def __init__(self,
activation,
checkpoint_dir,
trainingset_dir,
w_reconstruction=[1., 1.],
w_decision=[1., 1.],
w_reg=1e-10,
w_rate=0,
beta0=None,
beta_steps=200000,
sampling_off=False,
dataset='attention_search_both2',
input_distribution_dim=-1,
batch_size=64,
dropout_prob=0.5,
regenerate_steps=1000,
dataset_size=5000):
if "attention" not in dataset:
raise Exception(
"{} is not an acceptable training task.".format(dataset))
self.name = "popout"
activation_dict = {
"relu": tf.nn.relu,
"sigmoid": tf.nn.sigmoid,
"tanh": tf.nn.tanh,
"elu": tf.nn.elu
}
self.dataset = dataset
self.dataset_size = int(dataset_size)
self.iterator = None
self.next_element = None
self.batch_size = batch_size
self.regenerate_steps = regenerate_steps
self.decision_dim = 2
# Activation function
self.activation = activation_dict[activation]
self.image_width = 32
self.RGB = True
self.image_channels = 3
self.latent_size = 500
# Size of hidden layer(s)
self.w_rate = w_rate # Initialize to this value, but possibly change
self.beta1 = w_rate # Final value for rate_loss_weight
self.beta0 = beta0 # Initial value for rate_loss_weight
if self.beta0 is not None and self.beta0 > self.beta1:
raise Exception("beta1 must be greater than beta0")
self.beta_steps = beta_steps
self.w_reconstruction_enc, self.w_reconstruction_dec = w_reconstruction
self.w_decision_enc, self.w_decision_dec = w_decision
self.w_reg = w_reg
self.dropout_prob = dropout_prob
self.sampling_off = sampling_off
self.checkpoint_top_dir = Path(checkpoint_dir)
self.trainingset_dir = Path(trainingset_dir)
self.checkpoint_dir = make_memnet_checkpoint_dir(checkpoint_dir, self)
print("Using {} activation".format(activation))
print("Loss weights (rate, reconstruction, decision): ",
(self.w_rate, w_reconstruction, w_decision))
self.sess = None
def visualize_network_samples_grid_1D(net,
start=5,
stop=100,
step=10,
f_ext="pdf"):
"""
[PLANTS STIMULI ONLY]
Like the 2D version above, but just a single line of images (just one
stimulus dimension).
"""
save_dir = make_memnet_checkpoint_dir(
Path("plots/vis_memnet_samples/grid/"), net)
if not save_dir.exists():
save_dir.makedirs()
relevant_vals = range(start, stop, step)
grid_width = len(relevant_vals)
print("Loading plants.")
inputs0, stim_vals0 = load_plants(net.image_width,
net.trainingset_dir,
layer_type=net.layer_type,
normalize=True)
inputs0 = inputs0 / 255. # Normalize
print("Finished.")
irrelevant_dim = int(net.input_distribution_dim == 0)
if irrelevant_dim == 0:
inds = list(product([50], relevant_vals))
else:
inds = list(product(relevant_vals, [50]))
inputs_valid = inputs0[stim_vals0[:, irrelevant_dim] == 50]
stim_vals = stim_vals0[stim_vals0[:, irrelevant_dim] == 50]
fig, axes = plt.subplots(1,
grid_width,
figsize=(50, 50 / len(relevant_vals)))
preds = []
inputs = []
print("Making plots...")
for i in range(len(stim_vals)):
if tuple(stim_vals[i]) in inds:
print("Making ", stim_vals[i])
inputs.append(inputs_valid[i])
preds = net.predict(inputs, keep_session=True)
for i in range(len(preds)):
ax = axes[i]
ax.imshow(preds[i:i + 1].reshape(net.image_width, net.image_width),
cmap="gray_r")
ax.set_xticklabels("")
ax.set_xticks([])
ax.set_xticklabels("")
ax.set_yticks([])
ax.set_yticklabels("")
plt.savefig(save_dir.joinpath('grid.' + f_ext))
print("Saved to ", save_dir.joinpath('grid.' + f_ext))
return
def plants_average_reconstruction(net,
plant_coords=[[0, 0], [0, 99], [99, 99],
[99, 0]],
n_samples=100,
f_ext="pdf"):
"""For the specified plants (given by `plant_coords`), plot
the target image alongside the mean reconstructed image, averaged
over `n_samples` samples.
