├── Code2Pix.ipynb ├── Encoder Decoder Visualization.ipynb ├── Hydra Autoencoder.ipynb ├── LICENSE └── README.md /Code2Pix.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "markdown", 5 | "metadata": {}, 6 | "source": [ 7 | "# DSL2Pix - Deep Learning Compiler for User Interfaces\n", 8 | "---\n", 9 | "##### Noah Gundotra | SUSA @ UC Berkeley\n", 10 | "This work was made possible by the coordination of UC Berkeley's Statistics Undergraduate Student Association [SUSA](https://susa.berkeley.edu/) and [Uizard](https://uizard.io/). This notebook is research sponsored by Uizard. \n", 11 | "\n", 12 | "This notebook encapsulates the model design process to create a differentiable compiler that can be used to effectively compile simple user interfaces for Android, iOS, and Web backends. Specifically, we are \"compiling\" from a domain specific language (DSL) introduced by Tony Beltramelli in his original [pix2code paper](https://arxiv.org/abs/1705.07962). In fact, the dataset and vocabulary assets come from the [public pix2code repository](https://github.com/tonybeltramelli/pix2code).\n", 13 | "\n", 14 | "The successful model design ended up being using 1D convolutional layers for analyzing the DSL tokens, paired with a convolutional decoder taken from autoencoders trained on the pix2code datasets. See the Hydra Autoencoder.ipynb to train these.\n", 15 | "\n", 16 | "In the near future, we hope to use this model to improve the accuracy of pix2code and related models by using dsl2pix in a GAN with pix2code." 17 | ] 18 | }, 19 | { 20 | "cell_type": "code", 21 | "execution_count": null, 22 | "metadata": {}, 23 | "outputs": [], 24 | "source": [ 25 | "__authors__ = 'Noah Gundotra, Samyak Parajuli, Luke Dai, Ajay Raj, Aismit Das, Dennis Yang, Japjot Singh'" 26 | ] 27 | }, 28 | { 29 | "cell_type": "code", 30 | "execution_count": null, 31 | "metadata": { 32 | "ExecuteTime": { 33 | "end_time": "2018-04-18T20:37:43.261847Z", 34 | "start_time": "2018-04-18T20:37:39.755871Z" 35 | } 36 | }, 37 | "outputs": [], 38 | "source": [ 39 | "import numpy as np\n", 40 | "import matplotlib.pyplot as plt \n", 41 | "% matplotlib inline\n", 42 | "import os\n", 43 | "from os.path import join\n", 44 | "\n", 45 | "import keras\n", 46 | "from keras.models import Sequential\n", 47 | "from keras.layers import (LSTM, Dense, Conv2D, Input, Reshape, concatenate, MaxPooling2D, Dropout, Flatten, \n", 48 | "RepeatVector, UpSampling2D, Conv1D, Permute, BatchNormalization, Activation, \n", 49 | " UpSampling2D, MaxPooling1D, GlobalAveragePooling1D, Embedding)\n", 50 | "from keras import Model\n", 51 | "from keras.engine.topology import Layer, InputSpec\n", 52 | "from keras import backend as K\n", 53 | "from keras.optimizers import RMSprop\n", 54 | "from keras.callbacks import EarlyStopping\n", 55 | "from keras.utils import to_categorical\n", 56 | "from keras.losses import mean_squared_error\n", 57 | "from keras.preprocessing import sequence\n", 58 | "\n", 59 | "from Vocabulary import Vocabulary\n", 60 | "from keras.backend import clear_session\n", 61 | "\n", 62 | "clear_session()" 63 | ] 64 | }, 65 | { 66 | "cell_type": "markdown", 67 | "metadata": {}, 68 | "source": [ 69 | "### Choose your backend" 70 | ] 71 | }, 72 | { 73 | "cell_type": "code", 74 | "execution_count": null, 75 | "metadata": { 76 | "ExecuteTime": { 77 | "end_time": "2018-04-18T20:37:43.267061Z", 78 | "start_time": "2018-04-18T20:37:43.263712Z" 79 | } 80 | }, 81 | "outputs": [], 82 | "source": [ 83 | "datasets = ['android', 'ios', 'web']\n", 84 | "BACKEND = datasets[2]\n", 85 | "traindir = \"{}/training_features\".format(BACKEND)\n", 86 | "vocab_path = '../pix2code/bin/{}/'.format(BACKEND)" 87 | ] 88 | }, 89 | { 90 | "cell_type": "markdown", 91 | "metadata": {}, 92 | "source": [ 93 | "# Load Data (tokens, images)" 94 | ] 95 | }, 96 | { 97 | "cell_type": "code", 98 | "execution_count": null, 99 | "metadata": {}, 100 | "outputs": [], 101 | "source": [ 102 | "def process_string(data_dir, file):\n", 103 | " \"\"\"Turns the gui descriptor files into a list of their 'words'\"\"\"\n", 104 | " f = open(join(data_dir, file))\n", 105 | " original_string = f.read()\n", 106 | " list_words = original_string.split()\n", 107 | " new_list = []\n", 108 | " for word in list_words:\n", 109 | " # Remove special characters\n", 110 | " special_characters = \" ,\\n\"\n", 111 | " for c in special_characters:\n", 112 | " word = word.replace(c, '')\n", 113 | " new_list.append(word)\n", 114 | " return new_list, original_string\n", 115 | "\n", 116 | "def load_data(data_dir):\n", 117 | " \"\"\"Loads in the sentence and image data together.\n", 118 | " Returns (sentences, images, original strings (in descriptor files))\"\"\"\n", 119 | " images = []\n", 120 | " sentences = []\n", 121 | " original_strings = []\n", 122 | " files = os.listdir(data_dir)\n", 123 | " for listing in files:\n", 124 | " # Load the sentence files, then the images\n", 125 | " # There are less likely to be user-made files in this dir that end with .gui\n", 126 | " if listing.endswith('.gui'):\n", 127 | " base = listing[:-4]\n", 128 | " try:\n", 129 | " img_f = join(data_dir, base + '.npz')\n", 130 | " img = np.load(img_f)['features']\n", 131 | " except Exception as e:\n", 132 | " print(e)\n", 133 | " print(\"Error trying to match img for {}\".format(listing))\n", 134 | " continue\n", 135 | " sent, original_string = process_string(data_dir, listing)\n", 136 | " sentences.append(sent)\n", 137 | " original_strings.append(original_string)\n", 138 | " images.append(img)\n", 139 | " return sentences, np.array(images), original_strings\n", 140 | "\n", 141 | "sentences, GUIS, original_strings = load_data(traindir)\n", 142 | "assert len(sentences) == len(GUIS) and len(sentences) == 1500\n", 143 | "print(\"Loaded succesfully\")" 144 | ] 145 | }, 146 | { 147 | "cell_type": "markdown", 148 | "metadata": {}, 149 | "source": [ 150 | "# One Hot\n", 151 | "\n", 152 | "Here we are drawing on files and a class from pix2code training. This Vocabulary class is from `pix2code/model/classes/Vocabulary.py`" 153 | ] 154 | }, 155 | { 156 | "cell_type": "code", 157 | "execution_count": null, 158 | "metadata": {}, 159 | "outputs": [], 160 | "source": [ 161 | "voc = Vocabulary()\n", 162 | "voc.retrieve(vocab_path)\n", 163 | "VOCAB_SIZE = len(voc.binary_vocabulary)" 164 | ] 165 | }, 166 | { 167 | "cell_type": "code", 168 | "execution_count": null, 169 | "metadata": { 170 | "ExecuteTime": { 171 | "end_time": "2018-04-18T20:38:24.701841Z", 172 | "start_time": "2018-04-18T20:38:23.574258Z" 173 | } 174 | }, 175 | "outputs": [], 176 | "source": [ 177 | "#Find max seq length\n", 178 | "max_length = 0\n", 179 | "for j in range(len(sentences)):\n", 180 | " if len(sentences) > max_length:\n", 181 | " max_length = len(sentences)\n", 182 | "\n", 183 | "def to_hot(sentences, max_length, voc):\n", 184 | " hot = np.zeros((len(sentences), max_length, len(voc.binary_vocabulary)))\n", 185 | " for i, sent in enumerate(sentences):\n", 186 | " for j, word in enumerate(sent):\n", 187 | " hot[i, j] = voc.binary_vocabulary[word]\n", 188 | " return hot\n", 189 | "\n", 190 | "hot = to_hot(sentences, max_length, voc)\n", 191 | "assert len(hot) == 1500\n", 192 | "print(\"Converted to one-hot succesfully\")" 193 | ] 194 | }, 195 | { 196 | "cell_type": "markdown", 197 | "metadata": {}, 198 | "source": [ 199 | "# Model Design\n", 200 | "\n", 201 | "Functions that help the model design process." 202 | ] 203 | }, 204 | { 205 | "cell_type": "code", 206 | "execution_count": null, 207 | "metadata": {}, 208 | "outputs": [], 209 | "source": [ 210 | "def load_frozen(model_path, freeze=False):\n", 211 | " \"\"\"Loads in the model from model_path. If freeze=True, then the model will be non-trainable.\"\"\"\n", 212 | " model = keras.models.load_model(model_path)\n", 213 | " for layer in model.layers:\n", 214 | " layer.trainable = False if freeze else True\n", 215 | " return model\n", 216 | "\n", 217 | "def make_callback(model, dataset):\n", 218 | " \"\"\"Creates a lambda callback that plots the models prediction on the first 20 pictures of the dataset.\"\"\"\n", 219 | " def callback(epoch, logs):\n", 220 | " gif_range(model, dataset, start=0, stop=20)\n", 221 | " return keras.callbacks.LambdaCallback(on_epoch_end=callback)\n", 222 | "\n", 223 | "def shuffle_weights(model, weights=None):\n", 224 | " \"\"\"\n", 225 | " @author: jklient\n", 226 | " @source: https://github.com/keras-team/keras/issues/341\n", 227 | "\n", 228 | " Randomly permute the weights in `model`, or the given `weights`.\n", 229 | "\n", 230 | " This is a fast approximation of re-initializing the weights of a model.\n", 231 | "\n", 232 | " Assumes weights are distributed independently of the dimensions of the weight tensors\n", 233 | " (i.e., the weights have the same distribution along each dimension).\n", 234 | "\n", 235 | " :param Model model: Modify the weights of the given model.\n", 236 | " :param list(ndarray) weights: The model's weights will be replaced by a random permutation of these weights.\n", 237 | " If `None`, permute the model's current weights.\n", 238 | " \"\"\"\n", 239 | " if weights is None:\n", 240 | " weights = model.get_weights()\n", 241 | " weights = [np.random.permutation(w.flat).reshape(w.shape) for w in weights]\n", 242 | " # Faster, but less random: only permutes along the first dimension\n", 243 | " # weights = [np.random.permutation(w) for w in weights]\n", 244 | " model.set_weights(weights)\n", 245 | " \n", 246 | "def gif_range(model, data, start=0, stop=1, save=True):\n", 247 | " \"\"\"\n", 248 | " Adapted from - \n", 249 | " @author: Eli Bendersky\n", 250 | " @source: https://eli.thegreenplace.net/2016/drawing-animated-gifs-with-matplotlib/\n", 251 | " Makes a GIF that reveals the model's predictions for the first (stop-start) images in data.