'position'] = 'BLL'\ntracking.loc[tracking['displayName']=='football', 'nflId'] = 99999.9","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:01:22.494956Z","iopub.execute_input":"2021-12-30T06:01:22.495636Z","iopub.status.idle":"2021-12-30T06:03:18.436016Z","shell.execute_reply.started":"2021-12-30T06:01:22.495564Z","shell.execute_reply":"2021-12-30T06:03:18.434984Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Filter for the punt data\npunt_plays = plays.loc[plays['specialTeamsPlayType']=='Punt']\npunt_tracking = tracking.loc[tracking['playUniqueId'].isin(punt_plays.playUniqueId.unique())]\nprint(f\"There are {len(punt_plays)} punt plays and {len(punt_tracking)} punt tracking records\")\n# Tidy up data that is no longer needed\ndel plays\ndel tracking","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:18.437521Z","iopub.execute_input":"2021-12-30T06:03:18.437869Z","iopub.status.idle":"2021-12-30T06:03:32.140286Z","shell.execute_reply.started":"2021-12-30T06:03:18.437828Z","shell.execute_reply":"2021-12-30T06:03:32.139411Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n\nThe 8 different punt outcomes that can be split into 2 groups, ones where the ball was successfully punted and made it's way down field, and ones where the ball was not successfully punted. The outcomes in the first group, successful punts, are all individually more numerous than any of the categories in the second group, unsuccessful punts.\n
\nSuccessful Punts (in order of most numerous):\n
\nThe length of a punt play can vary for example, a kick out of bounds would be much shorter than a zig-zagging return run. The length of a play is measured in the number of frames it occupies.\n
\nMost plays (5,981) have at least 57 frames, at this point a slow decline in the number of plays still happening starts to happen. A steep decline occurs between 87 frames, where the number of plays drops below 5,000, and 92 frames where the number of plays drops below 4,000. From here there is a more gradual decline with the below 3,000 occuring at 105 frames, below 2,000 at 119 frames and below 1,000 at 132 frames. At 140 frames 90% of the plays have finished leaving a long tail that continues out to 272 frames.\n
\nTo help the network learn I organise the data to make it as consistent as possible:\n
\nA generator is created, which feeds random data through a network to create a punt play. A discriminator is created, which determines whether a punt play is real or fake from a mix of examples containing generated and real punt plays. The output is then used to tune both networks so that the generator becomes better at generating convincing punt plays and the discriminator becomes better at detecting fake plays.\n
","metadata":{}},{"cell_type":"code","source":"BUFFER_SIZE = 12000\nBATCH_SIZE = 128\n\nos.environ['KMP_WARNINGS'] = '0'\n\ncross_entropy = tf.keras.losses.BinaryCrossentropy(from_logits=True)\n\ntrain_dataset = tf.data.Dataset.from_tensor_slices(loaded_training_examples).shuffle(BUFFER_SIZE).batch(BATCH_SIZE)","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:09:35.037968Z","iopub.execute_input":"2021-12-30T06:09:35.038360Z","iopub.status.idle":"2021-12-30T06:09:35.625975Z","shell.execute_reply.started":"2021-12-30T06:09:35.038286Z","shell.execute_reply":"2021-12-30T06:09:35.624880Z"},"_kg_hide-input":true,"_kg_hide-output":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def generator_model():\n model = tf.keras.Sequential()\n model.add(layers.Dense(35*6*8, use_bias=False, input_shape=(60,)))\n model.add(layers.BatchNormalization())\n model.add(layers.LeakyReLU())\n\n model.add(layers.Reshape((35, 6, 8)))\n assert model.output_shape == (None, 35, 6, 8)\n\n model.add(layers.Conv2DTranspose(8, (5, 5), strides=(1, 1), padding='same', use_bias=False, data_format='channels_last'))\n assert model.output_shape == (None, 35, 6, 8)\n model.add(layers.BatchNormalization())\n model.add(layers.LeakyReLU())\n\n model.add(layers.Conv2DTranspose(8, (5, 5), strides=(1, 1), padding='same', use_bias=False, data_format='channels_last'))\n assert