"""
images, stim_vals = load_plants(net.image_width,
net.trainingset_dir,
layer_type=net.layer_type,
normalize=True)
save_dir = make_memnet_checkpoint_dir(
Path("plots/plants_average_reconstruction/"), net)
if not save_dir.exists():
save_dir.mkdir()
for coord in plant_coords:
# Iterate over all stimulus values for target image
i_target = np.logical_and(coord[0] == stim_vals[:, 0],
coord[1] == stim_vals[:, 1])
x = images[i_target]
X = np.repeat(x, n_samples, axis=0) # Take n samples
Xhat = net.predict(X, keep_session=True)
mean_reconstruction = Xhat.mean(axis=0) # Take mean over samples
fig, ax = plt.subplots(nrows=1, ncols=2, figsize=(5.6, 3.1))
ax[0].imshow(x.reshape(net.image_width, net.image_width),
cmap="gray_r")
ax[1].imshow(mean_reconstruction.reshape(net.image_width,
net.image_width),
cmap="gray_r")
ax[0].set_xlabel("Target", fontsize=14)
ax[1].set_xlabel("Mean reconstruction", fontsize=14)
ax[0].set_xticks([])
ax[1].set_xticks([])
ax[0].set_xticklabels("")
ax[1].set_xticklabels("")
ax[0].set_yticks([])
ax[1].set_yticks([])
ax[0].set_yticklabels("")
ax[1].set_yticklabels("")
fig.suptitle("Leaf width {}, leaf angle {}".format(coord[0], coord[1]))
plt.tight_layout()
# plt.subplots_adjust(top=0.9)
save_pth = save_dir.joinpath("width{}_angle{}.{}".format(
coord[0], coord[1], f_ext))
plt.savefig(save_pth, dpi=300)
print("Saved to " + save_dir)
return
def visualize_network_samples_grid(net,
start=5,
stop=100,
step=10,
f_ext="pdf"):
"""
[PLANTS STIMULI ONLY]
Plot grid of images, where (x, y) position corresponds to the stimulus
values (width, droop) of the input to the network. Images are the outputs
of the network given the input.
"""
save_dir = make_memnet_checkpoint_dir(
Path("plots/vis_memnet_samples/grid/"), net)
if not save_dir.exists():
save_dir.makedirs()
inds = list(product(range(start, stop, step), range(start, stop, step)))
grid_width = len(range(start, stop, step))
print("Loading plants.")
inputs, stim_vals = load_plants(net.image_width,
net.trainingset_dir,
layer_type=net.layer_type,
normalize=True)
print("Finished.")
fig, axes = plt.subplots(grid_width, grid_width, figsize=(50, 50))
preds = []
print("Making plots...")
for i in range(len(stim_vals)):
if tuple(stim_vals[i]) in inds:
print("Making ", stim_vals[i])
preds.append(
net.predict(inputs[i:i + 1],
keep_session=True).reshape(net.image_width,
net.image_width))
grid_coord = tuple(stim_vals[i] / step)
ax = axes[int(grid_coord[1]), int(grid_coord[0])]
# ax = axes[tuple(stim_vals[i] / step)]
ax.imshow(preds[-1], cmap="gray_r")
ax.set_xticklabels("")
ax.set_xticks([])
ax.set_xticklabels("")
ax.set_yticks([])
ax.set_yticklabels("")
plt.savefig(save_dir.joinpath('grid.' + f_ext))
print("Saved to ", save_dir.joinpath('grid.' + f_ext))
return
def correlation(net, start=0, stop=100, step=10, n_samples=2, f_ext="pdf"):
"""An analysis for networks trained with non-uniform input distributions.
(Designed for 'plants' experiment)
Correlate reconstructed images with all original stimulus images.
Since one plant dimension will be uniform and one will be non-uniform,
marginalize over the non-uniform dimension
n_samples: Number of network samples to average over
"""
from scipy.stats import pearsonr
from sklearn.metrics import auc
# Save directory
save_dir = make_memnet_checkpoint_dir(Path("plots/correlation/"), net)
if not save_dir.exists():
save_dir.mkdir()
dim = net.input_distribution_dim
if dim == -1:
raise Exception
other_dim = 1 if dim == 0 else 0
images, stim_vals = load_plants(net.image_width,
net.trainingset_dir,
layer_type=net.layer_type,
normalize=True)
target_selection = range(start, stop, step) # Select only subset of values
probe_selection = range(0, 100, 1)
unique_stim_vals = 100
corrs = np.zeros((len(target_selection), len(probe_selection)))
xhats = np.zeros((len(target_selection), len(probe_selection),
unique_stim_vals, images.shape[1]))
A_left = []
A_right = []
for i, dim_t in enumerate(target_selection):
# Iterate over all stimulus values for target image
inds_x = [
np.logical_and(dim_t == stim_vals[:, dim],
d == stim_vals[:, other_dim])
for d in range(unique_stim_vals)
]
x = np.vstack([images[ix] for ix in inds_x])
