def set_platform(self, **write):
""" Sets fabric.api.env attributes from the selected platform
The setup follows the sequence:
fabfile.env.platforms > tests_integration.platforms.common > tests_integration.platforms.<selected_platform>
:param write: dictionary to inject into fabric.api.env
"""
module = import_module('.platforms.' + self.platform.platform_name, ROOT)
getattr(module, self.platform.platform_name)(**self.platform.get_hosts(True))
self.platform.user = getattr(self.env, 'default_ssh_user', 'root')
for k, v in write.iteritems():
setattr(self.env, k, v)
# wait sshd daemon running on host platform
self.platform.wait_process('/usr/sbin/sshd')
# save and replace all references to fabric.api.env
self.org_api_env = pyjack.replace_all_refs(api.env, _AttributeDict(api.env))
self.org_instances = pyjack.replace_all_refs(api.env.instances, dict(api.env.instances))
return self
print "SomeCls.someiterable:", SomeCls.someiterable
print "SomeCls.anotheriterable:", SomeCls.anotheriterable
print "Contents of innerfun:"
#>And inner fun:
innerfun_gen = innerfun()
print "First yield:", innerfun_gen.next()
print "Second yield:", innerfun_gen.next()
#>Now, let's replace iterable with some new data
new_iterable = (
'new',
'data',
'set',
)
org_iterable = pyjack.replace_all_refs(iterable, new_iterable)
#>Then look at the new results
print "iterable:", iterable
print "SomeCls.someiterable:", SomeCls.someiterable
print "SomeCls.anotheriterable:", SomeCls.anotheriterable
#>And inner fun, notice the function closure was updated:
innerfun_gen = innerfun()
print "First yield:", innerfun_gen.next()
print "Second yield:", innerfun_gen.next()
#>Then, reverse:
new_iterable = pyjack.replace_all_refs(new_iterable, org_iterable)
#>Then look at the new results
datum,
)
return inner
inner = outer(item)
print item
print data
print Foobar.the_item
print inner()
#>Then replace them:
new = (101, 'one hundred and one')
org_item = pyjack.replace_all_refs(item, new)
print item
print data
print Foobar.the_item
print inner()
#>But you still have the org data:
print org_item
#>So the process is reversible:
new = pyjack.replace_all_refs(new, org_item)
print item
print data
print Foobar.the_item
def prune_resnet101_conv_layer(model, layer_index, filter_index): #layer_to_prune = [] #layers_affected = [] layer_affected = [] layer = 0 activation_index = 0 for block in model.features: if block == model.conv1 or block == model.bn1: # print "Adding initial conv1 or bn1 to graph" # print block if block == model.conv1: pass #print "conv1" else: pass #print "bn1" if isinstance(block, torch.nn.modules.conv.Conv2d): if layer_index == layer: layer_to_prune = block print model.layer1 layer_affected.append(model.bn1) for (bottleneck_name, bottleneck) in model.layer1._modules.items(): if bottleneck_name == '0': for (layer_name, module) in bottleneck._modules.items(): if layer_name == 'conv1': layer_affected.append(module) activation_index += 1 layer += 1 elif block == model.layer1 or block == model.layer2 or block == model.layer3 or block == model.layer4: # print "Adding Residual Block to graph" for (block_name, bottleneck) in block._modules.items(): # print block_name for (name, module) in bottleneck._modules.items(): # print name,module if isinstance(module, torch.nn.modules.Sequential): pass #print "Sequential Block" #out += module(x) #x = F.relu(out) else: # print "Conv or BN Inside Bottleneck" # print name, module if name == 'conv1': pass #print "bottleneck conv1" #out = module(x) elif name == 'bn1' or name == 'bn2': pass #print "bottleneck bn1 or bn2" #out = F.relu(module(out)) elif name == 'conv2' or name == 'conv3': pass #print "bottleneck conv2 or conv3" #out = module(out) elif name == 'bn3': pass #print "bottleneck bn3" #out = module(out) else: pass #out = module(out) if isinstance(module, torch.nn.modules.conv.Conv2d): if layer_index == layer: layer_to_prune = module activation_index += 1 layer += 1 conv = layer_to_prune print "Pruning This Conv Block" print conv print "Layers