Skip to content

Commit

Permalink
interpretability changed to use h3.6m. Also added umap code to visualise
Browse files Browse the repository at this point in the history
  • Loading branch information
@bouracha
bouracha committed Oct 23, 2020
1 parent 2954461 commit df5a1fe
Showing 1 changed file with 82 additions and 48 deletions.
130 changes: 82 additions & 48 deletions interpretability.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -19,7 +19,6 @@
import utils.model as nnmodel
import utils.data_utils as data_utils


import argparse

parser = argparse.ArgumentParser()
Expand All @@ -29,47 +28,45 @@

opt = parser.parse_args()


is_cuda = torch.cuda.is_available()




input_n = 10
output_n = 25
dct_n = 35
sample_rate = 2

cartesian = False

if opt.dataset=='h3.6m':
if opt.dataset == 'h3.6m':
node_n = 48
n_z = 384
actions = data_utils.define_actions('all', 'h3.6m', out_of_distribution=False)
elif opt.dataset=='cmu_mocap':
elif opt.dataset == 'cmu_mocap':
node_n = 64
actions = data_utils.define_actions('all', 'cmu_mocap', out_of_distribution=False)
n_z = 512


model = nnmodel.GCN(input_feature=dct_n, hidden_feature=256, p_dropout=0.3,
num_stage=12, node_n=node_n, variational=True, n_z=8, num_decoder_stage=6)
num_stage=12, node_n=node_n, variational=True, n_z=8, num_decoder_stage=6)
if is_cuda:
model.cuda()
print(">>> total params: {:.2f}M".format(sum(p.numel() for p in model.parameters()) / 1000000.0))
#optimizer = torch.optim.Adam(model.parameters(), lr=opt.lr)
# optimizer = torch.optim.Adam(model.parameters(), lr=opt.lr)

if opt.model_path == None:
model_path_len = 'checkpoint/test/ckpt_main_cmu_mocap_in50_out25_dctn35_dropout_0.3_var_lambda_0.003_nz_8_lr_0.0005_n_layers_6_last.pth.tar'
#model_path_len = 'checkpoint/test/ckpt_main_h3.6m_in10_out10_dctn20_dropout_0.0_var_lambda_0.003_nz_8_lr_0.0005_n_layers_6_last.pth.tar'
else:
model_path_len = opt.model_path
# model_path_len = 'checkpoint/test/ckpt_main_h3.6m_in10_out10_dctn20_dropout_0.0_var_lambda_0.003_nz_8_lr_0.0005_n_layers_6_last.pth.tar'
print(">>> loading ckpt len from '{}'".format(model_path_len))
if is_cuda:
ckpt = torch.load(model_path_len)
else:
ckpt = torch.load(model_path_len, map_location='cpu')

batch_size=512
batch_size = 512
train_data = dict()
val_data = dict()
test_data = dict()
for act in actions:
print("Loading action {} for train set".format(act))
Expand All @@ -79,7 +76,7 @@
elif opt.dataset == 'cmu_mocap':
train_dataset = CMU_Motion(path_to_data='cmu_mocap/', actions=[act], input_n=input_n, output_n=output_n,
split=0, dct_n=dct_n)
#print(train_dataset.__len__())
# print(train_dataset.__len__())
data_std = train_dataset.data_std
data_mean = train_dataset.data_mean
dim_used = train_dataset.dim_used
Expand All @@ -89,32 +86,38 @@
shuffle=False,
num_workers=10,
pin_memory=True)
print("Loading action {} for test set".format(act))
print("Loading action {} for val and test set".format(act))
if opt.dataset == 'h3.6m':
val_dataset = H36motion(path_to_data='h3.6m/', actions=[act], input_n=input_n, output_n=output_n,
split=2, sample_rate=2, data_mean=data_mean, data_std=data_std, dct_n=dct_n)
val_data[act] = DataLoader(
dataset=val_dataset,
batch_size=128,
shuffle=False,
num_workers=10,
pin_memory=True)
test_dataset = H36motion(path_to_data='h3.6m/', actions=[act], input_n=input_n, output_n=output_n,
sample_rate=sample_rate, split=1, data_mean=data_mean, data_std=data_std, dct_n=dct_n)
elif opt.dataset == 'cmu_mocap':
test_dataset = CMU_Motion(path_to_data='cmu_mocap/', actions=[act], input_n=input_n, output_n=output_n,
split=1, data_mean=data_mean, data_std=data_std, dim_used=dim_used, dct_n=dct_n)
#print(test_dataset.__len__())
# print(test_dataset.__len__())
test_data[act] = DataLoader(
dataset=test_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=10,
pin_memory=True)



if output_n >= 25:
eval_frame = [1, 3, 7, 9, 13, 24]
elif output_n == 10:
eval_frame = [1, 3, 7, 9]