for j, dim_p in enumerate(probe_selection):
# Iterate over all stimulus values for probe image
inds_y = [
np.logical_and(dim_p == stim_vals[:, dim],
d == stim_vals[:, other_dim])
for d in range(unique_stim_vals)
]
y = np.vstack([images[iy] for iy in inds_y])
xhats[i, j] = np.mean(
[net.predict(x, keep_session=True) for _ in range(n_samples)],
axis=0)
corrs[i, j] = np.mean([
pearsonr(xhats[i, j][k], y[k])[0]
for k in range(unique_stim_vals)
])
print(i, j)
i_true_target_val = probe_selection.index(
dim_t
# Note, probe_selection must include all the target values for this
# to work
)
rAleft = corrs[i, max(0, i_true_target_val - 10):i_true_target_val + 1]
rAright = corrs[i, i_true_target_val:i_true_target_val + 11]
if len(rAleft) > 1:
A_left.append(auc(range(len(rAleft)), rAleft))
else:
A_left.append(0)
if len(rAright) > 1:
A_right.append(auc(range(len(rAright)), rAright))
else:
A_right.append(0)
cutoff = 100
fig, ax = plt.subplots()
for i in range(len(target_selection)):
ind_t = probe_selection.index(target_selection[i])
line = ax.plot(probe_selection[max(0, ind_t - cutoff):ind_t + cutoff],
corrs[i][max(0, ind_t - cutoff):ind_t + cutoff])
color = line[0].get_color()
ax.axvline(x=target_selection[i], color=color, linestyle="--")
ax.set_ylim(0.6, 1)
ax.set_xlim(0, 100)
ax.set_xlabel("stimulus value", fontsize=16)
ax.set_ylabel("correlation coefficient", fontsize=16)
save_pth = save_dir.joinpath("correlation_dim{}_mean{}_std{}.{}".format(
net.input_distribution_dim, net.input_mean, net.input_std, f_ext))
plt.savefig(save_pth)
print("Saved to ", save_pth)
return
def visualize_network_samples(net,
outputs_per_example=5,
color_map="gray",
setsize=[1, 6],
f_ext="pdf"):
"""Make grid of sample images from decoder
"""
n_examples = net.batch_size
if net.name == "cifar":
inputs, targets = generate_training_data(
None,
None,
dataset=net.dataset,
train_test="test",
conv=True,
data_dir=net.trainingset_dir)[0:2]
inputs, targets = net.get_batch(inputs, targets, None)[0:2]
sig = True # For predicting decisions
elif net.name == "fruits":
inputs, recall_targets = net.generate_batch("test")
targets = inputs
sig = True
elif net.name == "popout":
inputs, recall_targets = net.generate_training_data(n_examples)
targets = inputs
sig = False
else:
inputs = generate_training_data(n_examples,
net.image_width,
task_weights=net.task_weights,
dataset=net.dataset,
conv=net.layer_type == "conv",
data_dir=net.trainingset_dir,
setsize=setsize,
mean=net.input_mean,
std=net.input_std,
dim=net.input_distribution_dim)[0]
targets = inputs
sig = False
if net.image_channels > 1:
shp = [net.image_width, net.image_width, net.image_channels]
else:
shp = [net.image_width, net.image_width]
preds0, dec0 = zip(*[
net.predict_both(
inputs, np.zeros_like(inputs), sigmoid=sig, keep_session=True)
for i in range(outputs_per_example)
])
preds1 = np.array(preds0)
preds2 = preds1.swapaxes(0, 1)
preds = preds2.reshape([n_examples, outputs_per_example] + shp)
preds = preds.clip(0, 1)
# dec = np.array(dec0).swapaxes(0, 1)
# print("decisions: ", dec)
# print(dec)
save_dir = make_memnet_checkpoint_dir(Path("plots/vis_memnet_samples/"),
net)
if not save_dir.exists():
save_dir.makedirs()
print("Saving images to: ", save_dir)
for i in range(n_examples):
fig, axes = plt.subplots(nrows=1,
ncols=outputs_per_example + 1,
figsize=(45, 10))
if net.RGB:
axes[0].imshow(targets[i], cmap=color_map)
else:
axes[0].imshow(targets[i].reshape(net.image_width,
net.image_width),
cmap=color_map)
[
axes[j + 1].imshow(preds[i][j], cmap=color_map)
for j in range(outputs_per_example)
]
[ax.set_aspect('equal') for ax in axes]
xlabs = ["Target"] + [
"Sample {}".format(j) for j in range(1, outputs_per_example + 1)
]
[ax.set_xlabel(l, size=50) for ax, l in zip(axes, xlabs)]
for ax in axes:
ax.set_xticks([])
ax.set_xticklabels("")
ax.set_yticks([])
ax.set_yticklabels("")
plt.tight_layout()
plt.savefig(
save_dir.joinpath('setsize_{}_sample{}.{}'.format(
"_".join([str(x) for x in setsize]),
str(i).zfill(2), f_ext)))
def __init__(self,
hidden_size,
latent_size,
activation,
checkpoint_dir,
trainingset_dir,
decision_size=100,
encoder_layers=1,
decoder_layers=1,
decision_layers=1,
kernel_size=3,
