def main():
model = efficientnetv2_s()
# option1
for name, para in model.named_parameters():
# 除head外，其他权重全部冻结
if "head" not in name:
para.requires_grad_(False)
else:
print("training {}".format(name))
complexity = model.complexity(224, 224, 3)
table = PrettyTable()
table.field_names = ["params", "freeze-params", "train-params", "FLOPs", "acts"]
table.add_row([complexity["params"],
complexity["freeze"],
complexity["params"] - complexity["freeze"],
complexity["flops"],
complexity["acts"]])
print(table)
# option2
tensor = (torch.rand(1, 3, 224, 224),)
flops = FlopCountAnalysis(model, tensor)
print(flops.total())
print(parameter_count_table(model))
def test_flop_count_empty(self) -> None:
model = nn.ReLU()
inputs = (torch.randn((1, 10)), )
table = flop_count_table(FlopCountAnalysis(model, inputs))
self.assertGreater(len(table), 0)
out = flop_count_str(FlopCountAnalysis(model, inputs))
self.assertGreater(len(out), 0)
def profile_fvcore(model, input_size=(3, 224, 224), batch_size=1, detailed=False, force_cpu=False):
if force_cpu:
model = model.to('cpu')
device, dtype = next(model.parameters()).device, next(model.parameters()).dtype
example_input = torch.ones((batch_size,) + input_size, device=device, dtype=dtype)
fca = FlopCountAnalysis(model, example_input)
aca = ActivationCountAnalysis(model, example_input)
if detailed:
fcs = flop_count_str(fca)
print(fcs)
return fca.total(), aca.total()
def test_flop_count_table(self) -> None:
model = TestNet()
inputs = (torch.randn((1, 10)), )
table = flop_count_table(FlopCountAnalysis(model, inputs))
self.assertFalse(" a1 " in table) # Wrapper skipping successful
self.assertFalse("a1.b1.c1.d1.bias" in table) # Didn't go to depth 4
self.assertTrue("a1.b1.c1.d1" in table) # Did go to depth 3
self.assertTrue(" a1.b1 " in table) # Didn't skip different stats
self.assertTrue("a2.b1.c1.weight" in table) # Weights incuded
self.assertTrue("(10, 10)" in table) # Shapes included
self.assertTrue(" a2.b1 "
in table) # Didn't skip through mod with >1 child
self.assertFalse("#activations" in table) # No activations
self.assertTrue(" 0.33K" in table) # Pretty stats, correct indentation
self.assertFalse(" 0.33K" in table) # Correct indentation
self.assertTrue("#parameters or shape" in table) # Correct header
# Expected:
# | module | #parameters or shape | #flops |
# |:-------------------|:-----------------------|:---------|
# | model | 0.33K | 0.3K |
# | a1.b1 | 0.11K | 100 |
# | a1.b1.c1 | 0.11K | N/A |
# | a1.b1.c1.d1 | 0.11K | 100 |
# | a2.b1 | 0.22K | 0.2K |
# | a2.b1.c1 | 0.11K | 100 |
# | a2.b1.c1.weight | (10, 10) | |
# | a2.b1.c1.bias | (10,) | |
# | a2.b1.c2 | 0.11K | 100 |
# | a2.b1.c2.weight | (10, 10) | |
# | a2.b1.c2.bias | (10,) | |
# Test activations and no parameter shapes
table = flop_count_table(
flops=FlopCountAnalysis(model, inputs),
activations=ActivationCountAnalysis(model, inputs),
show_param_shapes=False,
)
self.assertTrue("#activations" in table) # Activation header
self.assertTrue(" 20"
in table) # Activation value with correct indent
self.assertFalse("#parameters or shape" in table) # Correct header
self.assertTrue("#parameters") # Correct header
self.assertFalse("a2.b1.c1.weight" in table) # Weights not included
self.assertFalse("(10, 10)" in table) # Shapes not included
self.assertFalse("a2.b1.c1.d2" in table) # Skipped empty
def test_detr_fbnet_export(self):
runner = create_runner("d2go.projects.detr.runner.DETRRunner")
cfg = runner.get_default_cfg()
cfg.MODEL.DEVICE = "cpu"
# DETR
self._set_detr_cfg(cfg, 3, 3, 50, 256)
# backbone
cfg.MODEL.BACKBONE.NAME = "FBNetV2C4Backbone"
cfg.MODEL.FBNET_V2.ARCH = "FBNetV3_A_dsmask_C5"
cfg.MODEL.FBNET_V2.WIDTH_DIVISOR = 8
cfg.MODEL.FBNET_V2.OUT_FEATURES = ["trunk4"]
# build model
model = runner.build_model(cfg).eval()
model = model.detr
print(model)
scripted_model = torch.jit.script(model)
self._assert_model_output(model, scripted_model)
# print flops
table = flop_count_table(
FlopCountAnalysis(model, ([torch.rand(3, 224, 320)], )))
print(table)
def main():
# Self-Attention
a1 = Attention(dim=512, num_heads=1)
a1.proj = torch.nn.Identity() # remove Wo
# Multi-Head Attention
a2 = Attention(dim=512, num_heads=8)
# [batch_size, num_tokens, total_embed_dim]
t = (torch.rand(32, 1024, 512), )
flops1 = FlopCountAnalysis(a1, t)
print("Self-Attention FLOPs:", flops1.total())
flops2 = FlopCountAnalysis(a2, t)
print("Multi-Head Attention FLOPs:", flops2.total())
def test_flop_count_str(self) -> None:
"""
Tests calculating model flops and outputing them in model print format.