\n", 252 | " Saves the GIF to `dsl_predictions.gif`\n", 253 | " \"\"\"\n", 254 | " import sys\n", 255 | " from matplotlib.animation import FuncAnimation\n", 256 | "\n", 257 | " fig, ax = plt.subplots()\n", 258 | " # fig.set_tight_layout(True)\n", 259 | "\n", 260 | " # Query the figure's on-screen size and DPI. Note that when saving the figure to\n", 261 | " # a file, we need to provide a DPI for that separately.\n", 262 | " # print('fig size: {0} DPI, size in inches {1}'.format(\n", 263 | " # fig.get_dpi(), fig.get_size_inches()))\n", 264 | " \n", 265 | " ax.set_xticks([])\n", 266 | " ax.set_yticks([])\n", 267 | " \n", 268 | " def update(i):\n", 269 | " label = 'wordembedding {0}'.format(i)\n", 270 | " # print(label)\n", 271 | " ax.imshow(model.predict(np.expand_dims(data[i], axis=0))[0])\n", 272 | " ax.set_xlabel(label)\n", 273 | " return ax\n", 274 | " \n", 275 | " fig.tight_layout()\n", 276 | " anim = FuncAnimation(fig, update, frames=np.arange(start, stop), interval=70)\n", 277 | " if save:\n", 278 | " anim.save('dsl_predictions.gif', dpi=80, writer='imagemagick')\n", 279 | " else:\n", 280 | " # plt.show() will just loop the animation forever.\n", 281 | " plt.show()\n", 282 | " plt.close(fig)\n", 283 | " \n", 284 | "def unroll_hydra(model, last):\n", 285 | " \"\"\"Unrolls the decoder head of the hydra model given. Each hydra model has only 1 head, so this\n", 286 | " should be relatively simple.\"\"\"\n", 287 | " return model.layers[-1](last)" 288 | ] 289 | }, 290 | { 291 | "cell_type": "markdown", 292 | "metadata": {}, 293 | "source": [ 294 | "# Training Code2Pix\n", 295 | "\n", 296 | "The following code cells create the code2pix model and train it.\n", 297 | "\n", 298 | "\n", 299 | "## Pro-Tip\n", 300 | "\n", 301 | "Instead of just looking at numbers while this trains - look at pretty pictures! The `val_plot` callback to this model will update `dsl_predictions.gif` with the model's predictions on the first 20 images in the validation data. It's pretty cool to watch. **Highly recommend.**" 302 | ] 303 | }, 304 | { 305 | "cell_type": "code", 306 | "execution_count": null, 307 | "metadata": {}, 308 | "outputs": [], 309 | "source": [ 310 | "def make_conv(autoencoder_path):\n", 311 | " \"\"\"1D conv model :/\"\"\"\n", 312 | " x_in = Input(shape=(max_length, VOCAB_SIZE))\n", 313 | " last = Conv1D(64, 7, padding='same', activation='relu')(x_in)\n", 314 | " last = MaxPooling1D()(last)\n", 315 | " last = Conv1D(64, 5, padding='same', activation='relu')(last)\n", 316 | " last = Conv1D(128, 4, padding='same', activation='relu')(last)\n", 317 | " last = MaxPooling1D()(last)\n", 318 | " last = Flatten()(last)\n", 319 | " last = Dense(1024, activation='relu')(last)\n", 320 | " last = Reshape((8, 8, 16))(last)\n", 321 | " last = UpSampling2D()(last)\n", 322 | " autoencoder = load_frozen(autoencoder_path, freeze=False)\n", 323 | " last = unroll_hydra(autoencoder, last)\n", 324 | " model = Model(x_in, last)\n", 325 | " model.compile(RMSprop(lr=0.01), loss='binary_crossentropy')\n", 326 | " return model\n", 327 | "\n", 328 | "clear_session()\n", 329 | "model = make_conv('{}-hydra-model'.format(BACKEND))\n", 330 | "\n", 331 | "# Uncomment to see the number of model parameters\n", 332 | "# model.summary()\n", 333 | "\n", 334 | "val_plot = make_callback(model, hot[1400:])\n", 335 | "\n", 336 | "# Early Stopping to prevent severe overfitting\n", 337 | "early_stop = EarlyStopping(monitor='val_loss', patience=7)\n", 338 | "\n", 339 | "model.fit(hot[:1400], GUIS[:1400], callbacks=[val_plot, early_stop],\n", 340 | " validation_data=(hot[1400:], GUIS[1400:]), epochs=65, batch_size=32, shuffle=False)" 341 | ] 342 | }, 343 | { 344 | "cell_type": "code", 345 | "execution_count": null, 346 | "metadata": {}, 347 | "outputs": [], 348 | "source": [ 349 | "model.save('dsl2pix-{}'.format(BACKEND))" 350 | ] 351 | }, 352 | { 353 | "cell_type": "markdown", 354 | "metadata": {}, 355 | "source": [ 356 | "# Visualizing Results" 357 | ] 358 | }, 359 | { 360 | "cell_type": "markdown", 361 | "metadata": {}, 362 | "source": [ 363 | "The following code cell chooses a random picture from the validation set and plots the input and output of the code2pix model and compares it to the ground-truth output." 364 | ] 365 | }, 366 | { 367 | "cell_type": "code", 368 | "execution_count": null, 369 | "metadata": {}, 370 | "outputs": [], 371 | "source": [ 372 | "from collections import Counter\n", 373 | "\n", 374 | "def compare(sentences, hot, pics, idx):\n", 375 | " \"\"\"Compares the results of the model.\n", 376 | "\n", 377 | " Displays the input to the model, output of the model on the input, and ground truth.\n", 378 | " Input to the model is chosen as the hot[idx].\n", 379 | " \"\"\"\n", 380 | " fig, axes = plt.subplots(1, 2)\n", 381 | " for ax in axes:\n", 382 | " ax.set_yticks([])\n", 383 | " ax.set_xticks([])\n", 384 | " \n", 385 | " axes[0].imshow(model.predict(np.expand_dims(hot[idx], 0))[0])\n", 386 | " axes[0].set_xlabel(\"Predicted GUI\")\n", 387 | " axes[1].imshow(pics[idx])\n", 388 | " axes[1].set_xlabel(\"Actual GUI\")\n", 389 | " \n", 390 | " print('gui descriptor:\\t', ' '.join(sentences[idx]))\n", 391 | " count = Counter(sentences[idx])\n", 392 | " print(count)\n", 393 | " fig.savefig(\"{}-code2pix-predict.png\".format(BACKEND))\n", 394 | " return count\n", 395 | "\n", 396 | "count = compare(sentences, hot, GUIS, np.random.choice(list(range(1400, 1500))))" 397 | ] 398 | }, 399 | { 400 | "cell_type": "markdown", 401 | "metadata": {}, 402 | "source": [ 403 | "The following code cells are left for additional experimentation and educational purposes." 404 | ] 405 | }, 406 | { 407 | "cell_type": "markdown", 408 | "metadata": {}, 409 | "source": [ 410 | "# Word2Vec Model - For fine tuned embeddings\n", 411 | "---\n", 412 | "\n", 413 | "Embedding the DSL sentence representation into condensed, semantically-infused vector representations. **Currently unused.**" 414 | ] 415 | }, 416 | { 417 | "cell_type": "code", 418 | "execution_count": null, 419 | "metadata": { 420 | "ExecuteTime": { 421 | "end_time": "2018-04-18T20:38:27.084892Z", 422 | "start_time": "2018-04-18T20:38:26.107041Z" 423 | } 424 | }, 425 | "outputs": [], 426 | "source": [ 427 | "#All word embeddings (size of training set x sentence length x 100(embedding dimension))\n", 428 | "def create_wordembeddings(sentences, EMBED_SIZE):\n", 429 | " \"\"\"Creates a Word2Vec model to compress the sentence data.\n", 430 | " Returns a np.array of (#sentences, sentence length, EMBED_SIZE)\"\"\"\n", 431 | " from gensim.models import Word2Vec\n", 432 | " \n", 433 | " model = Word2Vec(sentences, min_count=1,size=EMBED_SIZE)\n", 434 | " words = list(model.wv.vocab)\n", 435 | " \n", 436 | " wordembeddings = []\n", 437 | " for sentence in sentences:\n", 438 | " result_array = np.empty((0, EMBED_SIZE))\n", 439 | " for word in sentence:\n", 440 | " result = model[word].reshape(1, EMBED_SIZE)\n", 441 | " result_array = np.append(result_array, result, axis=0)\n", 442 | " result_array = np.pad(result_array, ((max_length - result_array.shape[0],0), (0,0)), 'constant', constant_values = 0)\n", 443 | " wordembeddings.append(result_array)\n", 444 | " wordembeddings = np.array(wordembeddings)\n", 445 | " return wordembeddings\n", 446 | "\n", 447 | "wordembeddings = create_wordembeddings(sentences, VOCAB_SIZE)\n", 448 | "assert wordembeddings.shape[0] == 1500" 449 | ] 450 | }, 451 | { 452 | "cell_type": "markdown", 453 | "metadata": {}, 454 | "source": [ 455 | "## Extra Model Parts\n", 456 | "\n", 457 | "**Currently unused.** Used in previous iterations to repeat 3D embedding data. For example repeating a (16, 16, 16) embedding." 458 | ] 459 | }, 460 | { 461 | "cell_type": "code", 462 | "execution_count": null, 463 | "metadata": { 464 | "ExecuteTime": { 465 | "end_time": "2018-04-18T20:38:29.800684Z", 466 | "start_time": "2018-04-18T20:38:29.769226Z" 467 | } 468 | }, 469 | "outputs": [], 470 | "source": [ 471 | "#Custom Layer to get from 3D -> 4D\n", 472 | "class RepeatVector4D(Layer):\n", 473 | " def __init__(self, n, **kwargs):\n", 474 | " self.n = n\n", 475 | " self.input_spec = [InputSpec(ndim=3)]\n", 476 | " super(RepeatVector4D, self).__init__(**kwargs)\n", 477 | " \n", 478 | " def get_output_shape_for(self, input_shape):\n", 479 | " return (input_shape[0], self.n, input_shape[1], input_shape[2])\n", 480 | " def compute_output_shape(self, input_shape):\n", 481 | " return (input_shape[0], self.n, input_shape[1], input_shape[2])\n", 482 | "\n", 483 | " def call(self, x, mask=None):\n", 484 | " x = K.expand_dims(x, 1)\n", 485 | " pattern = K.stack([1, self.n, 1, 1])\n", 486 | " return K.tile(x, pattern)" 487 | ] 488 | }, 489 | { 490 | "cell_type": "markdown", 491 | "metadata": {}, 492 | "source": [ 493 | "# LSTM-Based Model\n", 494 | "\n", 495 | "This model is based off the pix2code model design. This model uses lstms to encode the token data, and loads in pretrained lstms from the pix2code models. Super slow to train ~3min/epoch. ***Deprecated.*** If you would like to run these, feel free to, but the rest of the code is not explicitly designed to support these models anymore." 