output_np = np.swapaxes(output_np, 1, 2) #(batch, frames, featureDim)
output_np = np.reshape(output_np, (-1, featureDim))
output_np = np.swapaxes(output_np, 0, 1) #(featureDim, frames)
inputData_np_ori = np.swapaxes(inputData_np_ori, 1,
2) #(batch, frames, featureDim)
inputData_np_ori = np.reshape(inputData_np_ori, (-1, featureDim))
inputData_np_ori = np.swapaxes(inputData_np_ori, 0, 1)
# outputData_np_ori = np.swapaxes(outputData_np_ori,1,2) #(batch, frames, 73)
# outputData_np_ori = np.reshape(outputData_np_ori,(-1,featureDim73))
# outputData_np_ori = np.swapaxes(outputData_np_ori,0,1)
faceData = [output_np, inputData_np_ori]
glViewer.SetFaceParmData(faceData, False)
glViewer.init_gl()
continue
#Save Speaking prediction
pred = output.data.cpu().numpy()
pred_all = np.concatenate((pred_all, pred[:, -1]), axis=0)
test_X_vis = np.concatenate((test_X_vis, inputData_np_ori), axis=0)
#Computing Accuracy
pred_binary = pred_all[:] >= 0.5
pred_binary = pred_binary[:, -1]
from sklearn.metrics import accuracy_score
test_Y_cropped = test_Y[:len(pred_binary), -1]
plt.title('Prediction (probability)')
#plt.ion()
plt.show()
#plt.pause(1)
bVisualize = False
if bVisualize:
#by jhugestar
sys.path.append('/ssd/codes/glvis_python/')
#from glViewer import SetFaceParmData,setSpeech,setSpeechGT,setSpeech_binary, setSpeechGT_binary, init_gl #opengl visualization
import glViewer
maxFrameNum = 2000
frameNum = test_X_vis.shape[0]
startIdx = 0
test_X_vis = np.swapaxes(
test_X_vis, 0,
1) #(frames, 200) ->(200, frames) where num can be thought as frames
while startIdx < frameNum:
endIdx = min(frameNum, startIdx + maxFrameNum)
glViewer.setSpeechGT_binary([test_Y_cropped[startIdx:endIdx]])
glViewer.setSpeech_binary([pred_binary[startIdx:endIdx]])
glViewer.SetFaceParmData([test_X_vis[:, startIdx:endIdx]])
glViewer.init_gl()
startIdx += maxFrameNum
sampleNum = 1
for _ in range(100):
sample_2 = torch.randn(sampleNum, 100) #0.5
# print(sample_2[:10])
faceData = []
for model in model_list:
for iter in [1.0, 3.0]:
sample_2_iter = Variable(sample_2 * iter).cuda()
output = model.decode(sample_2_iter)
#De-standardaize
output_np = output.data.cpu().numpy() #(batch, featureDim, frames)
output_np = output_np * Xstd + Xmean
output_np = np.swapaxes(output_np, 1,
2) #(batch, frames, featureDim)
output_np = np.reshape(output_np, (-1, featureDim))
output_np = np.swapaxes(output_np, 0, 1) #(featureDim, frames)
#faceData = [output_np]
faceData.append(output_np)
glViewer.SetFaceParmData(faceData)
glViewer.init_gl()
output_np = output_np * Xstd + Xmean
output_np = np.swapaxes(output_np, 1, 2) #(batch, frames, featureDim)
output_np = np.reshape(output_np, (-1, featureDim))
output_np = np.swapaxes(output_np, 0, 1) #(featureDim, frames)
faceData.append(output_np)
inputData_np_ori = np.swapaxes(inputData_np_ori, 1,
2) #(batch, frames, featureDim)
inputData_np_ori = np.reshape(inputData_np_ori, (-1, featureDim))
inputData_np_ori = np.swapaxes(inputData_np_ori, 0, 1)
faceData.append(inputData_np_ori) #recon, flipped, original
#faceData = [output_np, inputData_np_ori]
glViewer.SetFaceParmData(faceData, bComputeNormal=False)