"""
model = TestNet()
inputs = (torch.randn((1, 10)), )
model_str = flop_count_str(FlopCountAnalysis(model, inputs))
self.assertTrue(
"N/A indicates a possibly missing statistic" in model_str)
self.assertTrue("n_params: 0.11K, n_flops: 100" in model_str)
self.assertTrue("ReLU()" in model_str) # Suppress trivial statistics
self.assertTrue("n_params: 0.11K, n_flops: N/A"
in model_str) # Uncalled stats
self.assertTrue("[[1, 10]]") # Input sizes
# Expected:
# "Input sizes (torch.Tensor only): [[1, 10]]\n"
# "N/A indicates a possibly missing statistic due to how the "
# "module was called. Missing values are still included in the "
# "parent's total.\n"
# "TestNet(\n"
# " n_params: 0.33K, n_flops: 0.3K\n"
# " (a1): A1(\n"
# " n_params: 0.11K, n_flops: 100\n"
# " (b1): A1B1(\n"
# " n_params: 0.11K, n_flops: 100\n"
# " (c1): A1B1C1(\n"
# " n_params: 0.11K, n_flops: N/A\n"
# " (d1): Linear(\n"
# " in_features=10, out_features=10, bias=True\n"
# " n_params: 0.11K, n_flops: 100\n"
# " )\n"
# " (d2): ReLU()\n"
# " )\n"
# " )\n"
# " )\n"
# " (a2): A2(\n"
# " n_params: 0.22K, n_flops: 0.2K\n"
# " (b1): A2B1(\n"
# " n_params: 0.22K, n_flops: 0.2K\n"
# " (c1): Linear(\n"
# " in_features=10, out_features=10, bias=True\n"
# " n_params: 0.11K, n_flops: 100\n"
# " )\n"
# " (c2): Linear(\n"
# " in_features=10, out_features=10, bias=True\n"
# " n_params: 0.11K, n_flops: 100\n"
# " )\n"
# " )\n"
# " )\n"
# ")"
# Test with activations
model_str = flop_count_str(
FlopCountAnalysis(model, inputs),
activations=ActivationCountAnalysis(model, inputs),
)
self.assertTrue(
"n_params: 0.33K, n_flops: 0.3K, n_acts: 30" in model_str)
self.assertTrue(
"n_params: 0.11K, n_flops: N/A, n_acts: N/A" in model_str)
model = MPViT(
img_size=224,
num_stages=4,
num_path=[2, 3, 3, 3],
num_layers=[1, 3, 8, 3],
embed_dims=[128, 224, 368, 480],
mlp_ratios=[4, 4, 4, 4],
num_heads=[8, 8, 8, 8],
**kwargs,
)
model.default_cfg = _cfg_mpvit()
return model
if __name__ == "__main__":
model = mpvit_xsmall()
model.eval()
inputs = torch.randn(1, 3, 224, 224)
model(inputs)
from fvcore.nn import FlopCountAnalysis, ActivationCountAnalysis
flops = FlopCountAnalysis(model, inputs)
param = sum(p.numel() for p in model.parameters() if p.requires_grad)
acts = ActivationCountAnalysis(model, inputs)
print(f"total flops : {flops.total()}")
print(f"total activations: {acts.total()}")
print(f"number of parameter: {param}")
def transformer(
save,
load,
half,
bt2fairseq,
):
xlarge = False
large = False
DEFAULT_PADDING_IDX = 1
avg_sequence_length = 128
max_sequence_length = 256
batch_size = 64
class FairseqEncoder(torch.nn.Module):
def __init__(
self,
embed_dim,
attention_heads,
ffn_embed_dim,
num_layers,
embedding_layer, # torch.nn.Embedding. Must have a padding_idx field
dropout=0,
normalize_before=False,
torch_encoder=None, # torch encoder that you can map weights from
activation="relu",
):
super().__init__()
cfg = FairseqTransformerConfig()
cfg.encoder.embed_dim = embed_dim
cfg.encoder.attention_heads = attention_heads
cfg.encoder.ffn_embed_dim = ffn_embed_dim
cfg.dropout = dropout
cfg.encoder.normalize_before = normalize_before
cfg.encoder.layers = num_layers
# make embedding behavior same as other encoders
cfg.no_token_positional_embeddings = True
cfg.no_scale_embedding = True
cfg.activation_fn = activation
dictionary = {} # TODO: verify what this is
self.encoder = FairseqTransformerEncoder(
cfg, dictionary, embedding_layer, return_fc=False
)
if torch_encoder is not None:
for src_layer, dst_layer in zip(