496 | ] 497 | }, 498 | { 499 | "cell_type": "code", 500 | "execution_count": null, 501 | "metadata": {}, 502 | "outputs": [], 503 | "source": [ 504 | "def make_model(encoder_path, lstm_decoder_path, autoencoder_path):\n", 505 | " # Freeze the pretrained encoder & decoder (respectively)\n", 506 | " pix2code_lstm = load_frozen(encoder_path, False)\n", 507 | " lstm_decoder = load_frozen(lstm_decoder_path, False)\n", 508 | " autoencoder = load_frozen(autoencoder_path, False)\n", 509 | "\n", 510 | " #Input DSL -> Intermediate Encoding through convolution\n", 511 | " x_in = Input(shape = (max_length, 19), name = 'x_in')\n", 512 | " drop_rate = 0.01\n", 513 | "# last = Dropout(drop_rate)(last)\n", 514 | "# #y_lstm = LSTM(128, return_sequences = True, input_shape=(max_length, 100))(y_in)\n", 515 | " \n", 516 | "# b1 = Dense(48*48, activation='relu')(x_flatten)\n", 517 | "# b1 = Reshape((48, 48, 1))(b1)\n", 518 | "# b1 = Conv2D(filters=64, kernel_size=(3,3), strides=(3,3), padding='valid', \n", 519 | "# activation='relu', name='b1conv_1')(b1)\n", 520 | "# b1 = Conv2D(16, (3,3), strides=(1,1), padding='same', activation='relu')(b1)\n", 521 | "# last = Reshape((48, 19))(last) # Input size of pix2code's first encoder\n", 522 | " last = pix2code_lstm(x_in)\n", 523 | " last = LSTM(512, return_sequences = True)(last)\n", 524 | " last = Dropout(drop_rate)(last)\n", 525 | " last = lstm_decoder.layers[-1](last)\n", 526 | " reshape = Reshape((8,8,8))(last)\n", 527 | " last = UpSampling2D((2,2), name='upsampler-trainable')(reshape)\n", 528 | " last = Conv2D(16, kernel_size=(3,3), padding='same', activation='relu')(last)\n", 529 | " last = Reshape((16,16,16))(last)\n", 530 | "\n", 531 | " # Load in the v0 autoencoder\n", 532 | "# last = keras.layers.Add()([b1, last])\n", 533 | "# last = Conv2D(16, kernel_size=(2,2), padding='same',activation='relu')(last)\n", 534 | " last = unroll_hydra(autoencoder, last)\n", 535 | " \n", 536 | " model = Model(x_in, last)\n", 537 | " opt = RMSprop(lr=0.001)\n", 538 | " model.compile(optimizer=opt, loss='binary_crossentropy')\n", 539 | " return model\n", 540 | "\n", 541 | "clear_session()\n", 542 | "model = make_model('ios-p2c-encoder-LSTM', 'ios-p2c-decoder-LSTM', 'ios-hydra-model')\n", 543 | "# model.summary()\n", 544 | "callback = make_callback(model, hot)\n", 545 | "# shuffle_weights(model)\n", 546 | "model.fit(hot, GUIS, callbacks=[callback], epochs=200, batch_size=32, validation_split=0.2, shuffle=False)\n", 547 | "\n", 548 | "\n", 549 | "def make_model_old(encoder_path, lstm_decoder_path, autoencoder_path='model/autoencoders-v0/ios.h5'):\n", 550 | " \"\"\"A deprecated iteration of the dsl2pix model design.\"\"\"\n", 551 | " # Freeze the pretrained encoder & decoder (respectively)\n", 552 | " pix2code_lstm = load_frozen(encoder_path, False)\n", 553 | " autoencoder = load_frozen(autoencoder_path, False)\n", 554 | "\n", 555 | " x_in = Input(shape = (max_length, VOCAB_SIZE), name = 'x_in')\n", 556 | " x_flatten = Flatten()(x_in)\n", 557 | " last = Dense(912, activation='relu')(x_flatten)\n", 558 | " drop_rate = 0.1\n", 559 | " last = Dropout(drop_rate)(last)\n", 560 | " #y_lstm = LSTM(128, return_sequences = True, input_shape=(max_length, 100))(y_in)\n", 561 | " \n", 562 | " b1 = Dense(48*48, activation='relu')(x_flatten)\n", 563 | " b1 = Reshape((48, 48, 1))(b1)\n", 564 | " b1 = Conv2D(filters=64, kernel_size=(3,3), strides=(3,3), padding='valid', \n", 565 | " activation='relu', name='b1conv_1')(b1)\n", 566 | " b1 = Conv2D(16, (3,3), strides=(1,1), padding='same', activation='relu')(b1)\n", 567 | " last = Reshape((48, 19))(last) # Input size of pix2code's first encoder\n", 568 | " last = LSTM(19, return_sequences=True)(last)\n", 569 | " last = LSTM(19, return_sequences=True)(last)\n", 570 | " last = pix2code_lstm(last)\n", 571 | " last = LSTM(128, return_sequences = True)(last)\n", 572 | " last = LSTM(256, return_sequences = True)(last)\n", 573 | " last = Dropout(drop_rate)(last)\n", 574 | " last = LSTM(512, return_sequences = False)(last)\n", 575 | " reshape = Reshape((8,8,8))(last)\n", 576 | " last = UpSampling2D((2,2), name='upsampler-trainable')(reshape)\n", 577 | " last = Conv2D(32, kernel_size=(4,4), padding='same', activation='relu')(last)\n", 578 | "# last = Dropout(drop_rate)(last)\n", 579 | " last = Conv2D(16, kernel_size=(3,3), padding='same', activation='relu')(last)\n", 580 | " last = Reshape((16,16,16))(last)\n", 581 | " \n", 582 | " # Load in the v0 autoencoder\n", 583 | " last = keras.layers.Add()([b1, last])\n", 584 | " last = Conv2D(16, kernel_size=(2,2), padding='same',activation='relu')(last)\n", 585 | " last = unroll_v1(autoencoder, last)\n", 586 | " \n", 587 | " model = Model(x_in, last)\n", 588 | " model.compile(optimizer='rmsprop', loss='binary_crossentropy')\n", 589 | " return model" 590 | ] 591 | } 592 | ], 593 | "metadata": { 594 | "kernelspec": { 595 | "display_name": "PyTorch Kernel", 596 | "language": "python", 597 | "name": "torchkern" 598 | }, 599 | "language_info": { 600 | "codemirror_mode": { 601 | "name": "ipython", 602 | "version": 3 603 | }, 604 | "file_extension": ".py", 605 | "mimetype": "text/x-python", 606 | "name": "python", 607 | "nbconvert_exporter": "python", 608 | "pygments_lexer": "ipython3", 609 | "version": "3.6.4" 610 | } 611 | }, 612 | "nbformat": 4, 613 | "nbformat_minor": 2 614 | } 615 | -------------------------------------------------------------------------------- /Hydra Autoencoder.ipynb: -------------------------------------------------------------------------------- 1 | { 2 | "cells": [ 3 | { 4 | "cell_type": "markdown", 5 | "metadata": {}, 6 | "source": [ 7 | "# Autoencoding Graphical User Interfaces\n", 8 | "\n", 9 | "In our initial approach to solving code2pix, we assumed we needed to pretrain a convolutional model that could decode textual data into images of graphical user interfaces. The purpose of training the following autoencoder is to train the decoder that we then use in the code2pix model. However, last minute experiments showed that training the autoencoder is **not strictly necessary** to train code2pix. If you would like to skip straight to code2pix, feel free to initialize the model, skip training, and just save the model for its architecture. Code2Pix will just needs the layer architecture constructed in this notebook" 10 | ] 11 | }, 12 | { 13 | "cell_type": "code", 14 | "execution_count": 1, 15 | "metadata": {}, 16 | "outputs": [ 17 | { 18 | "name": "stderr", 19 | "output_type": "stream", 20 | "text": [ 21 | "/home/ngundotra/.conda/envs/torchenv/lib/python3.6/site-packages/h5py/__init__.py:36: FutureWarning: Conversion of the second argument of issubdtype from `float` to `np.floating` is deprecated. In future, it will be treated as `np.float64 == np.dtype(float).type`.\n", 22 | " from ._conv import register_converters as _register_converters\n", 23 | "Using TensorFlow backend.\n" 24 | ] 25 | } 26 | ], 27 | "source": [ 28 | "import numpy as np\n", 29 | "import matplotlib.pyplot as plt \n", 30 | "% matplotlib inline\n", 31 | "import os\n", 32 | "from os.path import join\n", 33 | "\n", 34 | "import keras\n", 35 | "from keras.models import Sequential\n", 36 | "from keras.layers import (LSTM, Dense, Conv2D, Input, Reshape, concatenate, MaxPooling2D,Dropout, Flatten, \n", 37 | "RepeatVector, UpSampling2D, Conv1D, Permute, BatchNormalization, Activation, UpSampling2D)\n", 38 | "from keras import Model\n", 39 | "from keras.engine.topology import Layer, InputSpec\n", 40 | "from keras import backend as K\n", 41 | "from keras.optimizers import RMSprop\n", 42 | "from keras.utils import to_categorical\n", 43 | "from keras.losses import mean_squared_error\n", 44 | "from keras.preprocessing import sequence" 45 | ] 46 | }, 47 | { 48 | "cell_type": "markdown", 49 | "metadata": {}, 50 | "source": [ 51 | "# Hydra Autoencoder\n", 52 | "\n", 53 | "Model that understands our synthetic datasets (android, ios, web) interfaces, using a shared CNN to encode each picture, and unique heads to decode for each platform (3 heads). It takes about 15 minutes to train all three heads and the encoder on two Nvidia 1080 Ti-s." 54 | ] 55 | }, 56 | { 57 | "cell_type": "code", 58 | "execution_count": 2, 59 | "metadata": {}, 60 | "outputs": [ 61 | { 62 | "name": "stdout", 63 | "output_type": "stream", 64 | "text": [ 65 | "Epoch: 1\n", 66 | "Train on 1200 samples, validate on 300 samples\n", 67 | "Epoch 1/1\n", 68 | "1200/1200 [==============================] - 13s 11ms/step - loss: 0.4835 - acc: 0.0046 - val_loss: 0.3131 - val_acc: 0.0046\n", 69 | "300/300 [==============================] - 1s 4ms/step\n", 70 | "Train on 1200 samples, validate on 300 samples\n", 71 | "Epoch 1/1\n", 72 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.5459 - acc: 0.4587 - val_loss: 0.6028 - val_acc: 0.3122\n", 73 | "300/300 [==============================] - 1s 3ms/step\n", 74 | "Train on 1200 samples, validate on 300 samples\n", 75 | "Epoch 1/1\n", 76 | "1200/1200 [==============================] - 11s 10ms/step - loss: 0.5404 - acc: 0.3138 - val_loss: 0.3864 - val_acc: 0.3527\n", 77 | "300/300 [==============================] - 1s 3ms/step\n", 78 | "Model had avg val loss of: 0.43409594668282403\n", 79 | "Best avg validation loss is 0.43409594668282403\n", 80 | "Epoch: 2\n", 81 | "Train on 1200 samples, validate on 300 samples\n", 82 | "Epoch 1/1\n", 83 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.2470 - acc: 0.0052 - val_loss: 0.2188 - val_acc: 0.0057\n", 84 | "300/300 [==============================] - 1s 3ms/step\n", 85 | "Train on 1200 samples, validate on 300 samples\n", 86 | "Epoch 1/1\n", 87 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.2138 - acc: 0.5194 - val_loss: 0.2263 - val_acc: 0.5017\n", 88 | "300/300 [==============================] - 1s 3ms/step\n", 89 | "Train on 1200 samples, validate on 300 samples\n", 90 | "Epoch 1/1\n", 91 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.2127 - acc: 0.3643 - val_loss: 0.2422 - val_acc: 0.3818\n", 92 | "300/300 [==============================] - 1s 3ms/step\n", 93 | "Model had avg val loss of: 0.22908878293302326\n", 94 | "Best avg validation loss is 0.43409594668282403\n", 95 | "Epoch: 3\n", 96 | "Train on 1200 samples, validate on 300 samples\n", 97 | "Epoch 1/1\n", 98 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.2051 - acc: 0.0054 - val_loss: 0.2715 - val_acc: 0.0055\n", 99 | "300/300 [==============================] - 1s 3ms/step\n", 100 | "Train on 1200 samples, validate on 300 samples\n", 101 | "Epoch 1/1\n", 102 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1650 - acc: 0.5195 - val_loss: 0.6579 - val_acc: 0.5018\n", 103 | "300/300 [==============================] - 1s 3ms/step\n", 104 | "Train on 1200 samples, validate on 300 samples\n", 105 | "Epoch 1/1\n", 106 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1603 - acc: 0.3649 - val_loss: 0.1679 - val_acc: 0.3788\n", 107 | "300/300 [==============================] - 1s 3ms/step\n", 108 | "Model had avg val loss of: 0.36575216551621753\n", 109 | "Best avg validation loss is 0.43409594668282403\n", 110 | "Epoch: 4\n", 111 | "Train on 1200 samples, validate on 300 samples\n", 112 | "Epoch 1/1\n", 113 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1978 - acc: 0.0060 - val_loss: 0.2098 - val_acc: 0.0048\n", 114 | "300/300 [==============================] - 1s 3ms/step\n", 115 | "Train on 1200 samples, validate on 300 samples\n", 116 | "Epoch 1/1\n", 117 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1586 - acc: 0.5195 - val_loss: 0.2254 - val_acc: 0.5012\n", 118 | "300/300 [==============================] - 1s 3ms/step\n", 119 | "Train on 1200 samples, validate on 300 samples\n", 120 | "Epoch 1/1\n", 121 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1562 - acc: 0.3649 - val_loss: 0.1543 - val_acc: 0.3793\n", 122 | "300/300 [==============================] - 1s 3ms/step\n", 123 | "Model had avg val loss of: 0.19651010354359943\n", 124 | "Best avg validation loss is 0.43409594668282403\n", 125 | "Epoch: 5\n", 126 | "Train on 1200 samples, validate on 300 samples\n", 127 | "Epoch 1/1\n", 128 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1932 - acc: 0.0063 - val_loss: 0.1945 - val_acc: 0.0063\n", 129 | "300/300 [==============================] - 1s 3ms/step\n", 130 | "Train on 1200 samples, validate on 300 samples\n", 131 | "Epoch 1/1\n", 132 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1545 - acc: 0.5195 - val_loss: 0.2026 - val_acc: 0.5013\n", 133 | "300/300 [==============================] - 1s 3ms/step\n", 134 | "Train on 1200 samples, validate on 300 samples\n", 135 | "Epoch 1/1\n", 136 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1527 - acc: 0.3650 - val_loss: 0.2117 - val_acc: 0.3818\n", 137 | "300/300 [==============================] - 1s 3ms/step\n", 138 | "Model had avg val loss of: 0.20296844952636295\n", 139 | "Best avg validation loss is 0.43409594668282403\n", 140 | "Epoch: 6\n", 141 | "Train on 1200 samples, validate on 300 samples\n", 142 | "Epoch 1/1\n", 143 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1908 - acc: 0.0064 - val_loss: 0.1912 - val_acc: 0.0058\n", 144 | "300/300 [==============================] - 1s 3ms/step\n", 145 | "Train on 1200 samples, validate on 300 samples\n", 146 | "Epoch 1/1\n", 147 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1526 - acc: 0.5195 - val_loss: 0.1866 - val_acc: 0.5017\n", 148 | "300/300 [==============================] - 1s 3ms/step\n", 149 | "Train on 1200 samples, validate on 300 samples\n", 150 | "Epoch 1/1\n", 151 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1504 - acc: 0.3651 - val_loss: 0.2096 - val_acc: 0.3818\n", 152 | "300/300 [==============================] - 1s 3ms/step\n", 153 | "Model had avg val loss of: 0.19579220493634542\n", 154 | "Best avg validation loss is 0.43409594668282403\n", 155 | "Epoch: 7\n", 156 | "Train on 1200 samples, validate on 300 samples\n", 157 | "Epoch 1/1\n", 158 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1896 - acc: 0.0065 - val_loss: 0.1895 - val_acc: 0.0065\n", 159 | "300/300 [==============================] - 1s 3ms/step\n", 160 | "Train on 1200 samples, validate on 300 samples\n", 161 | "Epoch 1/1\n", 162 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1508 - acc: 0.5195 - val_loss: 0.1769 - val_acc: 0.5017\n", 163 | "300/300 [==============================] - 1s 3ms/step\n", 164 | "Train on 1200 samples, validate on 300 samples\n", 165 | "Epoch 1/1\n", 166 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1486 - acc: 0.3652 - val_loss: 0.2134 - val_acc: 0.3818\n", 167 | "300/300 [==============================] - 1s 3ms/step\n", 168 | "Model had avg val loss of: 0.19328342676162721\n", 169 | "Best avg validation loss is 0.43409594668282403\n", 170 | "Epoch: 8\n", 171 | "Train on 1200 samples, validate on 300 samples\n", 172 | "Epoch 1/1\n", 173 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1886 - acc: 0.0066 - val_loss: 0.1924 - val_acc: 0.0061\n", 174 | "300/300 [==============================] - 1s 3ms/step\n", 175 | "Train on 1200 samples, validate on 300 samples\n", 176 | "Epoch 1/1\n", 177 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1497 - acc: 0.5195 - val_loss: 0.1609 - val_acc: 0.5018\n", 178 | "300/300 [==============================] - 1s 3ms/step\n", 179 | "Train on 1200 samples, validate on 300 samples\n", 180 | "Epoch 1/1\n", 181 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1468 - acc: 0.3652 - val_loss: 0.1897 - val_acc: 0.3818\n", 182 | "300/300 [==============================] - 1s 3ms/step\n", 183 | "Model had avg val loss of: 0.18097938040892284\n", 184 | "Best avg validation loss is 0.43409594668282403\n", 185 | "Epoch: 9\n", 186 | "Train on 1200 samples, validate on 300 samples\n", 187 | "Epoch 1/1\n", 188 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1880 - acc: 0.0067 - val_loss: 0.1889 - val_acc: 0.0066\n", 189 | "300/300 [==============================] - 1s 3ms/step\n", 190 | "Train on 1200 samples, validate on 300 samples\n", 191 | "Epoch 1/1\n", 192 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1487 - acc: 0.5195 - val_loss: 0.2267 - val_acc: 0.5012\n", 193 | "300/300 [==============================] - 1s 3ms/step\n", 194 | "Train on 1200 samples, validate on 300 samples\n", 195 | "Epoch 1/1\n", 196 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1456 - acc: 0.3653 - val_loss: 0.1457 - val_acc: 0.3810\n", 197 | "300/300 [==============================] - 1s 3ms/step\n", 198 | "Model had avg val loss of: 0.1870827837785085\n", 199 | "Best avg validation loss is 0.43409594668282403\n", 200 | "Epoch: 10\n", 201 | "Train on 1200 samples, validate on 300 samples\n", 202 | "Epoch 1/1\n", 203 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1872 - acc: 0.0067 - val_loss: 0.1890 - val_acc: 0.0065\n", 204 | "300/300 [==============================] - 1s 3ms/step\n", 205 | "Train on 1200 samples, validate on 300 samples\n", 206 | "Epoch 1/1\n", 207 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1480 - acc: 0.5196 - val_loss: 0.1767 - val_acc: 0.5015\n", 208 | "300/300 [==============================] - 1s 3ms/step\n", 209 | "Train on 1200 samples, validate on 300 samples\n", 210 | "Epoch 1/1\n", 211 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1446 - acc: 0.3653 - val_loss: 0.1472 - val_acc: 0.3815\n", 212 | "300/300 [==============================] - 1s 3ms/step\n", 213 | "Model had avg val loss of: 0.17098331650098164\n", 214 | "Best avg validation loss is 0.43409594668282403\n", 215 | "Epoch: 11\n", 216 | "Train on 1200 samples, validate on 300 samples\n", 217 | "Epoch 1/1\n" 218 | ] 219 | }, 220 | { 221 | "name": "stdout", 222 | "output_type": "stream", 223 | "text": [ 224 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1867 - acc: 0.0068 - val_loss: 0.1879 - val_acc: 0.0069\n", 225 | "300/300 [==============================] - 1s 3ms/step\n", 226 | "Train on 1200 samples, validate on 300 samples\n", 227 | "Epoch 1/1\n", 228 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1471 - acc: 0.5196 - val_loss: 0.1584 - val_acc: 0.5018\n", 229 | "300/300 [==============================] - 1s 3ms/step\n", 230 | "Train on 1200 samples, validate on 300 samples\n", 231 | "Epoch 1/1\n", 232 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1441 - acc: 0.3654 - val_loss: 0.4554 - val_acc: 0.3818\n", 233 | "300/300 [==============================] - 1s 3ms/step\n", 234 | "Model had avg val loss of: 0.2672271705998315\n", 235 | "Best avg validation loss is 0.43409594668282403\n", 236 | "Epoch: 12\n", 237 | "Train on 1200 samples, validate on 300 samples\n", 238 | "Epoch 1/1\n", 239 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1864 - acc: 0.0068 - val_loss: 0.1864 - val_acc: 0.0069\n", 240 | "300/300 [==============================] - 1s 3ms/step\n", 241 | "Train on 1200 samples, validate