Image Classification with TFLearn
In this article you'll learn how to train a neural network to classify images using the TFLearn framework.
Courtesy of PublicDomainPictures.net.
That there is a dog. Most humans would know it's a dog, right? But can we train a computer to recognize that it's a dog?
Computer: I think that's a nil↩.
Well hey now.
labels = setup_cifar_labels() img = tf.read_file(Image↩) img = tf.image.decode_jpeg(img, channels=3) img.set_shape([None, None, 3]) img = tf.image.resize_images(img, (32, 32)) img = img.eval(session=sess) # convert to numpy array img = np.expand_dims(img, 0) # make 'batch' of 1 img = img/255.0 pred = model.predict(img) pred = labels["label_names"][np.argmax(pred)] pred
To train a computer to identify a photograph of a dog, we've made use of the the TFLearn API for the Tensorflow framework.
The TFLearn project is a higher-level API on top of Tensorflow, a symbolic math library which is widely used for machine learning and neural network tasks. TFLearn facilitates faster and more natural construction of neural networks, as well as easy training, evaluation, and prediction. In this article, we'll use TFLearn to classify the CIFAR-10 dataset, which is 10% dog pictures.
1. Setup
To run TFLearn we'll transclude a base environment with Tensorflow, and use conda to add in the Python HDF5 package, cython, pillow, and dill. We'll get TFLearn via pip, but we want the development version so we'll also need to install git for pip to use.
apt-get -qq update DEBIAN_FRONTEND=noninteractive apt-get install --no-install-recommends git apt-get clean rm -r /var/lib/apt/lists/* conda install -y -c anaconda -c conda-forge \ h5py dill cython pillow ldconfig pip install git+https://github.com/tflearn/tflearn.git
The CIFAR-10 dataset is a collection of 60,000 color, 32x32-pixel images in ten classes, 10,000 of which are in a test batch. TFLearn can automatically download the dataset, but let's save some time by doing it ourselves and uploading the file to the article's permanent storage.
Then we'll just have to copy it to our local filesystem and extract it so that TFLearn will pick it up.
A lot of initilization code will be needed multiple times, so we'll put it in its own runtime and then import the desired code cells where we need it.
We'll run 128-image batches and set up two training runs: a long, 500-epoch run to do the main work, and a short, 5-epoch run as an example.
# Residual blocks # 32 layers: n=5, 56 layers: n=9, 110 layers: n=18 n = 5 batch_size = 1000 num_classes = 10 epochs_shortrun = 5 epochs_longrun = 200 save_dir = "/files" res_dir = "/results" model_name = 'resnet_cifar10' import os save_fn = model_name + ".tfsave" save_file = os.path.join(save_dir, save_fn) if not os.path.isdir(res_dir): os.makedirs(res_dir) ckpt_dir = os.path.join(save_dir,"checkpoints") if not os.path.isdir(ckpt_dir): os.makedirs(ckpt_dir) tblog_dir = os.path.join(save_dir,"tflogs") if not os.path.isdir(tblog_dir): os.makedirs(tblog_dir) event_dir = os.path.join(tblog_dir,model_name)
Load the data and get it into a reasonable shape. Also set up a function to find the best checkpoint file, another to give us a look at the images we're analyzing, and finally set up to do real-time input-data augmentation.
#from __future__ import division, print_function, absolute_import import numpy as np import dill as pickle from math import * def setup_tf(): # set random seeds for reproducibility from tflearn import init_graph import tensorflow as tf tf.reset_default_graph() init_graph(seed=343) tf.set_random_seed(343) np.random.seed(343) def setup_cifar_data(verbose=False): from tflearn.datasets import cifar10 from tflearn.data_utils import to_categorical datadir = "/cifar/" datafile = datadir+"cifar-10-python.tar.gz" if not os.path.isfile(datafile): import shutil,tarfile os.makedirs(datadir, exist_ok=True) shutil.copyfile(cifar-10-python.tar.gz↩, datafile) with tarfile.open(datafile, "r:gz") as f: f.extractall(datadir) # Data loading (X, Y), (testX, testY) = cifar10.load_data(dirname="/cifar") Y = to_categorical(Y, num_classes) testY = to_categorical(testY, num_classes) return X, Y, testX, testY def setup_cifar_labels(): with open("/cifar/cifar-10-batches-py/batches.meta", 'rb') as fo: labels = pickle.load(fo) return labels # Function to find latest checkpoint file def last_ckpt(dir): fl = os.listdir(dir) fl = [x for x in fl if x.endswith(".index")] cf = "" if len(fl) > 0: steps = [float(x.split("-")[1][0:-6]) for x in fl] m = max(steps) cf = fl[steps.index(m)] cf = os.path.join(dir,cf) return(cf) def load_model_from_file(model,file): # load data from tarfile to model import tarfile with tarfile.open(file, "r:bz2") as tar: try: tar.getmember(save_fn+".index") tar.getmember(save_fn+".meta") tar.getmember(save_fn+".data-00000-of-00001") except KeyError: print("Minimum training results files not found!