first_instance = True
first_instance_train = True
first_instance_val = True
first_instance_test = True
for act in actions:
new_action = True
for i, (inputs, targets, all_seq) in enumerate(train_data[act]):
if is_cuda:
inputs = Variable(inputs.cuda()).float()
Expand All @@ -123,24 +126,49 @@
outputs, reconstructions, log_var, z = model(inputs.float())

print("For action {} the train z is: {}".format(act, z.shape))
for sample in z:
sample = sample.reshape(n_z).cpu().detach().numpy()
#print(sample.shape)
df = pd.DataFrame(np.expand_dims(sample, axis=0))
if first_instance:
df.to_csv('latents/all_train_z.csv', header=False, index=False)
first_instance = False
else:
with open('latents/all_train_z.csv', 'a') as f:
df.to_csv(f, header=False, index=False)
if new_action:
df.to_csv('latents/'+str(act)+'_train_z.csv', header=False, index=False)
new_action = False
else:
with open('latents/'+str(act)+'_train_z.csv', 'a') as f:
df.to_csv(f, header=False, index=False)

new_action = True
for i in range(z.shape[0]):
sample = z[i].reshape(n_z).cpu().detach().numpy()
sample_input = inputs[i].reshape(node_n * dct_n).cpu().detach().numpy()
df = pd.DataFrame(np.expand_dims(sample, axis=0))
df_input = pd.DataFrame(np.expand_dims(sample_input, axis=0))
df['label'] = act
df_input['label'] = act
if first_instance_train:
df.to_csv('latents/all_train_z.csv', header=False, index=False)
df_input.to_csv('inputs/all_train.csv', header=False, index=False)
first_instance_train = False
else:
with open('latents/all_train_z.csv', 'a') as f:
df.to_csv(f, header=False, index=False)
with open('inputs/all_train.csv', 'a') as f:
df_input.to_csv(f, header=False, index=False)

if opt.dataset == 'h3.6m':
for i, (inputs, targets, all_seq) in enumerate(val_data[act]):
if is_cuda:
inputs = Variable(inputs.cuda()).float()
all_seq = Variable(all_seq.cuda(non_blocking=True)).float()

outputs, reconstructions, log_var, z = model(inputs.float())

print("For action {} the val z is: {}".format(act, z.shape))
for i in range(z.shape[0]):
sample_z = z[i].reshape(n_z).cpu().detach().numpy()
sample_input = inputs[i].reshape(node_n * dct_n).cpu().detach().numpy()
df = pd.DataFrame(np.expand_dims(sample_z, axis=0))
df_input = pd.DataFrame(np.expand_dims(sample_input, axis=0))
df['label'] = act
df_input['label'] = act
if first_instance_val:
df.to_csv('latents/all_val_z.csv', header=False, index=False)
df_input.to_csv('inputs/all_val.csv', header=False, index=False)
first_instance_val = False
else:
with open('latents/all_val_z.csv', 'a') as f:
df.to_csv(f, header=False, index=False)
with open('inputs/all_val.csv', 'a') as f:
df_input.to_csv(f, header=False, index=False)

for i, (inputs, targets, all_seq) in enumerate(test_data[act]):
if is_cuda:
inputs = Variable(inputs.cuda()).float()
Expand All @@ -149,16 +177,22 @@
outputs, reconstructions, log_var, z = model(inputs.float())

print("For action {} the test z is: {}".format(act, z.shape))
for sample in z:
sample = sample.reshape(n_z).cpu().detach().numpy()
#print(sample.shape)
df = pd.DataFrame(np.expand_dims(sample, axis=0))
if new_action:
df.to_csv('latents/'+str(act)+'_test_z.csv', header=False, index=False)
new_action = False
else:
with open('latents/'+str(act)+'_test_z.csv', 'a') as f:
df.to_csv(f, header=False, index=False)
for i in range(z.shape[0]):
sample = z[i].reshape(n_z).cpu().detach().numpy()
sample_input = inputs[i].reshape(node_n * dct_n).cpu().detach().numpy()
df = pd.DataFrame(np.expand_dims(sample, axis=0))
df_input = pd.DataFrame(np.expand_dims(sample_input, axis=0))
df['label'] = act
df_input['label'] = act
if first_instance_test:
df.to_csv('latents/all_test_z.csv', header=False, index=False)
df_input.to_csv('inputs/all_test.csv', header=False, index=False)
first_instance_test = False
else:
with open('latents/all_test_z.csv', 'a') as f:
df.to_csv(f, header=False, index=False)
with open('inputs/all_test.csv', 'a') as f:
df_input.to_csv(f, header=False, index=False)



Expand Down

0 comments on commit df5a1fe

Please sign in to comment.