In [1]:
import sys
sys.path.append('..')
from deepgraph.utils.logging import log
from deepgraph.utils.common import batch_parallel, ConfigMixin, shuffle_in_unison_inplace, pickle_dump
from deepgraph.utils.image import batch_pad_mirror, rotate_transformer_scalar_float32, rotate_transformer_rgb_uint8
from deepgraph.constants import *
from deepgraph.conf import rng
from deepgraph.nn.core import Dropout
from deepgraph.pipeline import *
In [2]:
import math
class Transformer(Processor):
"""
Apply online random augmentation.
"""
def __init__(self, name, shapes, config, buffer_size=10):
super(Transformer, self).__init__(name, shapes, config, buffer_size)
self.mean = None
def init(self):
if self.conf("mean_file") is not None:
self.mean = np.load(self.conf("mean_file"))
else:
log("Transformer - No mean file specified.", LOG_LEVEL_WARNING)
def process(self):
packet = self.pull()
# Return if no data is there
if not packet:
return False
# Unpack
data, label = packet.data
# Do processing
log("Transformer - Processing data", LOG_LEVEL_VERBOSE)
i_h = 228
i_w = 304
d_h = 228
d_w = 304
start = time.time()
# Mean
if packet.phase == PHASE_TRAIN or packet.phase == PHASE_VAL:
data = data.astype(np.float32)
if self.mean is not None:
for idx in range(data.shape[0]):
# Subtract mean
data[idx] = data[idx] - self.mean.astype(np.float32)
if self.conf("offset") is not None:
label -= self.conf("offset")
if packet.phase == PHASE_TRAIN:
# Do elementwise operations
data_old = data
label_old = label
data = np.zeros((data_old.shape[0], data_old.shape[1], i_h, i_w), dtype=np.float32)
label = np.zeros((label_old.shape[0], d_h, d_w), dtype=np.float32)
for idx in range(data.shape[0]):
# Rotate
# We rotate before cropping to be able to get filled corners
# Maybe even adjust the border after rotating
deg = np.random.randint(-5,6)
# Operate on old data. Careful - data is already in float so we need to normalize and rescale afterwards
# data_old[idx] = 255. * rotate_transformer_rgb_uint8(data_old[idx] * 0.003921568627, deg).astype(np.float32)
# label_old[idx] = rotate_transformer_scalar_float32(label_old[idx], deg)
# Take care of any empty areas, we crop on a smaller surface depending on the angle
# TODO Remove this once loss supports masking
shift = 0 #np.tan((deg/180.) * math.pi)
# Random crops
cy = rng.randint(data_old.shape[2] - d_h - shift, size=1)
cx = rng.randint(data_old.shape[3] - d_w - shift, size=1)
data[idx] = data_old[idx, :, cy:cy+i_h, cx:cx+i_w]
label[idx] = label_old[idx, cy:cy+d_h, cx:cx+d_w]
# Flip horizontally with probability 0.5
p = rng.randint(2)
if p > 0:
data[idx] = data[idx, :, :, ::-1]
label[idx] = label[idx, :, ::-1]
# RGB we mult with a random value between 0.8 and 1.2
r = rng.randint(80,121) / 100.
g = rng.randint(80,121) / 100.
b = rng.randint(80,121) / 100.
data[idx, 0] = data[idx, 0] * r
data[idx, 1] = data[idx, 1] * g
data[idx, 2] = data[idx, 2] * b
# Shuffle
# data, label = shuffle_in_unison_inplace(data, label)
elif packet.phase == PHASE_VAL:
# Center crop
cy = (data.shape[2] - i_h) // 2
cx = (data.shape[3] - i_w) // 2
data = data[:, :, cy:cy+i_h, cx:cx+i_w]
label = label[:, cy:cy+d_h, cx:cx+d_w]
end = time.time()
log("Transformer - Processing took " + str(end - start) + " seconds.", LOG_LEVEL_VERBOSE)
# Try to push into queue as long as thread should not terminate
self.push(Packet(identifier=packet.id, phase=packet.phase, num=2, data=(data, label)))
return True
def setup_defaults(self):
super(Transformer, self).setup_defaults()
self.conf_default("mean_file", None)
self.conf_default("offset", None)
In [ ]:
from theano.tensor.nnet import relu
from deepgraph.graph import *
from deepgraph.nn.core import *
from deepgraph.nn.conv import *
from deepgraph.nn.loss import *
from deepgraph.solver import *
from deepgraph.nn.init import *
from deepgraph.pipeline import Optimizer, H5DBLoader, Pipeline
def build_graph():
graph = Graph("unet")
data = Data(graph, "data", T.ftensor4, shape=(-1, 3, 228, 304))
label = Data(graph, "label", T.ftensor3, shape=(-1, 1, 228, 304), config={
"phase": PHASE_TRAIN
})
conv_c1_1 = Conv2D(graph, "conv_c1_1", config={
"channels": 64,
"kernel": (3, 3),
"border_mode": 1,
"activation": relu,
"weight_filler": xavier(gain="relu"),
"bias_filler": constant(0)
}
)
conv_c1_2 = Conv2D(graph, "conv_c1_2", config={
"channels": 64,