model.output_shape == (None, 35, 6, 8)\n model.add(layers.BatchNormalization())\n model.add(layers.LeakyReLU())\n\n model.add(layers.Conv2DTranspose(4, (5, 5), strides=(2, 2), padding='same', use_bias=False, data_format='channels_last'))\n assert model.output_shape == (None, 70, 12, 4)\n model.add(layers.BatchNormalization())\n model.add(layers.LeakyReLU())\n\n model.add(layers.Conv2DTranspose(4, (5, 5), strides=(1, 1), padding='same', use_bias=False, data_format='channels_last'))\n assert model.output_shape == (None, 70, 12, 4)\n model.add(layers.BatchNormalization())\n model.add(layers.LeakyReLU())\n\n model.add(layers.Conv2DTranspose(2, (5, 5), strides=(2, 2), padding='same', use_bias=False, activation='tanh', data_format='channels_last'))\n assert model.output_shape == (None, 140, 24, 2)\n\n model.add(layers.Cropping2D(cropping=((0,0), (0,1))))\n assert model.output_shape == (None, 140, 23, 2)\n\n return model\n\ndef generator_loss(fake_output):\n return cross_entropy(tf.ones_like(fake_output), fake_output)","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:42.184251Z","iopub.execute_input":"2021-12-30T06:03:42.184514Z","iopub.status.idle":"2021-12-30T06:03:42.201511Z","shell.execute_reply.started":"2021-12-30T06:03:42.184483Z","shell.execute_reply":"2021-12-30T06:03:42.200803Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"def make_discriminator_model():\n model = tf.keras.Sequential()\n model.add(layers.Conv2D(64, (5, 5), strides=(2, 2), padding='same',\n input_shape=[140, 23, 2]))\n model.add(layers.LeakyReLU())\n model.add(layers.Dropout(0.3))\n\n model.add(layers.Conv2D(128, (5, 5), strides=(1, 1), padding='same'))\n model.add(layers.LeakyReLU())\n model.add(layers.Dropout(0.3))\n\n model.add(layers.Conv2D(128, (5, 5), strides=(1, 1), padding='same'))\n model.add(layers.LeakyReLU())\n model.add(layers.Dropout(0.3))\n\n model.add(layers.Conv2D(128, (5, 5), strides=(2, 2), padding='same'))\n model.add(layers.LeakyReLU())\n model.add(layers.Dropout(0.3))\n\n model.add(layers.Flatten())\n model.add(layers.Dense(1))\n\n return model\n\ndef discriminator_loss(real_output, fake_output):\n real_loss = cross_entropy(tf.ones_like(real_output), real_output)\n fake_loss = cross_entropy(tf.zeros_like(fake_output), fake_output)\n total_loss = real_loss + fake_loss\n return total_loss","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:42.202852Z","iopub.execute_input":"2021-12-30T06:03:42.203868Z","iopub.status.idle":"2021-12-30T06:03:42.220701Z","shell.execute_reply.started":"2021-12-30T06:03:42.203794Z","shell.execute_reply":"2021-12-30T06:03:42.219747Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Training setup\ngenerator = generator_model()\ndiscriminator = make_discriminator_model()\n\n# Set the optimizers\ngenerator_optimizer = tf.keras.optimizers.Adam(learning_rate=2e-4, beta_1=0.5)\ndiscriminator_optimizer = tf.keras.optimizers.Adam(learning_rate=2e-4, beta_1=0.5)\n\n# Checkpoints code\ncheckpoint_dir = '../input/nfl-gan-data/checkpoints'\ncheckpoint_prefix = os.path.join(checkpoint_dir, \"ckpt\")\ncheckpoint = tf.train.Checkpoint(generator_optimizer=generator_optimizer,\n discriminator_optimizer=discriminator_optimizer,\n generator=generator,\n discriminator=discriminator)\n\n# Training loop\nEPOCHS = 500\nnoise_dim = 60\nnum_examples_to_generate = 256\n\nseed = tf.random.normal([num_examples_to_generate, noise_dim])","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:42.222133Z","iopub.execute_input":"2021-12-30T06:03:42.223012Z","iopub.status.idle":"2021-12-30T06:03:42.570799Z","shell.execute_reply.started":"2021-12-30T06:03:42.222975Z","shell.execute_reply":"2021-12-30T06:03:42.569928Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"@tf.function\ndef train_step(images):\n noise = tf.random.normal([BATCH_SIZE, noise_dim])\n\n with tf.GradientTape() as gen_tape, tf.GradientTape() as disc_tape:\n generated_images = generator(noise, training=True)\n real_output = discriminator(images, training=True)\n fake_output = discriminator(generated_images, training=True)\n\n gen_loss = generator_loss(fake_output)\n