load_decision_weights=False,
task_weights=None,
w_reconstruction=[1., 1.],
w_decision=[1., 1.],
w_reg=1e-10,
w_rate=0,
beta0=None,
beta_steps=200000,
sampling_off=False,
encode_probe=False,
dataset='rectangle',
image_width=100,
RGB=False,
layer_type="MLP",
decision_target="same_different",
loss_func_dec="squared_error",
loss_func_recon=None,
dataset_size=int(1e4),
regenerate_steps=None,
input_std=10,
input_mean=50,
input_distribution_dim=-1,
sample_distribution="gaussian",
decision_dim=1,
image_channels=3,
batch_size=20,
dropout_prob=1.0):
if "float" in dataset and layer_type == "conv":
raise NotImplementedError
if dataset == "letter" and len(task_weights) != 4:
raise Exception
if layer_type == "conv" and hidden_size % 2:
raise Exception("Kernel size must be odd number")
self.name = "VAE"
activation_dict = {
"relu": tf.nn.relu,
"sigmoid": tf.nn.sigmoid,
"tanh": tf.nn.tanh,
"elu": tf.nn.elu
}
self.dataset = dataset
self.dataset_size = int(dataset_size)
self.regenerate_steps = regenerate_steps
self.batch_size = batch_size
self.input_mean = input_mean
self.input_std = input_std
self.input_distribution_dim = input_distribution_dim
if task_weights is None:
if decision_dim is None:
raise Exception(
"Need to specify either task_weights or decision_dim")
else:
task_weights = [1] * decision_dim
else:
if hasattr(task_weights, '__len__'):
if decision_dim is None:
decision_dim = len(task_weights)
elif not decision_dim == len(task_weights):
raise Exception(
"decision_dim and task_weights mismatched for length")
else:
raise Exception("task_weights must be list or tuple")
self.task_weights = np.array(task_weights)
self.decision_dim = decision_dim
self.layer_type = layer_type
self.loss_func_dec = loss_func_dec
self.decision_target = decision_target
self.logistic_decision = (False if self.decision_target == "tp_dist"
else True)
if dataset in ["gabor_array", "plants_setsize"]:
# Hardcode for now to avoid errors
self.decision_dim = 6
self.decision_target = "recall"
for i in range(1, 7):
# Hardcode for now to avoid errors
if dataset in [
"gabor_array{}".format(i), "plants_setsize{}".format(i)
]:
self.decision_dim = i
self.decision_target = "recall"
if loss_func_recon is None:
self.loss_func_recon = "squared_error"
else:
self.loss_func_recon = loss_func_recon
if dataset == "episodic_shape_race":
self.decision_dim = 10
elif dataset == "episodic_setsize":
self.decision_dim = 12
# Activation function
self.activation = activation_dict[activation]
self.image_width = image_width
self.RGB = RGB
if self.RGB:
self.image_channels = 3
else:
self.image_channels = 1
# Size of hidden layer(s)
self.hidden_size = hidden_size
self.kernel_size = kernel_size # Used if layers are convolutional
self.encoder_layers = encoder_layers
self.decoder_layers = decoder_layers
self.decision_layers = decision_layers
self.latent_size = latent_size
self.decision_size = decision_size
self.w_rate = float(
w_rate) # Initialize to this value, but possibly change
self.beta1 = w_rate # Final value for rate_loss_weight
self.beta0 = beta0 # Initial value for rate_loss_weight
if self.beta0 is not None and self.beta0 > self.beta1:
raise Exception("beta1 must be greater than beta0")
self.beta_steps = beta_steps
self.w_reconstruction_enc, self.w_reconstruction_dec = w_reconstruction
self.w_reconstruction_enc = float(self.w_reconstruction_enc)
self.w_reconstruction_dec = float(self.w_reconstruction_dec)
self.w_decision_enc, self.w_decision_dec = w_decision
self.w_decision_enc = float(self.w_decision_enc)
self.w_decision_dec = float(self.w_decision_dec)
self.w_reg = float(w_reg)
self.dropout_prob = dropout_prob
self.sampling_off = sampling_off
self.sample_distribution = sample_distribution
self.encode_probe = encode_probe
self.checkpoint_top_dir = Path(checkpoint_dir)
self.trainingset_dir = Path(trainingset_dir)
self.checkpoint_dir = make_memnet_checkpoint_dir(checkpoint_dir, self)
self.decision_checkpoint_dir = None
print("Using {} activation".format(activation))
print("Using {} loss function for reconstruction error".format(
self.loss_func_recon))
print("Loss weights (rate, reconstruction, decision): ",
(self.w_rate, w_reconstruction, w_decision))
self.sess = None