on 300 samples\n", 242 | "Epoch 1/1\n", 243 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1467 - acc: 0.5196 - val_loss: 0.1566 - val_acc: 0.5018\n", 244 | "300/300 [==============================] - 1s 3ms/step\n", 245 | "Train on 1200 samples, validate on 300 samples\n", 246 | "Epoch 1/1\n", 247 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1442 - acc: 0.3654 - val_loss: 0.2594 - val_acc: 0.3818\n", 248 | "300/300 [==============================] - 1s 3ms/step\n", 249 | "Model had avg val loss of: 0.20082141041755674\n", 250 | "Best avg validation loss is 0.43409594668282403\n", 251 | "Epoch: 13\n", 252 | "Train on 1200 samples, validate on 300 samples\n", 253 | "Epoch 1/1\n", 254 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1861 - acc: 0.0068 - val_loss: 0.1882 - val_acc: 0.0062\n", 255 | "300/300 [==============================] - 1s 3ms/step\n", 256 | "Train on 1200 samples, validate on 300 samples\n", 257 | "Epoch 1/1\n", 258 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1462 - acc: 0.5196 - val_loss: 0.1613 - val_acc: 0.5018\n", 259 | "300/300 [==============================] - 1s 3ms/step\n", 260 | "Train on 1200 samples, validate on 300 samples\n", 261 | "Epoch 1/1\n", 262 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1426 - acc: 0.3654 - val_loss: 0.1401 - val_acc: 0.3810\n", 263 | "300/300 [==============================] - 1s 3ms/step\n", 264 | "Model had avg val loss of: 0.16323615067534977\n", 265 | "Best avg validation loss is 0.43409594668282403\n", 266 | "Epoch: 14\n", 267 | "Train on 1200 samples, validate on 300 samples\n", 268 | "Epoch 1/1\n", 269 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1857 - acc: 0.0068 - val_loss: 0.1856 - val_acc: 0.0068\n", 270 | "300/300 [==============================] - 1s 3ms/step\n", 271 | "Train on 1200 samples, validate on 300 samples\n", 272 | "Epoch 1/1\n", 273 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1462 - acc: 0.5196 - val_loss: 0.1519 - val_acc: 0.5018\n", 274 | "300/300 [==============================] - 1s 3ms/step\n", 275 | "Train on 1200 samples, validate on 300 samples\n", 276 | "Epoch 1/1\n", 277 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1419 - acc: 0.3655 - val_loss: 0.1375 - val_acc: 0.3815\n", 278 | "300/300 [==============================] - 1s 3ms/step\n", 279 | "Model had avg val loss of: 0.1583106345600552\n", 280 | "Best avg validation loss is 0.43409594668282403\n", 281 | "Epoch: 15\n", 282 | "Train on 1200 samples, validate on 300 samples\n", 283 | "Epoch 1/1\n", 284 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1853 - acc: 0.0069 - val_loss: 0.1871 - val_acc: 0.0070\n", 285 | "300/300 [==============================] - 1s 3ms/step\n", 286 | "Train on 1200 samples, validate on 300 samples\n", 287 | "Epoch 1/1\n", 288 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1452 - acc: 0.5196 - val_loss: 0.1520 - val_acc: 0.5018\n", 289 | "300/300 [==============================] - 1s 3ms/step\n", 290 | "Train on 1200 samples, validate on 300 samples\n", 291 | "Epoch 1/1\n", 292 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1414 - acc: 0.3655 - val_loss: 0.1371 - val_acc: 0.3809\n", 293 | "300/300 [==============================] - 1s 3ms/step\n", 294 | "Model had avg val loss of: 0.15869325379530588\n", 295 | "Best avg validation loss is 0.43409594668282403\n", 296 | "Epoch: 16\n", 297 | "Train on 1200 samples, validate on 300 samples\n", 298 | "Epoch 1/1\n", 299 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1851 - acc: 0.0069 - val_loss: 0.1854 - val_acc: 0.0069\n", 300 | "300/300 [==============================] - 1s 3ms/step\n", 301 | "Train on 1200 samples, validate on 300 samples\n", 302 | "Epoch 1/1\n", 303 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1452 - acc: 0.5196 - val_loss: 0.1526 - val_acc: 0.5018\n", 304 | "300/300 [==============================] - 1s 3ms/step\n", 305 | "Train on 1200 samples, validate on 300 samples\n", 306 | "Epoch 1/1\n", 307 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1410 - acc: 0.3655 - val_loss: 0.1695 - val_acc: 0.3818\n", 308 | "300/300 [==============================] - 1s 3ms/step\n", 309 | "Model had avg val loss of: 0.16915018637975057\n", 310 | "Best avg validation loss is 0.43409594668282403\n", 311 | "Epoch: 17\n", 312 | "Train on 1200 samples, validate on 300 samples\n", 313 | "Epoch 1/1\n", 314 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1849 - acc: 0.0069 - val_loss: 0.1860 - val_acc: 0.0069\n", 315 | "300/300 [==============================] - 1s 3ms/step\n", 316 | "Train on 1200 samples, validate on 300 samples\n", 317 | "Epoch 1/1\n", 318 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1447 - acc: 0.5196 - val_loss: 0.1510 - val_acc: 0.5018\n", 319 | "300/300 [==============================] - 1s 3ms/step\n", 320 | "Train on 1200 samples, validate on 300 samples\n", 321 | "Epoch 1/1\n", 322 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1404 - acc: 0.3656 - val_loss: 0.1365 - val_acc: 0.3807\n", 323 | "300/300 [==============================] - 1s 3ms/step\n", 324 | "Model had avg val loss of: 0.15781751480367448\n", 325 | "Best avg validation loss is 0.43409594668282403\n", 326 | "Epoch: 18\n", 327 | "Train on 1200 samples, validate on 300 samples\n", 328 | "Epoch 1/1\n", 329 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1846 - acc: 0.0069 - val_loss: 0.1842 - val_acc: 0.0070\n", 330 | "300/300 [==============================] - 1s 3ms/step\n", 331 | "Train on 1200 samples, validate on 300 samples\n", 332 | "Epoch 1/1\n", 333 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1445 - acc: 0.5196 - val_loss: 0.1505 - val_acc: 0.5018\n", 334 | "300/300 [==============================] - 1s 3ms/step\n", 335 | "Train on 1200 samples, validate on 300 samples\n", 336 | "Epoch 1/1\n", 337 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1401 - acc: 0.3656 - val_loss: 0.1357 - val_acc: 0.3812\n", 338 | "300/300 [==============================] - 1s 3ms/step\n", 339 | "Model had avg val loss of: 0.15678167786863118\n", 340 | "Best avg validation loss is 0.43409594668282403\n", 341 | "Epoch: 19\n", 342 | "Train on 1200 samples, validate on 300 samples\n", 343 | "Epoch 1/1\n", 344 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1845 - acc: 0.0069 - val_loss: 0.1856 - val_acc: 0.0071\n", 345 | "300/300 [==============================] - 1s 3ms/step\n", 346 | "Train on 1200 samples, validate on 300 samples\n", 347 | "Epoch 1/1\n", 348 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1443 - acc: 0.5196 - val_loss: 0.1494 - val_acc: 0.5018\n", 349 | "300/300 [==============================] - 1s 3ms/step\n", 350 | "Train on 1200 samples, validate on 300 samples\n", 351 | "Epoch 1/1\n", 352 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1398 - acc: 0.3656 - val_loss: 0.1360 - val_acc: 0.3804\n", 353 | "300/300 [==============================] - 1s 3ms/step\n", 354 | "Model had avg val loss of: 0.15701211445861393\n", 355 | "Best avg validation loss is 0.43409594668282403\n", 356 | "Epoch: 20\n", 357 | "Train on 1200 samples, validate on 300 samples\n", 358 | "Epoch 1/1\n", 359 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1843 - acc: 0.0069 - val_loss: 0.1850 - val_acc: 0.0070\n", 360 | "300/300 [==============================] - 1s 3ms/step\n", 361 | "Train on 1200 samples, validate on 300 samples\n", 362 | "Epoch 1/1\n", 363 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1441 - acc: 0.5196 - val_loss: 0.1496 - val_acc: 0.5019\n", 364 | "300/300 [==============================] - 1s 3ms/step\n", 365 | "Train on 1200 samples, validate on 300 samples\n", 366 | "Epoch 1/1\n", 367 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1394 - acc: 0.3656 - val_loss: 0.1352 - val_acc: 0.3815\n", 368 | "300/300 [==============================] - 1s 3ms/step\n", 369 | "Model had avg val loss of: 0.15658335076438057\n", 370 | "Best avg validation loss is 0.43409594668282403\n", 371 | "Epoch: 21\n", 372 | "Train on 1200 samples, validate on 300 samples\n", 373 | "Epoch 1/1\n" 374 | ] 375 | }, 376 | { 377 | "name": "stdout", 378 | "output_type": "stream", 379 | "text": [ 380 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1842 - acc: 0.0069 - val_loss: 0.1859 - val_acc: 0.0071\n", 381 | "300/300 [==============================] - 1s 3ms/step\n", 382 | "Train on 1200 samples, validate on 300 samples\n", 383 | "Epoch 1/1\n", 384 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1439 - acc: 0.5196 - val_loss: 0.1499 - val_acc: 0.5018\n", 385 | "300/300 [==============================] - 1s 3ms/step\n", 386 | "Train on 1200 samples, validate on 300 samples\n", 387 | "Epoch 1/1\n", 388 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1392 - acc: 0.3657 - val_loss: 0.1400 - val_acc: 0.3806\n", 389 | "300/300 [==============================] - 1s 3ms/step\n", 390 | "Model had avg val loss of: 0.15858489149146607\n", 391 | "Best avg validation loss is 0.43409594668282403\n", 392 | "Epoch: 22\n", 393 | "Train on 1200 samples, validate on 300 samples\n", 394 | "Epoch 1/1\n", 395 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1840 - acc: 0.0069 - val_loss: 0.1854 - val_acc: 0.0070\n", 396 | "300/300 [==============================] - 1s 3ms/step\n", 397 | "Train on 1200 samples, validate on 300 samples\n", 398 | "Epoch 1/1\n", 399 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1437 - acc: 0.5196 - val_loss: 