glViewer.init_gl()
continue
#Save Speaking prediction
pred = output.data.cpu().numpy()
pred_all = np.concatenate((pred_all, pred[:, -1]), axis=0)
test_X_vis = np.concatenate((test_X_vis, inputData_np_ori), axis=0)
#Computing Accuracy
pred_binary = pred_all[:] >= 0.5
pred_binary = pred_binary[:, -1]
from sklearn.metrics import accuracy_score
test_Y_cropped = test_Y[:len(pred_binary), -1]
# bodyData_pred = body_mean.copy()[0,:66,:] #Mirroring (buyer) (73)
# bodyData_pred = np.repeat(bodyData_pred,bodyData[0].shape[1],axis=1)*HOLDEN_DATA_SCALING
# """
# bodyData_gt = bodyData[1][:-7,:]*HOLDEN_DATA_SCALING #GT
# bodyData_gt = bodyData_gt[:,:bodyData_pred.shape[1]]
# skelErr = ( bodyData_pred - bodyData_gt)**2 #66, frames
# skelErr = np.reshape(skelErr, (3,22,-1)) #3,22, frames
# skelErr = np.sqrt(np.sum(skelErr, axis=0)) #22,frames
# skelErr = np.mean(skelErr,axis=0) #frames
# skeletonErr_list.append(skelErr)
if bVisualize == False:
continue
glViewer.SetFaceParmData(faceData, False)
glViewer.init_gl()
# Compute error
##Draw Error Figure
avg_skelErr_list = []
total_avg_skelErr = 0
cnt = 0
for p in skeletonErr_list:
avgValue = np.mean(p)
avg_skelErr_list.append(avgValue)
print(avgValue)
total_avg_skelErr += avgValue * len(p)
cnt += len(p)
bad_test = train_X_raw[pIdx]
bad_test = np.max(np.max(bad_test, axis=2), axis=1)
import matplotlib.pyplot as plt
plt.plot(bad_test)
plt.show()
##Debug: visualize
sys.path.append('/ssd/codes/glvis_python/')
import glViewer
testParam = train_X_raw[2,
[9482, 11381, 15670, 15671, 15672, 17707], :, :100]
testParam = np.reshape(testParam, (-1, 100))
testParam = np.swapaxes(testParam, 0, 1)
glViewer.SetFaceParmData([testParam])
glViewer.init_gl()
# train_X = train_X[:-1:10,:,:]
# train_Y = train_Y[:-1:10,:]
test_data = np.load(datapath + test_dblist[0] + '.npz')
test_X_raw = test_data[
'clips'] #Input (3, numClip, chunkLengh, featureDim:200)
test_speech_raw = test_data['speech'] #Input (3, numClip, chunkLengh)
#Attach speech signal
test_X_raw = test_X_raw[:, :, :, :1]
test_speech_raw = np.expand_dims(test_speech_raw, 3)
test_X_raw = np.concatenate((test_X_raw, test_speech_raw), axis=3)
pred = pred.data.cpu().numpy()
pred = (pred * Ystd[:, :, :featureDim]) + Ymean[:, :, :featureDim]
pred = pred[:, -1, :] #(batch, feature)
vis_data = np.reshape(pred, (-1, featureDim))
vis_data = np.swapaxes(vis_data, 0, 1)
#GT visualization
vis_data_gt = outputData_np_ori[:, -1, :] #(batch, features)
vis_data_gt = np.swapaxes(vis_data_gt, 0, 1) #(features, batch)
#GT visualization (input)
vis_data_input = inputData_np_ori[:, -1, :] #(batch, features)
vis_data_input = np.swapaxes(vis_data_input, 0, 1) #(features, batch)
glViewer.SetFaceParmData([vis_data_input, vis_data_gt, vis_data])
glViewer.init_gl()
continue
#Save Speaking prediction
pred = output.data.cpu().numpy()
pred_all = np.concatenate((pred_all, pred[:, -1]), axis=0)
test_X_vis = np.concatenate((test_X_vis, inputData_np_ori), axis=0)
#Computing Accuracy
pred_binary = pred_all[:] >= 0.5
pred_binary = pred_binary[:, -1]
from sklearn.metrics import accuracy_score
test_Y_cropped = test_Y[:len(pred_binary), -1]