disc_loss = discriminator_loss(real_output, fake_output)\n\n gradients_of_generator = gen_tape.gradient(gen_loss, generator.trainable_variables)\n gradients_of_discriminator = disc_tape.gradient(disc_loss, discriminator.trainable_variables)\n\n generator_optimizer.apply_gradients(zip(gradients_of_generator, generator.trainable_variables))\n discriminator_optimizer.apply_gradients(zip(gradients_of_discriminator, discriminator.trainable_variables))\n\ndef train(dataset, epochs):\n\n checkpoint.restore(tf.train.latest_checkpoint(checkpoint_dir))\n for epoch in range(epochs):\n start = time.time()\n\n for image_batch in dataset:\n train_step(image_batch)\n\n # Save generated plays every 100 epochs\n if epoch % 100 == 0:\n generate_and_save_plays(generator,\n epoch + 1,\n seed)\n\n # Save the model every 15 epochs\n if (epoch + 1) % 15 == 0:\n checkpoint.save(file_prefix = checkpoint_prefix)\n\n print ('Time for epoch {} is {} sec'.format(epoch + 1, time.time()-start))\n\n # Generate after the final epoch\n generate_and_save_plays(generator,\n epochs,\n seed)\n\ngenerated_plays_during_training = []\n\ndef generate_and_save_plays(model, epoch, test_input):\n predictions = model(test_input, training=False)\n if len(predictions) > 0:\n generated_plays_during_training.append(predictions)","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:42.572433Z","iopub.execute_input":"2021-12-30T06:03:42.572710Z","iopub.status.idle":"2021-12-30T06:03:42.588274Z","shell.execute_reply.started":"2021-12-30T06:03:42.572678Z","shell.execute_reply":"2021-12-30T06:03:42.587364Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Train the network, only run number of epochs from latest checkpoint\nlast_ckpt = tf.train.latest_checkpoint(\n checkpoint_dir, latest_filename=None\n)\nlast_ckpt = int(last_ckpt.split(\"-\")[-1])\nnet_epochs = EPOCHS - (last_ckpt*15) if last_ckpt < (EPOCHS/15) else 0\n \ntrain(train_dataset, net_epochs)","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:42.589999Z","iopub.execute_input":"2021-12-30T06:03:42.590640Z","iopub.status.idle":"2021-12-30T06:03:43.358775Z","shell.execute_reply.started":"2021-12-30T06:03:42.590553Z","shell.execute_reply":"2021-12-30T06:03:43.357702Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"# Save the generated examples\nif len(generated_plays_during_training) > 1:\n generated_examples = np.array(generated_plays_during_training)\n\n # Check if previous examples created and add the new ones to it.\n generated_examples_path = \"../input/nfl-gan-data/generated_examples.npy\"\n\n if os.path.exists(generated_examples_path):\n # Add the examples generated this run to ones generated previously\n all_loaded_generated_examples = np.load('../input/nfl-gan-data/generated_examples.npy')\n all_examples = np.concatenate((all_loaded_generated_examples, generated_examples), axis = 0)\n np.save('../input/nfl-gan-data/generated_examples', all_examples)\n else:\n np.save('../input/nfl-gan-data/generated_examples', generated_examples)\n all_examples = generated_examples","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:43.359881Z","iopub.execute_input":"2021-12-30T06:03:43.360097Z","iopub.status.idle":"2021-12-30T06:03:43.366160Z","shell.execute_reply.started":"2021-12-30T06:03:43.360070Z","shell.execute_reply":"2021-12-30T06:03:43.365614Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\nThe animation of a play is adapted from a post created by JARON_MICHAL. See the post here - https://www.kaggle.com/jaronmichal/tracking-data-visualization