0.1492 - val_acc: 0.5019\n", 400 | "300/300 [==============================] - 1s 3ms/step\n", 401 | "Train on 1200 samples, validate on 300 samples\n", 402 | "Epoch 1/1\n", 403 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1390 - acc: 0.3657 - val_loss: 0.1351 - val_acc: 0.3816\n", 404 | "300/300 [==============================] - 1s 3ms/step\n", 405 | "Model had avg val loss of: 0.1565568486849467\n", 406 | "Best avg validation loss is 0.43409594668282403\n", 407 | "Epoch: 23\n", 408 | "Train on 1200 samples, validate on 300 samples\n", 409 | "Epoch 1/1\n", 410 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1839 - acc: 0.0070 - val_loss: 0.1854 - val_acc: 0.0071\n", 411 | "300/300 [==============================] - 1s 3ms/step\n", 412 | "Train on 1200 samples, validate on 300 samples\n", 413 | "Epoch 1/1\n", 414 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1436 - acc: 0.5196 - val_loss: 0.1484 - val_acc: 0.5020\n", 415 | "300/300 [==============================] - 1s 3ms/step\n", 416 | "Train on 1200 samples, validate on 300 samples\n", 417 | "Epoch 1/1\n", 418 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1387 - acc: 0.3657 - val_loss: 0.1376 - val_acc: 0.3812\n", 419 | "300/300 [==============================] - 1s 3ms/step\n", 420 | "Model had avg val loss of: 0.15717394212881722\n", 421 | "Best avg validation loss is 0.43409594668282403\n", 422 | "Epoch: 24\n", 423 | "Train on 1200 samples, validate on 300 samples\n", 424 | "Epoch 1/1\n", 425 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1837 - acc: 0.0070 - val_loss: 0.1877 - val_acc: 0.0069\n", 426 | "300/300 [==============================] - 1s 3ms/step\n", 427 | "Train on 1200 samples, validate on 300 samples\n", 428 | "Epoch 1/1\n", 429 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1435 - acc: 0.5196 - val_loss: 0.1482 - val_acc: 0.5018\n", 430 | "300/300 [==============================] - 1s 3ms/step\n", 431 | "Train on 1200 samples, validate on 300 samples\n", 432 | "Epoch 1/1\n", 433 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1384 - acc: 0.3657 - val_loss: 0.1344 - val_acc: 0.3810\n", 434 | "300/300 [==============================] - 1s 3ms/step\n", 435 | "Model had avg val loss of: 0.15679121070437962\n", 436 | "Best avg validation loss is 0.43409594668282403\n", 437 | "Epoch: 25\n", 438 | "Train on 1200 samples, validate on 300 samples\n", 439 | "Epoch 1/1\n", 440 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1837 - acc: 0.0070 - val_loss: 0.1850 - val_acc: 0.0071\n", 441 | "300/300 [==============================] - 1s 3ms/step\n", 442 | "Train on 1200 samples, validate on 300 samples\n", 443 | "Epoch 1/1\n", 444 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1432 - acc: 0.5196 - val_loss: 0.1482 - val_acc: 0.5019\n", 445 | "300/300 [==============================] - 1s 3ms/step\n", 446 | "Train on 1200 samples, validate on 300 samples\n", 447 | "Epoch 1/1\n", 448 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1383 - acc: 0.3658 - val_loss: 0.1343 - val_acc: 0.3813\n", 449 | "300/300 [==============================] - 1s 3ms/step\n", 450 | "Model had avg val loss of: 0.1558398103713989\n", 451 | "Best avg validation loss is 0.43409594668282403\n", 452 | "Epoch: 26\n", 453 | "Train on 1200 samples, validate on 300 samples\n", 454 | "Epoch 1/1\n", 455 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1836 - acc: 0.0070 - val_loss: 0.1855 - val_acc: 0.0070\n", 456 | "300/300 [==============================] - 1s 3ms/step\n", 457 | "Train on 1200 samples, validate on 300 samples\n", 458 | "Epoch 1/1\n", 459 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1431 - acc: 0.5196 - val_loss: 0.1478 - val_acc: 0.5018\n", 460 | "300/300 [==============================] - 1s 3ms/step\n", 461 | "Train on 1200 samples, validate on 300 samples\n", 462 | "Epoch 1/1\n", 463 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1380 - acc: 0.3658 - val_loss: 0.1344 - val_acc: 0.3814\n", 464 | "300/300 [==============================] - 1s 3ms/step\n", 465 | "Model had avg val loss of: 0.15590417948034074\n", 466 | "Best avg validation loss is 0.43409594668282403\n", 467 | "Epoch: 27\n", 468 | "Train on 1200 samples, validate on 300 samples\n", 469 | "Epoch 1/1\n", 470 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1834 - acc: 0.0070 - val_loss: 0.1862 - val_acc: 0.0071\n", 471 | "300/300 [==============================] - 1s 3ms/step\n", 472 | "Train on 1200 samples, validate on 300 samples\n", 473 | "Epoch 1/1\n", 474 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1430 - acc: 0.5196 - val_loss: 0.1504 - val_acc: 0.5018\n", 475 | "300/300 [==============================] - 1s 3ms/step\n", 476 | "Train on 1200 samples, validate on 300 samples\n", 477 | "Epoch 1/1\n", 478 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1379 - acc: 0.3658 - val_loss: 0.1337 - val_acc: 0.3814\n", 479 | "300/300 [==============================] - 1s 3ms/step\n", 480 | "Model had avg val loss of: 0.15675139950381384\n", 481 | "Best avg validation loss is 0.43409594668282403\n", 482 | "Epoch: 28\n", 483 | "Train on 1200 samples, validate on 300 samples\n", 484 | "Epoch 1/1\n", 485 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1834 - acc: 0.0070 - val_loss: 0.1843 - val_acc: 0.0070\n", 486 | "300/300 [==============================] - 1s 3ms/step\n", 487 | "Train on 1200 samples, validate on 300 samples\n", 488 | "Epoch 1/1\n", 489 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1430 - acc: 0.5196 - val_loss: 0.1477 - val_acc: 0.5019\n", 490 | "300/300 [==============================] - 1s 3ms/step\n", 491 | "Train on 1200 samples, validate on 300 samples\n", 492 | "Epoch 1/1\n", 493 | "1200/1200 [==============================] - 11s 9ms/step - loss: 0.1377 - acc: 0.3658 - val_loss: 0.1344 - val_acc: 0.3814\n", 494 | "300/300 [==============================] - 1s 3ms/step\n", 495 | "Model had avg val loss of: 0.15547384520371757\n", 496 | "Best avg validation loss is 0.43409594668282403\n" 497 | ] 498 | } 499 | ], 500 | "source": [ 501 | "def load_guis(data_dir):\n", 502 | " \"\"\"Loads in target data (picture data) in a given directory.\n", 503 | " Returns a np.array of the data loaded.\"\"\"\n", 504 | " GUIS = []\n", 505 | " training_GUI_listing = os.listdir(data_dir)\n", 506 | " for listing in training_GUI_listing:\n", 507 | " if listing.endswith('.npz'):\n", 508 | " temp = np.load(join(data_dir, listing))['features']\n", 509 | " # reshape to be flat\n", 510 | " GUIS += [temp]\n", 511 | " GUIS = np.array(GUIS)\n", 512 | " return GUIS\n", 513 | "\n", 514 | "def make_head():\n", 515 | " \"\"\"Returns a standard decoder model\"\"\"\n", 516 | " input_layer = Input(shape=(16, 16, 16))\n", 517 | " last = Conv2D(128, (4, 4), padding='same', activation='relu')(input_layer)\n", 518 | "\n", 519 | " last = BatchNormalization()(last)\n", 520 | " last = UpSampling2D(size=(2, 2))(last)\n", 521 | " last = Conv2D(96, (3, 3), padding='same', activation='relu')(last)\n", 522 | " # model.add(Conv2DTranspose(96, (2, 2), strides=(2, 2)))\n", 523 | " last = BatchNormalization()(last)\n", 524 | " last = UpSampling2D(size=(4, 4))(last)\n", 525 | " last = Conv2D(32, (3, 3), padding='same', activation='relu')(last)\n", 526 | " # model.add(Conv2DTranspose(32, (2, 2), strides=(2, 2)))\n", 527 | " last = BatchNormalization()(last)\n", 528 | " last = UpSampling2D(size=(2, 2))(last)\n", 529 | " # model.add(Conv2D(16, (3, 3), padding='same', activation='relu'))\n", 530 | " last = Conv2D(3, (3, 3), padding='same', activation='sigmoid')(last)\n", 531 | " # model.add(Conv2DTranspose(color, (3, 3), strides=(2, 2)))\n", 532 | " model = Model(input_layer, last)\n", 533 | " return model\n", 534 | " \n", 535 | "def make_hydra(loss='binary_crossentropy', optimizer='rmsprop'):\n", 536 | " \"\"\"Creates 3 headed autoencoder model to handle the different datasets.\n", 537 | " Returns [android, ios, web] models\"\"\"\n", 538 | " keras.backend.clear_session()\n", 539 | "\n", 540 | " input_layer = Input(shape=(256, 256, 3))\n", 541 | " drop_rate = 0.25\n", 542 | " last = Conv2D(64, (5, 5), strides=(2,2), activation='relu')(input_layer)\n", 543 | " last = Conv2D(128, (4, 4), padding='same', strides=(1,1), activation='relu')(last)\n", 544 | " last = MaxPooling2D(pool_size=(2,2))(last)\n", 545 | " last = Dropout(drop_rate)(last)\n", 546 | " last = Conv2D(128, (4, 4), padding='same', strides=(2,2), activation='relu')(last)\n", 547 | " last = Conv2D(256, (3, 3), padding='same', strides=(1,1), activation='relu')(last)\n", 548 | " last = Conv2D(128, (3, 3), padding='same', strides=(1,1), activation='relu')(last)\n", 549 | " last = MaxPooling2D(pool_size=(2, 2))(last)\n", 550 | " last = Dropout(drop_rate)(last)\n", 551 | "\n", 552 | " last = Conv2D(128, (3, 3), padding='same', strides=(1,1), activation='relu')(last)\n", 553 | " last = Conv2D(64, (3, 3), padding='same', strides=(1,1), activation='relu')(last)\n", 554 | " last = Conv2D(16, (3, 3), padding='same', strides=(1,1), activation='relu')(last)\n", 555 | "\n", 556 | " encoder = Model(input_layer, last)\n", 557 | " android = make_head()\n", 558 | " ios = make_head()\n", 559 | " web = make_head()\n", 560 | " \n", 561 | " models = []\n", 562 | " for model in [android, ios, web]:\n", 563 | " head_spout = model(encoder.output)\n", 564 | " head = Model(input_layer, head_spout)\n", 565 | " head.compile(optimizer=optimizer, loss=loss, metrics=['accuracy'])\n", 566 | " models.append(head)\n", 567 | " return models\n", 568 | "\n", 569 | "def train_hydra(epochs, models, datasets, val_split=0.2):\n", 570 | " \"\"\"Assumes models=[android, ios, web], and datasets ordered alphabetically as well.