self.sig1 = self.sig1.to(device,torch.float)
self.miu2 = self.miu2.to(device,torch.float)
self.sig2 = self.sig2.to(device,torch.float)
def forward(self,input1,input2):
out1 = self.m1(input1)
out2 = self.m2(input2)
#unnormalize the predictions
out1 = (out1*self.sig1)+self.miu1
out2 = (out2*self.sig2)+self.miu2
x = out1+out2
return x
model1 = mdl.StateModel(54, 3)
model1.load_state_dict(torch.load("best_modelweights_S_PSM1.dat"))
model2 = mdl.StateModel(54, 3)
model2.load_state_dict(torch.load("best_modelweights_S.dat"))
full_model = dual_model(model1,model2,mean1[1:4],std1[1:4],mean2[1:4],std2[1:4])
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
full_model.transfer(device)
#%%
import torch.optim as opt
weight_file1 = "best_modelweights_S_PSM1_ft.dat"
weight_file2 = "best_modelweights_S_PSM2_ft.dat"
criterion = nn.MSELoss(reduction='sum')
# define optimization method
optimizer = opt.Adam(full_model.parameters(),lr=0.001,weight_decay=0)
state_input = np.array(msg.data)
state_input = (state_input - mean1[7:61]) / std1[7:61]
def teleop_callback(data):
global teleop
butt = data.buttons[0]
if butt > 0.5:
teleop = True
else:
teleop = False
#load model
model = mdl.StateModel(54, 3)
print("loading the model...")
weight_file = "best_modelweights_S.dat" #"best_modelweights_S_PSM1.dat"
model.load_state_dict(torch.load(weight_file))
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
model = model.to(device, torch.float)
model.eval()
# create rosnode
rospy.init_node("NN_node")
# create the state vector for input as a global variable
global state_input
global force_pred
global force_pred_out
global mean1
global std1
def run_ablations(model, num_ablations):
'''set up some persistent tracking variables'''
remove_features = [] # list of features we are removing
metrics_list = [] # list storing dictionary of performance metrics
# feature indexes
full_state_index = np.arange(7, 61)
input_state = 54
# create loss function
criterion = nn.MSELoss(reduction='sum')
# define optimization method
optimizer = opt.Adam(model.parameters(), lr=0.01)
param_count = []
param_count.append(count_params(model))
current_feature_list = np.array(qty)
# create the dataloader
dataloaders, dataset_sizes = dat.init_dataset(train_list, val_list,
val_list, model_type,
config_dict)
print('evaluating full model predictions...')
predictions = mdl.evaluate_model(model,
dataloaders['test'],
model_type=model_type,
no_pbar=True)
# compute the loss statistics
print('computing full model performance metrics...')
metrics = model_eval.compute_loss_metrics(
predictions, dataloaders['test'].dataset.label_array[:, 1:4])
metrics_list.append(metrics)
print('Performance Summary of Full Model:')
print(metrics)
print('Running ablation study on model type:' + model_type)
for iteration in range(num_ablations):
print('-' * 10)
print('Begin ablation run: {}/{}'.format(iteration + 1, num_ablations))
print('-' * 10)
# compute the backprop values:
gbp_data = model_eval.compute_GBP(model,
dataloaders['test'],
num_state_inputs=input_state,
model_type=model_type,
no_pbar=True)
# evaluate means
df_gbp_means = model_eval.compute_and_plot_gbp(gbp_data,
current_feature_list,
True,
suppress_plots=True)
# group by feature type and rank by value
df_gbp_means = df_gbp_means.groupby('feature').mean().sort_values(
by='gbp', ascending=False).reset_index()
# get top ranking value and append to removal list
feature_to_remove = df_gbp_means.iloc[0, 0]
print("removing " + feature_to_remove + "...")
remove_features.append(feature_to_remove)
# create the mask
mask = np.isin(qty, remove_features, invert=True)
# mask the full state vector in config_dict global variable
config_dict['custom_state'] = full_state_index[mask]
current_feature_list = np.array(qty)[
mask] #update the current feature list
# decrease the input dimension of the model by one
input_state = input_state - 1
# redefine the models
print('redefining model with input state dims: {}'.format(input_state))
if model_type == "VS":
model = mdl.StateVisionModel(30,
input_state,
3,
feature_extract=feat_extract)
elif model_type == "S":
model = mdl.StateModel(input_state, 3)
# recalculate the number of parameters
param_count.append(count_params(model))
# redefine the optimizer
optimizer = opt.Adam(model.parameters(), lr=0.01)
# redefine the dataloader
dataloaders, dataset_sizes = dat.init_dataset(train_list, val_list,
val_list, model_type,
config_dict)
# retrain the model
model, train_history, val_history = mdl.train_model(
model,
criterion,
optimizer,
dataloaders,
dataset_sizes,
num_epochs=50,
model_type=model_type,
weight_file=weight_file,
no_pbar=True)
print('retraining completed')
# do inference
print('evaluating model predictions...')
predictions = mdl.evaluate_model(model,
dataloaders['test'],
model_type=model_type,
no_pbar=True)
# compute the loss statistics
print('computing performance metrics...')
metrics = model_eval.compute_loss_metrics(
predictions, dataloaders['test'].dataset.label_array[:, 1:4])
metrics_list.append(metrics)
print('Performance Summary:')
print(metrics)
return remove_features, param_count, metrics_list