","metadata":{}},{"cell_type":"code","source":"# THIS CODE CREATES THE FOOTBALL PITCH\n\nimport matplotlib.patches as patches\nfrom matplotlib.patches import Arc\nfrom matplotlib import pyplot as plt\nimport matplotlib.patches as mpatches\n\n# Change size of the figure\nplt.rcParams['figure.figsize'] = [12, 8]\ndef drawPitch(width, height, color=\"w\"):\n fig = plt.figure()\n ax = plt.axes(xlim=(-10, width + 30), ylim=(-15, height + 5))\n plt.axis('off')\n\n # Grass around pitch\n rect = patches.Rectangle((-10, -5), width + 40, height + 10, linewidth=1, facecolor='#3f995b', capstyle='round')\n ax.add_patch(rect)\n ###################\n\n # Pitch boundaries\n rect = plt.Rectangle((0, 0), width + 20, height, ec=color, fc=\"None\", lw=2)\n ax.add_patch(rect)\n ###################\n\n # vertical lines - every 5 yards\n for i in range(21):\n plt.plot([10 + 5 * i, 10 + 5 * i], [0, height], c=\"w\", lw=2)\n ###################\n \n # distance markers - every 10 yards\n for yards in range(10, width, 10):\n yards_text = yards if yards <= width / 2 else width - yards\n # top markers\n plt.text(10 + yards - 2, height - 7.5, yards_text, size=15, c=\"w\", weight=\"bold\")\n # botoom markers\n plt.text(10 + yards - 2, 7.5, yards_text, size=15, c=\"w\", weight=\"bold\", rotation=180)\n ###################\n\n # yards markers - every yard\n # bottom markers\n for x in range(20):\n for j in range(1, 5):\n plt.plot([10 + x * 5 + j, 10 + x * 5 + j], [1, 3], color=\"w\", lw=2)\n\n # top markers\n for x in range(20):\n for j in range(1, 5):\n plt.plot([10 + x * 5 + j, 10 + x * 5 + j], [height - 1, height - 3], color=\"w\", lw=2)\n\n # middle bottom markers\n y = (height - 18.5) / 2\n for x in range(20):\n for j in range(1, 5):\n plt.plot([10 + x * 5 + j, 10 + x * 5 + j], [y, y + 2], color=\"w\", lw=2)\n\n # middle top markers\n for x in range(20):\n for j in range(1, 5):\n plt.plot([10 + x * 5 + j, 10 + x * 5 + j], [height - y, height - y - 2], color=\"w\", lw=2)\n ###################\n\n # draw home end zone\n plt.text(2.5, (height - 15) / 2, \"HOME\", size=30, c=\"w\", weight=\"bold\", rotation=90)\n rect = plt.Rectangle((0, 0), 10, height, ec=color, fc=\"#0064dc\", lw=2)\n ax.add_patch(rect)\n\n # draw away end zone \n plt.text(111, (height - 15) / 2, \"AWAY\", size=30, c=\"w\", weight=\"bold\", rotation=-90)\n rect = plt.Rectangle((width + 10, 0), 10, height, ec=color, fc=\"#c80014\", lw=2)\n ax.add_patch(rect)\n ###################\n \n # draw extra spot point\n # left\n y = (height - 3) / 2\n plt.plot([10 + 2, 10 + 2], [y, y + 3], c=\"w\", lw=2)\n \n # right\n plt.plot([width + 10 - 2, width + 10 - 2], [y, y + 3], c=\"w\", lw=2)\n ###################\n \n # draw goalpost\n goal_width = 6 # yards\n y = (height - goal_width) / 2\n # left\n plt.plot([0, 0], [y, y + goal_width], \"-\", c=\"y\", lw=10, ms=20)\n # right\n plt.plot([width + 20, width + 20], [y, y + goal_width], \"-\", c=\"y\", lw=10, ms=20)\n \n return fig, ax","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:43.367235Z","iopub.execute_input":"2021-12-30T06:03:43.367842Z","iopub.status.idle":"2021-12-30T06:03:43.396920Z","shell.execute_reply.started":"2021-12-30T06:03:43.367796Z","shell.execute_reply":"2021-12-30T06:03:43.395936Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"code","source":"%matplotlib ipympl\nfrom matplotlib import animation\nfrom IPython.display import HTML\ndef animate_generated_play(df):\n fig, ax = drawPitch(100, 53.3)\n play = df\n team_left, = ax.plot([], [], 'o', markersize=20, markerfacecolor=\"r\", markeredgewidth=2, markeredgecolor=\"white\", zorder=7)\n team_right, = ax.plot([], [], 'o', markersize=20, markerfacecolor=\"b\", markeredgewidth=2, markeredgecolor=\"white\", zorder=7)\n ball, = ax.plot([], [], 'o', markersize=10, markerfacecolor=\"black\", markeredgewidth=2, markeredgecolor=\"white\", zorder=7)\n drawings = [team_left, team_right, ball]\n\n def init():\n team_left.set_data([], [])\n team_right.set_data([], [])\n ball.set_data([], [])\n return drawings\n \n # Re-scale the generated examples\n def upscale_coordinates(coords):\n x,y = coords\n x = (x * 60) + 60\n y = (y * 26.65) + 26.65\n return [x, y]\n\n def draw_teams(i):\n X = []\n Y = []\n for coords in play[i][1:12]:\n x, y = upscale_coordinates(coords)\n X.append(x)\n Y.append(y)\n team_left.set_data(X, Y)\n \n X = []\n Y = []\n for coords in play[i][12:23]:\n x, y = upscale_coordinates(coords)\n X.append(x)\n Y.append(y)\n team_right.set_data(X, Y)\n\n def animate(i):\n draw_teams(i)\n x, y = upscale_coordinates(play[i][0])\n ball.set_data([x, y])\n return drawings\n \n anim = animation.FuncAnimation(fig, animate, init_func=init,\n frames=140, interval=100, blit=True)\n\n return HTML(anim.to_html5_video())","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:43.398345Z","iopub.execute_input":"2021-12-30T06:03:43.398861Z","iopub.status.idle":"2021-12-30T06:03:43.436159Z","shell.execute_reply.started":"2021-12-30T06:03:43.398827Z","shell.execute_reply":"2021-12-30T06:03:43.435171Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n\nIn this section I compare a real punt play with plays generated by the model during the early, mid and final stages of training. \n