"kernel": (3, 3),
"border_mode": 1,
"activation": relu,
"weight_filler": xavier(gain="relu"),
"bias_filler": constant(0)
}
)
pool_c1 = Pool(graph, "pool_c0", config={
"kernel": (2, 2)
})
conv_c2_1 = Conv2D(graph, "conv_c2_1", config={
"channels": 128,
"kernel": (3, 3),
"border_mode": 1,
"activation": relu,
"weight_filler": xavier(gain="relu"),
"bias_filler": constant(0)
}
)
conv_c2_2 = Conv2D(graph, "conv_c2_2", config={
"channels": 128,
"kernel": (3, 3),
"border_mode": 1,
"activation": relu,
"weight_filler": xavier(gain="relu"),
"bias_filler": constant(0)
}
)
up_e2 = Upsample(graph, "up_e2", config={
"kernel": (2, 2)
})
up_conv_e2 = Conv2D(graph, "up_conv_e2", config={
"channels": 64,
"kernel": (3, 3),
"border_mode": 1,
"activation": None,
"weight_filler": xavier(),
"bias_filler": constant(0)
}
)
concat_1 = Concatenate(graph, "concat_1", config={
"axis": 1
})
conv_e1_1 = Conv2D(graph, "conv_e1_1", config={
"channels": 64,
"kernel": (3, 3),
"border_mode": 1,
"activation": relu,
"weight_filler": xavier(gain="relu"),
"bias_filler": constant(0)
}
)
conv_e1_2 = Conv2D(graph, "conv_e1_2", config={
"channels": 64,
"kernel": (3, 3),
"border_mode": 1,
"activation": relu,
"weight_filler": xavier(gain="relu"),
"bias_filler": constant(0)
}
)
conv_e_f= Conv2D(graph, "conv_e_f", config={
"channels": 1,
"kernel": (1, 1),
"activation": None,
"weight_filler": xavier(),
"bias_filler": constant(0)
}
)
loss = EuclideanLoss(graph, "loss")
error = MSE(graph, "mse", config={
"root": True,
"is_output": True,
"phase": PHASE_TRAIN
})
# Connect
data.connect(conv_c1_1)
conv_c1_1.connect(conv_c1_2)
conv_c1_2.connect(concat_1)
conv_c1_2.connect(pool_c1)
pool_c1.connect(conv_c2_1)
conv_c2_1.connect(conv_c2_2)
conv_c2_2.connect(up_e2)
up_e2.connect(up_conv_e2)
up_conv_e2.connect(concat_1)
concat_1.connect(conv_e1_1)
conv_e1_1.connect(conv_e1_2)
conv_e1_2.connect(conv_e_f)
conv_e_f.connect(loss)
conv_e_f.connect(error)
label.connect(loss)
label.connect(error)
return graph
In [ ]:
if __name__ == "__main__":
batch_size = 16
chunk_size = 10*batch_size
transfer_shape = ((chunk_size, 3, 228, 304), (chunk_size, 228, 304))
g = build_graph()
# Build the training pipeline
db_loader = H5DBLoader("db", ((chunk_size, 3, 480, 640), (chunk_size, 1, 480, 640)), config={
"db": '/home/ga29mix/nashome/data/nyu_depth_v2_combined_50.hdf5',
# "db": '../data/nyu_depth_unet_large.hdf5',
"key_data": "images",
"key_label": "depths",
"chunk_size": chunk_size
})
transformer = Transformer("tr", transfer_shape, config={
# Measured for the data-set
# "offset": 2.7321029
"mean_file": "/home/ga29mix/nashome/data/nyu_depth_v2_combined_50.npy"
})
optimizer = Optimizer("opt", g, transfer_shape, config={
"batch_size": batch_size,
"chunk_size": chunk_size,
"learning_rate": 0.0001,
"momentum": 0.9,
"weight_decay": 0.0005,
"print_freq": 50,
"save_freq": 10000,
"weights": "../data/unet_test_two_paths_only_iter_20000.zip",
"save_prefix": "../data/unet_test_two_paths_only"
})
p = Pipeline(config={
"validation_frequency": 50,
"cycles": 6200
})
p.add(db_loader)
p.add(transformer)
p.add(optimizer)
p.run()
In [ ]:
g = build_graph()
g.load_weights("../data/alexnet_scale_1_iter_19000.zip")
g.compile()
In [ ]:
import h5py, numpy as np
f = h5py.File("/home/ga29mix/nashome/data/nyu_depth_v2_combined_50.hdf5")
b = int(f["images"].shape[0] * 0.9)
images = np.array(f["images"][b:])
depths = np.array(f["depths"][b:])
print images.shape
mean = np.load("/home/ga29mix/nashome/data/nyu_depth_v2_combined_50.npy")
In [ ]:
%matplotlib inline
import matplotlib.pyplot as plt
from deepgraph.nn.core import Dropout
w = 304
h = 228
plot = True
idx = 100
diffs = []
Dropout.set_dp_off()
for image in images[100:120]:
tmp = image.astype(np.float32)
tmp -= mean
cy = (tmp.shape[1] - h) // 2
cx = (tmp.shape[2] - w) // 2
crop = tmp[:,cy:cy+h, cx:cx+w]
res = g.infer([crop.reshape((1,3,228,304))])["reshape_0"]
res = res.squeeze()
depth = depths[idx][cy:cy + h, cx:cx + w]
depth = depth[::4,::4]
if plot and idx % 5 == 0:
plt.imshow(image.transpose((1,2,0)).astype(np.uint8))
plt.show()
plt.imshow(depth)
plt.show()
plt.imshow(res)
plt.show()
print "RMSE: " + str(np.sqrt(np.mean((res-depth)**2)))
diffs.append(res - depth)
idx += 1
diffs = np.array(diffs)
rmse = np.sqrt(np.mean(diffs ** 2))
print "Accumulated RMSE: " + str(rmse)
In [ ]: