deepdrive.py

import os
import math
import numpy as np
import h5py
import glob
from tqdm import tqdm
import scipy

import keras
from keras.models import Sequential, Model
from keras.layers.core import Dense, Dropout, Activation, Flatten, Reshape
from keras.layers import Embedding, Input, merge, ELU
from keras.layers.recurrent import SimpleRNN, LSTM
from keras.layers.convolutional import Convolution2D, MaxPooling2D
from keras.optimizers import SGD, Adam, RMSprop
import sklearn.metrics as metrics

import matplotlib.pyplot as plt
plt.ion()

# Read in deep drive data
dfiles = glob.glob('data/*.h5')

# 999 data points, images are 3x227x227
# vehicle_states are just target at previous time
# no distance finder, but this is enough to work with
dfile = dfiles[9]
h5f = h5py.File(dfile,'r')
A = dict(h5f.items()) 

# view image
plt.imshow(A['images'].value[0].swapaxes(0,2).swapaxes(0,1))
AA = A['images'].value

# look at targets
A['targets'].value[:2]
A['vehicle_states'].value[:2]

at = A['targets'].value[:-1]
av = A['vehicle_states'].value[1:]

# slim inputs
B = np.array(A['images'].value,dtype=np.uint8)
plt.imshow(B[0][::-1,:,:].transpose((1,2,0)))

B = B[:,::-1,:,:]

bb = scipy.misc.imresize(B[0],(128,128),'cubic','RGB')
plt.imshow(bb)

bb = scipy.misc.imresize(B[0],(64,64),'cubic','RGB')
plt.imshow(bb)


# determine scaling parameters
# speed and accel
speedmax = None
speedmin = None
accelmax = None
accelmin = None
for dfile in dfiles:
    with h5py.File(dfile,'r') as h5f:
        # raw data
        A = dict(h5f.items()) 
        smx = np.max(A['vehicle_states'].value[:,2])
        smn = np.min(A['vehicle_states'].value[:,2])
        amx = np.max(A['vehicle_states'].value[:,3])
        amn = np.min(A['vehicle_states'].value[:,3])
        if speedmax is None or smx > speedmax:
            speedmax = smx
        if speedmin is None or smn < speedmin:
            speedmin = smn
        if accelmax is None or amx > accelmax:
            accelmax = amx
        if accelmin is None or amn < accelmin:
            accelmin = amn
        #plt.plot(A['targets'].value[:,3],A['targets'].value[:,5],'.')


# throttle was supposed to be zero to 1, but has negative values, probably -1 to 1

# steering is nominally -1 to 1, but doesn't go below zero?
throttlemax = None
throttlemin = None
steermax = None
steermin = None
for dfile in dfiles:
    with h5py.File(dfile,'r') as h5f:
        # raw data
        A = dict(h5f.items()) 
        smx = np.max(A['targets'].value[:,4])
        smn = np.min(A['targets'].value[:,4])
        tmx = np.max(A['targets'].value[:,5])
        tmn = np.min(A['targets'].value[:,5])
        if steermax is None or smx > steermax:
            steermax = smx
        if steermin is None or smn < steermin:
            steermin = smn
        if throttlemax is None or tmx > throttlemax:
            throttlemax = tmx
        if throttlemin is None or tmn < throttlemin:
            throttlemin = tmn


ndata = 0
imgsize = 64
# frame size
nrows = 64
ncols = 64

# speed, accel, distance, angle
real_in = Input(shape=(2,), name='real_input')

# video frame in, grayscale
frame_in = Input(shape=(3,nrows,ncols))

# convolution for image input
conv1 = Convolution2D(8,5,5,border_mode='same')
conv_l1 = conv1(frame_in)
Econv_l1 = ELU()(conv_l1)
pool_l1 = MaxPooling2D(pool_size=(2,2))(Econv_l1)

conv2 = Convolution2D(8,5,5,border_mode='same')
conv_l2 = conv2(pool_l1)
Econv_l2 = ELU()(conv_l2)
pool_l2 = MaxPooling2D(pool_size=(2,2))(Econv_l2)

flat = Flatten()(pool_l2)

M = merge([flat,real_in], mode='concat', concat_axis=1)