\n\nA real play taken from the training data showing an actual punt play from the NFL. Things that define this as a real punt play include:\n
\nThe early output of the generator looks nothing like a real punt play. All players and the ball clustered around the centere of the pitch making seemingly random movements. None of the elements listed above that define a real punt play are present in this example.\n
","metadata":{}},{"cell_type":"code","source":"animate_generated_play(all_loaded_generated_examples[0][0])","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:44.489687Z","iopub.execute_input":"2021-12-30T06:03:44.490041Z","iopub.status.idle":"2021-12-30T06:03:44.955317Z","shell.execute_reply.started":"2021-12-30T06:03:44.489993Z","shell.execute_reply":"2021-12-30T06:03:44.954300Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\nBy mid-training the output is starting to look more convincing. The initial position of the players and the ball look realistic, which covers point 1 of our checklist. However, the ball does not appear to travel from the snapper to the punter, instead it hangs around the line of scrimmage for a few frames before making it's way down field. There is movement of all the players towards the ball once it makes it's way down field though. Overall 2 out of 3 elements are already present midway through training.
","metadata":{}},{"cell_type":"code","source":"animate_generated_play(all_loaded_generated_examples[4][0])","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:44.957134Z","iopub.execute_input":"2021-12-30T06:03:44.957553Z","iopub.status.idle":"2021-12-30T06:03:45.212841Z","shell.execute_reply.started":"2021-12-30T06:03:44.957504Z","shell.execute_reply":"2021-12-30T06:03:45.212065Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\nIn the final output all initial player positions looks accurate, the ball travels in a realistic manner, and the players move as expected during a punt play.
","metadata":{}},{"cell_type":"code","source":"predictions = generator(seed, training=False)\npredictions = np.array(predictions)\nanimate_generated_play(predictions[0])","metadata":{"execution":{"iopub.status.busy":"2021-12-30T06:03:45.214125Z","iopub.execute_input":"2021-12-30T06:03:45.214360Z","iopub.status.idle":"2021-12-30T06:03:45.566441Z","shell.execute_reply.started":"2021-12-30T06:03:45.214331Z","shell.execute_reply":"2021-12-30T06:03:45.565644Z"},"_kg_hide-input":true,"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\n\nThe aim of this notebook was to find out if a GAN could be trained to generate a realistic looking punt play. I believe this was achieved as output from the final model showed realistic starting positions and expected movements of both the players and the ball.\n
\nThe practical applications for this work in it's current state are limited. A coach could generate and watch plays in the hope of seeing a new and unique formation or it could be used as a training tool to give players exposure to more punt plays in the hope they will obtain a greater intuition for how they play out.\n
\nThis model was a proof of concept but, if the model was developed further, it could provide greater practical application. For example, if the starting position of the kicking team players were used as an input, the model could generate numerous examples from which monte carlo simulations could help inform which is the optimal running lane that will provide the punt returner with the greatest chance of the most yards.\n
\nThanks for reading my analysis.\n