\n") tar.extractall(path=save_dir) print("Loading {}...".format(save_file)) model.load(save_file, weights_only=False) def cifar_grid(Xset,Yset,inds,n_col, predictions=None): #Visualizing CIFAR 10, takes indicides and shows in a grid import matplotlib.pyplot as plt if predictions is not None: if Yset.shape != predictions.shape: print("Predictions must equal Yset in length!") return(None) N = len(inds) n_row = int(ceil(1.0*N/n_col)) fig, axes = plt.subplots(n_row,n_col,figsize=(10,10)) clabels = labels["label_names"] for j in range(n_row): for k in range(n_col): i_inds = j*n_col+k i_data = inds[i_inds] axes[j][k].set_axis_off() if i_inds < N: axes[j][k].imshow(Xset[i_data,...], interpolation='nearest') label = clabels[np.argmax(Yset[i_data,...])] axes[j][k].set_title(label) if predictions is not None: pred = clabels[np.argmax(predictions[i_data,...])] if label != pred: label += " n" axes[j][k].set_title(pred, color='red') fig.set_tight_layout(True) return fig
Let's look at a random selection of the dataset images.
x_train, y_train, x_test, y_test = setup_cifar_data(verbose=True) labels = setup_cifar_labels() indices = [np.random.choice(range(len(x_train))) for i in range(36)] cifar_grid(x_train,y_train,indices,6)
The model we're using is a deep residual network design taken from the TFLearn example scripts.
from tflearn import ImagePreprocessing, ImageAugmentation from tflearn import input_data, DNN from tflearn import conv_2d, residual_block from tflearn import batch_normalization, activation, global_avg_pool from tflearn import fully_connected, Momentum, regression from tflearn.callbacks import Callback print("Using real-time data augmentation.\n") # Real-time data preprocessing img_prep = ImagePreprocessing() img_prep.add_featurewise_zero_center(per_channel=True, mean=[ 0.49139968, 0.48215841, 0.44653091 ]) # Real-time data augmentation img_aug = ImageAugmentation() img_aug.add_random_flip_leftright() img_aug.add_random_crop([32, 32], padding=4) # Building Residual Network net = input_data(shape=[None, 32, 32, 3], data_preprocessing=img_prep, data_augmentation=img_aug) net = conv_2d(net, 16, 3, regularizer='L2', weight_decay=0.0001) net = residual_block(net, n, 16) net = residual_block(net, 1, 32, downsample=True) net = residual_block(net, n-1, 32) net = residual_block(net, 1, 64, downsample=True) net = residual_block(net, n-1, 64) net = batch_normalization(net) net = activation(net, 'relu') net = global_avg_pool(net) # Regression net = fully_connected(net, num_classes, activation='softmax') mom = Momentum(0.1, lr_decay=0.1, decay_step=32000, staircase=True) net = regression(net, optimizer=mom, loss='categorical_crossentropy') # Initialize model ckpt_file = os.path.join(ckpt_dir,"model.ckpt") model = DNN(net, checkpoint_path=ckpt_file, max_checkpoints=10, clip_gradients=0., tensorboard_dir=tblog_dir,tensorboard_verbose=0) # disabled until directories can be written to /results #cff = last_ckpt(ckpt_dir) #if cff != "": # print("Loading ",cff,"...") # model.load(cff) # define the early-stop callback class EarlyStoppingCallback(Callback): def __init__(self, val_loss_thresh, val_loss_patience): """ minimum loss improvement setup """ self.val_loss_thresh = val_loss_thresh self.val_loss_last = float('inf') self.val_loss_patience = val_loss_patience self.val_loss_squint = 0 def on_batch_end(self, training_state, snapshot=False): """ loss improvement threshold w/ patience """ # Apparently this can happen. if training_state.val_loss is None: return if (self.val_loss_last - training_state.val_loss) < self.val_loss_thresh: # unacceptable! if self.val_loss_squint >= self.val_loss_patience: raise StopIteration else: self.val_loss_squint += 1 else: # we good again - reset self.val_loss_last = training_state.val_loss self.val_loss_squint = 0 # Initializae our callback. early_stopping_cb = EarlyStoppingCallback( val_loss_thresh=0.001, val_loss_patience=25 )
2. Training
Now we're ready to train using the GPU. The training cell will save to /results/, and then for analysis and visualization we'll just need to load that data.