\n", 571 | " Trains the models sequentially, 1 epoch after the other. Validation split is last\n", 572 | " val_split % of the data.\n", 573 | " \n", 574 | " ~16 epochs seems to give good perf, >30 pretty good\"\"\"\n", 575 | " train = [data[:int((1 - val_split)*len(data))] for data in datasets]\n", 576 | " validation = [data[int((1-val_split)*len(data)):] for data in datasets]\n", 577 | " val_loss = 20\n", 578 | " val_best = 20\n", 579 | " for i in range(epochs):\n", 580 | " print(\"Epoch: {}\".format(i+1))\n", 581 | " val_loss = 0\n", 582 | " for model, data, val in zip(models, train, validation):\n", 583 | " model.fit(data, data, batch_size=16, validation_data=(val, val))\n", 584 | " val_loss += model.evaluate(val, val)[0]\n", 585 | " print(\"Model had avg val loss of: {}\".format(val_loss / 3))\n", 586 | " if val_loss < val_best:\n", 587 | " val_best = val_loss / 3\n", 588 | " print(\"Best avg validation loss is {}\".format(val_best))\n", 589 | " \n", 590 | "android_data = load_guis('android/training_features/')\n", 591 | "ios_data = load_guis('ios/training_features/')\n", 592 | "web_data = load_guis('web/training_features/')\n", 593 | "models = make_hydra()\n", 594 | "train_hydra(28, models, [android_data, ios_data, web_data])" 595 | ] 596 | }, 597 | { 598 | "cell_type": "markdown", 599 | "metadata": {}, 600 | "source": [ 601 | "Visualizing the models' performance. Example visualizations are saved in `android_hydra.png`, `ios_hydra.png`, and `web_hydra.png` respectively. To change which pictures are shown, modify the `idx` argument passed to predict. Should be between \\[0, len(data) = 1500)." 602 | ] 603 | }, 604 | { 605 | "cell_type": "code", 606 | "execution_count": 3, 607 | "metadata": {}, 608 | "outputs": [ 609 | { 610 | "data": { 611 | "image/png": "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\n", 612 | "text/plain": [ 613 | "" 614 | ] 615 | }, 616 | "metadata": {}, 617 | "output_type": "display_data" 618 | }, 619 | { 620 | "data": { 621 | "image/png": "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\n", 622 | "text/plain": [ 623 | "" 624 | ] 625 | }, 626 | "metadata": {}, 627 | "output_type": "display_data" 628 | }, 629 | { 630 | "data": { 631 | "image/png": "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\n", 632 | "text/plain": [ 633 | "" 634 | ] 635 | }, 636 | "metadata": {}, 637 | "output_type": "display_data" 638 | } 639 | ], 640 | "source": [ 641 | "def predict(model, dataset, idx, name):\n", 642 | " fig, axes = plt.subplots(1, 2)\n", 643 | " axes[0].imshow(dataset[idx])\n", 644 | " axes[1].imshow(model.predict(np.expand_dims(dataset[idx], 0))[0])\n", 645 | " for ax in axes:\n", 646 | " ax.set_yticks([])\n", 647 | " ax.set_xticks([])\n", 648 | " \n", 649 | " fig.savefig(name, dpi=70)\n", 650 | " plt.show()\n", 651 | "\n", 652 | "for model, data, name in zip(models, [android_data, ios_data, web_data], ['android', 'ios', 'web']):\n", 653 | " predict(model, data, 100, '{}_hydra.png'.format(name))\n", 654 | " model.save('{}-hydra-model'.format(name))" 655 | ] 656 | }, 657 | { 658 | "cell_type": "markdown", 659 | "metadata": {}, 660 | "source": [ 661 | "Verifying that the models saved properly" 662 | ] 663 | }, 664 | { 665 | "cell_type": "code", 666 | "execution_count": 4, 667 | "metadata": {}, 668 | "outputs": [ 669 | { 670 | "name": "stdout", 671 | "output_type": "stream", 672 | "text": [ 673 | "android-hydra-model ios-hydra-model web-hydra-model\r\n" 674 | ] 675 | } 676 | ], 677 | "source": [ 678 | "% ls *-model" 679 | ] 680 | } 681 | ], 682 | "metadata": { 683 | "kernelspec": { 684 | "display_name": "PyTorch Kernel", 685 | "language": "python", 686 | "name": "torchkern" 687 | }, 688 | "language_info": { 689 | "codemirror_mode": { 690 | "name": "ipython", 691 | "version": 3 692 | }, 693 | "file_extension": ".py", 694 | "mimetype": "text/x-python", 695 | "name": "python", 696 | "nbconvert_exporter": "python", 697 | "pygments_lexer": "ipython3", 698 | "version": "3.6.4" 699 | } 700 | }, 701 | "nbformat": 4, 702 | "nbformat_minor": 2 703 | } 704 | -------------------------------------------------------------------------------- /LICENSE: -------------------------------------------------------------------------------- 1 | Apache License 2 | Version 2.0, January 2004 3 | http://www.apache.org/licenses/ 4 | 5 | TERMS AND CONDITIONS FOR USE, REPRODUCTION, AND DISTRIBUTION 6 | 7 | 1. Definitions. 8 | 9 | "License" shall mean the terms and conditions for use, reproduction, 10 | and distribution as defined by Sections 1 through 9 of this document. 11 | 12 | "Licensor" shall mean the copyright owner or entity authorized by 13 | the copyright owner that is granting the License. 14 | 15 | "Legal Entity" shall mean the union of the acting entity and all 16 | other entities that control, are controlled by, or are under common 17 | control with that entity. For the purposes of this definition, 18 | "control" means (i) the power, direct or indirect, to cause the 19 | direction or management of such entity, whether by contract or 20 | otherwise, or (ii) ownership of fifty percent (50%) or more of the 21 | outstanding shares, or (iii) beneficial ownership of such entity. 22 | 23 | "You" (or "Your") shall mean an individual or Legal Entity 24 | exercising permissions granted by this License. 25 | 26 | "Source" form shall mean the preferred form for making modifications, 27 | including but not limited to software source code, documentation 28 | source, and configuration files. 29 | 30 | "Object" form shall mean any form resulting from mechanical 31 | transformation or translation of a Source form, including but 32 | not limited to compiled object code, generated documentation, 33 | and conversions to other media types. 34 | 35 | "Work" shall mean the work of authorship, whether in Source or 36 | Object form, made available under the License, as indicated by a 37 | copyright notice that is included in or attached to the work 38 | (an example is provided in the Appendix below). 39 | 40 | "Derivative Works" shall mean any work, whether in Source or Object 41 | form, that is based on (or derived from) the Work and for which the 42 | editorial revisions, annotations, elaborations, or other modifications 43 | represent, as a whole, an original work of authorship. For the purposes 44 | of this License, Derivative Works shall not include works that remain 45 | separable from, or merely link (or bind by name) to the interfaces of, 46 | the Work and Derivative Works thereof. 47 | 48 | "Contribution" shall mean any work of authorship, including 49 | the original version of the Work and any modifications or additions 50 | to that Work or Derivative Works thereof, that is intentionally 51 | submitted to Licensor for inclusion in the Work by the copyright owner 52 | or by an individual or Legal Entity authorized to submit on behalf of 53 | the copyright owner. For the purposes of this definition, "submitted" 54 | means any form of electronic, verbal, or written communication sent 55 | to the Licensor or its representatives, including but not limited to 56 | communication on electronic mailing lists, source code control systems, 57 | and issue tracking systems that are managed by, or on behalf of, the 58 | Licensor for the purpose of discussing and improving the Work, but 59 | excluding communication that is conspicuously marked or otherwise 60 | designated in writing by the copyright owner as "Not a Contribution." 61 | 62 | "Contributor" shall mean Licensor and any individual or Legal Entity 63 | on behalf of whom a Contribution has been received by Licensor and 64 | subsequently incorporated within the Work. 65 | 66 | 2. Grant of Copyright License. Subject to the terms and conditions of 67 | this License, each Contributor hereby grants to You a perpetual, 68 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 69 | copyright license to reproduce, prepare Derivative Works of, 70 | publicly display, publicly perform, sublicense, and distribute the 71 | Work and such Derivative Works in Source or Object form. 72 | 73 | 3. Grant of Patent License. Subject to the terms and conditions of 74 | this License, each Contributor hereby grants to You a perpetual, 75 | worldwide, non-exclusive, no-charge, royalty-free, irrevocable 76 | (except as stated in this section) patent license to make, have made, 77 | use, offer to sell, sell, import, and otherwise transfer the Work, 78 | where such license applies only to those patent claims licensable 79 | by such Contributor that are necessarily infringed by their 80 | Contribution(s) alone or by combination of their Contribution(s) 81 | with the Work to which such Contribution(s) was submitted. If You 82 | institute patent litigation against any entity (including a 83 | cross-claim or counterclaim in a lawsuit) alleging that the Work 84 | or a Contribution incorporated within the Work constitutes direct 85 | or contributory patent infringement, then any patent licenses 86 | granted to You under this License for that Work shall terminate 87 | as of the date such litigation is filed. 