Affected" print layer_affected #next_conv = None #offset = 1 # while layer_index + offset < len(model.features._modules.items()): # res = model.features._modules.items()[layer_index+offset] # if isinstance(res[1], torch.nn.modules.conv.Conv2d): # next_name, next_conv = res # break # offset = offset + 1 # new_conv = \ torch.nn.Conv2d(in_channels = conv.in_channels, \ out_channels = conv.out_channels - 1, kernel_size = conv.kernel_size, \ stride = conv.stride, padding = conv.padding, dilation = conv.dilation, groups = conv.groups, bias = conv.bias) old_weights = conv.weight.data.cpu().numpy() new_weights = new_conv.weight.data.cpu().numpy() new_weights[: filter_index, :, :, :] = old_weights[: filter_index, :, :, :] new_weights[filter_index : , :, :, :] = old_weights[filter_index + 1 :, :, :, :] new_conv.weight.data = torch.from_numpy(new_weights).cuda() if conv.bias: bias_numpy = conv.bias.data.cpu().numpy() bias = np.zeros(shape = (bias_numpy.shape[0] - 1), dtype = np.float32) bias[:filter_index] = bias_numpy[:filter_index] bias[filter_index : ] = bias_numpy[filter_index + 1 :] new_conv.bias.data = torch.from_numpy(bias).cuda() print "Adding new Conv" print new_conv conv = pyjack.replace_all_refs(conv,new_conv) new_in_channels = new_conv.out_channels for layers in layer_affected: if isinstance(layers, torch.nn.modules.conv.Conv2d): new_layer = \ torch.nn.Conv2d(in_channels = new_in_channels, \ out_channels = layers.out_channels, kernel_size = layers.kernel_size, \ stride = layers.stride, padding = layers.padding, dilation = layers.dilation, groups = layers.groups, bias = layers.bias) old_weights = layers.weight.data.cpu().numpy() new_weights = new_layer.weight.data.cpu().numpy() new_weights[:, : filter_index, :, :] = old_weights[:, : filter_index, :, :] new_weights[:, filter_index :, :, :] = old_weights[:, filter_index + 1 :, :, :] new_layer.weight.data = torch.from_numpy(new_weights).cuda() if layers.bias: bias_numpy = layers.bias.data.cpu().numpy() new_layer.bias.data = torch.from_numpy(bias_numpy).cuda() if isinstance(layers, torch.nn.modules.batchnorm.BatchNorm2d): new_layer = torch.nn.BatchNorm2d(new_in_channels) old_weights = layers.weight.data.cpu().numpy() new_in_channels = new_layer.out_channels layers = pyjack.replace_all_refs(layers,new_layer) new_in_channels = new_conv.out_channels #print "Conv1 New" #print model.conv1 return model
innerfun = myfun(iterable)
#>So look at the org results
print "iterable:", iterable
print "SomeCls.someiterable:", SomeCls.someiterable
print "SomeCls.anotheriterable:", SomeCls.anotheriterable
print "Contents of innerfun:"
#>And inner fun:
innerfun_gen = innerfun()
print "First yield:", innerfun_gen.next()
print "Second yield:", innerfun_gen.next()
#>Now, let's replace iterable with some new data
new_iterable = ('new', 'data', 'set',)
org_iterable = pyjack.replace_all_refs(iterable, new_iterable)
#>Then look at the new results
print "iterable:", iterable
print "SomeCls.someiterable:", SomeCls.someiterable
print "SomeCls.anotheriterable:", SomeCls.anotheriterable
#>And inner fun, notice the function closure was updated:
innerfun_gen = innerfun()
print "First yield:", innerfun_gen.next()
print "Second yield:", innerfun_gen.next()
#>Then, reverse:
new_iterable = pyjack.replace_all_refs(new_iterable, org_iterable)
#>Then look at the new results
def outer(datum):
def inner():
return ("Here is the datum:", datum,)
return inner
inner = outer(item)
print item
print data
print Foobar.the_item
print inner()
#>Then replace them:
new = (101, 'one hundred and one')
org_item = pyjack.replace_all_refs(item, new)
print item
print data
print Foobar.the_item
print inner()
#>But you still have the org data:
print org_item
#>So the process is reversible:
new = pyjack.replace_all_refs(new, org_item)
print item
print data
print Foobar.the_item
def unset_platform(self):
# restore fabric.api.env
pyjack.replace_all_refs(api.env, self.org_api_env)
pyjack.replace_all_refs(api.env.instances, self.org_instances)
return self