import tarfile, glob, shutil # makes Tensorflow shush about SSE and such os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2" setup_tf() x_train, y_train, x_test, y_test = setup_cifar_data(verbose=True) print("Starting to train...") # checkpoints disabled until directories can be written to /results try: model.fit(x_train, y_train, n_epoch=epochs_longrun, validation_set=(x_test, y_test), snapshot_epoch=True, snapshot_step=None, show_metric=True, batch_size=batch_size, shuffle=True, run_id='resnet_cifar10', callbacks=early_stopping_cb) except StopIteration: print("Got bored, stopping early.") print("Training complete.") model.save(save_file) # copy events file to /results for history plotting evfiles = list(filter(os.path.isfile, glob.glob(os.path.join(event_dir, "events.out.tfevents.*")))) evfiles.sort(key=lambda x: os.path.getmtime(x)) shutil.copyfile(os.path.join(event_dir,evfiles[-1]), os.path.join(res_dir,model_name+".tfevents")) # can only save single files to /results, so let's tar the saves tar_file = os.path.join(res_dir,model_name)+".tar.bz2" with tarfile.open(tar_file, "w:bz2") as tar: for name in [x for x in os.listdir(save_dir) if x.startswith(save_fn)]: tar.add(os.path.join(save_dir, name), arcname=name)
3. Results
Alrighty, now we can take a look at the trained model. We'll use the definitions from earlier, initialize the model, then load weights from a save file for evaluation and prediction. Currently this requires a GPU instance simply for the increased RAM allocation.
import tensorflow as tf import tflearn # makes Tensorflow shush about SSE and such os.environ["TF_CPP_MIN_LOG_LEVEL"] = "2" os.environ["CUDA_VISIBLE_DEVICES"] = "-1" tf.device('/cpu:0') setup_tf() x_train, y_train, x_test, y_test = setup_cifar_data(verbose=True) sess = tf.Session() tflearn.is_training(False, session=sess) load_model_from_file(model, resnet_cifar10.tar.bz2train↩)
acc = model.evaluate(x_test, y_test) # While we've got it set up, run predictions # for the test batch and save to file x_test_copy = np.copy(x_test) # copy because predict() modifies input (bug?) y_pred = model.predict(x_test_copy) with open("/results/test_predictions.dat","wb") as f: pickle.dump(y_pred,f) print("Average accuracy: {:.2f}%.".format(acc[0]*100))
We can sample the prediction results with images.
with open(test_predictions.dat↩,"rb") as f: y_pred = pickle.load(f) indices = [np.random.choice(range(len(x_test))) for i in range(36)] labels = setup_cifar_labels() cifar_grid(x_test,y_test,indices,6, predictions=y_pred)
TFLearn tells Tensorflow to write to tfevents files during training. These are geared towards the interactive analysis and visualization program tensorboard, but with a bit of work we can pull the training history out of these files and plot it ourselves.
from tensorboard.backend.event_processing import event_accumulator import shutil # need to copy out to get .tfevents extension, because...raisins shutil.copy(resnet_cifar10.tfeventstrain↩,"/tmp/hist.tfevents") ea = event_accumulator.EventAccumulator("/tmp/hist.tfevents", size_guidance={ # see below regarding this argument event_accumulator.SCALARS: 0 }) ea.Reload() # loads events from file # fiddly stuff to inspect tags/scalar data entries #print(ea.Tags()) #print([x for x in ea.Tags()['scalars'] if not x.startswith("Momentum")]) # pull out four metrics, plot hist = { 'Accuracy' : [x.value for x in ea.Scalars('Accuracy')], 'Validation Accuracy' : [x.value for x in ea.Scalars('Accuracy/Validation')], 'Loss' : [x.value for x in ea.Scalars('Loss')], 'Validation Loss' : [x.value for x in ea.Scalars('Loss/Validation')] } import matplotlib import matplotlib.pyplot as plt fig = plt.figure() keys = ['Accuracy', 'Loss', 'Validation Accuracy', 'Validation Loss'] for i,thing in enumerate(keys): trace = hist[thing] plt.subplot(2,2,i+1) plt.plot(range(len(trace)),trace) plt.title(thing) fig.set_tight_layout(True) fig
Finally, with a trained network we can bring in new images and see how the model does classifying those. We go back to our opening image of a fluffy white dog.
labels = setup_cifar_labels() img = tf.read_file(Image↩) img = tf.image.decode_jpeg(img, channels=3) img.set_shape([None, None, 3]) img = tf.image.resize_images(img, (32, 32)) img = img.eval(session=sess) # convert to numpy array img = np.expand_dims(img, 0) # make 'batch' of 1 img = img/255.0 pred = model.predict(img) pred = labels["label_names"][np.argmax(pred)] pred
Our model assures us once again that this is a nil↩.