torch_encoder.layers, self.encoder.layers
):
w_q, w_k, w_v = src_layer.self_attn.in_proj_weight.chunk(3, dim=0)
b_q, b_k, b_v = src_layer.self_attn.in_proj_bias.chunk(3, dim=0)
dst_layer.self_attn.q_proj.weight = torch.nn.Parameter(w_q)
dst_layer.self_attn.q_proj.bias = torch.nn.Parameter(b_q)
dst_layer.self_attn.k_proj.weight = torch.nn.Parameter(w_k)
dst_layer.self_attn.k_proj.bias = torch.nn.Parameter(b_k)
dst_layer.self_attn.v_proj.weight = torch.nn.Parameter(w_v)
dst_layer.self_attn.v_proj.bias = torch.nn.Parameter(b_v)
dst_layer.self_attn.out_proj.weight = (
src_layer.self_attn.out_proj.weight
)
dst_layer.self_attn.out_proj.bias = (
src_layer.self_attn.out_proj.bias
)
dst_layer.fc1.weight = src_layer.linear1.weight
dst_layer.fc1.bias = src_layer.linear1.bias
# fairseq may use fusedlayernorm from nvidia apex - diff properties
dst_layer.self_attn_layer_norm.load_state_dict(
src_layer.norm1.state_dict()
)
dst_layer.fc2.weight = src_layer.linear2.weight
dst_layer.fc2.bias = src_layer.linear2.bias
dst_layer.final_layer_norm.load_state_dict(
src_layer.norm2.state_dict()
)
# self.encoder = self.encoder.eval().cuda().half()
def forward(self, tokens, src_lengths=None):
return self.encoder(
tokens,
src_lengths=src_lengths,
return_all_hiddens=False,
token_embeddings=None,
)
def get_layers_embedding_dim_num_heads_for_configuration(xlarge, large):
if xlarge:
# XLM-R extra large (no BERT-XL exists)
L = 24 # Layers
D = 2560 # Embedding Dim
H = 32 # Number of Heads
FD = 10240 # Feed-forward network dim
V = 30000 # Vocab Size
elif large:
# BERT-large
L = 24
D = 1024
H = 16
FD = 4096
V = 30000
else:
# BERT-base
L = 12
D = 768
H = 12
FD = 3072
V = 30000
return (L, D, H, FD, V)
# Better transformer
class PTTransformer(torch.nn.Module):
def __init__(self, transformer, embedding):
super().__init__()
self.transformer = transformer
self.embedding = embedding
self.padding_idx = DEFAULT_PADDING_IDX
def forward(self, x):
padding_mask = None
if not x.is_nested:
padding_mask = x.eq(self.padding_idx)
x = self.embedding(x)
return self.transformer(x, src_key_padding_mask=padding_mask)
def make_transformer():
return (
PTTransformer(
torch.nn.TransformerEncoder(
torch.nn.TransformerEncoderLayer(
d_model=D,
nhead=H,
dim_feedforward=FD,
batch_first=True,
activation="relu",
),
num_layers=L,
enable_nested_tensor=False,
),
embedding_layer,
)
.eval()
.cuda()
)
def copy_weights(layers_fairseq, layers_bt):
for src_layer, dst_layer in zip(layers_fairseq, layers_bt):
w_q = src_layer.self_attn.q_proj.weight
b_q = src_layer.self_attn.q_proj.bias
w_k = src_layer.self_attn.k_proj.weight
b_k = src_layer.self_attn.k_proj.bias
w_v = src_layer.self_attn.v_proj.weight
b_v = src_layer.self_attn.v_proj.bias
dst_layer.self_attn.in_proj_weight = torch.nn.Parameter(
torch.cat((w_q, w_k, w_v), dim=0)
)
dst_layer.self_attn.in_proj_bias = torch.nn.Parameter(
torch.cat((b_q, b_k, b_v), dim=0)
)
dst_layer.self_attn.out_proj.weight = src_layer.self_attn.out_proj.weight
dst_layer.self_attn.out_proj.bias = src_layer.self_attn.out_proj.bias
dst_layer.linear1.weight = src_layer.fc1.weight
dst_layer.linear1.bias = src_layer.fc1.bias
dst_layer.linear2.weight = src_layer.fc2.weight
dst_layer.linear2.bias = src_layer.fc2.bias
dst_layer.norm1.weight = src_layer.self_attn_layer_norm.weight
dst_layer.norm1.bias = src_layer.self_attn_layer_norm.bias
dst_layer.norm2.weight = src_layer.final_layer_norm.weight
dst_layer.norm2.bias = src_layer.final_layer_norm.bias
(L, D, H, FD, V) = get_layers_embedding_dim_num_heads_for_configuration(
