def initializePruning():
global blockList #ip.getBlockList('vgg16')
global featureList
global convIdx
global module
global prune_count
blockList = ip.createBlockList(newModel) #ip.getBlockList('vgg16')
featureList = ip.createFeatureList(newModel)
convIdx = ip.findConvIndex(newModel)
module = ip.getPruneModule(newModel)
prune_count = ip.getPruneCount(module=module, blocks=blockList, maxpr=.1)
global newList
global layer_number
global st
global en
global candidateConvLayer
newList = []
layer_number = 0
st = 0
en = 0
candidateConvLayer = []
print(f"Block List = {blockList}\n"
f"Feature List = {featureList}\n"
f"Conv Index = {convIdx}\n"
f"Prune Count = {prune_count}\n"
f"Start Index = {st}\n"
f"End Index = {en}\n"
f"Initial Layer Number = {layer_number}\n"
f"Empy candidate layer list = {candidateConvLayer}")
def initializePruning():
global blockList #ip.getBlockList('vgg16')
global featureList
global convIdx
global module
global prune_count
with open(outLogFile, "a") as f:
blockList = ip.createBlockList(newModel) #ip.getBlockList('vgg16')
featureList = ip.createFeatureList(newModel)
convIdx = ip.findConvIndex(newModel)
module = ip.getPruneModule(newModel)
prune_count = ip.getPruneCount(module=module,
blocks=blockList,
maxpr=.1)
global newList
global layer_number
global st
global en
global candidateConvLayer
newList = []
layer_number = 0
st = 0
en = 0
candidateConvLayer = []
f.write(f"\nBlock List = {blockList}"
f"\nFeature List = {featureList}"
f"\nConv Index = {convIdx}"
f"\nPrune Count = {prune_count}"
f"\nStart Index = {st}"
f"\nEnd Index = {en}"
f"\nInitial Layer Number = {layer_number}"
f"\nEmpy candidate layer list = {candidateConvLayer}")
f.close()
epochs=1, max_lr=0.01, weight_decay=0, L1=0, grad_clip=.1, logFile=logFile)
#Save the trained model
SavePath = '/home/pragnesh/Model/vgg16-v2'
torch.save(newModel, SavePath)
# ## Pruning
# #### Pruning Initialization
# In[4]:
blockList = ip.createBlockList(newModel) #ip.getBlockList('vgg16')
featureList = ip.createFeatureList(newModel)
convIdx = ip.findConvIndex(newModel)
module = ip.getPruneModule(newModel)
prune_count = ip.getPruneCount(module=module,blocks=blockList,maxpr=.1)
print(f"Block List = {blockList}\n"
f"Feature List = {featureList}\n"
f"Conv Index = {convIdx}\n"
f"Prune Count = {prune_count}"
)
# for i in range(len(module)):
# print(f"{i+1} :{module[i]}")
# #### Implementing custom pruning process
# 1. Initialization: blockList,featureList,convidx,prune_count,module
# 2. ComputeCandidateLayer
# 3. ComputenewList
# 4. Call CustomPruning
# 5. Commit Pruning
# 6. Update feature list
# 7. Create new temp model with updated feature list
# 8. Perform deep copy
# 9. Train pruned model
# 10. Evalute the pruned model
# 11. Continue another iteration if required and accepted
# In[4]: Pruning Initialization
blockList = ip.createBlockList(newModel) # ip.getBlockList('vgg16')
featureList = ip.createFeatureList(
newModel) # Number of feature map in each convolution layer
convIdx = ip.findConvIndex(
newModel
) # get the index of convolution parameter in newModel.features list
module = ip.getPruneModule(
newModel) # extract all the conv layer and store them in the module list
prune_count = ip.getPruneCount(
module=module, blocks=blockList,
maxpr=.1) # set amount to be pruned from each layer
print(f"Block List = {blockList}\n"
f"Feature List = {featureList}\n"
f"Conv Index = {convIdx}\n"
f"Prune Count = {prune_count}")
# In[11]:Implementing custom pruning process
newList = []