88 | 89 | 4. Redistribution. You may reproduce and distribute copies of the 90 | Work or Derivative Works thereof in any medium, with or without 91 | modifications, and in Source or Object form, provided that You 92 | meet the following conditions: 93 | 94 | (a) You must give any other recipients of the Work or 95 | Derivative Works a copy of this License; and 96 | 97 | (b) You must cause any modified files to carry prominent notices 98 | stating that You changed the files; and 99 | 100 | (c) You must retain, in the Source form of any Derivative Works 101 | that You distribute, all copyright, patent, trademark, and 102 | attribution notices from the Source form of the Work, 103 | excluding those notices that do not pertain to any part of 104 | the Derivative Works; and 105 | 106 | (d) If the Work includes a "NOTICE" text file as part of its 107 | distribution, then any Derivative Works that You distribute must 108 | include a readable copy of the attribution notices contained 109 | within such NOTICE file, excluding those notices that do not 110 | pertain to any part of the Derivative Works, in at least one 111 | of the following places: within a NOTICE text file distributed 112 | as part of the Derivative Works; within the Source form or 113 | documentation, if provided along with the Derivative Works; or, 114 | within a display generated by the Derivative Works, if and 115 | wherever such third-party notices normally appear. The contents 116 | of the NOTICE file are for informational purposes only and 117 | do not modify the License. You may add Your own attribution 118 | notices within Derivative Works that You distribute, alongside 119 | or as an addendum to the NOTICE text from the Work, provided 120 | that such additional attribution notices cannot be construed 121 | as modifying the License. 122 | 123 | You may add Your own copyright statement to Your modifications and 124 | may provide additional or different license terms and conditions 125 | for use, reproduction, or distribution of Your modifications, or 126 | for any such Derivative Works as a whole, provided Your use, 127 | reproduction, and distribution of the Work otherwise complies with 128 | the conditions stated in this License. 129 | 130 | 5. Submission of Contributions. Unless You explicitly state otherwise, 131 | any Contribution intentionally submitted for inclusion in the Work 132 | by You to the Licensor shall be under the terms and conditions of 133 | this License, without any additional terms or conditions. 134 | Notwithstanding the above, nothing herein shall supersede or modify 135 | the terms of any separate license agreement you may have executed 136 | with Licensor regarding such Contributions. 137 | 138 | 6. Trademarks. This License does not grant permission to use the trade 139 | names, trademarks, service marks, or product names of the Licensor, 140 | except as required for reasonable and customary use in describing the 141 | origin of the Work and reproducing the content of the NOTICE file. 142 | 143 | 7. Disclaimer of Warranty. Unless required by applicable law or 144 | agreed to in writing, Licensor provides the Work (and each 145 | Contributor provides its Contributions) on an "AS IS" BASIS, 146 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or 147 | implied, including, without limitation, any warranties or conditions 148 | of TITLE, NON-INFRINGEMENT, MERCHANTABILITY, or FITNESS FOR A 149 | PARTICULAR PURPOSE. You are solely responsible for determining the 150 | appropriateness of using or redistributing the Work and assume any 151 | risks associated with Your exercise of permissions under this License. 152 | 153 | 8. Limitation of Liability. In no event and under no legal theory, 154 | whether in tort (including negligence), contract, or otherwise, 155 | unless required by applicable law (such as deliberate and grossly 156 | negligent acts) or agreed to in writing, shall any Contributor be 157 | liable to You for damages, including any direct, indirect, special, 158 | incidental, or consequential damages of any character arising as a 159 | result of this License or out of the use or inability to use the 160 | Work (including but not limited to damages for loss of goodwill, 161 | work stoppage, computer failure or malfunction, or any and all 162 | other commercial damages or losses), even if such Contributor 163 | has been advised of the possibility of such damages. 164 | 165 | 9. Accepting Warranty or Additional Liability. While redistributing 166 | the Work or Derivative Works thereof, You may choose to offer, 167 | and charge a fee for, acceptance of support, warranty, indemnity, 168 | or other liability obligations and/or rights consistent with this 169 | License. However, in accepting such obligations, You may act only 170 | on Your own behalf and on Your sole responsibility, not on behalf 171 | of any other Contributor, and only if You agree to indemnify, 172 | defend, and hold each Contributor harmless for any liability 173 | incurred by, or claims asserted against, such Contributor by reason 174 | of your accepting any such warranty or additional liability. 175 | 176 | END OF TERMS AND CONDITIONS 177 | 178 | APPENDIX: How to apply the Apache License to your work. 179 | 180 | To apply the Apache License to your work, attach the following 181 | boilerplate notice, with the fields enclosed by brackets "[]" 182 | replaced with your own identifying information. (Don't include 183 | the brackets!) The text should be enclosed in the appropriate 184 | comment syntax for the file format. We also recommend that a 185 | file or class name and description of purpose be included on the 186 | same "printed page" as the copyright notice for easier 187 | identification within third-party archives. 188 | 189 | Copyright [yyyy] [name of copyright owner] 190 | 191 | Licensed under the Apache License, Version 2.0 (the "License"); 192 | you may not use this file except in compliance with the License. 193 | You may obtain a copy of the License at 194 | 195 | http://www.apache.org/licenses/LICENSE-2.0 196 | 197 | Unless required by applicable law or agreed to in writing, software 198 | distributed under the License is distributed on an "AS IS" BASIS, 199 | WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 200 | See the License for the specific language governing permissions and 201 | limitations under the License. 202 | -------------------------------------------------------------------------------- /README.md: -------------------------------------------------------------------------------- 1 | # Code2Pix: Generating Graphical User Interfaces from Code 2 | 3 | See the original blog post [here](https://medium.com/@ngundotra/code2pix-deep-learning-compiler-for-graphical-user-interfaces-1256c346950b). This is original research for [Uizard](https:/uizard.io) in collaboration with UC Berkeley's Statistics Undergraduate Student Association [(SUSA)](https://github.com/SUSA-org). 4 | 5 | The purpose of this project is to train a differentiable system capable of turning textual description of user interfaces into images of the actual thing. Yes, standard compilers already do this, yes they also go from "code" to "pix". However, this is a useful project because our system, unlike standard compilers, can be used to train other deep learning models. 6 | 7 | Specifically, we believe that code2pix can be used to create a GAN to replace pix2code. This is a direction for future research that is within reach now having completed this project. 8 | 9 | ## Contributions 10 | 11 | All code presented is written by myself. However, the following members contributed to the pre-production code and project: Samyak Parajuli, Luke Dai, Ajay Raj, Aismit Das, Dennis Yang, and Japjot Singh. 12 | 13 | ## Requirements 14 | 15 | All that's needed to run these files is `keras` (2.1.4) and `tensorflow` (1.4.1). 16 | 17 | ## Data 18 | 19 | The data used in this repo comes from the [pix2code repo](https://github.com/tonybeltramelli/pix2code). Our directory structure has the dataset folders in our root directory. 20 | 21 | ``` 22 | code2pix/ 23 | ios/ 24 | training_features/ 25 | eval_set/ 26 | android/ 27 | ... 28 | web/ 29 | ... 30 | ``` 31 | 32 | ## Contents 33 | 34 | * Code2Pix.ipynb 35 | 36 | This notebook has the code necessary to train code2pix and load in the data. It requires that you run Hydra Autoencoder.ipynb first and save a model structure. Although short, it draws upon work that the previous notebooks provided. For the curious, there are portions of this notebook that draw upon pretrained pix2code models, and pretrained autoencoders. To train the pix2code models, you will have to download/clone the original [pix2code repo](https://github.com/tonybeltramelli/pix2code) and go through the setup steps. 37 | 38 | * Hydra Autoencoder.ipynb 39 | 40 | This is where we train our multiheaded autoencoder models that we use in the code2pix model. It's important to run this notebook first before trying out Code2Pix.ipynb. 41 | 42 | * Encoder Decoder Visualization.ipynb 43 | 44 | This is a notebook that allows the creation of GIFs and images that illustrate the inner workings of the autoencoders. These figures were not included in the final blog post. However, they are interesting enough to constitute a short post on their own. If you've never seen latent space visualizations before, I would **highly recommend** checking this notebook out. I'm biased, but I think the figures are really interesting. 45 | --------------------------------------------------------------------------------