D1 = Dense(64)(M)
ED1 = ELU()(D1)
D2 = Dense(32)(ED1)
ED2 = ELU()(D2)
D3 = Dense(32)(ED2)
ED3 = ELU()(D3)

A1 = Dense(32)(ED3)
EA1 = ELU()(A1)
S1 = Dense(32)(ED3)
ES1 = ELU()(S1)

Accel = Dense(1, activation='sigmoid')(EA1)
Steer = Dense(1, activation='sigmoid')(ES1)

model = Model(input=[real_in, frame_in], output=[Accel,Steer])

model.compile(loss='mean_squared_error',
              optimizer='adam',
              metrics=['accuracy'])

# Fake data
#nsamples = 1000
#fake_real = np.random.random((nsamples,2))
#fake_frame = np.random.random((nsamples,3,nrows,ncols))

#fake_A = np.random.random(nsamples)
#fake_P = np.random.random(nsamples)

#h = model.fit([fake_real, fake_frame], [fake_A, fake_P],
#        batch_size = 32, nb_epoch=10, verbose=1,
#        validation_split=0.1)
#h = model.predict([fake_real, fake_frame],
#        batch_size = 32, verbose=1)
# this produces distinct values

# batch process (fitting, really) one file at a time
for dfile in dfiles:
    with h5py.File(dfile,'r') as h5f:
        # raw data
        A = dict(h5f.items()) 
        # extract images in 1-byte format
        B = np.array(A['images'].value,dtype=np.float16)/255.
        # change BGR to RGB
        B = B[:,::-1,:,:]
        # Scale down image size
        imgs = np.zeros((len(B),3,64,64),dtype=np.float16)
        for i,b in enumerate(B):
            imgs[i] = scipy.misc.imresize(b,(64,64),'cubic','RGB').transpose((2,0,1))
        # speed and accel scale
        speedx = A['vehicle_states'].value[:,2:4]
        speedx[:,0] = (speedx[:,0] - speedmin) / (speedmax-speedmin)
        speedx[:,1] = (speedx[:,1] - accelmin) / (accelmax-accelmin)
        # throttle and steering scale
        targets = (A['targets'].value[:,4:] + 1) / 2.
        # Train while we have this file open
        h = model.fit([speedx, imgs], [targets[:,1], targets[:,0]],
                batch_size = 32, nb_epoch=10, verbose=1,
                validation_split=0.1)

W = model.get_weights()

# look at conv filters separately in color channel
f, con = plt.subplots(4,3, sharex='col', sharey='row')
for row in range(4):
    for col in range(3):
        con[row,col].pcolormesh(W[0][row,col],cmap=plt.cm.hot)


# combine color channels into on filter image
f, con = plt.subplots(4,1, sharex='col', sharey='row')
for row in range(4):
    con[row].imshow(W[0][row].transpose((1,2,0)),
                    interpolation="none")

# View the correct and predicted steering angle and accel
# Draw a line from center to outside, length is accel, angle is steering
all_pred = list()
for dfile in dfiles:
    with h5py.File(dfile,'r') as h5f:
        # raw data
        A = dict(h5f.items()) 
        # extract images in 1-byte format
        B = np.array(A['images'].value,dtype=np.uint8)/255.
        # change BGR to RGB
        B = B[:,::-1,:,:]
        # Scale down image size
        imgs = np.zeros((len(B),3,64,64),dtype=np.uint8)
        for i,b in enumerate(B):
            imgs[i] = scipy.misc.imresize(b,(64,64),'cubic','RGB').transpose((2,0,1))
        # speed and accel
        speedx = A['vehicle_states'].value[:,2:4]
        speedx[:,0] = (speedx[:,0] - speedmin) / (speedmax-speedmin)
        speedx[:,1] = (speedx[:,1] - accelmin) / (accelmax-accelmin)
        # throttle and steering scale
        targets = (A['targets'].value[:,4:] + 1) / 2.
        # Train while we have this file open
        h = model.predict([speedx, imgs],
                batch_size = 32, verbose=1)
        all_pred.append(h)

# Make sure we get a variety of values
plt.plot([np.array([p[0][i] for p in all_pred]).reshape(11) for i in range(999)],'.')
plt.plot([np.array([p[1][i] for p in all_pred]).reshape(11) for i in range(999)],'.')



# Should look at intermediate values to see where this goes wrong
# throttle and steering may have been mixed up