xlarge, large
)
embedding_layer = torch.nn.Embedding(V, D, DEFAULT_PADDING_IDX)
# True means BT as source and fairseq is target, False means the other way
# mode1 = False
if bt2fairseq:
# BT as source and fairseq is target, copy BT's weight to fairseq
transformer = make_transformer()
fairseq_transformer = (
FairseqEncoder(
D,
H,
FD,
L,
embedding_layer,
dropout=0,
normalize_before=False,
torch_encoder=transformer.transformer,
activation="relu",
)
.eval()
.cuda()
)
if half:
transformer.half()
fairseq_transformer.half()
if not bt2fairseq:
# the other way around, fairseq is source and BT is target,copy fairseq's weight to BT
transformer = make_transformer()
fairseq_transformer = (
FairseqEncoder(
D,
H,
FD,
L,
embedding_layer,
dropout=0,
normalize_before=False,
torch_encoder=None,
activation="relu",
)
.eval()
.cuda()
)
# for the test where we need to load existing ckpt. It is tested that after loading
# the ckpt, the results between fairseq_transformer(BT kernel) equals BT
if half:
transformer.half()
fairseq_transformer.half()
if save:
torch.save(fairseq_transformer.state_dict(), "./fairseq.pt")
sys.exit(0)
if load:
fairseq_transformer.load_state_dict(torch.load("./fairseq.pt"))
# copy
copy_weights(fairseq_transformer.encoder.layers, transformer.transformer.layers)
device = "cuda"
lengths = (avg_sequence_length,) * batch_size
tokens = torch.full(
(batch_size, max_sequence_length),
DEFAULT_PADDING_IDX,
device=device,
dtype=torch.long,
)
for i in range(batch_size):
tokens[i, : lengths[i]] = torch.randint(
DEFAULT_PADDING_IDX + 1,
V - 1,
size=(lengths[i],),
device=device,
dtype=torch.long,
)
# mask
if half:
lengths_tensor = torch.Tensor(lengths).cuda().half()
else:
lengths_tensor = torch.Tensor(lengths).cuda()
with torch.inference_mode():
fs_output = fairseq_transformer(tokens, lengths_tensor)["encoder_out"][0]
fs_output = fs_output.transpose(0, 1)
with torch.inference_mode():
t_output = transformer(tokens)
test_lst = [
# (name, output, relative tolerance, absolute tolerance)
("FS", fs_output, 1e-4, 9e-3),
]
numerical_test(lengths, t_output, test_lst)
iters = 100
t = benchmark_torch_function(iters, transformer, tokens)
def bert_flops(B, T, D, L):
mlp = 2 * (B * T * D * 4 * D) + 2 * (B * T * D * 4 * D)
qkv = 3 * 2 * B * T * D * D
attn = 2 * B * D * T * T + 2 * B * D * T * T + 2 * B * T * D * D
return L * (mlp + qkv + attn)
flops = bert_flops(batch_size, avg_sequence_length, D, L)
flops_e = (
FlopCountAnalysis(transformer, (tokens[:, :avg_sequence_length])).total() * 2
)
with torch.inference_mode():
bt = benchmark_torch_function(iters, transformer, tokens)
fst = benchmark_torch_function(
iters, fairseq_transformer, tokens, lengths_tensor
)
def metrics(tt, baseline=None):
if baseline:
return metrics(tt) + f", Speedup: {baseline / tt:.2f}x"
return f"{tt * 1.0e3:.2f} ms/iter, {flops_e / tt / 1.0e12:.2f} TFLOP/s"
results = [
f"Seed: {seed}",
f"Padded tokens: {(1-sum(lengths)/(tokens.numel()))*100:.2f}%",
f"Batch shape: {tokens.shape}",
f"Analytical flops per batch: {flops/ batch_size / 1e9:.2f} GFLOPS",
f"Empirical flops per batch: {flops_e/ batch_size / 1e9:.2f} GFLOPS",
f"B: {batch_size}",
f"T: {avg_sequence_length}",
f"TMax: {max_sequence_length}",
f"Eager Time: {metrics(t)}",
f"BetterTransformer: {metrics(bt, t)}",
f"FST: {metrics(fst, t)}",
]
print("===========Speedup Results")
print("; ".join(results))
def flop_count_analysis(
self, inputs: Union[torch.Tensor, Tuple[torch.Tensor, ...]]
) -> FlopCountAnalysis:
return FlopCountAnalysis(self, inputs)
